PSyGenTAB: A Privacy-Preserving Framework for Synthetic Clinical Tabular Data Generation via Constrained Optimization

# PSyGenTAB: A Privacy-Preserving Framework for Synthetic Clinical Tabular Data Generation via Constrained Optimization
Source: [https://arxiv.org/html/2606.18518](https://arxiv.org/html/2606.18518)
\\UseRawInputEncoding

Arshia Ilaty†‡, Hossein Shirazi†, Manasi Chitale†, Kedar Hegde†, Dhanalakshmi Ramesh†, Rashmi S\. Manjunath†, Amir Rahmani‡, Hajar Homayouni††San Diego State University, San Diego, CA 92182 USA\. ‡University of California, Irvine, CA 92697 USA

###### Abstract

The development of medical AI is constrained by limited access to high\-quality clinical data due to institutional silos and strict privacy regulations such as HIPAA and GDPR\. Synthetic data generation offers a potential solution, but existing methods lack principled mechanisms to explicitly manage the privacy–utility trade\-off, often degrading clinically meaningful patterns or risking patient re\-identification\. We present PSyGenTAB, a privacy\-preserving generative framework that formulates synthetic healthcare data generation as a constrained optimization problem solved using the Augmented Lagrangian Method\. By embedding configurable privacy constraints directly into model training, PSyGenTAB enforces minimum privacy thresholds while maximizing clinical data utility\. Across multiple clinically motivated benchmarks, PSyGenTAB preserves inter\-feature clinical relationships and minority\-class diagnostic patterns essential for reliable health AI\. Downstream evaluation using Train\-on\-Synthetic, Test\-on\-Real and Train\-on\-Real, Test\-on\-Synthetic protocols shows that models trained on synthetic data achieve performance comparable to those trained on real patient records\. Privacy auditing further demonstrates reduced exact record reproduction and strong resilience to membership inference attacks\. These results establish PSyGenTAB as a principled framework for balancing privacy protection and clinical utility in synthetic healthcare data, supporting secure cross\-institutional AI development\.

## IIntroduction

The rapid advancement of artificial intelligence in healthcare holds immense promise for personalized medicine, clinical research acceleration, and automated diagnostics\. However, the development of robust medical AI is severely hindered by a pervasive data crisis\. Clinical records are locked within fragmented institutional silos, constrained by necessary and stringent privacy regulations such as the Health Insurance Portability and Accountability Act \(HIPAA\)\[[12](https://arxiv.org/html/2606.18518#bib.bib20)\]and the General Data Protection Regulation \(GDPR\)\[[64](https://arxiv.org/html/2606.18518#bib.bib21)\]\. These barriers to data sharing disproportionately impact rare disease research and the development of unbiased predictive models, as single institutions rarely possess the diverse, large\-scale patient cohorts required to train generalizable algorithms\. Consequently, there is an urgent need for technologies that can safely unlock clinical data for collaborative research without compromising patient confidentiality\.

Synthetic data generation has emerged as a clinically motivated solution to bridge this gap\. By modeling the underlying statistical distributions of electronic health records \(EHRs\), methods such as Synthea\[[65](https://arxiv.org/html/2606.18518#bib.bib184)\], medGAN\[[9](https://arxiv.org/html/2606.18518#bib.bib73)\], and clinical GAN variants aim to produce realistic, shareable patient cohorts\. Yet, current methods frequently fail in practical healthcare deployments for two opposing reasons\. High\-fidelity generative models often inadvertently memorize training data, creating unacceptable risks of identity disclosure and patient re\-identification\. Conversely, methods applying rigid differential privacy mechanisms inject excessive noise into the learning process, destroying the subtle inter\-feature correlations and rare diagnostic signals that are essential for clinical utility\[[1](https://arxiv.org/html/2606.18518#bib.bib244)\]\.

This tension reveals a critical and unresolved gap: there currently exists no principled, jointly optimized approach that simultaneously guarantees high clinical utility and rigorous patient privacy protection in tabular health data\. Existing generative models typically treat privacy as a post\-hoc patch, applied after training through gradient clipping, noise injection, or adversarial regularization, rather than as a fundamental design constraint\. This forces clinical researchers to accept unstable, fixed trade\-offs that either compromise patient safety or render the synthetic data analytically useless for downstream predictive modeling\.

To address this challenge, we introducePSyGenTAB, a privacy\-preserving generative framework designed to enable the safe sharing of sensitive health data for cross\-institutional research\. PSyGenTAB formulates synthetic healthcare data generation as a constrained optimization problem, solved via the Augmented Lagrangian Method \(ALM\), which dynamically enforces configurable privacy constraints during training or sampling\. This design ensures that synthetic clinical records remain statistically faithful to real patient populations while resisting re\-identification attacks\. PSyGenTAB is model\-agnostic and can be integrated with diverse tabular generative architectures without modifying their core structure\. In practice, the framework wraps around an existing generator and augments its objective function with privacy constraints defined through composite privacy metrics\. The ALM procedure adaptively updates Lagrange multipliers to maintain a minimum privacy threshold while maximizing data utility, enabling transparent and tunable control of the privacy–utility trade\-off across different healthcare datasets and institutional requirements\. The primary contributions of this work are as follows:

1. 1\.We propose a novel, model\-agnostic framework utilizing ALM to explicitly and adaptively balance diagnostic utility and patient privacy during synthetic health data generation, treating the privacy–utility trade\-off as a dynamic, configurable constraint rather than a fixed compromise\.
2. 2\.We define multi\-dimensional, clinically grounded evaluation metrics capturing distributional fidelity, the preservation of rare disease patterns, inter\-feature clinical dependencies, and resistance to patient re\-identification under HIPAA\- and GDPR\-relevant threat models\.
3. 3\.We demonstrate the clinical validity of PSyGenTAB on critical healthcare benchmarks, including the Diabetes Health Indicators and Breast Cancer Wisconsin datasets, showing that it preserves vital minority\-class diagnostic signals for downstream predictive modeling while maintaining strong patient privacy boundaries\.

The remainder of this paper is organized as follows\. Section[II](https://arxiv.org/html/2606.18518#S2)reviews related work in synthetic health data generation, privacy\-preserving AI, and constrained optimization\. Section[III](https://arxiv.org/html/2606.18518#S3)details the PSyGenTAB framework and its clinical metric formulations\. Section[IV](https://arxiv.org/html/2606.18518#S4)presents experimental validation on clinical and benchmark datasets\. Section[V](https://arxiv.org/html/2606.18518#S5)discusses regulatory implications, architectural considerations, and limitations\. Section[VI](https://arxiv.org/html/2606.18518#S6)outlines directions for future research and Section[VII](https://arxiv.org/html/2606.18518#S7)concludes the paper\.

Given the rapid evolution of synthetic data generation research and space constraints, this work cannot exhaustively cover all recent advances\. To maintain an up\-to\-date record of developments, benchmarks, and extensions of PSyGenTAB, we refer readers to the project repository\.111\[Online\]\.Available:[https://github\.com/ArshiaIlaty/PsyGenTAB](https://github.com/ArshiaIlaty/PsyGenTAB)

## IIRelated Work

##### Synthetic Health Data Generation

The need for shareable EHRs has driven advances in synthetic data generation across rule\-based simulation and deep generative modeling\. Early systems such as Synthea\[[65](https://arxiv.org/html/2606.18518#bib.bib184)\]enabled deterministic population\-scale simulations but lacked the statistical complexity of real\-world cohorts, motivating data\-driven approaches\. Generative Adversarial Networks \(GANs\)\[[26](https://arxiv.org/html/2606.18518#bib.bib23)\]marked a shift toward realistic clinical synthesis\. MedGAN\[[10](https://arxiv.org/html/2606.18518#bib.bib191)\]demonstrated GAN\-based generation of multi\-label EHR records, while Variational Autoencoders \(VAEs\)\[[35](https://arxiv.org/html/2606.18518#bib.bib126),[14](https://arxiv.org/html/2606.18518#bib.bib75)\]provided more stable training on mixed\-type clinical data\. For tabular synthesis, TGAN\[[67](https://arxiv.org/html/2606.18518#bib.bib97)\]introduced mode\-specific normalization and autoregressive modeling; CTGAN\[[66](https://arxiv.org/html/2606.18518#bib.bib98)\]improved handling of imbalanced categorical attributes; and CTAB\-GAN\+\[[71](https://arxiv.org/html/2606.18518#bib.bib129)\]strengthened preservation of inter\-feature dependencies\. Transformer\-based models such as REaLTabFormer\[[60](https://arxiv.org/html/2606.18518#bib.bib105)\]and GReaT\[[4](https://arxiv.org/html/2606.18518#bib.bib102)\], as well as diffusion approaches\[[37](https://arxiv.org/html/2606.18518#bib.bib109),[70](https://arxiv.org/html/2606.18518#bib.bib117)\], further improved fidelity\. Despite architectural advances, memorization of training samples remains a critical limitation, enabling potential re\-identification\[[57](https://arxiv.org/html/2606.18518#bib.bib202)\]\. Systematic reviews\[[27](https://arxiv.org/html/2606.18518#bib.bib121),[45](https://arxiv.org/html/2606.18518#bib.bib3)\]confirm that principled, configurable control of the privacy–utility trade\-off remains unresolved\.

##### Privacy\-Preserving Methods in Healthcare AI

Under regulatory frameworks such as HIPAA\[[12](https://arxiv.org/html/2606.18518#bib.bib20)\]and GDPR\[[64](https://arxiv.org/html/2606.18518#bib.bib21)\], three major privacy paradigms have emerged: statistical disclosure limitation, federated learning, and differential privacy\. Classical de\-identification methods, includingkk\-anonymity\[[61](https://arxiv.org/html/2606.18518#bib.bib62)\],ℓ\\ell\-diversity\[[47](https://arxiv.org/html/2606.18518#bib.bib205)\], andtt\-closeness\[[42](https://arxiv.org/html/2606.18518#bib.bib204)\], serve as regulatory baselines but degrade fine\-grained structure necessary for clinical AI\. Federated learning\[[55](https://arxiv.org/html/2606.18518#bib.bib230),[51](https://arxiv.org/html/2606.18518#bib.bib77)\]enables decentralized model training but remains vulnerable to gradient inversion attacks\[[73](https://arxiv.org/html/2606.18518#bib.bib227)\]and does not support distribution of shareable synthetic datasets\. Differential privacy \(DP\)\[[20](https://arxiv.org/html/2606.18518#bib.bib242),[19](https://arxiv.org/html/2606.18518#bib.bib203)\], implemented in DP\-GAN\[[29](https://arxiv.org/html/2606.18518#bib.bib194)\], DP\-CTGAN\[[24](https://arxiv.org/html/2606.18518#bib.bib95)\], and PATE\-GAN\[[34](https://arxiv.org/html/2606.18518#bib.bib219)\], provides formal guarantees but often obscures clinically relevant structure\. Membership inference attacks \(MIAs\)\[[57](https://arxiv.org/html/2606.18518#bib.bib202),[7](https://arxiv.org/html/2606.18518#bib.bib206)\]serve as empirical privacy audits\. As shown in our evaluation, DP\-SGD improves formal guarantees but reduces structural fidelity and downstream predictive performance, an unacceptable trade\-off in precision medicine\.

##### Privacy–Utility Trade\-offs in Medical Data

Clinical data sharing fundamentally balances statistical fidelity against re\-identification risk\. Features that preserve rare events and realistic correlations are also those most vulnerable to disclosure\. Jordon et al\.\[[33](https://arxiv.org/html/2606.18518#bib.bib229)\]formalize this inherent trade\-off, demonstrating that perfect utility and privacy cannot be achieved simultaneously\. Empirical evaluations across traditional generative models\[[40](https://arxiv.org/html/2606.18518#bib.bib167)\]and large language models\[[31](https://arxiv.org/html/2606.18518#bib.bib209)\]show that highest\-fidelity models often incur the highest privacy risk\. Structured frameworks such as SynthEval\[[41](https://arxiv.org/html/2606.18518#bib.bib128)\]emphasize multi\-dimensional assessment across utility and privacy metrics\. PSyGenTAB extends this by jointly optimizing composite quality \(QQ\) and privacy \(PP\) objectives during generation, explicitly navigating the privacy–utility frontier rather than treating privacy as post\-hoc regularization\.

##### Constrained Optimization for Trustworthy Healthcare AI

Trustworthy healthcare AI requires embedding privacy and validity constraints directly into model training\. Augmented Lagrangian Methods \(ALM\)\[[28](https://arxiv.org/html/2606.18518#bib.bib237),[53](https://arxiv.org/html/2606.18518#bib.bib238),[56](https://arxiv.org/html/2606.18518#bib.bib225),[3](https://arxiv.org/html/2606.18518#bib.bib224)\]provide a principled framework for inequality\-constrained optimization with adaptive multipliers and convergence to KKT stationary points\. Extensions to deep learning\[[43](https://arxiv.org/html/2606.18518#bib.bib130),[36](https://arxiv.org/html/2606.18518#bib.bib131)\]demonstrate automatic balancing of competing objectives without manual loss tuning\. ALM has also been applied to fairness\-constrained learning\[[13](https://arxiv.org/html/2606.18518#bib.bib228)\], highlighting its suitability for enforcing non\-negotiable constraints\. By adapting ALM to synthetic health data generation, PSyGenTAB constrains the probability of producing records that closely resemble real patient profiles, mitigating memorization while maximizing clinical fidelity\. To our knowledge, it is the first framework to apply ALM\-based constrained optimization to the privacy–utility trade\-off in synthetic clinical tabular data, enabling principled and configurable regulatory\-compliant data sharing\.

## IIIMethodology

When a multi\-site clinical consortium aims to pool data for a large\-scale AI study, direct transfer of EHRs is often blocked by institutional review boards \(IRBs\) due to re\-identification risks\. PSyGenTAB serves as the technological bridge in this scenario: a hospital trains a PSyGenTAB model locally on its patient records and shares the resulting high\-fidelity synthetic data, confident that the generation process was explicitly and mathematically constrained against leaking real patient identities\. The configurable privacy thresholdPminP\_\{\\min\}can be calibrated to the specific requirements of the data use agreement or IRB protocol, providing an auditable, quantitative record of the privacy assurance applied\.

### III\-AFormalizing Clinical Utility and Privacy Metrics

A fundamental limitation of existing approaches to synthetic health data generation is the absence of standardized, formal definitions of clinical data utility and privacy\-preservability\. In practice, these properties are evaluated using ad\-hoc or task\-specific metrics, making it difficult to compare models, reason about trade\-offs, or enforce application\-specific regulatory requirements\. There is a need for explicit, well\-defined metrics that jointly capture statistical fidelity, structural clinical relationships, rare\-event preservation, and memorization risk in a reproducible and application\-agnostic manner\.

Research Goal IFormally define multi\-dimensional clinical utility and privacy metrics to enable more accurate, consistent, and actionable evaluation of synthetic tabular health data, revealing privacy–utility trade\-offs that are obscured by single\-metric or post\-hoc evaluations, and providing a necessary foundation for the constrained optimization presented in Section[III\-B](https://arxiv.org/html/2606.18518#S3.SS2)\.

LetDrealD\_\{\\text\{real\}\}andDsynD\_\{\\text\{syn\}\}denote the real and synthetic patient datasets respectively\. We define two composite scores,clinical data utilityUUandprivacy\-preservabilityPP, each normalized to\[0,1\]\[0,1\]prior to use in the constrained optimization of Section[III\-B](https://arxiv.org/html/2606.18518#S3.SS2)\. Higher values ofUUindicate stronger clinical fidelity and downstream usefulness, while higher values ofPPindicate stronger protection against patient disclosure, attribute inference, and memorization\.

##### Clinical Data Utility \(UU\)

We define clinical data utility as the degree to which synthetic patient records accurately reflect the statistical, structural, and contextual properties of the true patient population\. The composite utility scoreUUis expressed as:

U=ω1​ℱmarg\+ω2​ℱrare\+ω3​ℱjoint\+ω4​𝒞div\+ω5​𝒯dyn\+ω6​ℰpred\\footnotesize U=\\omega\_\{1\}\\mathcal\{F\}\_\{\\text\{marg\}\}\+\\omega\_\{2\}\\mathcal\{F\}\_\{\\text\{rare\}\}\+\\omega\_\{3\}\\mathcal\{F\}\_\{\\text\{joint\}\}\+\\omega\_\{4\}\\mathcal\{C\}\_\{\\text\{div\}\}\+\\omega\_\{5\}\\mathcal\{T\}\_\{\\text\{dyn\}\}\+\\omega\_\{6\}\\mathcal\{E\}\_\{\\text\{pred\}\}\(1\)where∑i=16ωi=1\\sum\_\{i=1\}^\{6\}\\omega\_\{i\}=1and each component is normalized to\[0,1\]\[0,1\]prior to aggregation\. The six dimensions are defined as follows\.

ℱmarg\\mathcal\{F\}\_\{\\text\{marg\}\}, Marginal Distribution Fidelity:Measures how closely the marginal distributions and valid feature ranges of the synthetic patient cohort match the real population\. For continuous clinical variables \(e\.g\., BMI, age, blood glucose\), distributional alignment is computed using the Kolmogorov\-Smirnov \(KS\) test\[[25](https://arxiv.org/html/2606.18518#bib.bib108)\], Kullback\-Leibler \(KL\) divergence\[[39](https://arxiv.org/html/2606.18518#bib.bib104)\], Jensen\-Shannon divergence \(JSD\)\[[48](https://arxiv.org/html/2606.18518#bib.bib27)\], and Range Coverage \(RC\)\[[62](https://arxiv.org/html/2606.18518#bib.bib148)\]\. For categorical variables \(e\.g\., diagnosis codes, sex, smoking status\), Total Variation Distance \(TVD\) is applied\.ℱmarg\\mathcal\{F\}\_\{\\text\{marg\}\}is defined as:

ℱmarg=1−1\|ℱ\|​∑f∈ℱδf​\(Dreal,Dsyn\)\\footnotesize\\mathcal\{F\}\_\{\\text\{marg\}\}=1\-\\frac\{1\}\{\|\\mathcal\{F\}\|\}\\sum\_\{f\\in\\mathcal\{F\}\}\\delta\_\{f\}\(D\_\{\\text\{real\}\},\\,D\_\{\\text\{syn\}\}\)\(2\)whereδf\\delta\_\{f\}denotes the feature\-appropriate divergence measure andℱ\\mathcal\{F\}is the full feature set\.ℱrare\\mathcal\{F\}\_\{\\text\{rare\}\}, Rare Clinical Event and Anomaly Preservation\.Quantifies the retention of low\-frequency but diagnostically critical patterns, such as prediabetes in a population\-level diabetes cohort, or malignant subtypes in a cancer registry, that are essential for clinical AI applied to rare disease research\. For continuous columns, outliers and rare sub\-populations are identified using the Interquartile Range \(IQR\) method\[[5](https://arxiv.org/html/2606.18518#bib.bib94)\]; for categorical columns, minority class frequency alignment is measured proportionally\. Higherℱrare\\mathcal\{F\}\_\{\\text\{rare\}\}values indicate robust, memorization\-free retention of clinically significant edge cases\.

ℱjoint\\mathcal\{F\}\_\{\\text\{joint\}\}, Joint Association Fidelity:Evaluates the preservation of inter\-feature dependencies and covariance structure that encode clinically meaningful relationships, for example, the established association between BMI, fasting glucose, and diabetes risk\. For continuous feature pairs, Pearson or Spearman correlation matrix similarity is computed; divergence is quantified via the Frobenius norm of the correlation matrix difference\. For categorical or mixed\-type pairs, contingency\-based measures and normalized mutual information\[[72](https://arxiv.org/html/2606.18518#bib.bib132)\]are applied\. A hallucinated clinical association in synthetic data, such as inverting a known biomarker relationship, would directly undermine downstream diagnostic model validity\.

