Cross-Domain Feature Expansion for Tabular Medical Data via Knowledge Graphs Injection

# Cross-Domain Feature Expansion for Tabular Medical Data via Knowledge Graphs Injection
Source: [https://arxiv.org/html/2606.31171](https://arxiv.org/html/2606.31171)
Mengying Zhou1Yongjie Yin2Haoyan Xin3Guoping Liu4Yang Chen2 1School of Computing and Artificial Intelligence, Shanghai University of Finance and Economics 2College of Computer Science and Artificial Intelligence, Fudan University 3Independent Researcher 4School of Basic Medical Sciences, Shanghai University of Traditional Chinese Medicine and Pharmacology

###### Abstract

Acquiring comprehensive cross\-domain biomedical profiles is often costly and time\-consuming, resulting in severe data scarcity in medical research\. To address this challenge, we propose MedKGTab, a knowledge\-injected framework specifically engineered for cross\-domain feature expansion in tabular medical data\. MedKGTab seeks to infer uncollected biomedical features from available ones by exploiting their inherent statistical dependencies and established medical correlations\. By employing a row\-column dual\-attention mechanism, MedKGTab operates directly on raw structured tabular data, inherently capturing exact numerical distributions without the structural loss caused by tokenization\. Crucially, MedKGTab integrates data\-driven statistical priors with the SPOKE biomedical knowledge graph, achieving an optimal synergy between the data and knowledge channels\. Within this synergy, the representations derived from the data channel are modulated by the injected biomedical knowledge, ensuring the final generated data are grounded in empirical medical research\. Experimental results demonstrate that MedKGTab achieves high data fidelity and realistic data representation in cross\-domain feature expansion\. It outperforms both SOTA medical large models \(e\.g\., Baichuan M3\-plus\) and specialized tabular models designed for medical data generation\. Furthermore, MedKGTab consistently delivers superior performance across various data generation scenarios, whether inferring missing features within the same dataset or generalizing across different medical cohorts\.

## 1Introduction

Biomedical profiles are essential for medical research since they provide important information for disease characterization, diagnosis, and treatment response analysis\. However, acquiring comprehensive cross\-domain profiles is prohibitively costly and time\-consuming, resulting in severe data scarcity, as depicted in Fig\.[1](https://arxiv.org/html/2606.31171#S1.F1)\(a\)\. Cross\-domain feature expansion in tabular medical data offers a promising generative solution to infer uncollected features from available ones by exploiting their inherent dependenciesSunet al\.\([2025](https://arxiv.org/html/2606.31171#bib.bib31)\); Wanget al\.\([2026](https://arxiv.org/html/2606.31171#bib.bib32)\)\. This generative workflow \(Fig\.[1](https://arxiv.org/html/2606.31171#S1.F1)\(b\)\) can rapidly and cost\-effectively synthesize a large volume of high\-fidelity biomedical profiles, which enables broader downstream applications, ranging from basic disease predictive modeling to complex tasks like disease\-drug interaction reasoning\. However, achieving reliable feature expansion across diverse biological domains remains exceptionally difficult due to small sample sizes, severe data sparsity, and complex underlying biological mechanismsAthieniti and Spyrou \([2023](https://arxiv.org/html/2606.31171#bib.bib29)\); Maniet al\.\([2025](https://arxiv.org/html/2606.31171#bib.bib28)\); Tarazonaet al\.\([2021](https://arxiv.org/html/2606.31171#bib.bib30)\)\.

Existing tabular data generation methods struggle to address these challenges, as they are primarily designed for intra\-domain synthesisChenet al\.\([2022](https://arxiv.org/html/2606.31171#bib.bib21)\); Gorishniyet al\.\([2021](https://arxiv.org/html/2606.31171#bib.bib9)\); Yeet al\.\([2025](https://arxiv.org/html/2606.31171#bib.bib19)\)\. To model complex cross\-domain dependencies, recent efforts have explored injecting external knowledge graphs \(KGs\) or developing specialized medical large language models \(LLMs\), but both paradigms face inherent limitations\. KG injection in tabular data generation models often serves merely as a superficial structural constraint\. The impact of this injected knowledge is easily diluted, and its strength lacks flexibility across varying tasksChoiet al\.\([2017](https://arxiv.org/html/2606.31171#bib.bib37)\)\. Meanwhile, although LLMs natively possess vast internal knowledge, they are fundamentally ill\-suited for numerical tabular data\. By relying on tokenization, LLMs fail to preserve the structural relationships between biological entitiesGrinsztajnet al\.\([2022](https://arxiv.org/html/2606.31171#bib.bib1)\), leading to severe loss of structural information and an inability to capture authentic numerical distributions\. Furthermore, the computational costs associated with updating model parameters in LLM\-based approaches remain prohibitively highLi and Liang \([2021](https://arxiv.org/html/2606.31171#bib.bib35)\); Wanget al\.\([2025](https://arxiv.org/html/2606.31171#bib.bib36)\)\.

![Refer to caption](https://arxiv.org/html/2606.31171v1/x1.png)Figure 1:Comparison between the traditional data collection workflow and our knowledge\-injected cross\-domain feature expansion workflow\.To overcome the limitations of both diluted KG injection and structure\-agnostic LLMs, we propose MedKGTab, a knowledge\-injected framework specifically designed for cross\-domain medical feature expansion\. By employing a row\-column dual\-attention mechanism, MedKGTab operates directly on raw structured tabular data, inherently capturing exact numerical distributions without tokenization\. More importantly, MedKGTab integrates data\-driven statistical priors with the SPOKE biomedical knowledge graph\. Unlike previous methods, which easily dilute knowledge, MedKGTab achieves optimal synergy between the data and knowledge channels\. By injecting feature\-level graph priors into the attention module as a target\-specific bias, it explicitly models inter\-domain biomedical associations\. This mechanism ensures the model receives vital biomedical guidance without overriding its intrinsic data representations, successfully preserving both tabular and graph structural information\. In summary, our main contributions are as follows:

- •We design a lightweight, controllable knowledge injection mechanism that overcomes tabular\-graph misalignments\. It effectively injects biomedical priors without the high computational cost of updating pre\-trained parameters\.
- •We propose MedKGTab, a framework that optimally synergizes data and knowledge channels to provide essential biomedical guidance without overriding intrinsic data representations\.
- •Extensive experiments demonstrate MedKGTab’s superior data fidelity and realistic data representation\. By capturing numerical distributions and preserving biomedical relationships, it outperforms classical models and specialized medical LLMs \(e\.g\., Baichuan M3\-plus\) across both intra\- and cross\-cohort settings\.

## 2Related Work

### 2\.1Tabular Data Generation

Synthetic tabular data generation has been widely studied for data augmentation, privacy preservation, and learning under limited supervision\. Early deep generative models, such as CTGAN, TVAEXuet al\.\([2019](https://arxiv.org/html/2606.31171#bib.bib2)\), and TabDDPMKotelnikovet al\.\([2023](https://arxiv.org/html/2606.31171#bib.bib4)\), effectively model mixed numerical and categorical features within a unified framework\. However, they primarily focus on within\-domain generation and struggle with cross\-domain medical features\. To improve the generalization of tabular models, recent studies have explored pretrained and transferable modeling methods\. TabPFNHollmannet al\.\([2025](https://arxiv.org/html/2606.31171#bib.bib5)\)learns transferable priors through pretraining on more than 100 million synthetic tabular tasks, enabling strong performance in low\-data settings\. FT\-TransformerGorishniyet al\.\([2021](https://arxiv.org/html/2606.31171#bib.bib9)\)and TabRGorishniyet al\.\([2024](https://arxiv.org/html/2606.31171#bib.bib6)\)further improve transferability through stronger architectures or retrieval\-based mechanisms\. MediTabWanget al\.\([2024](https://arxiv.org/html/2606.31171#bib.bib7)\)improves transferability through data consolidation and task alignment across heterogeneous medical tables, while MedTransTabChenet al\.\([2025](https://arxiv.org/html/2606.31171#bib.bib8)\)leverages medical LLM\.

Despite these advances, existing methods heavily emphasize predictive performance or statistical fidelity\. They do not explicitly incorporate expert biomedical knowledge to reliably govern cross\-domain feature dependencies, making them suboptimal for rigorous medical feature expansion\.

