Leakage-Controlled Multimodal Learning for Diagnosis and Progression Prediction in Alzheimer's Disease Research

# Leakage-Controlled Multimodal Learning for Diagnosis and Progression Prediction in Alzheimer’s Disease Research
Source: [https://arxiv.org/html/2610.10648](https://arxiv.org/html/2610.10648)
Journal:Knowledge\-Based SystemsAkeem Temitope OtapoEmail:[akeem\.otapo@u\-pec\.fr](mailto:[email protected])Affiliation:Laboratoire Images, Signaux et Systèmes Intelligents \(LiSSi\), Université Paris\-Est Créteil \(UPEC\), 122 rue Paul Armangot, Vitry\-sur\-Seine, 94400, FranceGhazaleh KhodabandelouEmail:[ghazaleh\.khodabandelou@u\-pec\.fr](mailto:[email protected])Corresponding author:Corresponding author\.Affiliation:Laboratoire Images, Signaux et Systèmes Intelligents \(LiSSi\), Université Paris\-Est Créteil \(UPEC\), 122 rue Paul Armangot, Vitry\-sur\-Seine, 94400, FranceZuheng MingEmail:[zuheng\.ming@univ\-paris13\.fr](mailto:[email protected])Affiliation:L2TI, Institut Galilée, Université Sorbonne Paris Nord \(USPN\), 99 avenue Jean\-Baptiste Clément, Villetaneuse, 93430, FranceAlice OthmaniEmail:[alice\.othmani@u\-pec\.fr](mailto:[email protected])Affiliation:Laboratoire Images, Signaux et Systèmes Intelligents \(LiSSi\), Université Paris\-Est Créteil \(UPEC\), 122 rue Paul Armangot, Vitry\-sur\-Seine, 94400, France

###### Abstract

Alzheimer’s disease prediction involves irregular visits, heterogeneous measurements and incomplete modalities\. This study presents a multimodal multitask framework combining an adapted SFCN MRI encoder, four causal clinical Transformers, shared fusion and task\-specific ODE\-GRU dynamics\. Fine\-tuning and LoRA adapt the final two MRI blocks\. Task\-DRO balances task losses, while Group\-CVaR targets cohort and comorbidity strata\. Branch\-specific input controls, subject\-grouped partitions and empirical causality checks support longitudinal evaluation\. Across 2,649 subjects and 17,317 visits from ADNI, OASIS\-2 and MIRIAD, internal validation yields diagnosis, stage\-1 progression and first\-stage\-1\-visit progression AUROCs of 0\.935±\\pm0\.002, 0\.884±\\pm0\.003 and 0\.870±\\pm0\.005, respectively \(mean±\\pmSD across three seeds\)\. Corresponding hybrid AUROCs are 0\.951, 0\.909 and 0\.896\. Next\-visit MMSE mean absolute error \(MAE\) is 1\.61 points; worst\-stratum diagnosis AUROC is 0\.827±\\pm0\.008\. Sampled ADNI explanations identify task\-specific input dependence\. OASIS\-3 external validation yields network and hybrid diagnosis AUROCs of 0\.763 and 0\.767, hybrid next\-visit progression AUROC of 0\.764, diagnosis calibration error decreasing from 0\.197 to 0\.052, and next\-visit MMSE MAE of 0\.86\. Seed\-42 paired ablations of six components yield pooled diagnosis and progression AUROC differences between−0\.004\-0\.004and\+0\.004\+0\.004; removing clinical encoder inputs lowers diagnosis AUROC by 0\.272\. The framework integrates longitudinal prediction, missing\-modality handling, auxiliary comorbidity modelling and subgroup evaluation within a common pipeline\.