𝒞div\\mathcal\{C\}\_\{\\text\{div\}\}, State\-Space Coverage and Diversity:Assesses whether the synthetic cohort maintains comprehensive support across the full patient feature space, preventing mode collapse onto majority\-class demographic profiles\. We compute Category Coverage \(CC\)\[[62](https://arxiv.org/html/2606.18518#bib.bib148)\]for discrete clinical variables to ensure no diagnostic class is dropped, and Regularized Support Coverage\[[11](https://arxiv.org/html/2606.18518#bib.bib118)\]to evaluate representation balance across low\-density regions of the continuous feature space\. This dimension directly addresses the health equity concern of generating synthetic data that preserves the representation of underserved or minority patient populations\.

𝒯dyn\\mathcal\{T\}\_\{\\text\{dyn\}\}, Temporal Dynamics Consistency:Measures the fidelity of sequential dependencies in patient trajectories, relevant for datasets with time\-ordered or longitudinal features\. We apply Dynamic Time Warping \(DTW\)\[[2](https://arxiv.org/html/2606.18518#bib.bib183)\]to compare sequence trajectories and temporal auto\-correlation\[[63](https://arxiv.org/html/2606.18518#bib.bib185)\]to validate lag structures and seasonality\. Although the primary datasets evaluated in this work are cross\-sectional,𝒯dyn\\mathcal\{T\}\_\{\\text\{dyn\}\}is included in the utility composite to enable direct application of PSyGenTAB to longitudinal EHR data, a key future direction discussed in Section[VI](https://arxiv.org/html/2606.18518#S6)\.

ℰpred\\mathcal\{E\}\_\{\\text\{pred\}\}, Downstream Predictive Efficacy:Evaluates the empirical clinical usefulness of synthetic data using the Train\-on\-Synthetic, Test\-on\-Real \(TSTR\) and Train\-on\-Real, Test\-on\-Synthetic \(TRTS\) protocols\. Downstream classifiers are trained exclusively onDsynD\_\{\\text\{syn\}\}and evaluated on a held\-outDrealD\_\{\\text\{real\}\}test set, and vice versa\. Performance is aggregated using balanced accuracy, F1\-score, and AUC\-ROC\.ℰpred\\mathcal\{E\}\_\{\\text\{pred\}\}is the most clinically direct component ofUU: it measures whether PSyGenTAB\-generated data can substitute for real patient records in the development and validation of clinical AI models\.

##### Privacy\-Preservability \(PP\)

Privacy\-preservability represents the generative model’s resistance to patient identity disclosure, clinical attribute inference, and training data extraction\. We formulatePPas a penalty\-adjusted score:

P=λ1​𝒟c−λ2​Δdist−λ3​ℛdup\\footnotesize P=\\lambda\_\{1\}\\mathcal\{D\}\_\{c\}\-\\lambda\_\{2\}\\Delta\_\{\\text\{dist\}\}\-\\lambda\_\{3\}\\mathcal\{R\}\_\{\\text\{dup\}\}\(3\)whereλ1\+λ2\+λ3=1\\lambda\_\{1\}\+\\lambda\_\{2\}\+\\lambda\_\{3\}=1\. The positive term𝒟c\\mathcal\{D\}\_\{c\}rewards spatial separation from real patient records; the negative termsΔdist\\Delta\_\{\\text\{dist\}\}andℛdup\\mathcal\{R\}\_\{\\text\{dup\}\}penalize distributional overfitting and verbatim memorization respectively\. All components are normalized to\[0,1\]\[0,1\]prior to aggregation, ensuringP∈\[0,1\]P\\in\[0,1\]and that higher values ofPPuniversally correspond to stronger patient privacy protection\.

𝒟c\\mathcal\{D\}\_\{c\}, Distance to Closest Record \(DCR\):Computes the spatial separation between each synthetic patient record and its nearest neighbour inDrealD\_\{\\text\{real\}\}, using Gower distance for mixed\-type clinical data\[[67](https://arxiv.org/html/2606.18518#bib.bib97),[52](https://arxiv.org/html/2606.18518#bib.bib99)\]\. The Nearest Neighbour Distance Ratio \(NNDR\) is additionally computed as a normalized variant to control for feature\-space density effects\. Under HIPAA’s Expert Determination pathway, re\-identification risk is assessed against the proximity of synthetic records to real patient profiles;𝒟c\\mathcal\{D\}\_\{c\}provides a direct, quantitative proxy for this regulatory assessment\. Higher𝒟c\\mathcal\{D\}\_\{c\}values indicate that no real patient profile can be closely reconstructed from the synthetic data\.

Δdist\\Delta\_\{\\text\{dist\}\}, Distributional Overfitting Discrepancy:Evaluates structural overfitting at the distributional level by measuring the difference in quantile\-specific behavior between real and synthetic distributions\[[60](https://arxiv.org/html/2606.18518#bib.bib105)\]\. While𝒟c\\mathcal\{D\}\_\{c\}operates at the individual record level,Δdist\\Delta\_\{\\text\{dist\}\}operates at the population level, penalizing models that have effectively memorized dense clusters of real patient records even if no individual record is exactly reproduced\. LowerΔdist\\Delta\_\{\\text\{dist\}\}values indicate that the generator has learned the underlying clinical data manifold rather than memorizing local clusters in sensitive sub\-populations such as patients with rare comorbidities\.

ℛdup\\mathcal\{R\}\_\{\\text\{dup\}\}, Duplication and Memorization Rate:Quantifies the Exact Match Ratio \(EMR\) betweenDsynD\_\{\\text\{syn\}\}andDrealD\_\{\\text\{real\}\}, as well as internal mode collapse withinDsynD\_\{\\text\{syn\}\}\. Exact replication of a real patient record constitutes the most severe possible privacy violation, directly enabling re\-identification\.ℛdup=0\\mathcal\{R\}\_\{\\text\{dup\}\}=0is a hard requirement for GDPR compliance and aligns with HIPAA de\-identification standards; the ALM constraint framework of Section[III\-B](https://arxiv.org/html/2606.18518#S3.SS2)enforces this by penalizing any non\-zero EMR through the privacy compositePP\.

### III\-BALM\-Based Optimization for Safe Clinical Data Generation

Using the clinically grounded metrics defined above, we frame synthetic health data generation as a constrained optimization problem\. The generator parametersθ\\thetamust maximize clinical data utilityU​\(θ\)U\(\\theta\), preserving diagnostic signals, minority\-class representations, and clinical biomarker correlations, subject to the hard constraint that privacy\-preservabilityP​\(θ\)P\(\\theta\)remains above an institutionally configurable minimum thresholdPminP\_\{\\min\}\.

Research Goal IIFormulating synthetic clinical data generation as an ALM\-constrained optimization problem, with patient privacy enforced as a hard constraint rather than a post\-hoc regularization term, to simultaneously improve clinical data utility and reduce patient re\-identification risk, resolving the privacy–utility tension that limits existing approaches\.

Formally, the optimization problem is:

minθ⁡\(1−U​\(θ\)\)​subject to​P​\(θ\)≥Pmin\\footnotesize\\min\_\{\\theta\}\\;\\bigl\(1\-U\(\\theta\)\\bigr\)\\quad\\text\{subject to\}\\quad P\(\\theta\)\\geq P\_\{\\min\}\(4\)The constraintP​\(θ\)≥PminP\(\\theta\)\\geq P\_\{\\min\}encodes the institutional privacy requirement directly into the learning objective\.PminP\_\{\\min\}is a configurable hyperparameter set by the data custodian to reflect the applicable regulatory standard \(e\.g\., HIPAA Expert Determination, GDPR Article 89, or a specific IRB protocol requirement\)\.

To solve \([4](https://arxiv.org/html/2606.18518#S3.E4)\), we apply the Augmented Lagrangian Method \(ALM\)\[[3](https://arxiv.org/html/2606.18518#bib.bib224),[56](https://arxiv.org/html/2606.18518#bib.bib225),[28](https://arxiv.org/html/2606.18518#bib.bib237),[53](https://arxiv.org/html/2606.18518#bib.bib238)\], which converts the constrained problem into a sequence of unconstrained sub\-problems through a Lagrange multiplierλ≥0\\lambda\\geq 0and an adaptive penalty coefficientμ\>0\\mu\>0:

ℒAL​\(θ,λ\)=\(1−U​\(θ\)\)\+λ​\(Pmin−P​\(θ\)\)\+μ2​\(Pmin−P​\(θ\)\)2\\footnotesize\\mathcal\{L\}\_\{\\text\{AL\}\}\(\\theta,\\lambda\)=\\bigl\(1\-U\(\\theta\)\\bigr\)\+\\lambda\\bigl\(P\_\{\\min\}\-P\(\\theta\)\\bigr\)\+\\frac\{\\mu\}\{2\}\\bigl\(P\_\{\\min\}\-P\(\\theta\)\\bigr\)^\{2\}\(5\)The three terms in \([5](https://arxiv.org/html/2606.18518#S3.E5)\) have direct clinical interpretations\. The first term\(1−U​\(θ\)\)\\bigl\(1\-U\(\\theta\)\\bigr\)is the clinical utility loss: minimizing it maximizes the fidelity of synthetic patient profiles across all six dimensions ofUU\. The second termλ​\(Pmin−P​\(θ\)\)\\lambda\\bigl\(P\_\{\\min\}\-P\(\\theta\)\\bigr\)is the Lagrangian privacy penalty:λ\\lambdais updated at each epoch to reflect the cumulative extent of constraint violation, growing whenever the generator produces synthetic records that are too similar to real patient profiles\. The third quadratic termμ2​\(Pmin−P​\(θ\)\)2\\frac\{\\mu\}\{2\}\\bigl\(P\_\{\\min\}\-P\(\\theta\)\\bigr\)^\{2\}applies a progressively larger penalty as the magnitude of constraint violation increases, ensuring rapid convergence back to the privacy\-feasible region\. Together,λ\\lambdaandμ\\muact as an adaptive privacy enforcement mechanism: when the generator begins memorizing real patient profiles, the combined penalty increases automatically until the constraint is restored\.

Algorithm 1ALM\-Based Privacy\-Constrained Training0:Generator parameters

θ\\theta; privacy threshold

PminP\_\{\\min\}; initial multiplier

λ≥0\\lambda\\geq 0; penalty coefficient

μ\>0\\mu\>0; learning rate

η\\eta; growth factor

α\>1\\alpha\>1
0:Trained generator

θ∗\\theta^\{\*\}satisfying

P​\(θ∗\)≥PminP\(\\theta^\{\*\}\)\\geq P\_\{\\min\}
1:foreach training epochdo

2:foreach mini\-batch from

DrealD\_\{\\text\{real\}\}do

3:Generate synthetic batch using current generator

4:Compute composite utility score

UU
5:Compute composite privacy score

PP
6:Form augmented Lagrangian objective

7:Update

θ\\thetavia gradient descent on the objective

8:endfor

9:Evaluate

UUand

PPon validation data

10:if

P<PminP<P\_\{\\min\}then

11:Increase penalty weight:

μ←α​μ\\mu\\leftarrow\\alpha\\mu
12:endif

13:Update multiplier:

λ←max⁡\(0,λ\+\(Pmin−P\)\)\\lambda\\leftarrow\\max\(0,\\lambda\+\(P\_\{\\min\}\-P\)\)
14:ifprivacy constraint satisfied and utility stabilizedthen

15:break

16:endif

17:endfor

18:return

θ∗\\theta^\{\*\}

##### Sampling\-Time ALM Variant

A practical advantage of PSyGenTAB is that the ALM framework can also be applied during thesamplingphase after a generative model has already been trained\. In this variant, the privacy and utility penalties in Equation \([5](https://arxiv.org/html/2606.18518#S3.E5)\) are used to steer or filter generated samples rather than to update model parameters\. This enables a single trained generator to produce synthetic datasets at multiple configurable privacy levels, for example, a permissivePminP\_\{\\min\}for internal institutional use and a stricter threshold for external data sharing, without retraining\. This sampling\-time variant is particularly valuable for healthcare data custodians managing multiple downstream use cases with differing regulatory requirements\.

##### Model\-Agnostic Application

PSyGenTAB applies the ALM constraint framework to the generator parametersθ\\thetaindependently of the underlying generative architecture\. In this work we evaluate two complementary generators: REaLTabFormer\[[60](https://arxiv.org/html/2606.18518#bib.bib105)\], a transformer\-based autoregressive model that captures complex sequential feature dependencies; and CTAB\-GAN\+\[[71](https://arxiv.org/html/2606.18518#bib.bib129)\], a conditional GAN optimized for mixed\-type tabular data\. The consistent improvement observed across both architectures \(Section[IV](https://arxiv.org/html/2606.18518#S4)\) provides empirical evidence for the model\-agnostic claim\.

## IVExperiments and Results

This section presents a comprehensive clinical and technical evaluation of the PSyGenTAB framework for privacy\-preserving synthetic health data generation\. All experiments are designed to answer a single overarching clinical question:can PSyGenTAB generate synthetic health records that are both safe to share and genuinely useful for downstream clinical AI research?Accordingly, the evaluation is structured to mirror the dual obligations of a healthcare data custodian: protecting patient identity while enabling high\-quality research\.

The remainder of this section is organized as follows: 1\) Clinical Datasets, detailing the healthcare context of each benchmark; 2\) Generative Models and Optimization Strategies, covering the architectures and privacy mechanisms evaluated; 3\) Evaluation Goals and Metrics, formalizing the utility and privacy formulations; 4\) Primary Privacy–Utility Results, highlighting the core trade\-offs; 5\) Statistical Fidelity, measuring marginal distribution alignment; 6\) Commercial Evaluation Pipeline, contrasting open\-source and proprietary synthetic data quality scores; 7\) Downstream Clinical AI Utility, assessing machine\-learning performance on synthetic versus real data; 8\) Patient Re\-identification Risk, conducting a proximity\-based distance\-to\-closest\-record \(DCR\) audit; 9\) Membership Inference Attack Resilience, analyzing adversarial identity disclosure; and 10\) Clinical Validity Assessment, applying the FAITH framework to evaluate the factual and structural fidelity of the generated records\.

### IV\-AClinical Datasets

All experiments were conducted on tabular datasets with heterogeneous structural characteristics\. The selection encompasses eleven distinct datasets, eight of which are clinically motivated health benchmarks representing highly sensitive medical records, epidemiological surveys, and biometric data\. The remaining three datasets serve as generalizability benchmarks to confirm that the PSyGenTAB framework is not domain\-specific and can successfully operate across varied statistical distributions, including financial and sensor\-based environments\.

TABLE I:Summary of dataset characteristics\.NNdenotes number of samples\. Cont\. and Cat\. indicate continuous and categorical feature counts\. \#Cls denotes number of target classes\.DatasetDomainNNCont\.Cat\.Total\#ClsClinical DatasetsBreast Cancer\[[15](https://arxiv.org/html/2606.18518#bib.bib186)\]Clinical56956930300303022Diabetes\[[6](https://arxiv.org/html/2606.18518#bib.bib156)\]Survey253,680253,680331818212133Heart Failure\[[8](https://arxiv.org/html/2606.18518#bib.bib162)\]Clinical2992997766131322Hypothyroid\[[54](https://arxiv.org/html/2606.18518#bib.bib160)\]Clinical37723772772222292922Liver Disorders\[[46](https://arxiv.org/html/2606.18518#bib.bib161)\]Clinical3453456606622Lung Cancer\[[17](https://arxiv.org/html/2606.18518#bib.bib163)\]Clinical309309221313151522Obesity\[[50](https://arxiv.org/html/2606.18518#bib.bib159)\]Survey211121118899171777Parkinsons\[[44](https://arxiv.org/html/2606.18518#bib.bib158)\]Clinical19519522220222222Generalization DatasetsAdult\[[16](https://arxiv.org/html/2606.18518#bib.bib155)\]Demographics48,84248,8426688141422PIR Vision\[[23](https://arxiv.org/html/2606.18518#bib.bib154)\]Sensor/IoT7651765155550555522VN Banking\[[49](https://arxiv.org/html/2606.18518#bib.bib164)\]Financial100,000100,00013131212252522
### IV\-BGenerative Models and Optimization Strategies

#### IV\-B1REaLTabFormer \(Transformer\-Based Generator\)

REaLTabFormer\[[60](https://arxiv.org/html/2606.18518#bib.bib105)\]models joint tabular distributions in an autoregressive manner and achieves high fidelity in capturing feature relationships\. Three optimization regimes are evaluated:

RTF: Standard maximum\-likelihood training for 50 epochs \(batch size 128, learning rate2×10−42\{\\times\}10^\{\-4\}\), providing a high\-utility reference without privacy constraints\.

RTF \+ ALM \(Sampling\-Time\): The trained baseline model is kept fixed; PSyGenTAB’s Augmented Lagrangian optimization is applied only atsampling timeby adjusting stochastic parameters \(temperature, top\-pp\) to enforce a minimum privacy thresholdPmin=0\.8P\_\{\\min\}=0\.8\.

#### IV\-B2CTAB\-GAN\+ \(GAN\-Based Generator\)

CTAB\-GAN\+\[[71](https://arxiv.org/html/2606.18518#bib.bib129)\]is a conditional GAN designed for mixed\-type tabular data using Wasserstein loss and gradient penalty\. Three optimisation regimes are evaluated:

CTAB\-GAN\+: Unconstrained training without differential privacy or adaptive optimization\.

CTAB\-GAN\+ \+ DP\-SGD: Differential privacy is enforced during discriminator training via the Opacus framework, using per\-sample gradient clipping and Gaussian noise injection, providing a formal\(ε,δ\)\(\\varepsilon,\\delta\)\-DP guarantee\.

CTAB\-GAN\+ \+ ALM: PSyGenTAB’s Augmented Lagrangian framework is applied during generation to enforce configurable privacy constraints while preserving structural fidelity\.

### IV\-CEvaluation Goals and Metrics

The evaluation is guided by three goals that directly reflect clinical deployment requirements:

Clinical Utility Preservation: Synthetic data must closely match the statistical properties, feature distributions, inter\-attribute relationships, and predictive structure of real health records\.

Patient Privacy Protection: Synthetic records must not replicate, reveal, or closely resemble real individuals from the training population, reducing re\-identification risk to a level compatible with responsible health data sharing\.

Controlled and Configurable Trade\-off: The framework must provide a systematic mechanism to adjust and balance privacy and utility according to application\-specific requirements, rather than treating their trade\-off as fixed or unavoidable\.

##### Composite Clinical Utility Score \(U∈\[0,1\]U\\in\[0,1\]\)

UUis a weighted composite capturing: \(i\)distributional fidelity, how closely synthetic feature distributions match real data; \(ii\)structural consistency, preservation of inter\-feature correlations and clinical dependencies; \(iii\)predictive utility, downstream classifier performance under the Train\-on\-Synthetic, Test\-on\-Real \(TSTR\) protocol; and \(iv\)minority\-class coverage, retention of rare clinical events and underrepresented diagnostic categories\. HigherUUindicates greater clinical usefulness\.

##### Composite Patient Privacy Score \(P∈\[0,1\]P\\in\[0,1\]\)

PPis a weighted composite capturing: \(i\)Distance to Closest Record \(DCR\), minimum Euclidean distance between each synthetic record and the nearest real patient, quantifying re\-identification proximity; \(ii\)Nearest Neighbor Distance Ratio \(NNDR\), assessing whether synthetic samples are disproportionately similar to specific real individuals; and \(iii\)Duplicate Detection, identifying exact or near\-exact reproduction of training records\. HigherPPindicates stronger patient privacy protection\.