### 2\.2Medical Large Language Models

Benefiting from pretraining on diverse clinical datasets, medical LLMs have demonstrated strong performance on biomedical reasoning\. Representative models such as Med\-PaLM 2Singhalet al\.\([2023](https://arxiv.org/html/2606.31171#bib.bib10)\)and Med\-PaLM MTuet al\.\([2023](https://arxiv.org/html/2606.31171#bib.bib11)\)demonstrate effectiveness on medical question answering, long\-form response generation, and multimodal biomedical understanding\. Open medical LLMs, including ChatDoctorLiet al\.\([2023](https://arxiv.org/html/2606.31171#bib.bib12)\), BioMistralLabraket al\.\([2024](https://arxiv.org/html/2606.31171#bib.bib13)\), MEDITRONChenet al\.\([2023](https://arxiv.org/html/2606.31171#bib.bib14)\), and Baichuan\-M3Douet al\.\([2026](https://arxiv.org/html/2606.31171#bib.bib15)\), further improve medical reasoning and complex clinical inference through instruction tuning and domain\-specific knowledge injection\.

Although rich in medical knowledge, medical LLMs are designed for unstructured or semi\-structured sequences\. Applying them directly to tabular feature expansion is severely constrained by limited context windows and a fundamental inability to model complex statistical distributions and quantitative inter\-feature correlations\.

## 3Methodology

In the real world, acquiring comprehensive biomedical data is often prohibitively time\-consuming and expensive\. Consequently, cross\-domain feature expansion seeks to infer uncollected target\-domain features from readily available source\-domain features\. Although recent tabular foundation models provide strong transferable priors for small\-sample tabular learning, they are prone to learning unstable source\-target correlations from limited paired data, often resulting in medically implausible outputs\. To address this issue, we propose MedKGTab, a knowledge\-injected tabular framework for cross\-domain medical feature expansion, as illustrated in Fig\.[2](https://arxiv.org/html/2606.31171#S3.F2)\. Specifically, we derive a feature\-level biomedical prior from an external biomedical knowledge graph and explicitly inject it into the feature attention mechanism of a pre\-trained tabular foundation model\. Through this synergistic integration, target\-domain feature expansion is governed not only by empirical data statistics but also modulated by established biomedical knowledge\.

![Refer to caption](https://arxiv.org/html/2606.31171v1/x2.png)Figure 2:Overall framework of MedKGTab### 3\.1Problem Formulation

For multi\-center cross\-domain biomedical feature expansion, let the complete global feature space be composed ofDDdistinct clinical variables\. In real\-world clinical practice, different institutions often record disjoint feature domains\. For instance, as illustrated in Fig\.[2](https://arxiv.org/html/2606.31171#S3.F2), the data from Hospital 1 may contain domain A but lack domain B, whereas Hospital 2 exhibits the inverse pattern\.

Suppose we are provided with datasets from multiple distinct cohorts\. For thekk\-th cohort comprisingN\(k\)N^\{\(k\)\}patients, only a specific subset ofdkd\_\{k\}features is collected, yielding the corresponding data matrix:

𝒟\(k\)=𝐗obs\(k\)∈ℝN\(k\)×dk\.\\mathcal\{D\}^\{\(k\)\}=\\mathbf\{X\}\_\{\\mathrm\{obs\}\}^\{\(k\)\}\\in\\mathbb\{R\}^\{N^\{\(k\)\}\\times d\_\{k\}\}\.The complementary feature domains for these patients remain uncollected, denoted as the missing data𝐗mis\(k\)\\mathbf\{X\}\_\{\\mathrm\{mis\}\}^\{\(k\)\}\.

The overall task is to leverage the aggregated available data across all centers to accurately infer the uncollected feature blocks𝐗^mis\(k\)\\hat\{\\mathbf\{X\}\}\_\{\\mathrm\{mis\}\}^\{\(k\)\}for all cohortskk, thereby systematically reconstructing a fully complete cross\-domain dataset\.

### 3\.2Data Channel

As depicted in the top\-left panel of Fig\.[2](https://arxiv.org/html/2606.31171#S3.F2), the data channel is designed to structure multi\-center datasets into a unified prior context and iterative generation targets\.

First, we aggregate all collected multi\-domain data across cohorts to construct a unified context\. Next, to cast the multi\-target feature expansion as a sequence of tractable inferences, this whole generation task is partitioned into feature\-level sub\-tasks \(e\.g\., features B1 to Bnnfor Hospital 1\)\. For thett\-th missing feature, we construct its corresponding input for inference:

𝒟tsub=\(𝐗context,𝐘:,tobs\),\\mathcal\{D\}^\{\\mathrm\{sub\}\}\_\{t\}=\(\\mathbf\{X\}^\{\\mathrm\{context\}\},\\mathbf\{Y\}^\{\\mathrm\{obs\}\}\_\{:,t\}\),where𝐗context\\mathbf\{X\}^\{\\mathrm\{context\}\}is the aggregated context and𝐘:,tobs\\mathbf\{Y\}^\{\\mathrm\{obs\}\}\_\{:,t\}represents the available data for this feature from other cohorts\. MedKGTab then traverses the uncollected feature list to sequentially infer data for each missing feature\.

### 3\.3Knowledge Channel

As illustrated in the lower\-left panel of Fig\.[2](https://arxiv.org/html/2606.31171#S3.F2), the knowledge channel derives a sample\-independent biomedical prior from an external knowledge graph\. In this study, we use SPOKEMorriset al\.\([2023](https://arxiv.org/html/2606.31171#bib.bib17)\), a large\-scale knowledge graph that integrates diverse entities and relations from curated biomedical resources, providing a structured and reliable basis for constraining feature interactions in medical tabular data\.

To align the tabular domain with this graph without semantic ambiguity, we strictly map source and target features to task\-relevant node types \(e\.g\., mapping metabolites toCompoundnodes and microbial genera toOrganismnodes\)\. Each tabular feature is uniquely mapped to at most one corresponding SPOKE node\.

Given the inherent sparsity of biomedical networks, relying solely on direct edges is insufficient\. Therefore, we perform a topology\-to\-matrix projection that incorporates closed one\-hop neighborhood information, denoted as𝒩​\(⋅\)\\mathcal\{N\}\(\\cdot\)\. We project this enriched topology into a feature\-level knowledge matrixA∈ℝ\(dx\+dy\)×\(dx\+dy\)A\\in\\mathbb\{R\}^\{\(d\_\{x\}\+d\_\{y\}\)\\times\(d\_\{x\}\+d\_\{y\}\)\}:

Ai​j=\{1,if​i≠j​and​𝒩​\(vi\)∩𝒩​\(vj\)≠∅,0,otherwise,A\_\{ij\}=\\begin\{cases\}1,&\\text\{if \}i\\neq j\\text\{ and \}\\mathcal\{N\}\(v\_\{i\}\)\\cap\\mathcal\{N\}\(v\_\{j\}\)\\neq\\varnothing,\\\\ 0,&\\text\{otherwise\},\\end\{cases\}whereviv\_\{i\}andvjv\_\{j\}denote the corresponding SPOKE nodes for successfully mapped featuresiiandjj\. Accordingly,Ai​j=1A\_\{ij\}=1if the two features share at least one conceptual node in their immediate neighborhoods\. For unmapped \(out\-of\-vocabulary\) features, their corresponding rows and columns inAAremain zero\. This projection ensures thatAArobustly captures feature dependencies fundamentally grounded in shared biomedical topology\.

To finalize the prior for model injection, we add self\-loops to preserve self\-information and apply symmetric normalization:

A~=D−12​\(A\+I\)​D−12,\\tilde\{A\}=D^\{\-\\frac\{1\}\{2\}\}\(A\+I\)D^\{\-\\frac\{1\}\{2\}\},whereIIis the identity matrix andDDis the degree matrix ofA\+IA\+I\. The resulting normalized matrixA~\\tilde\{A\}serves as an interpretable and robust structural prior, mitigating the instability of empirical correlations estimated solely from limited paired tabular data\.