###### Keywords:

Alzheimer’s disease , Multimodal learning , Multitask learning , Progression prediction , Neural ODE , Comorbidity , Leakage control

††graphicalabstract:![[Uncaptioned image]](https://arxiv.org/html/2610.10648v1/figures/figures/graphic_arch.jpg)Graphical abstract\.Top: multi\-cohort harmonisation and leakage\-controlled preprocessing with subject\-grouped five\-fold cross\-validation\. Bottom: SFCN–ODE\-GRU architecture jointly predicting diagnosis, progression, MMSE, atrophy and comorbidities, trained with Task\-DRO and Group\-CVaR\.††highlights:Task\-specific input controls support multimodal longitudinal AD prediction\. ODE\-GRU dynamics model irregular visits across multiple prediction tasks\. Three\-seed internal diagnosis and progression AUROCs reach 0\.935 and 0\.884\. Comorbidity\-aware training integrates auxiliary supervision and subgroup auditing\. OASIS\-3 hybrid diagnosis and next\-visit progression AUROCs are 0\.767 and 0\.764\.## 1Introduction

Alzheimer’s disease \(AD\) is a progressive neurodegenerative disorder and a major cause of dementia\. Global dementia burden and its associated societal costs motivate improved diagnosis and prognosis\[[1](https://arxiv.org/html/2610.10648#bib.bib1),[2](https://arxiv.org/html/2610.10648#bib.bib2)\]\. These estimates concern dementia overall and should not be attributed exclusively to AD\. Reliable forecasts of progression may support patient management, clinical trial design and therapeutic research\.

However, AD progression is highly heterogeneous, influenced by genetic factors, lifestyle and comorbidities such as vascular disease, diabetes and depression\. Lamar et al\.\[[3](https://arxiv.org/html/2610.10648#bib.bib3)\]demonstrated that vascular comorbidities significantly affect qualitative error patterns in executive function tests among AD patients, suggesting that vascular pathology modifies the cognitive presentation of AD rather than simply accelerating progression\. Using latent Dirichlet allocation, Rajamaki et al\.\[[4](https://arxiv.org/html/2610.10648#bib.bib4)\]identified distinct comorbidity patterns in individuals with Alzheimer’s disease compared to those without, revealing that certain comorbidities co\-occur more frequently in AD patients\. Duara and Barker\[[5](https://arxiv.org/html/2610.10648#bib.bib5)\]reviewed heterogeneity in AD diagnosis and progression rates, emphasising implications for therapeutic trials\. Maciejewska et al\.\[[6](https://arxiv.org/html/2610.10648#bib.bib6)\]provided a comprehensive review of the mechanisms underlying selected comorbidities in Alzheimer’s disease\. Moreover, clinical visits occur at irregular intervals, and data modalities \(e\.g\., structural MRI, cognitive tests\) are often missing\. These challenges render many existing machine learning models impractical for real\-world clinical decision support\. As highlighted by Li et al\.\[[7](https://arxiv.org/html/2610.10648#bib.bib7)\], building models for longitudinal clinical data faces additional hurdles, including mismatched dimensionality between high\-dimensional images and low\-dimensional tabular variables, as well as inconsistent intervals in follow\-up data\.

Existing work addresses subsets of these challenges: irregular sampling, shared prediction tasks, comorbidity information and subgroup performance\. The proposed model addresses these challenges jointly\.

##### Irregular sampling and discrete\-time assumption

Some forecasting methods place irregular observations on a fixed grid or require missing\-value handling\[[8](https://arxiv.org/html/2610.10648#bib.bib8)\]\. Continuous\-time neural models provide an alternative representation of elapsed time\[[9](https://arxiv.org/html/2610.10648#bib.bib9),[10](https://arxiv.org/html/2610.10648#bib.bib10),[11](https://arxiv.org/html/2610.10648#bib.bib11),[12](https://arxiv.org/html/2610.10648#bib.bib12)\]; the experimental protocol therefore compares an ODE formulation with a time\-aware discrete GRU\.

##### Disjoint task formulation and static weighting

Diagnosis \(current visit status\) and prognosis \(future cognitive decline\) are typically treated separately, missing the opportunity to learn shared representations that benefit both tasks\. Multitask learning has been applied sporadically, but often with fixed, heuristic task weights that ignore varying task difficulty, label noise and data availability\[[13](https://arxiv.org/html/2610.10648#bib.bib13),[14](https://arxiv.org/html/2610.10648#bib.bib14),[15](https://arxiv.org/html/2610.10648#bib.bib15)\]\. Multitask learning has shown promise in clinical settings\[[16](https://arxiv.org/html/2610.10648#bib.bib16)\]\. Nevertheless, these approaches do not address comorbidity interactions\.