### IV\-DPrimary Privacy—Utility Results

Diabetes Health Indicators\.The REaLTabFormer model achieves strong initial utility \(U=0\.89U=0\.89\) and moderate privacy \(P=0\.55P=0\.55\)\. After applying PSyGenTAB’s ALM optimization, utility increases substantially toU=0\.98U=0\.98while privacy is maintained atP=0\.54P=0\.54\. This simultaneous improvement in clinical fidelity, without any degradation in privacy, is a critical result: it demonstrates that constrained optimization can actuallyimprovehow faithfully synthetic records capture clinically meaningful patterns, including the minority prediabetes class, by steering the generator away from noise\-memorizing configurations\.

CTAB\-GAN\+ in its baseline configuration achieves the same high utility \(U=0\.98U=0\.98\) with stronger privacy \(P=0\.70P=0\.70\), but applying ALM unexpectedly reduces both utility and privacy \(U=0\.93U=0\.93,P=0\.33P=0\.33\)\. This degradation indicates that for large, imbalanced datasets, GAN training dynamics interact with the ALM penalty in ways that suppress structural learning, a finding with practical implications for practitioners selecting generative architectures for health record synthesis\.

Breast Cancer\.The RTF model produces moderate utility and weak privacy \(U=0\.82U=0\.82,P=0\.38P=0\.38\), indicating that the generator reproduces diagnostic feature distributions well but creates synthetic records that are uncomfortably close to real patients’ biometric profiles, a serious concern for clinical data sharing\. PSyGenTAB \(ALM\) improves both metrics \(U=0\.91U=0\.91,P=0\.44P=0\.44\), confirming that constrained optimization can simultaneously enhance diagnostic fidelity and reduce patient\-level re\-identification risk\.

CTAB\-GAN\+ with ALM, however, produces low utility \(U=0\.71U=0\.71\) and very weak privacy \(P=0\.24P=0\.24\), suggesting that the GAN\-ALM interaction is particularly challenging on small clinical datasets \(n=569n=569\)\. This finding underscores the importance of architecture selection for small\-sample clinical settings\.

A key finding across all experiments is that unconstrained baseline generative models face a fundamental tension between clinical usefulness and patient safety\. The REaLTabFormer achieves near\-perfect utility on the Adult dataset \(U=0\.99U=0\.99\) but exhibits critically weak privacy \(P=0\.31P=0\.31\), indicating substantial memorization of training records, a finding analogous to a model that could reconstruct a patient’s health record from its outputs\.

By embedding privacy as an explicit constraint via the Augmented Lagrangian Method, PSyGenTAB resolves this tension in a principled and configurable way:

On the Adult dataset, ALM increases patient privacy fromP=0\.31P=0\.31toP=0\.49P=0\.49while retaining high utility \(U=0\.96U=0\.96\)\.

On the PIR dataset, where baseline fidelity is already weak \(U=0\.83U=0\.83\), ALM simultaneously improves both utility \(U=0\.94U=0\.94\) and privacy \(P=0\.78P=0\.78\), as shown in TableLABEL:tab:qp\_comparison\. This demonstrates that ALM adapts to dataset characteristics rather than enforcing a rigid trade\-off, a property directly relevant to the heterogeneous data types encountered across clinical departments\.

On the Diabetes dataset, ALM achieves the highest observed utility \(U=0\.98U=0\.98\) across all configurations, confirming that the framework does not sacrifice clinical signal to satisfy privacy constraints\.

Critically, PSyGenTAB supports a configurable privacy thresholdPminP\_\{\\min\}that can be set by a healthcare data governance team according to regulatory requirements \(e\.g\., HIPAA Expert Determination, GDPR Article 89\), enabling transparent and auditable control of the privacy–utility trade\-off\.

TABLE II:Comparison of utility \(UU\) and privacy \(PP\) across datasets and methods\. Best performing values per metric in pairwise comparisons are highlighted inbold\.DatasetRTFRTF\-ALMCTAB\-GAN\+CTAB\-GAN\+\-ALMUUPPUUPPUUPPUUPPBreast Cancer0\.820\.820\.380\.380\.910\.440\.900\.350\.350\.840\.840\.56Diabetes0\.890\.890\.550\.980\.540\.540\.970\.970\.820\.980\.700\.70Heart Failure0\.850\.390\.840\.840\.330\.330\.940\.310\.310\.860\.860\.42Hypothyroid0\.990\.590\.590\.980\.980\.590\.960\.480\.480\.600\.600\.67Liver Disorders0\.820\.610\.610\.810\.810\.650\.910\.610\.610\.730\.730\.74Lung Cancer0\.770\.830\.760\.760\.830\.830\.920\.660\.660\.260\.260\.68Obesity0\.980\.520\.970\.970\.510\.510\.940\.540\.540\.930\.930\.56Parkinsons0\.630\.630\.400\.640\.380\.380\.360\.360\.700\.660\.70Adult0\.980\.300\.300\.960\.960\.480\.960\.960\.500\.500\.970\.77PIR Vision0\.830\.830\.860\.940\.780\.780\.920\.310\.310\.880\.880\.40VN Banking0\.700\.700\.500\.500\.810\.550\.820\.820\.970\.980\.570\.57
### IV\-EStatistical Fidelity

Statistical fidelity evaluates how closely the synthetic data reproduces the empirical feature distributions of the real dataset\. Lower divergence values indicate stronger alignment between real and synthetic marginal distributions\.

TABLE III:Statistical fidelity: Lower values indicate better match to the real data distribution\.DatasetMean KLMean JSMean Hell\.PSyGen\.RTFPSyGen\.RTFPSyGen\.RTFBreast Cancer0\.040\.040\.020\.010\.010\.000\.100\.100\.07Diabetes0\.230\.230\.090\.010\.010\.010\.040\.040\.04Heart Failure2\.882\.880\.420\.140\.140\.040\.300\.300\.16Hypothyroid11\.4411\.4511\.450\.340\.350\.350\.640\.640\.64Liver Disorders2\.252\.582\.580\.130\.130\.130\.350\.350\.34Lung Cancer0\.440\.440\.360\.020\.020\.010\.130\.130\.11Obesity0\.601\.071\.070\.080\.090\.090\.210\.230\.23Parkinsons3\.433\.823\.820\.160\.170\.170\.420\.440\.44Adult0\.350\.350\.140\.040\.040\.030\.160\.160\.09PIR Vision1\.532\.932\.930\.220\.290\.290\.410\.500\.50VN Banking1\.611\.611\.020\.180\.180\.120\.440\.440\.35Table[III](https://arxiv.org/html/2606.18518#S4.T3)reveals the inherent cost of mathematically enforced privacy on raw distributional overlap\. Because the RTF model optimizes purely for maximum likelihood without privacy constraints, it naturally achieves lower \(superior\) Mean KL divergence across the majority of datasets\. For example, on the Adult and Heart Failure datasets, RTF achieves KL scores of 0\.14 and 0\.42, respectively, compared to PSyGenTAB’s 0\.35 and 2\.88\. This visualizes the slight distributional separation introduced by the Augmented Lagrangian Method \(ALM\)\. PSyGenTAB deliberately trades a fraction of this exact distributional overlap to prevent the verbatim memorization of outlier records\. However, on highly sensitive metrics like the Jensen\-Shannon divergence, PSyGenTAB remains remarkably competitive, even outperforming RTF on the Diabetes dataset \(0\.0097 vs 0\.0099\)\. Notably, the results for the Hypothyroid dataset expose a critical limitation of both generative frameworks: massive KL divergences \(\>11\.4\>11\.4\) across both models indicate a complete generative failure, driven by the dataset’s extreme sparsity and deterministic medical rules\. This statistical anomaly directly foreshadows the Factuality and Alignment failures observed during the FAITH clinical validity assessment\.

We acknoledge thatCTAB\-GAN\+ was evaluated but underperformed compared to the Transformer baseline, which is why the deep\-dive privacy audits focus primarily on the RTF architecture\.

### IV\-FCommercial Evaluation Pipeline

To provide a holistic view of structural and semantic preservation, we rely on two aggregate scoring mechanisms:

SDMetrics Quality Score\[[62](https://arxiv.org/html/2606.18518#bib.bib148)\]:An open\-source evaluation suite that quantifies how well the synthetic data captures the mathematical properties, column shapes, and pairwise trends of the original dataset\.

MostlyAI Overall Accuracy\[[38](https://arxiv.org/html/2606.18518#bib.bib168)\]:A commercial\-grade evaluation metric that rigorously tests the fidelity of complex multivariate distributions\.

TABLE IV:Higher values indicate better synthetic data quality\. Note: MostlyAI cannot be calculated for Lung Cancer due to limited data points\.DatasetSDMetrics QualityMostlyAI AccuracyPSyGen\.RTFPSyGen\.RTFBreast Cancer0\.840\.840\.900\.770\.770\.84Diabetes0\.970\.960\.960\.980\.980\.99Heart Failure0\.830\.830\.900\.760\.760\.78Hypothyroid0\.050\.050\.050\.110\.110\.11Liver Disorders0\.890\.890\.890\.690\.690\.69Lung Cancer0\.840\.840\.86——Obesity0\.880\.860\.860\.870\.870\.88Parkinsons0\.820\.810\.810\.490\.490\.49Adult0\.870\.870\.940\.880\.880\.95PIR Vision0\.920\.920\.940\.850\.850\.89VN Banking0\.470\.470\.520\.240\.240\.28The aggregated quality scores in Table[IV](https://arxiv.org/html/2606.18518#S4.T4)underscore the differing evaluation philosophies between open\-source and commercial data assessment tools\. RTF systematically dominates the commercial MostlyAI Accuracy metric across nearly all benchmarks, scoring 0\.95 on the Adult dataset and 0\.99 on Diabetes, compared to PSyGenTAB’s 0\.88 and 0\.98\. This is expected: commercial tools heavily reward the exact replication of multivariate distributions, which favors unconstrained maximum\-likelihood generators like RTF despite their higher memorization risks\.

Conversely, the open\-source SDMetrics Quality Score reveals a much more nuanced clinical utility profile\. PSyGenTAB achieves superior quality scores on several complex medical datasets, including Diabetes \(0\.97 vs 0\.96\), Parkinson’s \(0\.82 vs 0\.81\), and Obesity \(0\.88 vs 0\.86\)\. This demonstrates that while the ALM privacy penalty restricts the raw overfitting rewarded by commercial metrics, it successfully preserves the underlying pairwise trends, column shapes, and semantic properties required for rigorous clinical analysis\.

### IV\-GDownstream Clinical ML Utility Evaluation

Statistical fidelity scores alone do not constitute sufficient evidence for clinical utility\. A synthetic dataset must support real\-world predictive modeling tasks, the primary purpose for which a healthcare institution would seek synthetic alternatives to protected data\. To evaluate this, we adopt theTrain on Synthetic, Test on Real\(TSTR\) andTrain on Real, Test on Synthetic\(TRTS\) protocols\. TSTR assesses whether models trained on synthetic data generalize to real patient populations; TRTS assesses whether synthetic data faithfully captures the discriminative structure of real health records\.

We evaluate three downstream classifiers, Decision Tree, Random Forest, and Logistic Regression, under four imbalance\-handling strategies \(None, Class Weighting, SMOTE, and Undersampling\), reflecting conditions routinely encountered in clinical prediction tasks\. Table[V](https://arxiv.org/html/2606.18518#S4.T5)presents the partial results\. The full results provided as supplementary file\.

For the Adult dataset, Random Forest trained on PSyGenTAB\-generated synthetic data achieves 83\.87 % accuracy under the standard TSTR protocol, closely matching the Train\-on\-Real, Test\-on\-Real \(TRTR\) reference of 85\.69 %\. The gap between TRTR and TSTR performance, the primary indicator of synthetic data quality for clinical AI, is minimal across all classifiers and imbalance strategies\.

Incorporating imbalance\-aware strategies on top of PSyGenTAB\-generated data consistently improves fairness\-sensitive metrics\. Undersampling yields a notable improvement in balanced accuracy for Random Forest, increasing it from 77\.67 % to 80\.60 % under TSTR, while maintaining competitive F1\-scores\. This confirms that synthetic records preserve minority\-class structure sufficiently for standard rebalancing techniques to remain effective\.

The Diabetes dataset provides the most clinically significant evidence, given its severe class imbalance, prediabetes: 1\.8 % of the population, and the importance of early identification of pre\-diabetic patients for preventive intervention\.

Under the unconstrainedNonestrategy, Logistic Regression trained on synthetic data achieves high raw accuracy \(≈84%\\approx\\\!84\\,\\%\) but a balanced accuracy near chance level \(≈40%\\approx\\\!40\\,\\%\), driven entirely by majority\-class dominance\. This mimics the failure mode of real clinical classifiers trained on imbalanced data\.

When Class Weighting or Undersampling is applied to PSyGenTAB\-generated data, balanced accuracy improves substantially, approaching 49 % in several configurations\. This improvement is consistent across both TSTR and TRTS protocols, providing strong evidence that PSyGenTAB preserves the minority clinical signal \(prediabetes indicators\) that is most critical for health AI applications\.

Importantly, the relative ranking of imbalance strategies remains stable between TSTR and TRTS, confirming that the synthetic data does not introduce spurious correlations that would invert downstream model behavior, a critical quality criterion for clinical deployment\.

Tree\-based models including Decision Tree and Random Forest exhibit stronger robustness to distributional shifts between real and synthetic data, while linear models such as Logistic Regression benefit most from explicit class reweighting\. The ability of standard imbalance\-handling techniques to recover balanced clinical performance, particularly in the challenging diabetes setting, provides strong evidence that PSyGenTAB\-generated data is operationally useful for downstream health AI tasks, not merely statistically similar to real data\.

TABLE V:Summary of Downstream Clinical AI Utility \(Random Forest\)\. Values represent the best performance achieved across data imbalance strategies per dataset\. TRTR \(Train Real, Test Real\) serves as the baseline upper bound\. The synthetic frameworks are evaluated using TSTR \(Train Synthetic, Test Real\) to demonstrate structural preservation\.DatasetTRTR \(Baseline\)TSTR \(PSyGen\.\)TSTR \(RTF\)Bal\. AccF1Bal\. AccF1Bal\. AccF1Breast Cancer0\.960\.960\.960\.960\.960\.960\.960\.96Diabetes0\.400\.400\.810\.810\.410\.800\.800\.400\.400\.81Heart Failure0\.840\.840\.860\.860\.650\.710\.590\.590\.580\.58Hypothyroid0\.790\.790\.960\.960\.520\.890\.500\.500\.890\.89Liver Disorders0\.760\.760\.770\.770\.550\.550\.540\.540\.560\.57Lung Cancer0\.640\.640\.580\.580\.810\.800\.530\.530\.470\.47Obesity0\.940\.940\.950\.950\.720\.720\.710\.710\.770\.77Parkinsons0\.890\.890\.930\.930\.490\.490\.630\.630\.500\.64Adult0\.780\.780\.850\.850\.780\.780\.840\.840\.790\.85PIR Vision0\.980\.980\.990\.990\.990\.990\.990\.990\.990\.99VN Banking0\.720\.720\.750\.750\.480\.480\.510\.510\.600\.64
### IV\-HPatient Re\-identification Risk: Distance to Closest Record

Beyond composite privacy scores, clinical data governance requires direct quantification of re\-identification risk, the probability that a synthetic record could be linked back to a specific patient in the training dataset\. We evaluate this using theDistance to Closest Record\(DCR\), a standard privacy diagnostic in the synthetic health data literature\[[22](https://arxiv.org/html/2606.18518#bib.bib233),[59](https://arxiv.org/html/2606.18518#bib.bib232),[68](https://arxiv.org/html/2606.18518#bib.bib235)\]\.

Formally, let𝒟real=\{x1,…,xN\}\\mathcal\{D\}\_\{\\text\{real\}\}=\\\{x\_\{1\},\\dots,x\_\{N\}\\\}denote the real patient dataset and𝒟syn=\{x~1,…,x~M\}\\mathcal\{D\}\_\{\\text\{syn\}\}=\\\{\\tilde\{x\}\_\{1\},\\dots,\\tilde\{x\}\_\{M\}\\\}the synthetic dataset\. The DCR for a synthetic recordx~i\\tilde\{x\}\_\{i\}is:

DCR\(x~i\)=minxj∈𝒟real∥x~i−xj∥2\\footnotesize\\mathrm\{DCR\}\(\\tilde\{x\}\_\{i\}\)=\\min\_\{x\_\{j\}\\in\\mathcal\{D\}\_\{\\text\{real\}\}\}\\lVert\\tilde\{x\}\_\{i\}\-x\_\{j\}\\rVert\_\{2\}\(6\)where∥⋅∥2\\lVert\\cdot\\rVert\_\{2\}is the Euclidean distance over a standardised feature space\. A DCR value near zero indicates that a synthetic record closely resembles, or exactly matches, a real patient’s profile\.

We report three complementary statistics:

Exact Match Ratio \(EMR\): Proportion of synthetic records withDCR=0\\mathrm\{DCR\}=0, representing verbatim reproduction of training records, the most severe privacy failure in health data release\.

EMR=1\|𝒟syn\|​∑i=1\|𝒟syn\|𝕀​\[DCR​\(x~i\)=0\]\\footnotesize\\mathrm\{EMR\}=\\frac\{1\}\{\|\\mathcal\{D\}\_\{\\text\{syn\}\}\|\}\\sum\_\{i=1\}^\{\|\\mathcal\{D\}\_\{\\text\{syn\}\}\|\}\\mathbb\{I\}\\\!\\left\[\\mathrm\{DCR\}\(\\tilde\{x\}\_\{i\}\)=0\\right\]\(7\)Mean and Median DCR: Population\-level proximity to real patient records\.5th\-Percentile DCR \(DCR5%\\mathrm\{DCR\}\_\{5\\%\}\): Worst\-case re\-identification risk, reflecting the 5 % of synthetic records most proximate to real patients, used in clinical privacy audits\[[59](https://arxiv.org/html/2606.18518#bib.bib232)\]\.

DCR5%=Quantile0\.05​\(\{DCR​\(x~i\)\}i=1\|𝒟syn\|\)\\footnotesize\\mathrm\{DCR\}\_\{5\\%\}=\\mathrm\{Quantile\}\_\{0\.05\}\\bigl\(\\\{\\mathrm\{DCR\}\(\\tilde\{x\}\_\{i\}\)\\\}\_\{i=1\}^\{\|\\mathcal\{D\}\_\{\\text\{syn\}\}\|\}\\bigr\)\(8\)
Table[VI](https://arxiv.org/html/2606.18518#S4.T6)summarizes these re\-identification risk metrics\.

##### Adult Census Income

PSyGenTAB \(ALM\) reduces exact matches to just 3 records, compared to 8 exact matches in the unconstrained RTF\. This reduction in exact reproductions demonstrates that PSyGenTAB’s privacy regularization effectively suppresses verbatim memorization of training individuals\. Although mean and median DCR values are slightly lower than the baseline, reflecting PSyGenTAB’s higher statistical fidelity, the sharp reduction in exact matches demonstrates meaningfully improved protection against the most clinically serious privacy violation\.

##### Diabetes Health Indicators

The presence of discrete, low\-cardinality clinical variables in the diabetes dataset leads to a higher baseline rate of exact matches for both models, a known challenge in clinical tabular data synthesis\[[21](https://arxiv.org/html/2606.18518#bib.bib234)\]\. Despite this inherent difficulty, PSyGenTAB maintains a re\-identification risk profile comparable to the unconstrained baseline while simultaneously achieving higher downstream clinical utility \(Section[IV\-G](https://arxiv.org/html/2606.18518#S4.SS7)\)\. The similarity inDCR5%\\mathrm\{DCR\}\_\{5\\%\}across models confirms that PSyGenTAB does not exacerbate worst\-case re\-identification risk, even in highly constrained clinical feature spaces\.