### 3\.4Knowledge\-Injected Dual Attention Inference

#### 3\.4\.1Dual Attention Mechanism

This module serves as the core inference engine, bridging the upstream data and knowledge channels\. We employ TabPFNHollmannet al\.\([2025](https://arxiv.org/html/2606.31171#bib.bib5)\)as the backbone, leveraging its strong transferable priors cultivated through extensive pre\-training, making it exceptionally well\-suited for medical scenarios with limited supervision\. Unlike standard sequence\-to\-sequence Transformers, TabPFN processes tabular context via a dual attention mechanism \(as shown in the center panel of Fig\.[2](https://arxiv.org/html/2606.31171#S3.F2)\): aninter\-feature attentionthat captures column\-wise variable dependencies, and aninter\-sample attentionthat captures row\-wise patient relationships\.

Specifically, each TabPFN layer consists of a three\-stage sequential process: \(1\) inter\-feature attention, \(2\) inter\-sample attention, and \(3\) feed\-forward MLP, formulated as follows:

𝐇feat\(t,l\)\\displaystyle\\mathbf\{H\}^\{\(t,l\)\}\_\{\\mathrm\{feat\}\}=𝐇\(t,l\)\+Attnfeat​\(𝐇\(t,l\)\),\\displaystyle=\\mathbf\{H\}^\{\(t,l\)\}\+\\mathrm\{Attn\}\_\{\\mathrm\{feat\}\}\(\\mathbf\{H\}^\{\(t,l\)\}\),𝐇sample\(t,l\)\\displaystyle\\mathbf\{H\}^\{\(t,l\)\}\_\{\\mathrm\{sample\}\}=𝐇feat\(t,l\)\+Attnsample​\(𝐇feat\(t,l\)\),\\displaystyle=\\mathbf\{H\}^\{\(t,l\)\}\_\{\\mathrm\{feat\}\}\+\\mathrm\{Attn\}\_\{\\mathrm\{sample\}\}\(\\mathbf\{H\}^\{\(t,l\)\}\_\{\\mathrm\{feat\}\}\),𝐇\(t,l\+1\)\\displaystyle\\mathbf\{H\}^\{\(t,l\+1\)\}=𝐇sample\(t,l\)\+MLP​\(𝐇sample\(t,l\)\)\.\\displaystyle=\\mathbf\{H\}^\{\(t,l\)\}\_\{\\mathrm\{sample\}\}\+\\mathrm\{MLP\}\(\\mathbf\{H\}^\{\(t,l\)\}\_\{\\mathrm\{sample\}\}\)\.where𝐇\(t,l\)∈ℝN×D×h\\mathbf\{H\}^\{\(t,l\)\}\\in\\mathbb\{R\}^\{N\\times D\\times h\}denotes the hidden table representation at thell\-th layer for thett\-th feature generation sub\-task, withNNandDDrepresenting the sample and feature dimensions of𝒟tsub\\mathcal\{D\}^\{\\mathrm\{sub\}\}\_\{t\}, andhhis the hidden dimension\.

Our adopted TabPFN v2\.5 stacksL=18L=18such layers\. The final representation𝐇\(t,L\)\\mathbf\{H\}^\{\(t,L\)\}is then passed to the prediction head to infer the missing values for this specific feature across the required clinical cohorts\.

#### 3\.4\.2Explicit Knowledge Injection

Since our derived biomedical prior knowledge explicitly encodes the relationships between biological variables, we structurally inject this knowledge exclusively into the inter\-feature attention\. This design choice compels the model to prioritize biologically corroborated feature interactions, while preserving the purely data\-driven nature of the inter\-sample attention\.

To execute this targeted injection, we dynamically construct a task\-specific prior for each feature generation\. During the task iteration for a given targetyty\_\{t\}, we extract its corresponding source–source, source–target, and target–target relations from the global priorA~\\tilde\{A\}, forming the sub\-matrixA~\(t\)\\tilde\{A\}^\{\(t\)\}:

A~\(t\)=\[A~1:dx,1:dxA~1:dx,dx\+tA~dx\+t,1:dxA~dx\+t,dx\+t\]∈ℝ\(dx\+1\)×\(dx\+1\)\.\\tilde\{A\}^\{\(t\)\}=\\begin\{bmatrix\}\\tilde\{A\}\_\{1:d\_\{x\},\\;1:d\_\{x\}\}&\\tilde\{A\}\_\{1:d\_\{x\},\\;d\_\{x\}\+t\}\\\\ \\tilde\{A\}\_\{d\_\{x\}\+t,\\;1:d\_\{x\}\}&\\tilde\{A\}\_\{d\_\{x\}\+t,\\;d\_\{x\}\+t\}\\end\{bmatrix\}\\in\\mathbb\{R\}^\{\(d\_\{x\}\+1\)\\times\(d\_\{x\}\+1\)\}\.To match the dimensionality of TabPFN v2\.5’s internal logits, we apply mean pooling with a window size ofw=3w=3toA~\(t\)\\tilde\{A\}^\{\(t\)\}, yielding the final injectable matrixA~w\(t\)\\tilde\{A\}^\{\(t\)\}\_\{w\}\.

As TabPFN’s default preprocessing disrupts the correspondence between input and original biomedical features, directly injecting the knowledge matrix into TabPFN is infeasible\. To preserve this crucial alignment, we disable its internal feature expansion, fingerprint generation, and feature shuffling\. These targeted modifications guarantee the backbone’s feature indices remain perfectly aligned with our biomedical prior\.

Finally, the knowledge matrixA~w\(t\)\\tilde\{A\}^\{\(t\)\}\_\{w\}is added as a structural bias to the pre\-softmax logits of the inter\-feature attention\. This matrix is shared identically across all attention heads and broadcast across the batch dimension in every attention layer\. Formally, the knowledge\-injected inter\-feature attention is computed as:

Attnfeat​\(Q,K,V\)=softmax​\(Q​K⊤d\+α​A~w\(t\)\)​V,\\mathrm\{Attn\}\_\{\\mathrm\{feat\}\}\(Q,K,V\)=\\mathrm\{softmax\}\\left\(\\frac\{QK^\{\\top\}\}\{\\sqrt\{d\}\}\+\\alpha\\tilde\{A\}^\{\(t\)\}\_\{w\}\\right\)V,whereQQ,KK, andVVrepresent the query, key, and value matrices of the inter\-feature attention,ddis the scaling dimension, andα\\alphais a tunable hyperparameter controlling the injection strength\. By applyingA~w\(t\)\\tilde\{A\}^\{\(t\)\}\_\{w\}as a targeted attention bias, the model is explicitly encouraged to assign larger attention weights to features corroborated by established biomedical knowledge, thereby mitigating the risk of overfitting to spurious correlations in the limited paired data\.

### 3\.5Output and Downstream

The output module aggregates the target\-wise inferences to construct a complete tabular profile\. Once alldyd\_\{y\}target features are sequentially generated via the task iterations, we concatenate the predictions along the feature dimension:

𝐘^single=\[𝐘^:,1single,𝐘^:,2single,…,𝐘^:,dysingle\]\.\\hat\{\\mathbf\{Y\}\}^\{\\mathrm\{single\}\}=\[\\hat\{\\mathbf\{Y\}\}^\{\\mathrm\{single\}\}\_\{:,1\},\\hat\{\\mathbf\{Y\}\}^\{\\mathrm\{single\}\}\_\{:,2\},\\ldots,\\hat\{\\mathbf\{Y\}\}^\{\\mathrm\{single\}\}\_\{:,d\_\{y\}\}\]\.The final output,𝒟^single=\(𝐗single,𝐘^single\)\\hat\{\\mathcal\{D\}\}^\{\\mathrm\{single\}\}=\(\\mathbf\{X\}^\{\\mathrm\{single\}\},\\hat\{\\mathbf\{Y\}\}^\{\\mathrm\{single\}\}\), yields a large\-size synthetic dataset containing fully populated cross\-domain biomedical profiles, which is ready for diverse downstream clinical analysis\.

## 4Experiments

In this section, we evaluate MedKGTab under intra\-cohort and cross\-cohort feature expansion tasks, and conduct ablation studies to assess the contribution of knowledge injection\.