##### Neglect of comorbidities and subgroup fairness

Comorbidities can influence the clinical presentation and progression of AD\[[3](https://arxiv.org/html/2610.10648#bib.bib3),[5](https://arxiv.org/html/2610.10648#bib.bib5),[17](https://arxiv.org/html/2610.10648#bib.bib17)\]\. Subgroup\-aware learning and distributionally robust optimisation provide tools for examining uneven performance\[[18](https://arxiv.org/html/2610.10648#bib.bib18),[19](https://arxiv.org/html/2610.10648#bib.bib19),[20](https://arxiv.org/html/2610.10648#bib.bib20)\]\. Recorded\-comorbidity strata inform the robust training objective and the subgroup evaluation\.

##### Temporal leakage and unverifiable explanations

Longitudinal AD benchmarks are especially exposed to information leakage: visits of one subject placed in both training and test sets, features that are imputed backwards from future visits, cognitive scores that are themselves part of the diagnostic definition, or preprocessing statistics computed on the whole cohort\[[21](https://arxiv.org/html/2610.10648#bib.bib21)\]\. Such leaks inflate reported performance and make post\-hoc explanations unreliable, because a model that has seen the answer has no reason to rely on disease biology\. Few studies test for leakage empirically, and fewer still relate their explanations to known clinical and biological markers of AD\.

To address these gaps, this study introduces aunified, leakage\-controlled frameworkthat integrates 3D structural MRI and longitudinal clinical data with continuous\-time dynamics, multitask learning and explicit comorbidity modelling\. The main contributions are:

1. 1\.Multimodal multitask model for irregular visits\.An SFCN brain\-age encoder, whose last two blocks are adapted by fine\-tuning, low\-rank adaptation \(LoRA\) and an image\-level objective, is fused with four task\-specific causal clinical Transformers through shared causal attention\. Each branch is evolved over irregular visit times with an ODE\-GRU\. A single model jointly predicts current diagnosis,3\-year stage\-1 progression, current and next\-visit MMSE, next\-visit atrophy, and nine comorbidities, with task balancing by a scale\-free, capped Task\-DRO\.
2. 2\.Comorbidity\-aware learning and subgroup audit\.Nine comorbidities derived from ADNI medical histories are predicted as an auxiliary task under an explicit availability mask; a Group\-CVaR penalty targets the worst\-performing cohort and comorbidity strata during training; and both primary tasks are evaluated within every comorbidity stratum\.
3. 3\.Partition controls and empirical leakage audits\.All visits of a subject stay in one fold; all inputs are computed causally \(last observation carried forward only, rates from past visits only\); preprocessing statistics are fitted on the training fold; branch\-specific prohibited inputs prevent circular use of diagnostic criteria; and three empirical checks \(partition, feature causality, model causality\) are run for every fold model\.
4. 4\.Task\-specific explanations\.Integrated Gradients and group occlusion describe model dependence on available clinical and biological inputs\. The explanations are interpreted in the light of the branch\-specific input rules\.
5. 5\.Systematic evaluation\.The evaluation comprises three training seeds, eight paired component and modality ablations with subject\-level bootstrap tests, calibration analysis, a per\-stratum subgroup audit and external validation on the held\-out OASIS\-3 cohort\.

The remainder of this paper is organised as follows\. Section[2](https://arxiv.org/html/2610.10648#S2)reviews related work in multimodal AD diagnosis, continuous\-time modelling, multitask learning and robust optimisation\. Section[3](https://arxiv.org/html/2610.10648#S3)describes the proposed architecture and training objectives\. Section[4](https://arxiv.org/html/2610.10648#S4)details the datasets, preprocessing pipeline, leakage controls and experimental setup\. Section[5](https://arxiv.org/html/2610.10648#S5)presents the main results, ablations, subgroup analyses and interpretability analyses\. Section[6](https://arxiv.org/html/2610.10648#S6)discusses the implications and limitations, and Section[7](https://arxiv.org/html/2610.10648#S7)concludes the paper\.