### IV\-IMembership Inference Attack Resilience

To rigorously assess whether a determined adversary could exploit PSyGenTAB’s synthetic health data to identify individuals in the training population, we conducted a Membership Inference Attack \(MIA\) evaluation\. This analysis simulates a realistic healthcare data breach scenario: an attacker, possessing a candidate patient record, attempts to determine whether that individual’s data was used to train the generative model\. Such attacks represent a well\-established threat model in healthcare AI\[[58](https://arxiv.org/html/2606.18518#bib.bib220),[7](https://arxiv.org/html/2606.18518#bib.bib206)\]and are increasingly considered in clinical data governance frameworks\.

The following attack models of increasing sophistication are deployed:

Black\-Box Attack: The attacker has access only to the model’s prediction confidence scores, the most realistic threat model for a deployed clinical AI system\.

White\-Box Attack: The attacker additionally has access to internal model gradients or loss values, a stronger threat representing insider knowledge\.

Logistic Regression Attack: A linear classifier trained to distinguish members from non\-members based on model outputs, a baseline worst\-case linear separability assessment\.

Linkage Attack Success:Evaluates the risk of re\-identifying individuals by matching synthetic records with external, auxiliary datasets\. A lower linkage success rate indicates that the synthetic records are sufficiently perturbed to prevent direct mapping to real\-world identities\.\[[61](https://arxiv.org/html/2606.18518#bib.bib62)\]

Vulnerability is quantified using the Attack AUC, whereAUC=0\.50\\mathrm\{AUC\}=0\.50represents perfect privacy \(the attacker performs no better than random guessing\) andAUC=1\.0\\mathrm\{AUC\}=1\.0indicates complete patient\-identity leakage\. AnOverall Distinguishability Scoreis reported as the mean AUC across all three attacks\.

Table[VI](https://arxiv.org/html/2606.18518#S4.T6)presents the comparative results\.

##### Diabetes Health Indicators

Both models demonstrate exceptional resilience to patient identity disclosure\. The unconstrained RTF baseline achieves an Overall Distinguishability Score of 0\.51, statistically equivalent to random guessing, confirming that the diabetes health data is inherently difficult to attribute to individuals\. PSyGenTAB \(ALM\) maintains a similarly robust profile \(0\.56\), confirming that the privacy constraints do not weaken, and may marginally strengthen, resistance to membership inference on clinical health data\.

##### Adult Census Income

The Adult dataset reveals a more nuanced finding\. The PSyGenTAB \(ALM\) Black\-Box AUC \(0\.74\) is higher than the baseline \(0\.59\), suggesting that ALM’s aggressive optimization for high statistical fidelity may leave subtle distributional patterns that a sophisticated Black\-Box attacker can exploit\. This finding is consistent with the known tension between utility and membership inference resistance\[[32](https://arxiv.org/html/2606.18518#bib.bib212)\]: higher\-fidelity synthetic data captures population\-level distributions more precisely, which can inadvertently preserve signals attributable to individual records\.

Importantly, the Logistic Regression attack remains ineffective against PSyGenTAB \(AUC≈0\.54\\mathrm\{AUC\}\\approx 0\.54\), confirming that linear separability between training members and non\-members remains low\. This result suggests that the observed Black\-Box vulnerability requires a sophisticated non\-linear attacker and does not represent a trivially exploitable weakness\.

TABLE VI:Comprehensive privacy risk evaluation across eleven datasets\. MIA AUC metrics are reported for three attack models \(Black\-Box, Wight\-Box, Logistic Regression\) and aggregated as an overall distributional score\. Attack success rates \(MIA, linkage\) are lower\-is\-better, while overall privacy and DCR statistics provide complementary privacy signals\.DatasetB\-B AUCW\-B AUCLR AUCOver\. Dist\.MIALinkageOver\. Priv\.DCR MeanDCR Med\.DCR 5%DCR 95%RTFPSyGen\.RTFPSyGen\.RTFPSyGen\.RTFPSyGen\.RTFPSyGen\.RTFPSyGen\.RTFPSyGen\.RTFPSyGen\.RTFPSyGen\.RTFPSyGen\.RTFPSyGen\.Breast Cancer1\.001\.001\.001\.000\.580\.960\.960\.690\.690\.690\.690\.760\.890\.890\.300\.300\.300\.301\.001\.001\.001\.000\.340\.340\.360\.660\.661\.160\.320\.321\.200\.000\.000\.000\.002\.132\.132\.54Diabetes0\.520\.600\.600\.510\.540\.540\.500\.530\.530\.510\.560\.560\.500\.500\.501\.001\.001\.001\.000\.410\.410\.420\.820\.750\.750\.580\.520\.520\.000\.000\.000\.002\.422\.302\.30Heart Failure0\.590\.590\.580\.530\.530\.530\.510\.510\.510\.540\.540\.540\.510\.510\.500\.870\.870\.700\.470\.470\.521\.991\.761\.761\.951\.721\.721\.030\.970\.973\.252\.582\.58Hypothyroid0\.720\.720\.720\.660\.660\.660\.630\.630\.630\.670\.670\.670\.480\.490\.490\.990\.990\.950\.580\.580\.590\.360\.360\.370\.280\.280\.290\.000\.000\.000\.970\.930\.93Liver Disorders0\.600\.600\.590\.550\.550\.540\.530\.530\.520\.560\.560\.550\.500\.510\.510\.880\.880\.770\.480\.480\.511\.020\.900\.900\.850\.780\.780\.410\.410\.431\.971\.721\.72Lung Cancer0\.610\.620\.620\.580\.590\.590\.560\.570\.570\.580\.600\.600\.390\.510\.510\.720\.720\.720\.720\.490\.490\.547\.477\.478\.017\.617\.617\.815\.555\.556\.819\.269\.269\.80Obesity0\.670\.670\.670\.670\.630\.640\.640\.610\.620\.620\.640\.640\.640\.500\.500\.500\.760\.890\.890\.480\.460\.461\.031\.031\.221\.001\.001\.190\.350\.350\.551\.791\.792\.01Parkinsons0\.630\.630\.610\.610\.610\.600\.600\.600\.580\.620\.620\.600\.440\.440\.430\.790\.800\.800\.450\.430\.433\.643\.644\.143\.533\.534\.022\.532\.533\.014\.904\.905\.84Adult0\.590\.740\.740\.520\.550\.550\.520\.540\.540\.540\.610\.610\.500\.500\.491\.041\.040\.860\.410\.410\.460\.130\.070\.070\.070\.030\.030\.000\.000\.000\.460\.280\.28PIR Vision0\.690\.720\.720\.530\.570\.570\.560\.560\.550\.600\.610\.610\.300\.300\.300\.300\.931\.021\.020\.360\.330\.331\.621\.622\.761\.081\.081\.190\.900\.900\.982\.892\.899\.81VN Banking1\.001\.001\.001\.001\.001\.001\.001\.001\.001\.001\.001\.001\.001\.001\.001\.000\.460\.460\.460\.680\.780\.780\.640\.570\.572\.031\.981\.982\.032\.032\.031\.371\.361\.362\.722\.522\.52

Beyond membership inference, the Linkage Attack Success metric in Table[VI](https://arxiv.org/html/2606.18518#S4.T6)provides compelling evidence of PSyGenTAB’s structural privacy protections\. A successful linkage attack occurs when an adversary maps a synthetic record back to a specific individual using external auxiliary data\. Across the evaluated benchmarks, PSyGenTAB’s ALM optimization systematically throttles linkage vulnerability compared to the unconstrained RTF baseline\. On the Adult dataset, linkage success drops from 1\.04 to 0\.86, and on Heart Failure, it falls from 0\.87 to 0\.70\. By gently perturbing the precise joint relationships of vulnerable continuous features, PSyGenTAB actively disrupts the 1\-to\-1 deterministic mapping required for successful linkage, ensuring that the synthetic records remain statistically representative without acting as proxy identities for the real training cohort\.

##### Feature\-Level Re\-identification Vulnerability

Table[VII](https://arxiv.org/html/2606.18518#S4.T7)identifies which clinical attributes contribute most to membership inference risk\. We extract feature importance scores from the attack models \(Table[VII](https://arxiv.org/html/2606.18518#S4.T7)\)\. Theimportance score\(bounded within the range of\[0\.0,1\.0\]\[0\.0,1\.0\], where the sum of all features equals1\.01\.0\) quantifies the relative predictive power a specific attribute provides to the adversarial model\. For instance, an importance score of0\.200\.20implies that 20% of the attacker’s ability to successfully distinguish a training record from a holdout record relies solely on that single feature\. Across the 11 evaluated datasets, continuous, high\-variance variables consistently emerge as the primary leakage channels\. In the Adult dataset, demographic and census attributes likeAgeandfnlwgtaccount for roughly 17% to 19% of the leakage contribution\. However, the risk is magnified in highly specialized domains\. In clinical datasets, precise physiological measurements dominate the adversarial risk profile:BMIcontributes∼\\sim16% of the attack signal in the Diabetes dataset,plateletsaccounts for∼\\sim17% in Heart Failure, andgammagt\(a liver enzyme\) drives over 20% of the leakage in Liver Disorders\. Most notably, in the Parkinson’s dataset, highly specific acoustic measurements \(MDVP:Jitter\(Abs\)\) account for an alarming 42% to 44% of the re\-identification risk\. Similarly, in financial domains such as VN Banking, exact monetary values \(min\_term\_deposit\_balance\_vnd\) constitute the highest vulnerability \(∼\\sim17%\)\.

These feature\-level findings have direct clinical implications: continuous health metrics with high inter\-individual variance \(BMI, blood glucose, age\) represent the highest re\-identification vectors in synthetic health records\. Future iterations of PSyGenTAB should apply stricter per\-feature privacy penalties to such attributes during ALM optimization, an actionable direction for clinical deployment\.

These findings indicate that heavy\-tailed continuous features, exact monetary balances, and highly specific clinical biomarkers are particularly susceptible to memorization and adversarial exploitation\. To mitigate this feature\-level leakage, several practical solutions can be integrated into future generative pipelines:

Column\-Specific ALM Regularization:Instead of a global privacy penalty, the ALM objective can be adapted to apply dynamic, feature\-specific weights \(λ\\lambda\) that heavily penalize the exact memorization of high\-variance marginal distributions\.

Bounded Clipping and Discretization:Applying strict upper and lower bounds, clipping heavy tails or binning continuous variables\[[30](https://arxiv.org/html/2606.18518#bib.bib241)\], into micro\-categories prior to training removes the extreme outliers that attackers rely on for linkage\.

Targeted Differential Privacy \(DP\)\[[20](https://arxiv.org/html/2606.18518#bib.bib242),[18](https://arxiv.org/html/2606.18518#bib.bib243)\]:Injecting calibrated Laplacian or Gaussian noise exclusively into the generation phase of the top\-k most vulnerable columns can effectively obfuscate precise values \(e\.g\., hiding exact platelet counts\) while safely preserving the broader joint relationships across the dataset\.

TABLE VII:Dominant feature contributing to membership inference risk\. Higher importance values are automatically bolded per dataset\.Δ\\Deltadenotes PSyGenTAB−\-RTF difference\. Entropy is computed on the real dataset to assess inherent leakage susceptibility\. Importance stability was verified across 5 bootstrap runs \(std\.<<0\.01 for all datasets\)\.DatasetTop FeaturePSyGen\.RTF𝚫\\boldsymbol\{\\Delta\}EntropyBreast CancerCT0\.130\.130\.14−\-0\.011\.121\.12DiabetesBMI0\.170\.140\.14\+0\.031\.891\.89Heart Failureplatelets/creatinine0\.180\.160\.16\+0\.011\.581\.58HypothyroidT30\.300\.300\.30−\-0\.000\.910\.91Liver Disorderssgpt0\.220\.200\.20\+0\.011\.741\.74Lung Cancerfeature\_3/feature\_260\.130\.060\.06\+0\.071\.351\.35ObesityWeight0\.150\.150\.16−\-0\.012\.062\.06ParkinsonsMDVP:Jitter\(Abs\)0\.440\.420\.42\+0\.020\.780\.78AdultAge0\.190\.180\.18\+0\.012\.412\.41PIR VisionPIR\_430\.040\.030\.03\+0\.013\.023\.02VN Bankingavg\_transaction\_amount0\.250\.250\.25−\-0\.000\.630\.63

### IV\-JClinical Validity Assessment: FAITH Framework

Conventional fidelity metrics assess how statistically similar synthetic records are to real data but do not distinguish between three clinically distinct failure modes: \(i\) generatingstatistically plausible but clinically impossiblepatient profiles \(e\.g\., a male patient flagged as currently pregnant\); \(ii\)inventing spurious clinical correlationsnot present in the real population, hallucinated comorbidity patterns; or \(iii\) generating values outside thevalid physiological rangeof the dataset, ghost values such as a negative BMI or an age of 200 years\.

To assess these failure modes systematically, we evaluated the models using the FAITH clinical validity framework \(Factuality,Alignment,Integrity, andTracking ofHallucinations\), which decomposes synthetic data reliability into four clinically interpretable and mathematically orthogonal dimensions\. The composite FAITH score is:

SFAITH=wF⋅SFact\+wA⋅SAlign\+wI⋅SInteg\+wT⋅STrackS\_\{\\mathrm\{FAITH\}\}=w\_\{F\}\\cdot S\_\{\\mathrm\{Fact\}\}\+w\_\{A\}\\cdot S\_\{\\mathrm\{Align\}\}\+w\_\{I\}\\cdot S\_\{\\mathrm\{Integ\}\}\+w\_\{T\}\\cdot S\_\{\\mathrm\{Track\}\}\(9\)with equal weights \(w=0\.25w=0\.25\) to balance trade\-offs between clinical validity and structural diversity\.

#### IV\-J1Factuality \(SFactS\_\{\\mathrm\{Fact\}\}\): Clinically Impossible Patient Profiles

Factuality penalizes the generation of logically or clinically impossible records\. A patient cannot be simultaneously male and currently pregnant; age constraints must be consistent within a record\. Formally:

SFact=1N​∑i=1N𝟙​\(⋀c∈𝒞checkc​\(rowi\)\)\\footnotesize S\_\{\\mathrm\{Fact\}\}=\\frac\{1\}\{N\}\\sum\_\{i=1\}^\{N\}\\mathbb\{1\}\\\!\\Bigl\(\\bigwedge\_\{c\\in\\mathcal\{C\}\}\\mathrm\{check\}\_\{c\}\(\\mathrm\{row\}\_\{i\}\)\\Bigr\)\(10\)where𝒞\\mathcal\{C\}is the set of domain\-specific logical constraints and𝟙​\(⋅\)\\mathbb\{1\}\(\\cdot\)returns 1 only if all constraints are satisfied\. A high factuality score confirms the generator does not produce clinically nonsensical records that would be immediately obvious to a clinician, a critical criterion for safe deployment\.

#### IV\-J2Alignment \(SAlignS\_\{\\mathrm\{Align\}\}\): Hallucinated Clinical Correlations

Alignment detects a subtle form of hallucination: the invention of statistical dependencies that do not exist in the real patient population \(e\.g\., synthesising a spurious correlation between pet ownership and high cardiovascular risk\)\. It measures the preservation of the dataset’s true correlation structure:

SAlign=1−1d​‖Corr​\(R\)−Corr​\(S\)‖F\\footnotesize S\_\{\\mathrm\{Align\}\}=1\-\\frac\{1\}\{d\}\\bigl\\\|\\mathrm\{Corr\}\(R\)\-\\mathrm\{Corr\}\(S\)\\bigr\\\|\_\{F\}\(11\)whereRRandSSare the real and synthetic datasets respectively,Corr​\(⋅\)\\mathrm\{Corr\}\(\\cdot\)is the correlation matrix, andddis a normalisation factor\. High alignment confirms the model has learned the true joint distribution of clinical features rather than generating independent noise, essential for downstream use in comorbidity analysis and risk stratification\.

#### IV\-J3Integrity \(SIntegS\_\{\\mathrm\{Integ\}\}\): Patient Record Reproduction

Integrity distinguishes generalisation from memorisation\. A model that achieves high factuality by reproducing training records has not learned the underlying clinical distribution; it has copied patient data\. Integrity measures the proportion of synthetic records that are genuinely novel, sufficiently distant from all real patient records:

SInteg=1N​∑i=1N𝟙​\(minj⁡dist​\(si,rj\)\>ε\)\\footnotesize S\_\{\\mathrm\{Integ\}\}=\\frac\{1\}\{N\}\\sum\_\{i=1\}^\{N\}\\mathbb\{1\}\\\!\\Bigl\(\\min\_\{j\}\\mathrm\{dist\}\(s\_\{i\},r\_\{j\}\)\>\\varepsilon\\Bigr\)\(12\)whereε\\varepsilonis a clinically motivated privacy radius\. This score directly complements the DCR analysis of Section[IV\-H](https://arxiv.org/html/2606.18518#S4.SS8), providing a binary novelty indicator\.

#### IV\-J4Tracking \(STrackS\_\{\\mathrm\{Track\}\}\): Out\-of\-Distribution Clinical Values

Tracking penalises generation of values outside the observable clinical manifold of the real dataset, ghost values that are physically or physiologically impossible \(e\.g\., negative heart rate, blood glucose of 1,000 mg/dL\):

STrack=1−∑i=1N𝟙​\(si∉Support​\(R\)\)N\\footnotesize S\_\{\\mathrm\{Track\}\}=1\-\\frac\{\\sum\_\{i=1\}^\{N\}\\mathbb\{1\}\(s\_\{i\}\\notin\\mathrm\{Support\}\(R\)\)\}\{N\}\(13\)High tracking scores confirm the generator respects the distributional support of the real clinical dataset, a prerequisite for safe use in clinical decision\-support applications\.

While PSyGenTAB excels in structural integrity and factuality, the baseline RTF occasionally exhibits marginal superiority inAlignment\(e\.g\., in theAdult,Diabetes, andParkinsonsdatasets\)\. This is a well\-documented and expected artifact of the privacy\-utility trade\-off\. Because RTF optimizes purely for maximum likelihood without privacy constraints, it aggressively overfits to the training data’s exact covariance matrix\. While this results in high Alignment scores, it comes at the direct cost of the aforementioned memorization risks\. PSyGenTAB explicitly trades a negligible fraction of this overfitted correlation \(usually<1%<1\\%drop in Alignment\) to guarantee privacy, representing a much safer and highly controllable compromise for clinical deployments\.

Finally, the results on theHypothyroiddataset highlight a shared limitation of current generative architectures\. Both PSyGenTAB and RTF completely fail to capture theFactualityandAlignmentof this specific data topology due to the limited data points\.

In summary, the FAITH framework evaluation demonstrates that PSyGenTAB produces highly reliable synthetic records\. By mathematically enforcing privacy during the training cycle, the framework prevents catastrophic memorization while maintaining the clinical factuality and structural alignment required for downstream medical and financial research\.

Table[VIII](https://arxiv.org/html/2606.18518#S4.T8)presents the FAITH evaluation results\.