### 4\.1Experimental Setup

##### Datasets\.

We evaluate MedKGTab on two chronic atrophic gastritis \(CAG\) datasets, denoted asCAG\-TongueandCAG\-Fecal, both comprising paired metabolite and microbiota profiles\.CAG\-Tongueutilizes tongue\-coating samples collected from A Hospital \(anonymized for double\-blind review\)\. After excluding individuals with gastric polyps, hemorrhage, tumors, prior gastrointestinal resection, or special types of gastritis, 114 valid samples remained \(78 CAG patients and 36 healthy controls\)\. Each profile includes 432 microbial taxa and 120 metabolites, obtained via 16S rRNA sequencing and UPLC\-MS/MS, respectively\. To evaluate cross\-cohort generalization, we further incorporateCAG\-Fecal, a public fecal dataset derived from Gai et al\.Gaiet al\.\([2023](https://arxiv.org/html/2606.31171#bib.bib24)\)\. This dataset contains 33 samples \(17 CAG patients and 16 healthy controls\), including 147 microbiota and 192 metabolite features\. For all datasets used in this work, we leveraged them solely for secondary analysis and were not involved in their collection or in any clinical experimentation\.

##### Baselines\.

We benchmark MedKGTab against ten baselines across two categories: dedicated tabular models and LLM\-based approaches\. The dedicated tabular baselines include a classical machine learning model, XGBoostChen and Guestrin \([2016](https://arxiv.org/html/2606.31171#bib.bib27)\); deep learning\-based tabular models, including FT\-TransformerGorishniyet al\.\([2021](https://arxiv.org/html/2606.31171#bib.bib9)\), DANetsChenet al\.\([2022](https://arxiv.org/html/2606.31171#bib.bib21)\), SwitchTabWuet al\.\([2024](https://arxiv.org/html/2606.31171#bib.bib20)\), and modernNCAYeet al\.\([2025](https://arxiv.org/html/2606.31171#bib.bib19)\); and a raw tabular foundation model, TabPFNHollmannet al\.\([2025](https://arxiv.org/html/2606.31171#bib.bib5)\)\. The LLM\-based baselines cover three sub\-categories: two tabular generation methods, including Curated LLMSeedatet al\.\([2024](https://arxiv.org/html/2606.31171#bib.bib26)\)and TabGen\-ICLFanget al\.\([2025](https://arxiv.org/html/2606.31171#bib.bib22)\); a general LLM, DeepSeek v3\.2Liuet al\.\([2025a](https://arxiv.org/html/2606.31171#bib.bib42)\); and a SOTA medical LLM, Baichuan M3\-plusDouet al\.\([2026](https://arxiv.org/html/2606.31171#bib.bib15)\), with domain\-specialized medical knowledge\.

##### Implementation\.

To accommodate the context window limitations of LLMs, we restrict the input feature space to the top 50 metabolites and top 50 microbial taxa, selected based on feature importance derived from Neural Additive Models \(NAM\)Agarwalet al\.\([2021](https://arxiv.org/html/2606.31171#bib.bib25)\)\.

For the intra\-cohort setting, we simulate a data\-scarce clinical environment by partitioning the patient samples:40%40\\%of the samples are retained as collected data, while the remaining60%60\\%are masked as uncollected data to serve as the generation target\. Moreover, we conduct bidirectional inference experiments on intra\-cohort settings, treating both microbiota and metabolites as the uncollected target domain in alternating trials\. For the cross\-cohort setting \(i\.e\., across different medical institutions\), the entirety ofCAG\-Tonguealong with one observed domain fromCAG\-Fecalis used to infer the remaining uncollected domain inCAG\-Fecal\.

MedKGTab and all tabular model baselines are implemented in PyTorch and run on a single NVIDIA RTX 4090 GPU \(24 GB memory\)\. FT\-Transformer, DANets, SwitchTab, and modernNCA are implemented based on TALENTLiuet al\.\([2025b](https://arxiv.org/html/2606.31171#bib.bib18)\)with default hyperparameter configurations\. For Curated LLM and TabGen\-ICL, we use DeepSeek v3\.2 as the inference model\. Raw LLM baselines are queried via their respective APIs\. To ensure fairness, all approaches share the same prompt templates and strict input constraints, with details provided in Appendix[A\.1](https://arxiv.org/html/2606.31171#A1.SS1)\.

##### Evaluation Metrics\.

For the intra\-cohort setting, we report Mean Squared Error \(MSE\) andR2R^\{2\}to assess overall generative accuracy and data fidelity, Median Absolute Error \(MedAE\) to capture typical\-sample deviation, and the 95th percentile of absolute errors \(P95\) to evaluate robustness against extreme outliers\. HigherR2R^\{2\}values indicate better performance, whereas lower values are preferred for MSE, MedAE, and P95\.

For the cross\-cohort setting, we evaluate generative data fidelity using the Synthetic Data Vault \(SDV\) suitePatkiet al\.\([2016](https://arxiv.org/html/2606.31171#bib.bib23)\), specifically reporting the Column Distribution and Inter\-Column Relationship scores\. These two metrics quantify how closely the generated target\-domain features match the ground\-truth data in terms of marginal per\-feature distributions and inter\-feature dependencies, respectively\. Furthermore, we introduce an overall Average Score, calculated as the mean of these two SDV metrics\.

### 4\.2Performance Evaluation

Table 1:Performance comparison of intra\-cohort feature expansion onCAG\-Tongue\(generating microbiota from metabolites\)\.Boldandunderlinedvalues denote the best and second\-best results\.#### 4\.2\.1Intra\-cohort Setting

Table[1](https://arxiv.org/html/2606.31171#S4.T1)reports the intra\-cohort feature expansion results onCAG\-Tongue, where microbiota features are generated from metabolite inputs\. The results demonstrate that MedKGTab achieves the best overall performance, outperforming all baselines\.

Among the tabular methods, TabPFN yields the strongest baseline performance, ranking second on both MSE andR2R^\{2\}\. This indicates that tabular foundation models are competitive for intra\-cohort feature expansion\. However, MedKGTab consistently improves over TabPFN across most metrics, underscoring the necessity and benefit of injecting biomedical knowledge into tabular foundation modeling\. In contrast, LLM\-based baselines generally underperform compared with tabular methods, suggesting that relying solely on broad semantic and medical knowledge is insufficient for this task without effectively modeling tabular structural dependencies\.

Furthermore, MedKGTab achieves the lowest MedAE, indicating a higher fidelity of the generated features to the ground\-truth observations on typical patient samples\. Regarding P95, although MedKGTab is suboptimal, it remains highly competitive and close to the best\-performing baseline\. This confirms that MedKGTab successfully maintains control over high\-error tail cases while securing a significant advantage on typical samples\.

To ensure comprehensive evaluation, we also conducted bidirectional inference experiments \(i\.e\., expanding metabolite features using microbiota inputs\)\. These results further validate the effectiveness and robustness of MedKGTab, with detailed findings provided in Appendix[A\.2](https://arxiv.org/html/2606.31171#A1.SS2)\.

#### 4\.2\.2Cross\-cohort Setting

Table 2:Performance comparison of cross\-cohort feature expansion fromCAG\-TonguetoCAG\-Fecal\(generating microbiota from metabolites\)\.Boldandunderlinedvalues denote the best and second\-best results\. The Avg\. Score is the mean of Column Distribution and Inter\-Column Relationship\.Table[2](https://arxiv.org/html/2606.31171#S4.T2)reports the cross\-cohort feature expansion results\. In this subsection, the entirety ofCAG\-Tonguealong with the observed metabolite domain fromCAG\-Fecalis used to infer the uncollected microbiota domain inCAG\-Fecal\. Given the universally worse performance of LLM\-based methods observed in the intra\-cohort evaluation, we exclude them from this setting\.

The results show that MedKGTab achieves the best overall generation quality with the highest Average Score of 0\.597, outperforming all tabular baselines under severe distribution shifts\. Specifically, MedKGTab obtains the best Column Distribution score of 0\.432, substantially surpassing the strongest baseline, SwitchTab\. This highlights that the synthesized microbiota features produced by MedKGTab more accurately reflect the marginal distribution of the real target dataset\. Although MedKGTab does not achieve the absolute highest Inter\-column Relationship score, it remains highly competitive\.