## 2Related Work

### 2\.1Multimodal Diagnosis and Prognosis

Reviews describe machine\-learning and deep\-learning approaches to AD diagnosis and conversion prediction\[[22](https://arxiv.org/html/2610.10648#bib.bib22),[23](https://arxiv.org/html/2610.10648#bib.bib23),[24](https://arxiv.org/html/2610.10648#bib.bib24)\]\. Residual imaging networks, joint prediction of clinical outcomes, multimodal recurrent networks, latent representations and contrastive fusion illustrate distinct modelling strategies\[[25](https://arxiv.org/html/2610.10648#bib.bib25),[26](https://arxiv.org/html/2610.10648#bib.bib26),[27](https://arxiv.org/html/2610.10648#bib.bib27),[28](https://arxiv.org/html/2610.10648#bib.bib28),[29](https://arxiv.org/html/2610.10648#bib.bib29),[30](https://arxiv.org/html/2610.10648#bib.bib30)\]\. Recent work includes explainable multimodal prediction, longitudinal survival models, clinical–MRI embeddings and cognitive\-decline prediction\[[31](https://arxiv.org/html/2610.10648#bib.bib31),[32](https://arxiv.org/html/2610.10648#bib.bib32),[33](https://arxiv.org/html/2610.10648#bib.bib33),[34](https://arxiv.org/html/2610.10648#bib.bib34),[35](https://arxiv.org/html/2610.10648#bib.bib35),[7](https://arxiv.org/html/2610.10648#bib.bib7)\]\. Their outcomes, horizons and validation protocols differ, so published scores are not treated as matched benchmarks here\. Bélanger et al\. studied dementia identification in home\-health assessment records\[[36](https://arxiv.org/html/2610.10648#bib.bib36)\]; their OASIS assessment system is distinct from the OASIS neuroimaging cohorts used in this study\.

### 2\.2Continuous\-Time Modelling

Fixed\-grid forecasting has been used in the TADPOLE setting\[[8](https://arxiv.org/html/2610.10648#bib.bib8),[37](https://arxiv.org/html/2610.10648#bib.bib37)\]\. Neural ODEs, latent ODEs, GRU\-ODE models and neural controlled differential equations provide representations for irregularly sampled sequences\[[9](https://arxiv.org/html/2610.10648#bib.bib9),[10](https://arxiv.org/html/2610.10648#bib.bib10),[38](https://arxiv.org/html/2610.10648#bib.bib38),[11](https://arxiv.org/html/2610.10648#bib.bib11),[12](https://arxiv.org/html/2610.10648#bib.bib12)\]\. Applications include longitudinal volumetric modelling and multidimensional AD progression models\[[39](https://arxiv.org/html/2610.10648#bib.bib39),[40](https://arxiv.org/html/2610.10648#bib.bib40)\]\. Sharma et al\.’s cLDM\-ODE, continuous\-time biomarker operators, progression\-guided MRI generation and stage\-aware mixture\-of\-experts models address related forecasting problems\[[41](https://arxiv.org/html/2610.10648#bib.bib41),[42](https://arxiv.org/html/2610.10648#bib.bib42),[43](https://arxiv.org/html/2610.10648#bib.bib43),[44](https://arxiv.org/html/2610.10648#bib.bib44)\]\.

### 2\.3Multitask Learning and Task Balancing

Uncertainty weighting, GradNorm and clinical time\-series benchmarks motivate shared representations and adaptive task weights\[[13](https://arxiv.org/html/2610.10648#bib.bib13),[14](https://arxiv.org/html/2610.10648#bib.bib14),[16](https://arxiv.org/html/2610.10648#bib.bib16)\]\. AD applications include longitudinal progression, correlated cognitive outcomes and task\-relationship learning\[[15](https://arxiv.org/html/2610.10648#bib.bib15),[45](https://arxiv.org/html/2610.10648#bib.bib45),[46](https://arxiv.org/html/2610.10648#bib.bib46),[47](https://arxiv.org/html/2610.10648#bib.bib47),[48](https://arxiv.org/html/2610.10648#bib.bib48),[49](https://arxiv.org/html/2610.10648#bib.bib49)\]\. Survival analysis provides a complementary formulation\[[50](https://arxiv.org/html/2610.10648#bib.bib50)\]\. The proposed Task\-DRO update adapts the weights to relative losses on top of fixed base weights\.