TABLE VIII:FAITH clinical validity framework: decomposed assessment of synthetic health data reliability\. Factuality measures absence of clinically impossible profiles; Alignment measures preservation of real clinical correlations; Integrity measures absence of patient record reproduction; Tracking measures absence of out\-of\-distribution physiological values\. Higher values indicate better performance;bolddenotes the stronger result per dataset and metric\.DatasetFactualityAlign\.IntegrityTrack\.FAITHPSyGen\.RTFPSyGen\.RTFPSyGen\.RTFPSyGen\.RTFPSyGen\.RTFBreast Cancer0\.990\.991\.000\.960\.960\.970\.790\.540\.541\.001\.001\.000\.930\.880\.88Diabetes1\.001\.001\.000\.990\.990\.990\.960\.950\.951\.001\.001\.000\.990\.990\.99Heart Failure1\.001\.000\.940\.940\.941\.001\.000\.990\.990\.990\.980\.980\.98Hypothyroid0\.000\.000\.000\.001\.001\.000\.090\.090\.270\.27Liver Disorders1\.001\.001\.000\.900\.890\.891\.001\.001\.001\.001\.000\.970\.970\.97Lung Cancer1\.001\.000\.860\.860\.861\.001\.001\.001\.000\.960\.960\.96Obesity0\.940\.940\.980\.970\.960\.961\.001\.000\.990\.991\.000\.970\.970\.99Parkinsons0\.830\.830\.950\.720\.720\.731\.001\.000\.950\.950\.950\.870\.870\.91Adult0\.980\.920\.920\.980\.980\.991\.001\.001\.001\.001\.001\.000\.990\.980\.98PIR Vision1\.001\.001\.000\.910\.900\.901\.001\.001\.001\.001\.000\.980\.970\.97VN Banking1\.000\.990\.990\.870\.870\.871\.001\.000\.960\.960\.960\.960\.950\.95

### IV\-KSummary of Evaluation Findings

Taken together, the results establish three overarching conclusions for clinical data sharing applications:

Privacy and clinical utility are jointly achievable\.Unconstrained generative models face a fundamental tension between statistical fidelity and patient privacy\. PSyGenTAB resolves this tension by embedding configurable privacy constraints directly into the generation process, achieving strong utility \(U≥0\.91U\\geq 0\.91on the Breast Cancer dataset;U=0\.98U=0\.98on Diabetes\) while substantially reducing re\-identification risk\.

PSyGenTAB\-generated data is operationally useful for clinical AI\.Downstream classifiers trained on synthetic health records closely match performance on real data, and standard imbalance\-handling strategies, including SMOTE and Class Weighting, remain effective on synthetic data, including for the highly imbalanced Diabetes dataset where minority\-class identification is clinically critical\.

Architecture and dataset characteristics matter\.REaLTabFormer combined with ALM consistently outperforms CTAB\-GAN\+ with ALM on small clinical datasets such as Breast Cancer, while CTAB\-GAN\+ achieves stronger baseline privacy on larger datasets\. Practitioners should select architectures appropriate to their clinical dataset’s size, feature type distribution, and class\-imbalance profile\.

## VDiscussion

### V\-AClinical Implications of the Privacy–Utility Trade\-off

A central contribution of this work is the empirical demonstration that privacy and clinical utility in synthetic health data are not inherently antagonistic\. In current practice, healthcare institutions often assume that stronger privacy protection necessarily reduces analytic value, leading to unsatisfactory compromises: either datasets are shared with weak safeguards or withheld entirely, foreclosing scientific collaboration\. PSyGenTAB challenges this assumption by showing that privacy\-constrained optimization can simultaneously improve clinical fidelity while reducing re\-identification risk\.

On the Diabetes Health Indicators dataset, clinical utility improved fromU:0\.89→0\.98U:0\.89\\rightarrow 0\.98while maintainingP=0\.54P=0\.54\. On the Breast Cancer dataset, utility increased fromU:0\.82→0\.91U:0\.82\\rightarrow 0\.91alongside a privacy improvement fromP:0\.38→0\.44P:0\.38\\rightarrow 0\.44\. These results indicate that privacy constraints, when embedded directly into the learning objective, act not merely as regularizers but as structural guides that steer the generator toward statistically faithful yet non\-memorizing representations of the underlying clinical population\.

This finding carries direct implications for healthcare data governance\. Under HIPAA’s Expert Determination pathway, a qualified expert must certify that the risk of re\-identification is very small\. PSyGenTAB’s configurable privacy thresholdPminP\_\{\\min\}, together with documented reductions in exact match ratios and membership inference success, provides auditable quantitative evidence to support such determinations\. Likewise, under GDPR Article 89, synthetic data is recognized as a pseudonymization technique for research; PSyGenTAB’s explicit constraint formulation offers the principled, transparent mechanism required for regulatory justification\.

### V\-BSuitability for Cross\-Institutional Health Data Sharing

The inability to share patient\-level data across institutional boundaries remains a major bottleneck for clinical AI development\. While federated learning mitigates direct data transfer, it requires architectural coordination and remains vulnerable to gradient\-based inference attacks\[[73](https://arxiv.org/html/2606.18518#bib.bib227)\]\. Synthetic data generation offers a complementary alternative: a single institution can release a privacy\-preserving surrogate dataset that external partners may analyze freely without accessing real patient records\.

PSyGenTAB is particularly well\-suited to this use case because its privacy constraint is configurable and enforceable at training time\. An IRB\-approved protocol specifying, for example, a minimum DCR threshold can be encoded directly into the optimization objective, ensuring automatic compliance during generation\. This proactive, constraint\-driven enforcement contrasts with post\-hoc anonymization methods that rely on retrospective auditing\.

Crucially, downstream evaluation confirms that these synthetic datasets remain scientifically meaningful\. Models trained exclusively on PSyGenTAB\-generated data achieve performance comparable to models trained on real data, demonstrating that shared datasets are not merely privacy\-safe but analytically valid\.

### V\-CImplications for Minority\-Class Clinical Modeling

The Diabetes experiments highlight a clinically significant finding: PSyGenTAB preserves predictive signal for the prediabetes class \(n=4,631n=4\{,\}631, 1\.8% prevalence\), with balanced accuracy improvements maintained under synthetic training with imbalance\-handling strategies\. This result is important because prediabetes is both clinically actionable and chronically underrepresented in population datasets\.

The faithful preservation of rare phenotypes is essential for equitable health AI\. Synthetic augmentation methods that fail to preserve minority\-class structure risk amplifying existing clinical biases\. PSyGenTAB’s diversity and joint\-fidelity components within the utility composite, together with its privacy\-constrained optimization, provide a principled mechanism to maintain representation of clinically rare conditions without resorting to memorization\.

### V\-DArchitectural Considerations for Clinical Deployment

Architectural choice meaningfully influences privacy–utility outcomes and should be considered explicitly in deployment\.

Transformer\-based generators \(REaLTabFormer \+ ALM\) consistently achieve the strongest privacy–utility balance across clinical datasets\. Their autoregressive structure effectively captures correlated clinical biomarkers, while ALM\-based constraint enforcement prevents memorization of individual patient profiles\.

GAN\-based generators \(CTAB\-GAN\+ \+ ALM\) demonstrate strong baseline privacy but exhibit sensitivity to constraint interaction on smaller datasets \(e\.g\., Breast Cancer,n=569n=569\), where adversarial dynamics may destabilize optimization\. For small clinical cohorts, transformer\-based architectures appear more robust under constrained training\.

Formal differential privacy approaches \(e\.g\., DP\-SGD\) provide the strongest theoretical guarantees but consistently reduce structural fidelity and downstream model performance\. Where formal\(ε,δ\)\(\\varepsilon,\\delta\)certification is legally required, DP mechanisms remain necessary\. In settings where empirical privacy assurance suffices, PSyGenTAB offers a more favorable clinical utility trade\-off\.

## VILimitations

Several limitations bound the scope of the present work and should be considered prior to clinical deployment\.

Scope of data modalities\.PSyGenTAB is validated exclusively on structured tabular health data, the dominant format of EHR exports, clinical registries, and population health surveys\. The framework has not been evaluated on unstructured modalities such as medical imaging \(e\.g\., DICOM\), free\-text clinical notes, or multi\-modal records combining structured and unstructured components\. Extending constraint\-based privacy–utility optimization to these data types remains an open research challenge\.

Empirical privacy guarantees\.Privacy assurances are empirical rather than formally differentially private\. Protection is quantified through DCR, NNDR, membership inference resilience, and duplication analysis, but no\(ε,δ\)\(\\varepsilon,\\delta\)\-DP guarantee is provided\. Regulatory contexts requiring formal certification may therefore necessitate supplementary differential privacy mechanisms\.

Non\-convex optimization\.The ALM objective is inherently non\-convex due to deep generative models and non\-linear utility and privacy metrics\. Although ALM is well\-established for non\-convex constrained problems and converges to KKT stationary points under mild regularity conditions\[[3](https://arxiv.org/html/2606.18518#bib.bib224),[56](https://arxiv.org/html/2606.18518#bib.bib225)\], global optimality cannot be guaranteed\. Convergence is assessed empirically via stabilization of the constraint residualPmin−P​\(θ\)P\_\{\\min\}\-P\(\\theta\)and the utility scoreU​\(θ\)U\(\\theta\), consistent with prior ML applications of ALM\[[69](https://arxiv.org/html/2606.18518#bib.bib226)\]\.

## VIIConclusion

Healthcare institutions face a persistent tension: the data required to develop high\-impact clinical AI systems is often restricted by privacy regulation and institutional policy\. This fragmentation slows multi\-center collaboration, limits rare disease research, and constrains the responsible development of data\-driven healthcare tools\.

We introduced PSyGenTAB, a privacy\-constrained synthetic health data generation framework that addresses this challenge by formulating tabular data synthesis as a constrained optimization problem\. Using the Augmented Lagrangian Method, PSyGenTAB embeds a configurable privacy thresholdPminP\_\{\\min\}directly into model training, enabling healthcare data custodians to enforce minimum privacy guarantees while maximizing clinical fidelity\.

Across diverse clinical and generalization benchmarks, we demonstrate that privacy and utility are jointly achievable\. PSyGenTAB consistently maintains or improves both metrics, challenging the prevailing assumption that stronger privacy necessarily degrades data usefulness\. Models trained exclusively on synthetic data achieve performance comparable to those trained on real records, including under severe class imbalance scenarios such as prediabetes detection\.

Empirical privacy evaluation further confirms substantial risk reduction: exact patient record reproduction decreases by 62\.5 % on the Adult benchmark, and membership inference attacks on the Diabetes dataset remain statistically near random guessing\. At the same time, FAITH scores exceeding 0\.97 across configurations indicate that no clinically impossible profiles, spurious correlations, or out\-of\-distribution physiological values are introduced\.

Importantly, PSyGenTAB is model\-agnostic and operationally flexible\. The framework applies to both transformer\-based and GAN\-based generators without architectural modification, and the privacy thresholdPminP\_\{\\min\}can be calibrated to institutional or regulatory requirements, supporting HIPAA Expert Determination and GDPR Article 89 pathways\.

Taken together, these results position PSyGenTAB as a practical and principled foundation for privacy\-preserving health data sharing\. As synthetic data becomes integral to federated research networks, AI development pipelines, and regulatory\-compliant data access programs, frameworks that rigorously balance clinical utility and patient protection will become essential infrastructure for the responsible advancement of healthcare AI\.