### 4\.3Ablation Study

#### 4\.3\.1Effectiveness of Knowledge Graph

To validate the knowledge injection mechanism, we evaluate MedKGTab against its backbone,raw TabPFN, alongside three modified graph variants:Complete Graph,Random Graph, andSingle\-domain Graph\. This setup allows us to discuss the utility of external knowledge from the impacts of information volume and topological selection\.

Table 3:Ablation study on knowledge injection mechanism of MedKGTab on intra\-cohort setting\.Table[3](https://arxiv.org/html/2606.31171#S4.T3)shows that MedKGTab achieves the best performance\. Compared with its base TabPFN, MedKGTab yields significant improvements, demonstrating the overall efficacy of incorporating curated biomedical priors\. Second, the performance degradation in both the complete and random graph settings confirms that the superiority of the full MedKGTab stems from meaningful biomedical structures rather than volume\. Most notably, the single\-domain graph variant suffers the most severe performance drop, performing even worse than the raw TabPFN\. This phenomenon indicates that injecting only single\-domain external information misguides the representation, causing the model to overfit to local intra\-domain noise\. It strongly underscores that explicitly modeling inter\-domain biomedical associations is the critical key to preventing negative transfer and enabling robust cross\-domain feature expansion\.

#### 4\.3\.2Robustness to Incomplete Knowledge Graphs

In practice, curated biomedical knowledge graphs may provide incomplete coverage for rare diseases or under\-studied populations\. To simulate this scenario, we evaluate MedKGTab’s robustness by randomly dropping 5%, 10%, 20%, and 50% of nodes or edges before constructing the biomedical knowledge injection\.

![Refer to caption](https://arxiv.org/html/2606.31171v1/x3.png)Figure 3:Robustness of MedKGTab to incomplete knowledge graphs on intra\-cohort setting\.As illustrated in Fig\.[3](https://arxiv.org/html/2606.31171#S4.F3), edge removal induces a negligible impact, with the MSE remaining relatively flat \(around9\.71∼9\.78×10−49\.71\\sim 9\.78\\times 10^\{\-4\}\) across all dropout ratios\. Conversely, while node removal is stable at low dropout ratios \(5%–10%\), it exhibits a pronounced degradation in data fidelity from 20% dropout ratio, where the predictive error reaches its maximum, with the MSE peaking at11\.12×10−411\.12\\times 10^\{\-4\}\. This phenomenon suggests that MedKGTab remains effective even with an incomplete knowledge graph\. Maintaining the direct mapping between tabular features and graph nodes is fundamentally more important than preserving every internal relational edge\.

### 4\.4Hyperparameter Sensitivity Analysis

We analyze the sensitivity of MedKGTab to the parameterα\\alpha, which explicitly controls the proportion of the external biomedical prior injected into the tabular data representations\. As Fig\.[4](https://arxiv.org/html/2606.31171#S4.F4)shows, MedKGTab achieves the lowest MSE on the intra\-cohort setting atα=4\\alpha=4\. Settings ofα\\alphabelow 4 result in an insufficient injection of knowledge guidance, whereas larger values degrade performance, as excessive external knowledge overrides the useful information from the data channel\.

![Refer to caption](https://arxiv.org/html/2606.31171v1/x4.png)Figure 4:Sensitivity of MedKGTab to the graph injection strengthα\\alphaon the intra\-cohort setting\.To understand this phenomenon, we investigate the alignment between the biomedical prior from the knowledge graph and the learned information from the data channel\. For each TabPFN layer, we compute the cosine similarity between the normalized knowledge matrixA~w\\tilde\{A\}\_\{w\}and the corresponding feature attention matrix\. This metric quantifies the structural agreement between the injected external knowledge and the intrinsic data\-driven feature dependencies\. We define a consistency score as the proportion of layers with non\-negative cosine similarity\. This score acts as an indicator of network\-wide agreement: higher values mean the injected graph more closely matches the data\-driven attention patterns\. Fig\.[4](https://arxiv.org/html/2606.31171#S4.F4)shows that the consistency score initially rises but then declines, reaching its peak atα=4\\alpha=4\. This demonstrates that while moderate injection helps provide meaningful relationships, excessive external knowledge could conflict with intrinsic data patterns\.α=4\\alpha=4achieves an optimal synergy, providing essential biomedical guidance without overriding core information from the data channel\. A more detailed parameter analysis is provided in Appendix[A\.4](https://arxiv.org/html/2606.31171#A1.SS4)\.

## 5Conclusion

We present MedKGTab, a knowledge\-injected framework for cross\-domain medical feature expansion\. Unlike previous methods that inject superficial knowledge information, MedKGTab achieves an optimal synergy between the data and knowledge channels\. It models inter\-domain biomedical associations, providing vital biomedical guidance without overriding the intrinsic data representations\. Built upon the TabPFN, MedKGTab operates directly on raw structured tabular data, inherently capturing numerical distributions and fundamentally outperforming LLMs that fail to preserve tabular structures and dependencies\. Extensive experiments validate MedKGTab’s superiority and robustness in both intra\- and cross\-cohort settings\.