### 2\.4Comorbidities and Robust Learning

Clinical studies and reviews examine vascular effects, comorbidity patterns and AD heterogeneity\[[3](https://arxiv.org/html/2610.10648#bib.bib3),[4](https://arxiv.org/html/2610.10648#bib.bib4),[5](https://arxiv.org/html/2610.10648#bib.bib5),[6](https://arxiv.org/html/2610.10648#bib.bib6),[17](https://arxiv.org/html/2610.10648#bib.bib17)\]\. Progression subtypes and longitudinal survival models address patient heterogeneity from complementary perspectives\[[51](https://arxiv.org/html/2610.10648#bib.bib51),[52](https://arxiv.org/html/2610.10648#bib.bib52)\]\. Group\-DRO, healthcare subgroup comparisons and under\-represented subgroup fine\-tuning motivate robust objectives\[[18](https://arxiv.org/html/2610.10648#bib.bib18),[19](https://arxiv.org/html/2610.10648#bib.bib19),[20](https://arxiv.org/html/2610.10648#bib.bib20)\]\. Site bias, demographic alignment and population\-specific dementia models also highlight transportability concerns\[[53](https://arxiv.org/html/2610.10648#bib.bib53),[54](https://arxiv.org/html/2610.10648#bib.bib54),[55](https://arxiv.org/html/2610.10648#bib.bib55)\]\. The Group\-CVaR penalty targets cohort and recorded\-comorbidity strata; its effect on the worst strata is assessed directly\.

### 2\.5Comparative Summary

Table[1](https://arxiv.org/html/2610.10648#S2.T1)compares selected studies by publication year, modality coverage, continuous\-time modelling, diagnosis/prognosis, comorbidity modelling, robustness objective, task and reported result\. Entries provide context and do not establish a matched performance ranking\.