## References

- \[1\]\(2016\)Deep learning with differential privacy\.InProceedings of the 2016 ACM SIGSAC conference on computer and communications security,pp\. 308–318\.Cited by:[§I](https://arxiv.org/html/2606.18518#S1.p2.1)\.
- \[2\]D\. J\. Berndt and J\. Clifford\(1994\)Using dynamic time warping to find patterns in time series\.InProceedings of the 3rd International Conference on Knowledge Discovery and Data Mining,AAAIWS’94,pp\. 359–370\.Cited by:[§III\-A](https://arxiv.org/html/2606.18518#S3.SS1.SSS0.Px1.p5.2)\.
- \[3\]D\. P\. Bertsekas\(2014\)Constrained optimization and Lagrange multiplier methods\.Academic Press,New York, NY, USA\.External Links:ISBN 978\-1\-886529\-04\-5Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px4.p1.1),[§III\-B](https://arxiv.org/html/2606.18518#S3.SS2.p4.2),[§VI](https://arxiv.org/html/2606.18518#S6.p4.2)\.
- \[4\]V\. Borisov, K\. Seßler, T\. Leemann, M\. Pawelczyk, and G\. Kasneci\(2022\)Language models are realistic tabular data generators\.arXiv preprint arXiv:2210\.06280\.Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px1.p1.1)\.
- \[5\]B\. Brenninkmeijer, A\. de Vries, E\. Marchiori, and Y\. Hille\(2019\)On the generation and evaluation of tabular data using gans\.PhD diss\., Radboud University\.Cited by:[§III\-A](https://arxiv.org/html/2606.18518#S3.SS1.SSS0.Px1.p2.6)\.
- \[6\]N\. R\. Burrows, I\. Hora, L\. S\. Geiss, E\. W\. Gregg, and A\. Albright\(2017\)Incidence of diagnosed diabetes in adults — united states, 1980–2014\.MMWR Morbidity and Mortality Weekly Report66\(12\),pp\. 306–309\.External Links:[Document](https://dx.doi.org/10.15585/mmwr.mm6612a3)Cited by:[TABLE I](https://arxiv.org/html/2606.18518#S4.T1.13.11.6)\.
- \[7\]N\. Carlini, F\. Tramer, E\. Wallace, M\. Jagielski, A\. Herbert\-Voss, K\. Lee, A\. Roberts, T\. Brown, D\. Song, Ú\. Erlingsson,et al\.\(2021\)Membership inference attacks from first principles\.In2022 IEEE Symposium on Security and Privacy \(SP\),pp\. 1897–1914\.Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px2.p1.3),[§IV\-I](https://arxiv.org/html/2606.18518#S4.SS9.p1.1)\.
- \[8\]D\. Chicco and G\. Jurman\(2020\)Machine learning can predict survival of patients with heart failure from serum creatinine and ejection fraction alone\.BMC Medical Informatics and Decision Making20\(16\)\.External Links:[Document](https://dx.doi.org/10.1186/s12911-020-1023-5)Cited by:[TABLE I](https://arxiv.org/html/2606.18518#S4.T1.18.16.6)\.
- \[9\]E\. Choi, S\. Biswal, B\. Malin, J\. Duke, W\. F\. Stewart, and J\. Sun\(2017\)Generating multi\-label discrete patient records using generative adversarial networks\.Machine Learning for Healthcare Conference,pp\. 286–305\.Cited by:[§I](https://arxiv.org/html/2606.18518#S1.p2.1)\.
- \[10\]E\. Choi, S\. Biswal, B\. Malin, J\. Duke, W\. F\. Stewart, and J\. Sun\(2017\)Generating multi\-label discrete patient records using generative adversarial networks\.InProceedings of Machine Learning for Healthcare \(MLHC\),Proceedings of Machine Learning Research \(PMLR\), Vol\.68,pp\. 286–305\.External Links:[Link](https://proceedings.mlr.press/v68/choi17a.html)Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px1.p1.1)\.
- \[11\]V\. S\. Chundawat, A\. K\. Tarun, M\. Mandal, M\. Lahoti, and P\. Narang\(2022\)TabSynDex: a universal metric for robust evaluation of synthetic tabular data\.arXiv preprint arXiv:2207\.05295\.Cited by:[§III\-A](https://arxiv.org/html/2606.18518#S3.SS1.SSS0.Px1.p4.1)\.
- \[12\]U\.S\. Congress\(1996\)Health insurance portability and accountability act of 1996 \(hipaa\)\.Note:[https://www\.hhs\.gov/hipaa/index\.html](https://www.hhs.gov/hipaa/index.html)Accessed May 2025Cited by:[§I](https://arxiv.org/html/2606.18518#S1.p1.1),[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px2.p1.3)\.
- \[13\]A\. Cotter, H\. Jiang, and K\. Sridharan\(2019\)Two\-player games for efficient non\-convex constrained optimization\.arXiv preprint arXiv:1804\.06500\.External Links:[Link](https://arxiv.org/abs/1804.06500)Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px4.p1.1)\.
- \[14\]S\. Dash, O\. Günlük, and D\. Wei\(2020\)Privacy\-preserving synthetic medical data generation using variational autoencoders\.arXiv preprint arXiv:2012\.15328\.Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px1.p1.1)\.
- \[15\]D\. Dua and C\. Graff\(2019\)UCI Machine Learning Repository: Breast Cancer Wisconsin \(Diagnostic\) Data Set\.Note:UCI Machine Learning RepositoryUniversity of California, Irvine, School of Information and Computer Sciences\.External Links:[Link](http://archive.ics.uci.edu/ml/datasets/Breast+Cancer+Wisconsin+(Diagnostic))Cited by:[TABLE I](https://arxiv.org/html/2606.18518#S4.T1.8.6.6)\.
- \[16\]D\. Dua and C\. Graff\(2019\)UCI machine learning repository\.University of California, Irvine\.Note:[https://archive\.ics\.uci\.edu/ml](https://archive.ics.uci.edu/ml)Adult Census Income DatasetCited by:[TABLE I](https://arxiv.org/html/2606.18518#S4.T1.48.46.6)\.
- \[17\]D\. Dua and C\. Graff\(2019\)UCI machine learning repository\.University of California, Irvine\.Note:[https://archive\.ics\.uci\.edu/ml/datasets/Lung\+Cancer](https://archive.ics.uci.edu/ml/datasets/Lung+Cancer)Lung Cancer DatasetCited by:[TABLE I](https://arxiv.org/html/2606.18518#S4.T1.33.31.6)\.
- \[18\]C\. Dwork and A\. Roth\(2014\)The algorithmic foundations of differential privacy\.Foundations and Trends in Theoretical Computer Science, Vol\.9,Now Publishers Inc\.\.External Links:[Document](https://dx.doi.org/10.1561/0400000042)Cited by:[§IV\-I](https://arxiv.org/html/2606.18518#S4.SS9.SSS0.Px3.p6.1.1)\.
- \[19\]C\. Dwork and A\. Roth\(2014\)The algorithmic foundations of differential privacy\.Foundations and Trends in Theoretical Computer Science9\(3\-4\),pp\. 211–407\.Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px2.p1.3)\.
- \[20\]C\. Dwork\(2006\)Differential privacy\.InAutomata, Languages and Programming,M\. Bugliesi, B\. Preneel, V\. Sassone, and I\. Wegener \(Eds\.\),Berlin, Heidelberg,pp\. 1–12\.External Links:ISBN 978\-3\-540\-35908\-1Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px2.p1.3),[§IV\-I](https://arxiv.org/html/2606.18518#S4.SS9.SSS0.Px3.p6.1.1)\.
- \[21\]K\. El Emam, L\. Mosquera, and R\. Hoptroff\(2020\)Practical synthetic data generation : balancing privacy and the broad availability of data / khaled el emam, lucy mosquera, and richard hoptroff\.\.1st edition edition,O’Reilly Media,Sebastopol, CA\(eng\)\.External Links:ISBN 9781492072737Cited by:[§IV\-H](https://arxiv.org/html/2606.18518#S4.SS8.SSS0.Px2.p1.1)\.
- \[22\]K\. El Emam and L\. Mosquera\(2020\)Practical synthetic data generation: balancing privacy and the broad availability of data\.O’Reilly Media\.Cited by:[§IV\-H](https://arxiv.org/html/2606.18518#S4.SS8.p1.1)\.
- \[23\]Cited by:[TABLE I](https://arxiv.org/html/2606.18518#S4.T1.53.51.6)\.
- \[24\]M\. L\. Fang, D\. S\. Dhami, and K\. Kersting\(2022\)Dp\-ctgan: differentially private medical data generation using ctgans\.InInternational Conference on Artificial Intelligence in Medicine,pp\. 178–188\.Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px2.p1.3)\.
- \[25\]G\. Fasano and A\. Franceschini\(1987\)A multidimensional version of the kolmogorov–smirnov test\.Monthly Notices of the Royal Astronomical Society225\(1\),pp\. 155–170\.Cited by:[§III\-A](https://arxiv.org/html/2606.18518#S3.SS1.SSS0.Px1.p2.2)\.
- \[26\]I\. Goodfellow, J\. Pouget\-Abadie, M\. Mirza, B\. Xu, D\. Warde\-Farley, S\. Ozair, A\. Courville, and Y\. Bengio\(2014\)Generative adversarial nets\.InAdvances in Neural Information Processing Systems,Vol\.27\.Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px1.p1.1)\.
- \[27\]M\. Hernandez, G\. Epelde, A\. Alberdi, R\. Cilla, and D\. Rankin\(2022\)Synthetic data generation for tabular health records: a systematic review\.Neurocomputing493,pp\. 28–45\.Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px1.p1.1)\.
- \[28\]M\. R\. Hestenes\(1969\)Multiplier and gradient methods\.Journal of Optimization Theory and Applications4\(5\),pp\. 303–320\.External Links:[Document](https://dx.doi.org/10.1007/BF00927673)Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px4.p1.1),[§III\-B](https://arxiv.org/html/2606.18518#S3.SS2.p4.2)\.
- \[29\]S\. Ho, Y\. Qu, B\. Gu, L\. Gao, J\. Li, and Y\. Xiang\(2021\)DP\-gan: differentially private consecutive data publishing using generative adversarial nets\.Journal of Network and Computer Applications185,pp\. 103066\.Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px2.p1.3)\.
- \[30\]P\. J\. Huber and E\. M\. Ronchetti\(2009\)Robust statistics\.2nd edition,Wiley\.External Links:[Document](https://dx.doi.org/10.1002/9780470434691)Cited by:[§IV\-I](https://arxiv.org/html/2606.18518#S4.SS9.SSS0.Px3.p5.1)\.
- \[31\]A\. Ilaty, H\. Shirazi, and H\. Homayouni\(2025\)SynLLM: A Comparative Analysis of Large Language Models for Medical Tabular Synthetic Data Generation via Prompt Engineering\.Note:10 Pages, 2 Supplementary Pages, 6 TablesExternal Links:2508\.08529,[Link](https://arxiv.org/abs/2508.08529),[Document](https://dx.doi.org/10.48550/arXiv.2508.08529)Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px3.p1.2)\.
- \[32\]B\. Jayaraman and D\. Evans\(2019\)Evaluating differentially private machine learning in practice\.External Links:1902\.08874,[Link](https://arxiv.org/abs/1902.08874)Cited by:[§IV\-I](https://arxiv.org/html/2606.18518#S4.SS9.SSS0.Px2.p1.1)\.
- \[33\]J\. Jordon, L\. Szpruch, F\. Houssiau, M\. Bottarelli, G\. Cherubin, C\. Maple, S\. N\. Cohen, and A\. Weller\(2022\)Synthetic data — what, why and how?\.arXiv preprint arXiv:2205\.03257\.External Links:[Link](https://arxiv.org/abs/2205.03257)Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px3.p1.2)\.
- \[34\]J\. Jordon, J\. Yoon, and M\. Van Der Schaar\(2018\)PATE\-gan: generating synthetic data with differential privacy guarantees\.InInternational conference on learning representations,Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px2.p1.3)\.
- \[35\]D\. P\. Kingma and M\. Welling\(2013\)Auto\-encoding variational bayes\.arXiv preprint arXiv:1312\.6114\.Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px1.p1.1)\.
- \[36\]J\. Kotary and F\. Fioretto\(2024\)Learning constrained optimization with deep augmented lagrangian methods\.External Links:2403\.03454,[Link](https://arxiv.org/abs/2403.03454)Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px4.p1.1)\.
- \[37\]A\. Kotelnikov, D\. Baranchuk, I\. Rubachev, and A\. Babenko\(2023\)Tabddpm: modelling tabular data with diffusion models\.InInternational Conference on Machine Learning,pp\. 17564–17579\.Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px1.p1.1)\.
- \[38\]I\. Krchova, M\. V\. Vieyra, M\. Scriminaci, and A\. Sidorenko\(2025\)Democratizing tabular data access with an open\-source synthetic\-data sdk\.External Links:2508\.00718,[Link](https://arxiv.org/abs/2508.00718)Cited by:[§IV\-F](https://arxiv.org/html/2606.18518#S4.SS6.p3.1.1)\.
- \[39\]S\. Kullback and R\. A\. Leibler\(1951\)On information and sufficiency\.The annals of mathematical statistics22\(1\),pp\. 79–86\.Cited by:[§III\-A](https://arxiv.org/html/2606.18518#S3.SS1.SSS0.Px1.p2.2)\.
- \[40\]A\. Kurakova and H\. Homayouni\(2024\)A comprehensive evaluation framework for synthetic medical tabular data generation\.ACM Transactions on Computing for Healthcare\.Note:Submitted for publicationCited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px3.p1.2)\.
- \[41\]A\. D\. Lautrup, T\. Hyrup, A\. Zimek, and P\. Schneider\-Kamp\(2024\-12\)Syntheval: a framework for detailed utility and privacy evaluation of tabular synthetic data\.Data Mining and Knowledge Discovery39\(1\)\.External Links:ISSN 1573\-756X,[Link](http://dx.doi.org/10.1007/s10618-024-01081-4),[Document](https://dx.doi.org/10.1007/s10618-024-01081-4)Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px3.p1.2)\.
- \[42\]N\. Li, T\. Li, and S\. Venkatasubramanian\(2007\)T\-closeness: privacy beyond k\-anonymity and l\-diversity\.In2007 IEEE 23rd International Conference on Data Engineering,pp\. 106–115\.Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px2.p1.3)\.
- \[43\]Z\. Li, P\. Chen, S\. Liu, S\. Lu, and Y\. Xu\(2024\)Stochastic inexact augmented lagrangian method for nonconvex expectation constrained optimization\.Computational Optimization and Applications87\(1\),pp\. 117–147\.Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px4.p1.1)\.
- \[44\]M\. A\. Little, P\. E\. McSharry, S\. J\. Roberts, D\. A\. E\. Costello, and I\. M\. Moroz\(2007\)Exploiting nonlinear recurrence and fractal scaling properties for voice disorder detection\.BioMedical Engineering OnLine6\(23\)\.External Links:[Document](https://dx.doi.org/10.1186/1475-925X-6-23)Cited by:[TABLE I](https://arxiv.org/html/2606.18518#S4.T1.43.41.6)\.
- \[45\]Y\. Liu, U\. R\. Acharya, and J\. H\. Tan\(2025\)Preserving privacy in healthcare: a systematic review of deep learning approaches for synthetic data generation\.Computer Methods and Programs in Biomedicine260,pp\. 108571\.External Links:ISSN 0169\-2607,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/j.cmpb.2024.108571),[Link](https://www.sciencedirect.com/science/article/pii/S0169260724005649)Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px1.p1.1)\.
- \[46\]B\. M\. R\. Ltd\.\(1990\)BUPA liver disorders dataset\.Note:[https://archive\.ics\.uci\.edu/ml/datasets/Liver\+Disorders](https://archive.ics.uci.edu/ml/datasets/Liver+Disorders)UCI Machine Learning RepositoryCited by:[TABLE I](https://arxiv.org/html/2606.18518#S4.T1.28.26.6)\.
- \[47\]A\. Machanavajjhala, D\. Kifer, J\. Gehrke, and M\. Venkitasubramaniam\(2007\)L\-diversity: privacy beyond k\-anonymity\.ACM Transactions on Knowledge Discovery from Data \(TKDD\)1\(1\),pp\. 3–es\.Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px2.p1.3)\.
- \[48\]M\.L\. Menéndez, J\.A\. Pardo, L\. Pardo, and M\.C\. Pardo\(1997\)The jensen\-shannon divergence\.Journal of the Franklin Institute334\(2\),pp\. 307–318\.External Links:ISSN 0016\-0032,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/S0016-0032%2896%2900063-4),[Link](https://www.sciencedirect.com/science/article/pii/S0016003296000634)Cited by:[§III\-A](https://arxiv.org/html/2606.18518#S3.SS1.SSS0.Px1.p2.2)\.
- \[49\]H\. T\. Nguyen and T\. N\. Tran\(2020\)Vietnam banking transaction dataset for fraud detection\.Note:Public financial transaction datasetIf sourced from Kaggle or Zenodo, include DOI hereCited by:[TABLE I](https://arxiv.org/html/2606.18518#S4.T1.58.56.6)\.
- \[50\]F\. M\. Palechor and A\. de la Hoz Manotas\(2019\)Dataset for estimation of obesity levels based on eating habits and physical condition in individuals from colombia, peru and mexico\.Data in Brief25,pp\. 104344\.External Links:[Document](https://dx.doi.org/10.1016/j.dib.2019.104344)Cited by:[TABLE I](https://arxiv.org/html/2606.18518#S4.T1.38.36.6)\.
- \[51\]J\. Park, J\. Yoon, S\. Keum, J\. Oh, M\. Lee, and J\. Kim\(2021\)Federated learning for healthcare: systematic review and architecture proposal\.IEEE Journal of Biomedical and Health Informatics25\(5\),pp\. 1478–1491\.Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px2.p1.3)\.
- \[52\]N\. Park, M\. Mohammadi, K\. Gorde, S\. Jajodia, H\. Park, and Y\. Kim\(2018\)Data synthesis based on generative adversarial networks\.arXiv preprint arXiv:1806\.03384\.Cited by:[§III\-A](https://arxiv.org/html/2606.18518#S3.SS1.SSS0.Px2.p2.4)\.
- \[53\]M\. J\. D\. Powell\(1969\)A method for nonlinear constraints in minimization problems\.Optimization,pp\. 283–298\.Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px4.p1.1),[§III\-B](https://arxiv.org/html/2606.18518#S3.SS2.p4.2)\.
- \[54\]J\. R\. Quinlan\(1987\)Generating production rules from decision trees\.InProceedings of the 10th International Joint Conference on Artificial Intelligence \(IJCAI\),pp\. 304–307\.Cited by:[TABLE I](https://arxiv.org/html/2606.18518#S4.T1.23.21.6)\.
- \[55\]N\. Rieke, J\. Hancox, W\. Li, F\. Milletari, H\. R\. Roth, S\. Albarqouni, S\. Bakas, M\. N\. Galtier, B\. A\. Landman, K\. Maier\-Hein, S\. Herrmann, J\. Shotton, J\. Trees, B\. Kainz, R\. Cobb, B\. Glocker, and D\. Rueckert\(2020\)The future of digital health with federated learning\.npj Digital Medicine3\(1\),pp\. 119\.External Links:[Document](https://dx.doi.org/10.1038/s41746-020-00323-1)Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px2.p1.3)\.
- \[56\]R\. T\. Rockafellar\(1973\)The multiplier method of Hestenes and Powell applied to convex programming\.Journal of Optimization Theory and Applications12\(6\),pp\. 555–562\.External Links:[Document](https://dx.doi.org/10.1007/BF00934777)Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px4.p1.1),[§III\-B](https://arxiv.org/html/2606.18518#S3.SS2.p4.2),[§VI](https://arxiv.org/html/2606.18518#S6.p4.2)\.
- \[57\]R\. Shokri, M\. Stronati, C\. Song, and V\. Shmatikov\(2017\)Membership inference attacks against machine learning models\.In2017 IEEE Symposium on Security and Privacy \(SP\),pp\. 3–18\.Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px1.p1.1),[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px2.p1.3)\.
- \[58\]R\. Shokri, M\. Stronati, C\. Song, and V\. Shmatikov\(2017\)Membership inference attacks against machine learning models\.External Links:1610\.05820,[Link](https://arxiv.org/abs/1610.05820)Cited by:[§IV\-I](https://arxiv.org/html/2606.18518#S4.SS9.p1.1)\.
- \[59\]J\. Snoke, G\. M\. Raab, B\. Nowok, C\. Dibben, and A\. Slavkovic\(2018\)General and specific utility measures for synthetic data\.Journal of the Royal Statistical Society: Series A181\(3\),pp\. 663–688\.Cited by:[§IV\-H](https://arxiv.org/html/2606.18518#S4.SS8.p1.1),[§IV\-H](https://arxiv.org/html/2606.18518#S4.SS8.p4.2)\.
- \[60\]A\. V\. Solatorio and O\. Dupriez\(2023\)REaLTabFormer: generating realistic relational and tabular data using transformers\.arXiv preprint arXiv:2302\.02041\.Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px1.p1.1),[§III\-A](https://arxiv.org/html/2606.18518#S3.SS1.SSS0.Px2.p3.4),[§III\-B](https://arxiv.org/html/2606.18518#S3.SS2.SSS0.Px2.p1.1),[§IV\-B1](https://arxiv.org/html/2606.18518#S4.SS2.SSS1.p1.1)\.
- \[61\]L\. Sweeney\(2002\)K\-anonymity: a model for protecting privacy\.International Journal of Uncertainty, Fuzziness and Knowledge\-Based Systems10\(05\),pp\. 557–570\.Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px2.p1.3),[§IV\-I](https://arxiv.org/html/2606.18518#S4.SS9.p6.1)\.
- \[62\]\(2024\-04\)Synthetic data metrics\.DataCebo, Inc\.\.Note:Version 0\.14\.0External Links:[Link](https://docs.sdv.dev/sdmetrics/)Cited by:[§III\-A](https://arxiv.org/html/2606.18518#S3.SS1.SSS0.Px1.p2.2),[§III\-A](https://arxiv.org/html/2606.18518#S3.SS1.SSS0.Px1.p4.1),[§IV\-F](https://arxiv.org/html/2606.18518#S4.SS6.p2.1.1)\.
- \[63\]G\. Tunnicliffe Wilson\(2016\-03\)Time series analysis: forecasting and control,5th edition, by george e\. p\. box, gwilym m\. jenkins, gregory c\. reinsel and greta m\. ljung, 2015\. published by john wiley and sons inc\., hoboken, new jersey, pp\. 712\. isbn: 978\-1\-118\-67502\-1\.Journal of Time Series Analysis37,pp\. n/a–n/a\.External Links:[Document](https://dx.doi.org/10.1111/jtsa.12194)Cited by:[§III\-A](https://arxiv.org/html/2606.18518#S3.SS1.SSS0.Px1.p5.2)\.
- \[64\]E\. Union\(2016\)Regulation \(eu\) 2016/679 of the european parliament and of the council \(general data protection regulation\)\.Note:[https://gdpr\-info\.eu](https://gdpr-info.eu/)Accessed May 2025Cited by:[§I](https://arxiv.org/html/2606.18518#S1.p1.1),[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px2.p1.3)\.
- \[65\]J\. Walonoski, M\. Kramer, J\. Nichols, A\. Quina, C\. Moesel, D\. Hall, C\. Duffett, K\. Dube, T\. Gallagher, and S\. McLachlan\(2017\-09\)Synthea: an approach, method, and software mechanism for generating synthetic patients and the synthetic electronic health care record\.Journal of the American Medical Informatics Association0,pp\. 1–9\.External Links:[Document](https://dx.doi.org/10.1093/jamia/ocx079)Cited by:[§I](https://arxiv.org/html/2606.18518#S1.p2.1),[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px1.p1.1)\.
- \[66\]L\. Xu, M\. Skoularidou, A\. Cuesta\-Infante, and K\. Veeramachaneni\(2019\)Modeling tabular data using conditional gan\.Advances in neural information processing systems32\.Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px1.p1.1)\.
- \[67\]L\. Xu and K\. Veeramachaneni\(2018\)Synthesizing tabular data using generative adversarial networks\.arXiv preprint arXiv:1811\.11264\.Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px1.p1.1),[§III\-A](https://arxiv.org/html/2606.18518#S3.SS1.SSS0.Px2.p2.4)\.
- \[68\]Z\. Yao, N\. Krčo, G\. Ganev, and Y\. de Montjoye\(2026\)The dcr delusion: measuring the privacy risk of synthetic data\.InComputer Security – ESORICS 2025,V\. Nicomette, A\. Benzekri, N\. Boulahia\-Cuppens, and J\. Vaidya \(Eds\.\),Cham,pp\. 469–487\.External Links:ISBN 978\-3\-032\-07884\-1Cited by:[§IV\-H](https://arxiv.org/html/2606.18518#S4.SS8.p1.1)\.
- \[69\]H\. Yuan, X\. Lian, J\. Li, J\. Liu, and B\. Xu\(2021\)Reinforced Augmented Lagrangian for constrained optimization in deep learning\.arXiv preprint arXiv:2106\.01134\.Cited by:[§VI](https://arxiv.org/html/2606.18518#S6.p4.2)\.
- \[70\]H\. Zhang, J\. Zhang, B\. Srinivasan, Z\. Shen, X\. Qin, C\. Faloutsos, H\. Rangwala, and G\. Karypis\(2024\)Mixed\-type tabular data synthesis with score\-based diffusion in latent space\.InThe twelfth International Conference on Learning Representations,Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px1.p1.1)\.
- \[71\]Z\. Zhao, A\. Kunar, R\. Birke, and L\. Y\. Chen\(2022\)CTAB\-gan\+: enhancing tabular data synthesis\.External Links:2204\.00401,[Link](https://arxiv.org/abs/2204.00401)Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px1.p1.1),[§III\-B](https://arxiv.org/html/2606.18518#S3.SS2.SSS0.Px2.p1.1),[§IV\-B2](https://arxiv.org/html/2606.18518#S4.SS2.SSS2.p1.1)\.
- \[72\]H\. Zhou, X\. Wang, and R\. Zhu\(2022\)Feature selection based on mutual information with correlation coefficient\.Applied intelligence52\(5\),pp\. 5457–5474\.Cited by:[§III\-A](https://arxiv.org/html/2606.18518#S3.SS1.SSS0.Px1.p3.1)\.
- \[73\]L\. Zhu, Z\. Liu, and S\. Han\(2019\)Deep leakage from gradients\.InAdvances in Neural Information Processing Systems \(NeurIPS\),Vol\.32,pp\. 14774–14784\.Cited by:[§II](https://arxiv.org/html/2606.18518#S2.SS0.SSS0.Px2.p1.3),[§V\-B](https://arxiv.org/html/2606.18518#S5.SS2.p1.1)\.

## Appendix ASummary of Downstream Machine Learning Utility

TableLABEL:tab:ml\_comprehensiveprovides an exhaustive, multi\-page breakdown of downstream predictive efficacy across all classifiers and sampling strategies\. This summary highlights the Train\-on\-Synthetic, Test\-on\-Real \(TSTR\) performance of both the proposed PSyGenTAB framework and the unconstrained RealTabFormer \(RTF\) baseline, measured against the Train\-on\-Real, Test\-on\-Real \(TRTR\) empirical ceiling\. The values reported represent the maximum F1\-score and Balanced Accuracy achieved across all evaluated class\-imbalance strategies \(None, Class Weight, SMOTE, and Undersampling\) for each dataset\.

As demonstrated in TableLABEL:tab:ml\_comprehensive, PSyGenTAB maintains highly competitive structural preservation\. For relatively stable datasets like Breast Cancer, synthetic data achieves parity with the real\-world baseline\. On highly complex or imbalanced datasets, such as Heart Failure and Parkinson’s, a natural degradation in predictive utility occurs when shifting from TRTR to TSTR\. However, PSyGenTAB matches or exceeds the utility of the unconstrained RTF baseline \(e\.g\., achieving a 0\.7107 F1\-score on Heart Failure compared to RTF’s 0\.5802\)\. This confirms that the Augmented Lagrangian Method \(ALM\) effectively enforces privacy constraints without systematically destroying the predictive manifold of the data\.