## References

- \[1\]\(2021\)Neural additive models: interpretable machine learning with neural nets\.InProc\. of NeurIPS,Cited by:[§4\.1](https://arxiv.org/html/2606.31171#S4.SS1.SSS0.Px3.p1.1)\.
- \[2\]C\. Atasoglu, C\. Valdés, N\. D\. Walker,et al\.\(1998\)De novo synthesis of amino acids by the ruminal bacteria Prevotella bryantii B14, Selenomonas ruminantium HD4, and Streptococcus bovis ES1\.Applied and Environmental Microbiology64\(8\),pp\. 2836–2843\.Cited by:[§A\.5](https://arxiv.org/html/2606.31171#A1.SS5.p3.1)\.
- \[3\]E\. Athieniti and G\. M\. Spyrou\(2023\)A guide to multi\-omics data collection and integration for translational medicine\.Computational and Structural Biotechnology Journal21,pp\. 134–149\.Cited by:[§1](https://arxiv.org/html/2606.31171#S1.p1.1)\.
- \[4\]J\. Chen, K\. Liao, Y\. Wan,et al\.\(2022\)DANets: deep abstract networks for tabular data classification and regression\.InProc\. of AAAI,Cited by:[Table 4](https://arxiv.org/html/2606.31171#A1.T4.13.10.3.1),[§1](https://arxiv.org/html/2606.31171#S1.p2.1),[§4\.1](https://arxiv.org/html/2606.31171#S4.SS1.SSS0.Px2.p1.1),[Table 1](https://arxiv.org/html/2606.31171#S4.T1.7.10.3.1),[Table 2](https://arxiv.org/html/2606.31171#S4.T2.3.6.3.1)\.
- \[5\]T\. Chen and C\. Guestrin\(2016\)XGBoost: a scalable tree boosting system\.InProc\. of SIGKDD,Cited by:[Table 4](https://arxiv.org/html/2606.31171#A1.T4.13.8.1.2),[§4\.1](https://arxiv.org/html/2606.31171#S4.SS1.SSS0.Px2.p1.1),[Table 1](https://arxiv.org/html/2606.31171#S4.T1.7.8.1.2),[Table 2](https://arxiv.org/html/2606.31171#S4.T2.3.4.1.1)\.
- \[6\]Y\. Chen, Q\. Guo, S\. You,et al\.\(2025\)MedTransTab: advancing medical cross\-table tabular data generation\.InProc\. of WSDM,Cited by:[§2\.1](https://arxiv.org/html/2606.31171#S2.SS1.p1.1)\.
- \[7\]Z\. Chen, A\. Hernández\-Cano, A\. Romanou,et al\.\(2023\)MEDITRON\-70b: scaling medical pretraining for large language models\.arXiv preprint arXiv:2311\.16079\.Cited by:[§2\.2](https://arxiv.org/html/2606.31171#S2.SS2.p1.1)\.
- \[8\]E\. Choi, M\. T\. Bahadori, L\. Song,et al\.\(2017\)GRAM: graph\-based attention model for healthcare representation learning\.InProc\. of KDD,Cited by:[§1](https://arxiv.org/html/2606.31171#S1.p2.1)\.
- \[9\]C\. Dou, F\. Yang, F\. Li,et al\.\(2026\)Baichuan\-m3: modeling clinical inquiry for reliable medical decision\-making\.arXiv preprint arXiv:2602\.06570\.Cited by:[Table 4](https://arxiv.org/html/2606.31171#A1.T4.13.17.10.1),[§2\.2](https://arxiv.org/html/2606.31171#S2.SS2.p1.1),[§4\.1](https://arxiv.org/html/2606.31171#S4.SS1.SSS0.Px2.p1.1),[Table 1](https://arxiv.org/html/2606.31171#S4.T1.7.17.10.1)\.
- \[10\]L\. Fang, A\. Liu, H\. Zhang,et al\.\(2025\)TABGEN\-ICL: residual\-aware in\-context example selection for tabular data generation\.InFindings of ACL,Cited by:[Table 4](https://arxiv.org/html/2606.31171#A1.T4.13.15.8.1),[§4\.1](https://arxiv.org/html/2606.31171#S4.SS1.SSS0.Px2.p1.1),[Table 1](https://arxiv.org/html/2606.31171#S4.T1.7.15.8.1)\.
- \[11\]F\. D\. Filippis, E\. Pasolli, A\. Tett,et al\.\(2019\)Distinct genetic and functional traits of human intestinal Prevotella copri strains are associated with different habitual diets\.Cell Host & Microbe25\(3\),pp\. 444–453\.e3\.Cited by:[§A\.5](https://arxiv.org/html/2606.31171#A1.SS5.p3.1)\.
- \[12\]X\. Gai, P\. Qian, B\. Guo,et al\.\(2023\)Heptadecanoic acid and pentadecanoic acid crosstalk with fecal\-derived gut microbiota are potential non\-invasive biomarkers for chronic atrophic gastritis\.Frontiers in Cellular and Infection Microbiology12,pp\. 1064737\.Cited by:[§4\.1](https://arxiv.org/html/2606.31171#S4.SS1.SSS0.Px1.p1.1)\.
- \[13\]Y\. Gorishniy, I\. Rubachev, N\. Kartashev,et al\.\(2024\)TabR: tabular deep learning meets nearest neighbors\.InProc\. of ICLR,Cited by:[§2\.1](https://arxiv.org/html/2606.31171#S2.SS1.p1.1)\.
- \[14\]Y\. Gorishniy, I\. Rubachev, V\. Khrulkov,et al\.\(2021\)Revisiting deep learning models for tabular data\.InProc\. of NeurIPS,Cited by:[Table 4](https://arxiv.org/html/2606.31171#A1.T4.13.9.2.1),[§1](https://arxiv.org/html/2606.31171#S1.p2.1),[§2\.1](https://arxiv.org/html/2606.31171#S2.SS1.p1.1),[§4\.1](https://arxiv.org/html/2606.31171#S4.SS1.SSS0.Px2.p1.1),[Table 1](https://arxiv.org/html/2606.31171#S4.T1.7.9.2.1),[Table 2](https://arxiv.org/html/2606.31171#S4.T2.3.5.2.1)\.
- \[15\]L\. Grinsztajn, E\. Oyallon, and G\. Varoquaux\(2022\)Why do tree\-based models still outperform deep learning on typical tabular data?\.InProc\. of NeurIPS,Cited by:[§1](https://arxiv.org/html/2606.31171#S1.p2.1)\.
- \[16\]N\. Hollmann, S\. Müller, L\. Purucker,et al\.\(2025\)Accurate predictions on small data with a tabular foundation model\.Nature637\(8044\),pp\. 319–326\.Cited by:[Table 4](https://arxiv.org/html/2606.31171#A1.T4.13.13.6.1),[§2\.1](https://arxiv.org/html/2606.31171#S2.SS1.p1.1),[§3\.4\.1](https://arxiv.org/html/2606.31171#S3.SS4.SSS1.p1.1),[§4\.1](https://arxiv.org/html/2606.31171#S4.SS1.SSS0.Px2.p1.1),[Table 1](https://arxiv.org/html/2606.31171#S4.T1.7.13.6.1),[Table 2](https://arxiv.org/html/2606.31171#S4.T2.3.9.6.1)\.
- \[17\]M\. Kanehisa and S\. Goto\(2000\)KEGG: kyoto encyclopedia of genes and genomes\.Nucleic acids research28\(1\),pp\. 27–30\.Cited by:[§A\.5](https://arxiv.org/html/2606.31171#A1.SS5.p3.1)\.
- \[18\]A\. Kotelnikov, D\. Baranchuk, I\. Rubachev,et al\.\(2023\)TabDDPM: modelling tabular data with diffusion models\.InProc\. of ICML,Cited by:[§2\.1](https://arxiv.org/html/2606.31171#S2.SS1.p1.1)\.
- \[19\]Y\. Labrak, A\. Bazoge, E\. Morin,et al\.\(2024\)BioMistral: a collection of open\-source pretrained large language models for medical domains\.arXiv preprint arXiv:2402\.10373\.Cited by:[§2\.2](https://arxiv.org/html/2606.31171#S2.SS2.p1.1)\.
- \[20\]X\. L\. Li and P\. Liang\(2021\)Prefix\-tuning: optimizing continuous prompts for generation\.InProc\. of ACL,Cited by:[§1](https://arxiv.org/html/2606.31171#S1.p2.1)\.
- \[21\]Y\. Li, Z\. Li, K\. Zhang,et al\.\(2023\)ChatDoctor: a medical chat model fine\-tuned on llama model using medical domain knowledge\.arXiv preprint arXiv:2303\.14070\.Cited by:[§2\.2](https://arxiv.org/html/2606.31171#S2.SS2.p1.1)\.
- \[22\]A\. Liu, A\. Mei, B\. Lin,et al\.\(2025\)Deepseek\-v3\.2: pushing the frontier of open large language models\.arXiv preprint arXiv:2512\.02556\.Cited by:[Table 4](https://arxiv.org/html/2606.31171#A1.T4.13.16.9.1),[§4\.1](https://arxiv.org/html/2606.31171#S4.SS1.SSS0.Px2.p1.1),[Table 1](https://arxiv.org/html/2606.31171#S4.T1.7.16.9.1)\.
- \[23\]S\. Liu, H\. Cai, Q\. Zhou,et al\.\(2025\)Talent: A tabular analytics and learning toolbox\.Journal of Machine Learning Research26\(226\),pp\. 226:1–226:16\.Cited by:[§4\.1](https://arxiv.org/html/2606.31171#S4.SS1.SSS0.Px3.p3.1)\.
- \[24\]S\. Mani, S\. R\. Lalani, and M\. Pammi\(2025\)Genomics and multiomics in the age of precision medicine\.Pediatric Research97\(4\),pp\. 1399–1410\.Cited by:[§1](https://arxiv.org/html/2606.31171#S1.p1.1)\.
- \[25\]J\. H\. Morris, K\. Soman, R\. E\. Akbas,et al\.\(2023\)The scalable precision medicine open knowledge engine \(SPOKE\): a massive knowledge graph of biomedical information\.Bioinformatics39\(2\)\.Cited by:[§3\.3](https://arxiv.org/html/2606.31171#S3.SS3.p1.1)\.
- \[26\]N\. Patki, R\. Wedge, and K\. Veeramachaneni\(2016\)The synthetic data vault\.InProc\. of DSAA,Cited by:[§4\.1](https://arxiv.org/html/2606.31171#S4.SS1.SSS0.Px4.p2.1)\.
- \[27\]N\. Seedat, N\. Huynh, B\. van Breugel,et al\.\(2024\)Curated LLM: synergy of LLMs and data curation for tabular augmentation in low\-data regimes\.InProc\. of ICML,Cited by:[Table 4](https://arxiv.org/html/2606.31171#A1.T4.13.14.7.2),[§4\.1](https://arxiv.org/html/2606.31171#S4.SS1.SSS0.Px2.p1.1),[Table 1](https://arxiv.org/html/2606.31171#S4.T1.7.14.7.2)\.
- \[28\]K\. Singhal, S\. Azizi, T\. Tu,et al\.\(2023\)Towards expert\-level medical question answering with large language models\.arXiv preprint arXiv:2305\.09617\.Cited by:[§2\.2](https://arxiv.org/html/2606.31171#S2.SS2.p1.1)\.
- \[29\]Y\. Sun, W\. Tan, Z\. Gu,et al\.\(2025\)A data\-efficient strategy for building high\-performing medical foundation models\.Nature Biomedical Engineering9,pp\. 539–551\.Cited by:[§1](https://arxiv.org/html/2606.31171#S1.p1.1)\.
- \[30\]S\. Tarazona, A\. Arzalluz\-Luque, and A\. Conesa\(2021\)Undisclosed, unmet and neglected challenges in multi\-omics studies\.Nature Computational Science1\(6\),pp\. 395–402\.Cited by:[§1](https://arxiv.org/html/2606.31171#S1.p1.1)\.
- \[31\]T\. Tu, S\. Azizi, H\. W\. Chung,et al\.\(2023\)Towards generalist biomedical AI\.arXiv preprint arXiv:2307\.14334\.Cited by:[§2\.2](https://arxiv.org/html/2606.31171#S2.SS2.p1.1)\.
- \[32\]J\. Wang, S\. Zhao, Z\. Luo,et al\.\(2026\)EEGDiffuser: label\-guided eeg signals synthesis via diffusion model for bci applications\.Neurocomputing670,pp\. 132636\.Cited by:[§1](https://arxiv.org/html/2606.31171#S1.p1.1)\.
- \[33\]S\. Wang, X\. Jin, Z\. Wang,et al\.\(2025\)Data whisperer: efficient data selection for task\-specific LLM fine\-tuning via few\-shot in\-context learning\.InProc\. of ACL,Cited by:[§1](https://arxiv.org/html/2606.31171#S1.p2.1)\.
- \[34\]Z\. Wang, C\. Gao, C\. Xiao,et al\.\(2024\)MediTab: scaling medical tabular data predictors via data consolidation, enrichment, and refinement\.InProc\. of IJCAI,Cited by:[§2\.1](https://arxiv.org/html/2606.31171#S2.SS1.p1.1)\.
- \[35\]D\. P\. Wright, C\. G\. Knight, S\. G\. Parkar,et al\.\(2000\)Cloning of a mucin\-desulfating sulfatase gene from Prevotella strain RS2 and its expression using a Bacteroides recombinant system\.Journal of Bacteriology182\(11\),pp\. 3002–3007\.Cited by:[§A\.5](https://arxiv.org/html/2606.31171#A1.SS5.p3.1)\.
- \[36\]D\. P\. Wright, D\. I\. Rosendale, and A\. M\. Roberton\(2000\)Prevotella enzymes involved in mucin oligosaccharide degradation and evidence for a small operon of genes expressed during growth on mucin\.FEMS Microbiology Letters190\(1\),pp\. 73–79\.Cited by:[§A\.5](https://arxiv.org/html/2606.31171#A1.SS5.p3.1)\.
- \[37\]J\. Wu, S\. Chen, Q\. Zhao,et al\.\(2024\)SwitchTab: switched autoencoders are effective tabular learners\.InProc\. of AAAI,Cited by:[Table 4](https://arxiv.org/html/2606.31171#A1.T4.13.11.4.1),[§4\.1](https://arxiv.org/html/2606.31171#S4.SS1.SSS0.Px2.p1.1),[Table 1](https://arxiv.org/html/2606.31171#S4.T1.7.11.4.1),[Table 2](https://arxiv.org/html/2606.31171#S4.T2.3.7.4.1)\.
- \[38\]L\. Xu, M\. Skoularidou, A\. Cuesta\-Infante,et al\.\(2019\)Modeling tabular data using conditional GAN\.InProc\. of NeurIPS,Cited by:[§2\.1](https://arxiv.org/html/2606.31171#S2.SS1.p1.1)\.
- \[39\]H\. Ye, H\. Yin, D\. Zhan,et al\.\(2025\)Revisiting nearest neighbor for tabular data: A deep tabular baseline two decades later\.InProc\. of ICLR,Cited by:[Table 4](https://arxiv.org/html/2606.31171#A1.T4.13.12.5.1),[§1](https://arxiv.org/html/2606.31171#S1.p2.1),[§4\.1](https://arxiv.org/html/2606.31171#S4.SS1.SSS0.Px2.p1.1),[Table 1](https://arxiv.org/html/2606.31171#S4.T1.7.12.5.1),[Table 2](https://arxiv.org/html/2606.31171#S4.T2.3.8.5.1)\.