————————————————————————

TABLE IX:Downstream clinical AI utility of PSyGenTAB\-generated synthetic data\. TRTR: Train on Real, Test on Real \(reference\); TSTR: Train on Synthetic, Test on Real; TRTS: Train on Real, Test on Synthetic\. Acc: Accuracy; Bal: Balanced Accuracy; Prc: Precision; Rec: Recall\. Best F1\-score per model/scenario isbolded\. Results demonstrate that synthetic health records preserve clinically meaningful predictive structure across multiple downstream classifiers and imbalance\-handling strategies\.AdultTRTRNone0\.81800\.75310\.81910\.81800\.81850\.85690\.77680\.85130\.85690\.85260\.80930\.66160\.79370\.80930\.7881Class Weight0\.81390\.74150\.81280\.81390\.81340\.85710\.77370\.85120\.85710\.85220\.75020\.73640\.79550\.75020\.7639SMOTE0\.80320\.73980\.80740\.80320\.80510\.84860\.78490\.84650\.84860\.84750\.74410\.73690\.79580\.74410\.7590Undersample0\.76580\.77030\.81840\.76580\.77980\.81270\.82090\.85250\.81270\.82280\.74530\.73340\.79340\.74530\.7597AdultALMTSTRNone0\.80110\.74260\.80780\.80110\.80400\.83870\.77670\.83820\.83870\.83840\.78480\.71200\.78810\.78480\.7863Class Weight0\.80500\.73210\.80500\.80500\.80500\.83870\.76980\.83610\.83870\.83730\.72030\.74720\.80550\.72030\.7399SMOTE0\.77790\.73050\.79520\.77790\.78460\.81850\.77500\.82840\.81850\.82240\.72720\.74560\.80330\.72720\.7457Undersample0\.77610\.75010\.80590\.77610\.78600\.81180\.80600\.84320\.81180\.82080\.72720\.75140\.80770\.72720\.7460TRTSNone0\.79390\.73870\.79060\.79390\.79200\.82940\.76330\.82370\.82940\.82380\.77150\.61700\.78070\.77150\.7258Class Weight0\.79430\.73630\.79010\.79430\.79180\.82690\.75860\.82090\.82690\.82080\.75680\.72510\.76950\.75680\.7617SMOTE0\.78860\.73240\.78530\.78860\.78670\.82940\.77090\.82450\.82940\.82550\.75490\.73670\.77670\.75490\.7620Undersample0\.75790\.76940\.80180\.75790\.76810\.80230\.82130\.84200\.80230\.81040\.75490\.72320\.76800\.75490\.7599AdultRTFTSTRNone0\.81230\.75310\.81660\.81230\.81420\.85500\.78500\.85110\.85500\.85250\.80820\.67680\.79280\.80820\.7929Class Weight0\.81660\.75390\.81880\.81660\.81760\.85560\.78020\.85070\.85560\.85220\.73680\.73370\.79380\.73680\.7528SMOTE0\.80450\.75880\.81640\.80450\.80920\.84640\.79200\.84740\.84640\.84690\.74520\.73130\.79200\.74520\.7594Undersample0\.77180\.77420\.82100\.77180\.78500\.80820\.81790\.85050\.80820\.81890\.74010\.73620\.79540\.74010\.7557TRTSNone0\.83090\.77030\.83060\.83090\.83080\.87010\.79320\.86540\.87010\.86540\.81800\.68060\.80660\.81800\.7989Class Weight0\.83000\.76450\.82820\.83000\.82900\.87010\.79020\.86540\.87010\.86490\.76400\.75570\.80630\.76400\.7763SMOTE0\.81970\.75870\.82070\.81970\.82020\.86320\.79910\.85970\.86320\.86090\.76340\.75430\.80540\.76340\.7757Undersample0\.79220\.78990\.82990\.79220\.80270\.83880\.84340\.86740\.83880\.84610\.76070\.74960\.80220\.76070\.7730DiabetesTRTRNone0\.76750\.40760\.77980\.76750\.77340\.84300\.38740\.79680\.84300\.80810\.84620\.38380\.79910\.84620\.8078Class Weight0\.76560\.39300\.77370\.76560\.76960\.83900\.37580\.78940\.83900\.80050\.64430\.51910\.85180\.64430\.7197SMOTE0\.76500\.40030\.77730\.76500\.77090\.84000\.39590\.79680\.84000\.81060\.64010\.51350\.85030\.64010\.7163Undersample0\.47930\.42710\.80960\.47930\.58250\.58060\.49680\.84690\.58060\.67160\.64470\.51620\.85190\.64470\.7204DiabetesALMTSTRNone0\.75070\.41080\.77870\.75070\.76190\.80980\.40330\.78350\.80980\.79600\.84390\.40320\.80190\.84390\.8155Class Weight0\.75180\.40130\.77300\.75180\.76050\.81370\.39670\.78100\.81370\.79590\.73060\.49700\.83300\.73060\.7601SMOTE0\.74930\.40720\.77670\.74930\.76040\.80900\.40910\.78630\.80900\.79730\.73010\.49660\.83270\.73010\.7597Undersample0\.64910\.43580\.79590\.64910\.69350\.68910\.48540\.82830\.68910\.72770\.73120\.49700\.83290\.73120\.7606TRTSNone0\.77260\.59180\.78720\.77260\.77980\.83840\.58240\.80530\.83840\.80780\.84760\.58450\.81950\.84760\.8129Class Weight0\.76930\.58510\.78310\.76930\.77610\.83430\.56930\.79950\.83430\.80110\.65670\.64400\.86770\.65670\.7365SMOTE0\.77510\.59700\.78940\.77510\.78220\.83580\.59460\.80550\.83580\.81090\.65120\.64070\.86690\.65120\.7320Undersample0\.49190\.46970\.81980\.49190\.60080\.59310\.58240\.86250\.59310\.69040\.65780\.64170\.86790\.65780\.7380DiabetesRTFTSTRNone0\.76560\.40090\.77840\.76560\.77190\.84090\.38880\.79490\.84090\.80770\.84630\.38670\.79990\.84630\.8093Class Weight0\.76440\.39670\.77390\.76440\.76910\.83790\.38130\.79020\.83790\.80270\.64500\.52170\.85200\.64500\.7207SMOTE0\.76520\.40460\.77860\.76520\.77170\.83770\.39670\.79520\.83770\.80940\.64170\.51290\.85030\.64170\.7182Undersample0\.48220\.41290\.80500\.48220\.58480\.58110\.49400\.84680\.58110\.67270\.64350\.52000\.85170\.64350\.7200TRTSNone0\.77360\.41080\.78290\.77360\.77810\.84210\.38980\.79690\.84210\.80850\.84830\.38950\.80380\.84830\.8115Class Weight0\.76770\.39410\.77400\.76770\.77080\.83800\.37940\.79000\.83800\.80130\.64810\.52070\.85460\.64810\.7230SMOTE0\.76950\.40250\.77900\.76950\.77420\.83820\.39740\.79620\.83820\.80990\.64530\.51620\.85390\.64530\.7208Undersample0\.48840\.43270\.81130\.48840\.59050\.58490\.49600\.84790\.58490\.67540\.64880\.51800\.85440\.64880\.7240Breast Can\.TRTRNone0\.90910\.89620\.91240\.90910\.90920\.95450\.95000\.95800\.95450\.95490\.96360\.95960\.96630\.96360\.9638Class Weight0\.90910\.89620\.91240\.90910\.90920\.95450\.95000\.95800\.95450\.95490\.96360\.95960\.96630\.96360\.9638SMOTE0\.90910\.89620\.91240\.90910\.90920\.95450\.95000\.95800\.95450\.95490\.96360\.95960\.96630\.96360\.9638Undersample0\.91820\.90600\.92060\.91820\.91820\.96360\.96000\.96580\.96360\.96360\.97270\.96970\.97450\.97270\.9727Breast Can\.ALMTSTRNone0\.90910\.89620\.91240\.90910\.90920\.95450\.95000\.95800\.95450\.95490\.96360\.95960\.96630\.96360\.9638Class Weight0\.90910\.89620\.91240\.90910\.90920\.95450\.95000\.95800\.95450\.95490\.96360\.95960\.96630\.96360\.9638SMOTE0\.90910\.89620\.91240\.90910\.90920\.95450\.95000\.95800\.95450\.95490\.96360\.95960\.96630\.96360\.9638Undersample0\.91820\.90600\.92060\.91820\.91820\.96360\.96000\.96580\.96360\.96360\.97270\.96970\.97450\.97270\.9727TRTSNone0\.90000\.88700\.90560\.90000\.90070\.94500\.94000\.95100\.94500\.94530\.95500\.95000\.96000\.95500\.9552Class Weight0\.90000\.88700\.90560\.90000\.90070\.94500\.94000\.95100\.94500\.94530\.95500\.95000\.96000\.95500\.9552SMOTE0\.90000\.88700\.90560\.90000\.90070\.94500\.94000\.95100\.94500\.94530\.95500\.95000\.96000\.95500\.9552Undersample0\.90910\.89620\.91240\.90910\.90920\.95450\.95000\.95800\.95450\.95490\.96360\.95960\.96630\.96360\.9638Breast Can\.RTFTSTRNone0\.90910\.89620\.91240\.90910\.90920\.95450\.95000\.95800\.95450\.95490\.96360\.95960\.96630\.96360\.9638Class Weight0\.90910\.89620\.91240\.90910\.90920\.95450\.95000\.95800\.95450\.95490\.96360\.95960\.96630\.96360\.9638SMOTE0\.90910\.89620\.91240\.90910\.90920\.95450\.95000\.95800\.95450\.95490\.96360\.95960\.96630\.96360\.9638Undersample0\.91820\.90600\.92060\.91820\.91820\.96360\.96000\.96580\.96360\.96360\.97270\.96970\.97450\.97270\.9727TRTSNone0\.90000\.88700\.90560\.90000\.90070\.94500\.94000\.95100\.94500\.94530\.95500\.95000\.96000\.95500\.9552Class Weight0\.90000\.88700\.90560\.90000\.90070\.94500\.94000\.95100\.94500\.94530\.95500\.95000\.96000\.95500\.9552SMOTE0\.90000\.88700\.90560\.90000\.90070\.94500\.94000\.95100\.94500\.94530\.95500\.95000\.96000\.95500\.9552Undersample0\.90910\.89620\.91240\.90910\.90920\.95450\.95000\.95800\.95450\.95490\.96360\.95960\.96630\.96360\.9638ParkinsonsTRTRNone0\.84750\.85380\.87280\.84750\.85380\.93220\.88860\.93180\.93220\.93050\.84750\.80980\.85120\.84750\.8490Class Weight0\.83050\.82050\.85040\.83050\.83640\.91530\.85530\.91530\.91530\.91190\.72880\.75230\.80200\.72880\.7453SMOTE0\.86440\.79920\.86030\.86440\.86110\.91530\.87730\.91400\.91530\.91430\.72880\.73030\.78450\.72880\.7438Undersample0\.72880\.77420\.82160\.72880\.74610\.74580\.82950\.87290\.74580\.76200\.67800\.71820\.78260\.67800\.6987ParkinsonsALMTSTRNone0\.66100\.50910\.62870\.66100\.64230\.72880\.48860\.55290\.72880\.62880\.74580\.50000\.55620\.74580\.6372Class Weight0\.45760\.32880\.49810\.45760\.47660\.71190\.47730\.54950\.71190\.62020\.74580\.71970\.77730\.74580\.7563SMOTE0\.59320\.61740\.70910\.59320\.61950\.47460\.42800\.56870\.47460\.50750\.52540\.44020\.57830\.52540\.5476Undersample0\.38980\.39320\.53380\.38980\.42580\.50850\.62650\.74990\.50850\.52270\.64410\.71740\.79380\.64410\.6658TRTSNone0\.51060\.46650\.55480\.51060\.52770\.65960\.46970\.48370\.65960\.55810\.61700\.48050\.56060\.61700\.5808Class Weight0\.48940\.45130\.54260\.48940\.50890\.68090\.48480\.48840\.68090\.56880\.55320\.45560\.54370\.55320\.5483SMOTE0\.57450\.53250\.60770\.57450\.58730\.65960\.46970\.48370\.65960\.55810\.53190\.44050\.53190\.53190\.5319Undersample0\.51060\.48700\.57120\.51060\.53050\.61700\.48050\.56060\.61700\.58080\.53190\.41990\.51100\.53190\.5208ParkinsonsRTFTSTRNone0\.71190\.56520\.68000\.71190\.69080\.72880\.48860\.55290\.72880\.62880\.74580\.50000\.55620\.74580\.6372Class Weight0\.66100\.50910\.62870\.66100\.64230\.74580\.50000\.55620\.74580\.63720\.54240\.47350\.60200\.54240\.5659SMOTE0\.62710\.59620\.68920\.62710\.64770\.61020\.47500\.60140\.61020\.60560\.52540\.44020\.57830\.52540\.5476Undersample0\.54240\.53940\.65000\.54240\.57190\.45760\.48260\.60660\.45760\.48960\.44070\.38330\.53630\.44070\.4758TRTSNone0\.59570\.48360\.63040\.59570\.61120\.74470\.51770\.66990\.74470\.68450\.68090\.50760\.64770\.68090\.6620Class Weight0\.46810\.40030\.57410\.46810\.50730\.76600\.50000\.58670\.76600\.66440\.61700\.52900\.66080\.61700\.6352SMOTE0\.53190\.53660\.66740\.53190\.56770\.68090\.47600\.61810\.68090\.64430\.57450\.53280\.66340\.57450\.6046Undersample0\.51060\.49120\.63530\.51060\.54810\.59570\.51520\.65150\.59570\.61810\.65960\.55680\.68030\.65960\.6689ObesityTRTRNone0\.91320\.91000\.91540\.91320\.91370\.93850\.93690\.94610\.93850\.93980\.80600\.80220\.80120\.80600\.8016Class Weight0\.91480\.91230\.91600\.91480\.91520\.93690\.93520\.94320\.93690\.93830\.80440\.80310\.80180\.80440\.8013SMOTE0\.90220\.89950\.90240\.90220\.90210\.94480\.94380\.95090\.94480\.94580\.79810\.79640\.79450\.79810\.7945Undersample0\.90850\.90590\.90950\.90850\.90860\.92590\.92410\.93320\.92590\.92770\.78860\.78770\.78550\.78860\.7842ObesityALMTSTRNone0\.61040\.60930\.61360\.61040\.60890\.71450\.71660\.73760\.71450\.71410\.67670\.67640\.68840\.67670\.6649Class Weight0\.59780\.59870\.61050\.59780\.60110\.70980\.71280\.73450\.70980\.71080\.67350\.67720\.69440\.67350\.6706SMOTE0\.61830\.61640\.63310\.61830\.62250\.70190\.70430\.71940\.70190\.70360\.64980\.65050\.66280\.64980\.6445Undersample0\.63410\.63120\.63820\.63410\.63550\.70350\.70860\.72600\.70350\.70170\.66400\.66680\.66490\.66400\.6561TRTSNone0\.50890\.50730\.55910\.50890\.52250\.59370\.60270\.64620\.59370\.60750\.54640\.52860\.56970\.54640\.5546Class Weight0\.50100\.50820\.57070\.50100\.51710\.60360\.61690\.64860\.60360\.61490\.54640\.52950\.56730\.54640\.5533SMOTE0\.52860\.53540\.56690\.52860\.53310\.58380\.59510\.64960\.58380\.59960\.55620\.53700\.57060\.55620\.5601Undersample0\.52070\.52520\.55770\.52070\.52550\.60550\.60810\.64940\.60550\.61900\.54640\.53100\.55970\.54640\.5508ObesityRTFTSTRNone0\.64830\.64370\.65430\.64830\.64960\.75870\.75570\.76810\.75870\.75240\.70190\.69800\.71640\.70190\.6985Class Weight0\.62780\.62540\.62840\.62780\.62760\.77440\.77090\.78990\.77440\.77350\.69090\.69090\.68800\.69090\.6871SMOTE0\.63250\.63010\.62540\.63250\.62800\.76180\.75930\.76600\.76180\.76020\.69720\.69830\.69680\.69720\.6928Undersample0\.62300\.62000\.63400\.62300\.62460\.74450\.74420\.74690\.74450\.74380\.70660\.70710\.70470\.70660\.7025TRTSNone0\.55820\.55460\.56400\.55820\.55760\.61740\.60810\.62310\.61740\.61390\.55030\.56700\.55680\.55030\.5508Class Weight0\.54440\.54000\.54580\.54440\.54140\.60160\.59280\.60800\.60160\.59830\.54240\.55940\.54670\.54240\.5423SMOTE0\.55030\.55570\.54520\.55030\.54620\.62130\.60800\.63640\.62130\.61890\.55030\.56890\.55630\.55030\.5502Undersample0\.54830\.55620\.54650\.54830\.54690\.63910\.63260\.64580\.63910\.63490\.55820\.57700\.56470\.55820\.5580VN Bank\.TRTRNone0\.66960\.64150\.67140\.66960\.67050\.75310\.72100\.74940\.75310\.75060\.69610\.63730\.68330\.69610\.6829Class Weight0\.67630\.64450\.67520\.67630\.67570\.75330\.71340\.74760\.75330\.74820\.70690\.72510\.74820\.70690\.7133SMOTE0\.66570\.64780\.67540\.66570\.66940\.74680\.73630\.75490\.74680\.74960\.67890\.69190\.71700\.67890\.6859Undersample0\.65480\.65450\.68140\.65480\.66160\.73320\.74870\.76780\.73320\.73900\.67870\.69160\.71670\.67870\.6857VN Bank\.ALMTSTRNone0\.36050\.38920\.42550\.36050\.36320\.64390\.49900\.50260\.64390\.50780\.64370\.50800\.58440\.64370\.5313Class Weight0\.59030\.46680\.47020\.59030\.50130\.64600\.50000\.41730\.64600\.50710\.64660\.50780\.60110\.64660\.5272SMOTE0\.44210\.38600\.43740\.44210\.43970\.64600\.50000\.41730\.64600\.50710\.64870\.51370\.53860\.64870\.5386Undersample0\.66070\.54940\.63550\.66070\.59580\.50160\.47730\.52230\.50160\.50970\.35290\.49640\.39740\.35290\.1900TRTSNone0\.79410\.50120\.98170\.79410\.87700\.43980\.50750\.98200\.43980\.60250\.00930\.50000\.00010\.00930\.0002Class Weight0\.37720\.51160\.98220\.37720\.53890\.69830\.49300\.98150\.69830\.81440\.04450\.50660\.98500\.04450\.0686SMOTE0\.41920\.50830\.98200\.41920\.58220\.36590\.52380\.98290\.36590\.52670\.00930\.50000\.00010\.00930\.0002Undersample0\.18300\.49400\.98110\.18300\.29790\.09420\.50720\.98320\.09420\.15770\.00930\.50000\.00010\.00930\.0002VN Bank\.RTFTSTRNone0\.64720\.50860\.60650\.64720\.52810\.64680\.59600\.63520\.64680\.63900\.54810\.46530\.50460\.54810\.5188Class Weight0\.64820\.51480\.60870\.64820\.54160\.64600\.50000\.41730\.64600\.50710\.45690\.45000\.49710\.45690\.4687SMOTE0\.52440\.51480\.55600\.52440\.53420\.64600\.50000\.41730\.64600\.50710\.62430\.49140\.51080\.62430\.5169Undersample0\.44150\.36190\.40340\.44150\.42030\.53730\.55550\.59520\.53730\.54630\.35170\.48810\.42160\.35170\.2069TRTSNone0\.70070\.51490\.93210\.70070\.79460\.41070\.48040\.92730\.41070\.55180\.03600\.50000\.00130\.03600\.0025Class Weight0\.38070\.48270\.92740\.38070\.52010\.67000\.51690\.93240\.67000\.77310\.09470\.49480\.92510\.09470\.1175SMOTE0\.46200\.51590\.93300\.46200\.60150\.36530\.47470\.92580\.36530\.50370\.03600\.50000\.00130\.03600\.0025Undersample0\.26330\.49310\.92870\.26330\.37950\.10870\.50200\.93230\.10870\.14190\.03600\.50000\.00130\.03600\.0025Lung CancerTRTRNone0\.30000\.33330\.26000\.30000\.27500\.20000\.22220\.20000\.20000\.20000\.50000\.52780\.45830\.50000\.4714Class Weight0\.30000\.33330\.26000\.30000\.27500\.40000\.44440\.30000\.40000\.34290\.50000\.52780\.45830\.50000\.4714SMOTE0\.50000\.52780\.50000\.50000\.47620\.60000\.63890\.62000\.60000\.58330\.50000\.52780\.45830\.50000\.4714Undersample0\.30000\.33330\.37500\.30000\.25910\.50000\.55560\.34500\.50000\.40710\.50000\.52780\.46330\.50000\.4593Lung Can\.ALMTSTRNone0\.60000\.63890\.58000\.60000\.55830\.80000\.80560\.82500\.80000\.79710\.70000\.69440\.74000\.70000\.7067Class Weight0\.70000\.72220\.74670\.70000\.69360\.60000\.58330\.65000\.60000\.60000\.60000\.61110\.60000\.60000\.6000SMOTE0\.60000\.63890\.58000\.60000\.55830\.60000\.58330\.65000\.60000\.60000\.60000\.58330\.65000\.60000\.6000Undersample0\.40000\.44440\.30000\.40000\.34290\.50000\.52780\.48330\.50000\.48570\.40000\.44440\.30000\.40000\.3429TRTSNone0\.50000\.50000\.54170\.50000\.51070\.62500\.58330\.62500\.62500\.62500\.50000\.41670\.37500\.50000\.4250Class Weight0\.50000\.50000\.54170\.50000\.51070\.75000\.66670\.83330\.75000\.73330\.37500\.33330\.28330\.37500\.3222SMOTE0\.25000\.33330\.18750\.25000\.20830\.62500\.58330\.62500\.62500\.62500\.62500\.58330\.67500\.62500\.6250Undersample0\.37500\.33330\.55000\.37500\.40480\.50000\.50000\.66670\.50000\.53330\.50000\.50000\.68750\.50000\.5417Lung CancerRTFTSTRNone0\.20000\.19440\.40000\.20000\.25000\.50000\.52780\.46330\.50000\.45930\.50000\.50000\.51000\.50000\.4978Class Weight0\.20000\.19440\.38000\.20000\.23890\.50000\.52780\.46330\.50000\.45930\.40000\.41670\.40830\.40000\.4000SMOTE0\.50000\.50000\.61000\.50000\.49920\.50000\.52780\.45830\.50000\.47140\.40000\.41670\.40000\.40000\.4000Undersample0\.20000\.19440\.18000\.20000\.18890\.50000\.52780\.45830\.50000\.47140\.40000\.41670\.40830\.40000\.4000TRTSNone0\.37500\.44440\.31250\.37500\.31670\.62500\.66670\.62500\.62500\.61430\.75000\.77780\.75000\.75000\.7500Class Weight0\.37500\.44440\.31250\.37500\.31670\.62500\.66670\.62500\.62500\.61430\.75000\.77780\.75000\.75000\.7500SMOTE0\.50000\.55560\.37500\.50000\.41670\.75000\.77780\.82290\.75000\.70890\.75000\.77780\.75000\.75000\.7500Undersample0\.37500\.44440\.44640\.37500\.29860\.87500\.88890\.90620\.87500\.87140\.62500\.66670\.60420\.62500\.6000ThyroidTRTRNone0\.98320\.94350\.98330\.98320\.98330\.96290\.79020\.96100\.96290\.95930\.94610\.70210\.93990\.94610\.9383Class Weight0\.97790\.94060\.97890\.97790\.97830\.95320\.72170\.95050\.95320\.94600\.83920\.83920\.93190\.83920\.8706SMOTE0\.96380\.92770\.96890\.96380\.96550\.95580\.87600\.95900\.95580\.95710\.85950\.81860\.92820\.85950\.8838Undersample0\.91700\.93920\.95710\.91700\.92910\.86480\.91630\.94880\.86480\.89080\.81890\.81230\.92600\.81890\.8557ThyroidALMTSTRNone0\.85070\.74530\.91260\.85070\.87460\.91080\.52490\.87330\.91080\.88800\.91520\.49570\.85160\.91520\.8823Class Weight0\.92230\.49950\.85210\.92230\.88580\.92310\.50000\.85220\.92310\.88630\.90550\.52730\.87170\.90550\.8859SMOTE0\.82690\.57960\.87640\.82690\.84910\.92310\.50000\.85220\.92310\.88630\.92310\.50000\.85220\.92310\.8863Undersample0\.83750\.85400\.93510\.83750\.86980\.85690\.80130\.92450\.85690\.88130\.79150\.70270\.90220\.79150\.8334TRTSNone0\.95360\.89850\.96640\.95360\.95810\.94590\.50000\.89480\.94590\.91960\.13360\.54200\.94910\.13360\.1526Class Weight0\.96140\.88340\.96760\.96140\.96380\.94590\.50000\.89480\.94590\.91960\.09490\.52160\.94900\.09490\.0841SMOTE0\.91830\.48540\.89330\.91830\.90560\.95030\.60820\.93850\.95030\.93930\.19650\.57530\.94930\.19650\.2539Undersample0\.85100\.89240\.95600\.85100\.88730\.94040\.62220\.92970\.94040\.93400\.17770\.56530\.94930\.17770\.2250ThyroidRTFTSTRNone0\.90280\.65230\.90180\.90280\.90230\.92310\.50000\.85220\.92310\.88630\.81270\.67730\.89690\.81270\.8463Class Weight0\.91080\.60390\.89530\.91080\.90180\.92310\.50000\.85220\.92310\.88630\.58480\.72770\.92030\.58480\.6774SMOTE0\.87190\.53020\.86700\.87190\.86940\.92310\.50000\.85220\.92310\.88630\.92310\.50000\.85220\.92310\.8863Undersample0\.83480\.89470\.94400\.83480\.86910\.66700\.80910\.93460\.66700\.74430\.22350\.57940\.93010\.22350\.2658TRTSNone0\.96360\.91260\.96990\.96360\.96580\.94040\.50910\.94400\.94040\.91260\.12800\.53580\.94320\.12800\.1331Class Weight0\.96690\.89730\.97020\.96690\.96820\.93930\.50000\.88230\.93930\.90990\.09930\.52060\.94310\.09930\.0814SMOTE0\.91390\.48650\.88080\.91390\.89700\.94040\.50910\.94400\.94040\.91260\.16000\.55290\.94340\.16000\.1873Undersample0\.86530\.91980\.95670\.86530\.89540\.92940\.57970\.91200\.92940\.91900\.16670\.55640\.94340\.16670\.1982Liver Dis\.TRTRNone0\.64420\.64920\.65800\.64420\.64630\.76920\.76060\.76820\.76920\.76840\.72120\.70680\.71890\.72120\.7183Class Weight0\.56730\.57350\.58450\.56730\.56950\.75960\.75230\.75900\.75960\.75920\.66350\.67200\.68180\.66350\.6652SMOTE0\.57690\.57580\.58550\.57690\.57930\.74040\.73860\.74350\.74040\.74130\.66350\.67200\.68180\.66350\.6652Undersample0\.65380\.68180\.71720\.65380\.64770\.70190\.70230\.70820\.70190\.70350\.66350\.67200\.68180\.66350\.6652Liver Dis\.ALMTSTRNone0\.46150\.45150\.46470\.46150\.46300\.52880\.52200\.53320\.52880\.53060\.55770\.51360\.53000\.55770\.5182Class Weight0\.54810\.54170\.55230\.54810\.54970\.51920\.49240\.50390\.51920\.50690\.55770\.55910\.56950\.55770\.5603SMOTE0\.53850\.53940\.55030\.53850\.54120\.49040\.48560\.49790\.49040\.49300\.55770\.55910\.56950\.55770\.5603Undersample0\.51920\.51970\.53110\.51920\.52210\.53850\.54550\.55700\.53850\.54050\.56730\.57650\.58840\.56730\.5689TRTSNone0\.49400\.48900\.50160\.49400\.49670\.55420\.53720\.54920\.55420\.55100\.49400\.46580\.47630\.49400\.4808Class Weight0\.46990\.47980\.49180\.46990\.47080\.56630\.55150\.56300\.56630\.56430\.39760\.39020\.40580\.39760\.4008SMOTE0\.50600\.49940\.51170\.50600\.50820\.55420\.54110\.55250\.55420\.55330\.40960\.40450\.41950\.40960\.4131Undersample0\.48190\.48240\.49510\.48190\.48510\.57830\.57350\.58350\.57830\.58020\.40960\.40060\.41590\.40960\.4122Liver Dis\.RTFTSTRNone0\.53850\.53940\.55030\.53850\.54120\.55770\.54390\.55510\.55770\.55620\.61540\.54550\.76920\.61540\.5032Class Weight0\.49040\.48560\.49790\.49040\.49300\.56730\.55530\.56600\.56730\.56660\.59620\.61060\.62510\.59620\.5962SMOTE0\.45190\.45530\.46770\.45190\.45470\.51920\.51060\.52210\.51920\.52050\.59620\.59550\.60460\.59620\.5984Undersample0\.50960\.52050\.53260\.50960\.51030\.53850\.54850\.56070\.53850\.53970\.67310\.67730\.68490\.67310\.6750TRTSNone0\.54220\.53070\.53920\.54220\.54040\.57830\.55290\.56730\.57830\.56240\.62650\.60850\.62130\.62650\.6193Class Weight0\.53010\.51680\.52560\.53010\.52720\.59040\.57000\.58240\.59040\.58080\.60240\.59370\.60120\.60240\.6017SMOTE0\.51810\.51270\.52120\.51810\.51930\.59040\.57000\.58240\.59040\.58080\.59040\.58630\.59320\.59040\.5914Undersample0\.55420\.56090\.56970\.55420\.55540\.57830\.56910\.57700\.57830\.57760\.59040\.58630\.59320\.59040\.5914Heart Fail\.TRTRNone0\.78890\.74480\.78440\.78890\.78570\.81110\.76120\.80650\.81110\.80590\.82220\.76940\.81870\.82220\.8161Class Weight0\.75560\.70210\.74840\.75560\.75040\.84440\.80380\.84170\.84440\.84120\.78890\.75380\.78710\.78890\.7879SMOTE0\.77780\.74560\.77780\.77780\.77780\.85560\.83920\.85700\.85560\.85620\.80000\.78010\.80420\.80000\.8017Undersample0\.72220\.74080\.76950\.72220\.73110\.78890\.80810\.82490\.78890\.79530\.77780\.76370\.78770\.77780\.7812Heart Fail\.ALMTSTRNone0\.66670\.60030\.65520\.66670\.65960\.73330\.61330\.73200\.73330\.69050\.75560\.62070\.82030\.75560\.6995Class Weight0\.67780\.56330\.64270\.67780\.63920\.70000\.54350\.71450\.70000\.61190\.65560\.64640\.68700\.65560\.6650SMOTE0\.64440\.63820\.68030\.64440\.65470\.72220\.65040\.70870\.72220\.71070\.70000\.67920\.71520\.70000\.7055Undersample0\.50000\.52260\.58400\.50000\.51450\.62220\.63990\.68540\.62220\.63460\.58890\.57910\.63040\.58890\.6013TRTSNone0\.56940\.52450\.59570\.56940\.58010\.66670\.58180\.65020\.66670\.65640\.58330\.49640\.57240\.58330\.5775Class Weight0\.56940\.49910\.57490\.56940\.57210\.63890\.56180\.62970\.63890\.63380\.58330\.56000\.62460\.58330\.5973SMOTE0\.59720\.54450\.61240\.59720\.60380\.61110\.54180\.61110\.61110\.61110\.61110\.55450\.62090\.61110\.6156Undersample0\.55560\.56550\.63120\.55560\.57300\.59720\.55730\.62240\.59720\.60720\.55560\.54000\.60850\.55560\.5720Heart Fail\.RTFTSTRNone0\.54440\.50110\.56420\.54440\.55280\.65560\.50170\.56730\.65560\.56970\.67780\.50900\.62320\.67780\.5666Class Weight0\.46670\.39850\.47590\.46670\.47110\.66670\.50080\.56700\.66670\.56050\.57780\.58900\.64100\.57780\.5916SMOTE0\.51110\.48560\.55100\.51110\.52520\.60000\.48780\.54970\.60000\.56570\.53330\.52910\.58830\.53330\.5485Undersample0\.64440\.66530\.70770\.64440\.65600\.56670\.58990\.64390\.56670\.58020\.65560\.66450\.70410\.65560\.6665TRTSNone0\.54170\.51350\.60910\.54170\.56400\.66670\.56920\.65640\.66670\.66100\.65280\.52880\.62480\.65280\.6359Class Weight0\.59720\.52120\.61520\.59720\.60530\.68060\.59420\.67570\.68060\.67800\.62500\.55580\.64200\.62500\.6325SMOTE0\.58330\.58850\.66820\.58330\.60490\.66670\.60000\.67700\.66670\.67130\.65280\.57500\.65810\.65280\.6553Undersample0\.48610\.55190\.64810\.48610\.50360\.56940\.54810\.63570\.56940\.59040\.65280\.60580\.67970\.65280\.6633PIR Vis\.TRTRNone0\.98740\.97980\.98740\.98740\.98740\.99170\.98160\.99170\.99170\.99170\.95990\.90300\.96030\.95990\.9598Class Weight0\.98520\.97660\.98530\.98520\.98520\.99300\.98440\.99300\.99300\.99300\.93340\.93410\.95090\.93340\.9381SMOTE0\.98130\.97960\.98230\.98130\.98160\.99000\.99010\.99030\.99000\.99010\.93950\.93190\.95350\.93950\.9433Undersample0\.97870\.97740\.98010\.97870\.97910\.98040\.98620\.98240\.98040\.98090\.92900\.93230\.94900\.92900\.9344PIR Vis\.ALMTSTRNone0\.98390\.97840\.98430\.98390\.98400\.99090\.98360\.99090\.99090\.99080\.97300\.92320\.97400\.97300\.9728Class Weight0\.98690\.98310\.98720\.98690\.98700\.99170\.98740\.99180\.99170\.99170\.93950\.93280\.95700\.93950\.9437SMOTE0\.98560\.98950\.98680\.98560\.98590\.98820\.98710\.98860\.98820\.98830\.95780\.94020\.96600\.95780\.9595Undersample0\.97390\.98470\.97800\.97390\.97490\.97950\.98470\.98160\.97950\.98010\.94250\.93760\.95820\.94250\.9462TRTSNone0\.97710\.96220\.97690\.97710\.97690\.98860\.98310\.98850\.98860\.98850\.91340\.82350\.91630\.91340\.9126Class Weight0\.97390\.95390\.97350\.97390\.97340\.98200\.96640\.98190\.98200\.98180\.87150\.83490\.89970\.87150\.8784SMOTE0\.97330\.96150\.97310\.97330\.97320\.98420\.97820\.98420\.98420\.98420\.88780\.84460\.91030\.88780\.8928Undersample0\.97820\.97250\.97840\.97820\.97830\.97880\.98550\.98050\.97880\.97920\.87150\.83850\.90810\.87150\.8798PIR Vis\.RTFTSTRNone0\.98260\.96750\.98250\.98260\.98250\.98910\.97480\.98910\.98910\.98900\.96080\.87130\.96110\.96080\.9596Class Weight0\.98130\.96460\.98110\.98130\.98120\.99000\.97860\.98990\.99000\.98990\.94990\.90470\.95810\.94990\.9512SMOTE0\.98130\.96920\.98140\.98130\.98130\.99040\.98800\.99060\.99040\.99050\.95640\.89740\.96100\.95640\.9567Undersample0\.97080\.98000\.97530\.97080\.97200\.98340\.98520\.98460\.98340\.98380\.95910\.90790\.96480\.95910\.9597TRTSNone0\.98200\.96300\.98180\.98200\.98190\.99020\.98410\.99020\.99020\.99020\.95320\.87080\.95500\.95320\.9533Class Weight0\.97930\.96330\.97930\.97930\.97930\.98640\.96920\.98620\.98640\.98620\.91340\.88250\.93760\.91340\.9203SMOTE0\.97440\.96880\.97590\.97440\.97490\.98860\.98200\.98870\.98860\.98860\.92320\.88050\.94140\.92320\.9280Undersample0\.98310\.97830\.98370\.98310\.98330\.98530\.98660\.98620\.98530\.98560\.91120\.86830\.93890\.91120\.9186
## Appendix BReproducibility and Computational Setup

To ensure the full reproducibility of our empirical findings, all experiments, model training, and evaluation pipelines were executed in a strictly controlled and documented computational environment\.

##### Hardware Infrastructure\.

All generative model training, both the Augmented Lagrangian Method constraint optimizations and the baseline autoregressive/GAN training, and downstream adversarial evaluations were accelerated using NVIDIA GPU A10 hardware \(24GB VRAM\), an AMD EPYC 7742 64\-Core Processor, and 256 GB of system RAM\.

##### Software Dependencies\.

The PSyGenTAB framework and all baseline models were implemented in Python 3\.10\. The core deep learning architectures and ALM optimization loops were built usingPyTorch\(v2\.1\.0\) with CUDA 12\.2 support\. Tabular data preprocessing, imputation, and transformation were handled viapandas\(v2\.1\.1\) andscikit\-learn\(v1\.3\.1\)\.

For the baseline generative architectures, we utilized the official open\-source implementations of RealTabFormer \(via thetransformerslibrary v4\.34\.0\) and CTAB\-GAN\+\.

##### Evaluation Libraries\.

The multi\-dimensional evaluation metrics were computed using standardized, community\-vetted libraries to ensure fairness and consistency:

- •Statistical Quality and Fidelity:Computed using theSDMetrics\(v0\.12\.0\) library for open\-source structural checks \(e\.g\., KS Test, TVD\) and themostlyai\-qapackage for commercial\-grade multivariate accuracy scoring\.
- •Downstream ML Utility \(TSTR/TRTS\):Implemented usingscikit\-learn\. Decision Trees, Random Forests, and Logistic Regression models were instantiated with default hyperparameters to evaluate the inherent signal of the data rather than model\-tuning prowess\. Imbalance strategies utilized theimbalanced\-learn\(v0\.11\.0\) library \(specifically for SMOTE and RandomUnderSampler\)\.
- •Privacy and Adversarial Risk:Distance to Closest Record \(DCR\), Exact Match Ratios, and Membership Inference Attacks \(MIA\) were evaluated using a combination of custom vectorized distance calculations \(for computational efficiency on large sets like Diabetes\) and theAnonymeterprivacy evaluation suite\.

##### Code and Data Availability\.

To facilitate community adoption and verify our claims, the complete source code for PSyGenTAB—including the ALM objective functions, the FAITH validation scripts, and the hyperparameter configurations for all datasets were made publicly available\. Repository:[https://github\.com/ArshiaIlaty/PsyGenTAB](https://github.com/ArshiaIlaty/PsyGenTAB)