## Appendix AAppendix

### A\.1Prompt Templates for LLM\-based Baselines

To facilitate reproducibility, we detail the prompt templates utilized for the LLM\-based baselines\. As discussed in the main text, we compare MedKGTab against four LLM\-based approaches: CLLM \(Curated LLM\), TabGen\-ICL, DeepSeek v3\.2, and Baichuan\-M3\-plus\. Notably, both CLLM and TabGen\-ICL employ DeepSeek v3\.2 as their backbone model\. To ensure a fair evaluation, we apply a unified system and user prompt templates across all LLM baselines\.

The system prompt used in our experiments is presented below\.

`System Prompt The user prompt template is shown below\. User Prompt These templates ensure that all LLM\-based baselines receive the same input context and output constraints, thereby improving the reproducibility and fairness of the comparison\. For the reverse\-direction supplementary experiment in Appendix A\.2, we use the same templates, with metabolite and microbiota fields swapped accordingly\.`

`A\.2 Reverse Feature Expansion: From Microbiota to Metabolites Table 4: Performance evaluation of cross\-domain feature expansion methods from microbiota to metabolites on CAG\-Tongue\. Lower MSE, MedAE, and P95 indicate better performance, whereas higher R2R^\{2\} indicates better performance\. Bold and underlined numbers denote the best and second\-best results, respectively\. ↑\\uparrow and ↑\(%\)\\uparrow\(\\%\) report the absolute and relative improvements of MedKGTab over the second\-best baseline when MedKGTab achieves the best result\. Methods MSE\(×102\\times 10^\{2\}\)↓\\downarrow R2↑R^\{2\}\\uparrow MedAE↓\\downarrow P95↓\\downarrow Tabular Model XGBoost \[5\] 9\.767 0\.415 111\.857 318\.271 FT\-Transformer \[14\] 16\.901 \-0\.013 80\.938 510\.080 DANets \[4\] 23\.102 \-0\.384 119\.033 507\.069 SwitchTab \[37\] 16\.861 \-0\.010 77\.891 506\.462 modernNCA \[39\] 10\.010 0\.400 114\.589 351\.722 TabPFN \[16\] 10\.041 0\.398 80\.697 369\.804 LLM Curated LLM \[27\] 11\.606 0\.304 108\.022 370\.197 TabGen\-ICL \[10\] 12\.666 0\.241 84\.398 404\.260 DeepSeek v3\.2 \[22\] 12\.734 0\.237 89\.224 393\.801 Baichuan M3\-plus \[9\] 13\.405 0\.197 141\.518 392\.842 MedKGTab \(Ours\) 9\.189 0\.449 100\.706 319\.612 ↑\\uparrow 0\.578 0\.034 / / ↑\(%\)\\uparrow\(\\%\) 5\.92 8\.19 / / In the main paper, we evaluate all methods on the task of cross\-domain feature expansion from metabolites to microbiota in CAG\-Tongue\. As a supplementary experiment, we further evaluate the reverse\-direction task, namely expanding metabolite features from microbiota features\. The experiment is intended to test whether the proposed method remains effective when the source and target domains are swapped\. We strictly follow the same experimental setting as in the main intra\-cohort setting, including the data split, baselines, and evaluation metrics\. In particular, the main paper uses a 40/60 collected\-uncollected split on CAG\-Tongue, retains the top 50 features from each domain, and evaluates intra\-cohort regression performance using MSE and R2R^\{2\}, while also reporting MedAE and P95 in the main results table\. Table 4 reports the results of the supplementary expansion task\. MedKGTab achieves the best performance on MSE and R2R^\{2\}, indicating the strongest overall expansion quality among baselines\. It also remains competitive on P95, suggesting relatively strong robustness on hard samples in the high\-error tail\. On MedAE, however, MedKGTab does not rank first, which indicates that there is still room for improvement on the typical\-sample error level\. Figure 5: Feature\-wise sparsity distributions of metabolite and microbiota features on CAG\-Tongue\. Sparsity is measured as the proportion of zero values in each feature\. The cumulative distributions show that microbiota features are substantially sparser than metabolite features, which helps explain the difficulty of the feature expansion task from microbiota to metabolites\. Notably, compared with the original metabolite\-to\-microbiota task, all methods perform worse on this reverse microbiota\-to\-metabolites feature expansion task\. This degradation stems from the inherent sparsity of microbiota data, which provides a fundamentally weaker source\-domain signal\. As illustrated in Fig\. 5, 19 microbiota features exhibit zero values in over 50% of patients, a high sparsity level entirely absent in the selected metabolite features\. This severe zero\-inflation substantially reduces the effective information available for learning reliable cross\-domain dependencies\. Despite this inherent sparsity challenge, MedKGTab consistently achieves the best overall MSE and R2R^\{2\}\. These results confirm that our knowledge\-injected framework effectively generalizes across varying biomedical feature\-expansion directions\. Table 5: Downstream evaluation on the disease detection task using raw versus expanded training data\. Balanced Accuracy, F1, and AUC are reported for Random Forest, MLP, and XGBoost\. Bold and underlined numbers denote the best and second\-best results, respectively\. Model Trainset Balanced ACC F1 AUC Random Forest Generated 0\.863 0\.863 0\.943 Raw 0\.776 0\.871 0\.909 MLP Generated 0\.741 0\.833 0\.798 Raw 0\.764 0\.821 0\.807 XGBoost Generated 0\.732 0\.634 0\.732 Raw 0\.618 0\.812 0\.873 A\.3 Downstream Utility of Expanded Features We conduct a downstream classification experiment to further evaluate whether the expanded features are functionally comparable to the real features\. Specifically, we consider a disease detection task, in which metabolite features and microbiota features are used to determine the health status of each subject\. We split the original dataset into 60% collected data and 40% held\-out uncollected data\. For the collected data, we construct an expanded dataset by replacing the target\-domain features with the generated ones, while keeping the real disease labels unchanged\. The held\-out 40% original data are used only for testing\. We evaluate three representative and architecture\-neutral classifiers, including Random Forest, MLP, and XGBoost, to avoid introducing evaluation bias from additional domain\-specific modules such as graph\-based structures or knowledge\-injection components\. Balanced Accuracy, F1 score, and AUC are adopted as evaluation metrics to provide a comprehensive assessment of classification performance\. The results are shown in Table 5\. Overall, classifiers trained on expanded data achieve performance broadly comparable to those trained on original data, indicating that the expanded data preserve the key discriminative information required for disease detection\. Moreover, the overall relative performance pattern across classifiers remains largely unchanged under the two training settings, suggesting that the expanded data do not disproportionately favor any particular model\. For MLP, the results under the two training settings are very close across all three metrics, further supporting the functional consistency between expanded and original data\. Overall, the downstream evaluation suggests that the expanded features are not merely numerically close to the original features, but also preserve disease\-relevant discriminative information for practical classification tasks\. A\.4 Effect of Knowledge Graph Injection Strength Figure 6: Effect of knowledge graph injection on the native attention ranking of the TabPFN backbone\. Top\-5 Overlap measures the consistency between the top\-5 attended features before and after graph injection, while Top\-1 Flip measures how often the top\-1 attended feature changes; larger injection strength leads to lower overlap and higher flip ratio, indicating stronger rewriting of the original attention organization\. The main paper shows that the performance of MedKGTab is sensitive to the injection strength α\\alpha\. Specifically, the best performance is achieved at α\\alpha = 4, whereas overly large α\\alpha decreases the consistency score and worsens MSE, indicating that excessively strong graph injection interferes with the model’s original data\-driven attention structure\. To better understand the phenomenon, we further analyze how knowledge graph injection changes the native attention ranking of the TabPFN backbone\. We use two ranking\-based metrics in the following analysis\. Top\-5 Overlap measures the overlap between the top\-5 attended features before and after graph injection, where a higher value indicates greater consistency with the original attention ranking\. Top\-1 Flip measures how often the top\-1 attended feature in each row changes after graph injection, where a higher value indicates more frequent rewriting of the original attention target\. Results in Fig\. 6 show a clear trend: as the injection strength increases, the enhanced attention increasingly departs from the original attention ranking\. Specifically, Top\-5 Overlap decreases from 89\.13% at α=1\\alpha=1 to 77\.63% at α=4\\alpha=4 and 69\.93% at α=9\\alpha=9, while Top\-1 Flip increases from 17\.94% to 34\.93% and 55\.29%, respectively\. When the injection strength is relatively small, the knowledge graph prior mainly acts as auxiliary structural guidance and preserves most of the native attention preference\. When the injection strength becomes large, however, the knowledge graph prior no longer provides only supplementary information; instead, it begins to rewrite the original attention organization over a broad range of rows\. Therefore, the best injection strength should strike a balance between incorporating external biomedical guidance and preserving useful native dependencies learned from data\. A\.5 Case Study: Attention Re\-ranking via KG Injection To elucidate how KG injection modulates feature preferences, we analyzed attention\-score re\-ranking for the Prevotellaceae family\. Following our intra\-cohort setting, metabolite source features were partitioned into 17 groups \(3 per group\) to infer the target feature\. We computed average attention scores across all layers and heads before and after injection; a higher rank shift indicates an increased contribution to feature fusion\. As shown in Fig\. 7, KG injection systematically prioritizes biologically relevant metabolites\. Groups mapped to the SPOKE KG exhibited the most significant upward shifts, notably G15∗G\_\{15\}^\{\*\} \(Picolinic acid, \+10\), G7∗G\_\{7\}^\{\*\} \(Tyrosine, \+9\), and G12∗G\_\{12\}^\{\*\} \(Leucine, \+8\)\. Other KG\-informed groups like G16∗G\_\{16\}^\{\*\} \(N\-Acetylglucosamine, \+2\) and G9∗G\_\{9\}^\{\*\} \(Valine, \+1\) also displayed positive shifts, whereas unmapped groups generally declined or remained stable\. This prioritization aligns with empirical evidence\. Prevotellaceae strains are key participants in amino\-acid metabolism, particularly involving branched\-chain amino acids like Leucine and Valine \[2, 11\]\. Furthermore, Prevotellaceae\-mediated mucin degradation involves enzymes acting on N\-acetylglucosamine substrates \[35, 36\], while the biomedical database KEGG \[17\] annotations confirm aromatic amino acid metabolism modules for Tyrosine\. The promotion of Picolinic acid further suggests alignment with tryptophan\-related pathways often altered in gut dysbiosis\. Consistent with Appendix A\.4, these results demonstrate that KG priors provide targeted guidance rather than uniform bias\. By reshaping attention toward biologically grounded features, MedKGTab fundamentally enhances model interpretability\. Figure 7: Case study of attention\-rank changes in metabolite feature groups after knowledge graph injection\. Bar tops indicate rank changes after injection\. Gi denotes the ii\-th metabolite feature group, and ∗\* marks groups containing metabolite features in SPOKE\. Red and gray bars indicate increased and non\-increased ranks, respectively\. A\.6 Limitations While MedKGTab demonstrates promising performance in cross\-domain feature expansion, we acknowledge two primary limitations regarding its knowledge\-injection mechanism\. • Its generalizability is bottlenecked by the semantic alignment between clinical variables and knowledge graph nodes\. Since real\-world tabular data often contain non\-standard abbreviations, automated mapping risks misalignment noise and negative transfer without expert curation\. • The framework is constrained by out\-of\-vocabulary features due to the inherent incompleteness of existing KGs\. For highly specialized or newly discovered markers absent from the graph, MedKGTab cannot derive effective structural priors, limiting its full potential on cutting\-edge datasets Future work will focus on addressing these challenges through LLM\-based semantic matching modules and dynamic graph expansion via literature knowledge mining\.`