Table 1:Comparison of related Alzheimer’s disease progression models: capabilities and reported results\.MethodYearMCTD\+PComRTask \(metric\)ReportedMultimodal Diagnosis and ConversionZhang & Shen\[[26](https://arxiv.org/html/2610.10648#bib.bib26)\]2012✓×\\times∘\\circ×\\times×\\timesMCI conv\. \(AUC\)0\.77†Lee et al\.\[[27](https://arxiv.org/html/2610.10648#bib.bib27)\]2019✓×\\times∘\\circ×\\times×\\timesMCI→\\toAD conv\. \(AUC\)0\.86Zhou et al\.\[[28](https://arxiv.org/html/2610.10648#bib.bib28)\]2019✓×\\times✓×\\times×\\timessMCI/pMCI \(AUC\)0\.755Huang\[[29](https://arxiv.org/html/2610.10648#bib.bib29)\]2023✓×\\times∘\\circ×\\times×\\timesDiagnosis, ADNI \(Acc\)0\.838Afifi et al\.\[[23](https://arxiv.org/html/2610.10648#bib.bib23)\]2026–––––Review–Simkhada et al\.\[[31](https://arxiv.org/html/2610.10648#bib.bib31)\]2026✓×\\times∘\\circ×\\times×\\timessMCI/pMCI, 2 y \(BA\)0\.820Lu et al\.\[[33](https://arxiv.org/html/2610.10648#bib.bib33)\]2026✓×\\times∘\\circ×\\times×\\timesProgression, 4 y \(AUC\)0\.91–0\.92Geng & Zhang\[[34](https://arxiv.org/html/2610.10648#bib.bib34)\]2026✓×\\times∘\\circ×\\times×\\timesMMSE decline, 12 mo \(AUC\)0\.773Zhang et al\.\[[32](https://arxiv.org/html/2610.10648#bib.bib32)\]2026✓∘\\circ∘\\circ×\\times×\\timesMCI→\\toAD, 1 y \(cdAUC\)0\.909Longitudinal and Continuous\-Time ModellingNguyen et al\.\[[8](https://arxiv.org/html/2610.10648#bib.bib8)\]2020✓×\\times✓×\\times×\\timesDx \(mAUC\) / ADAS\-Cog13 \(MAE\)0\.944 / 4\.30Rubanova et al\.\[[10](https://arxiv.org/html/2610.10648#bib.bib10)\]2019×\\times✓––×\\timesNon\-AD \(method\)–Bossa & Sahli\[[40](https://arxiv.org/html/2610.10648#bib.bib40)\]2023∘\\circ✓✓×\\times×\\timesMCI→\\toAD, 3 y \(AUC\)0\.86Sharma et al\. \(cLDM\-ODE\)\[[41](https://arxiv.org/html/2610.10648#bib.bib41)\]2026✓✓∘\\circ×\\times×\\timesAD conv\. \(Acc\) / MMSE \(MAE\)0\.88 / 1\.5He et al\. \(stage\-aware MoE\)\[[44](https://arxiv.org/html/2610.10648#bib.bib44)\]2026×\\times✓∘\\circ×\\times×\\timesRegional tau forecast \(rr\)0\.717Wang et al\.\[[42](https://arxiv.org/html/2610.10648#bib.bib42)\]2026✓✓∘\\circ×\\times×\\timesBiomarker maps \(rel\.L2L^\{2\}acc\.\)≈\\approx0\.90Zhao et al\.\[[43](https://arxiv.org/html/2610.10648#bib.bib43)\]2026×\\times×\\times∘\\circ×\\times×\\times4\-y MRI forecast \(SSIM\)0\.931Multitask LearningKendall et al\.\[[13](https://arxiv.org/html/2610.10648#bib.bib13)\]2018×\\times×\\times––×\\timesNon\-AD \(method\)–Chen et al\.\[[14](https://arxiv.org/html/2610.10648#bib.bib14)\]2018×\\times×\\times––×\\timesNon\-AD \(method\)–Zhou et al\.\[[15](https://arxiv.org/html/2610.10648#bib.bib15)\]2011✓×\\times∘\\circ×\\times×\\timesFuture MMSE/ADAS\-Cogn\.r\.Wang et al\.\[[46](https://arxiv.org/html/2610.10648#bib.bib46)\]2019×\\times×\\times∘\\circ×\\times×\\timesFuture MMSE / ADAS\-Cog \(rr\)0\.625 / 0\.664Liang et al\.\[[47](https://arxiv.org/html/2610.10648#bib.bib47)\]2023×\\times×\\times∘\\circ×\\times×\\times18 cognitive scores \(wR\)0\.416Jahani et al\.\[[50](https://arxiv.org/html/2610.10648#bib.bib50)\]2025✓∘\\circ∘\\circ×\\times×\\timesMCI→\\toAD \(C\-index\)0\.878Ma et al\.\[[48](https://arxiv.org/html/2610.10648#bib.bib48)\]2026∘\\circ×\\times✓×\\times×\\timesDx \(Acc\) / ADAS\-Cog \(R2R^\{2\}\)0\.928 / 0\.82Zhou et al\.\[[49](https://arxiv.org/html/2610.10648#bib.bib49)\]2026×\\times×\\times×\\times×\\times×\\times7 phenotypes, ADNI \(rr\)0\.158Comorbidity, Heterogeneity & Robust LearningLamar et al\.\[[3](https://arxiv.org/html/2610.10648#bib.bib3)\]2010–––✓–Clinical study–Maciejewska et al\.\[[6](https://arxiv.org/html/2610.10648#bib.bib6)\]2021–––✓–Review–Duara & Barker\[[5](https://arxiv.org/html/2610.10648#bib.bib5)\]2022–––✓–Review–Wachinger et al\.\[[53](https://arxiv.org/html/2610.10648#bib.bib53)\]2021×\\times×\\times×\\times×\\times∘\\circSite\-bias detection \(Acc\)0\.715Ren et al\.\[[51](https://arxiv.org/html/2610.10648#bib.bib51)\]2026✓×\\times∘\\circ×\\times×\\timesTrajectory, 48 mo \(Acc\)0\.791Karaman et al\.\[[52](https://arxiv.org/html/2610.10648#bib.bib52)\]2026✓∘\\circ∘\\circ×\\times×\\timesFuture AD dx \(AUROC\)n\.r\.Wang et al\.\[[20](https://arxiv.org/html/2610.10648#bib.bib20)\]2025∘\\circ×\\times∘\\circ×\\times✓Dementia accuracy\-gap reduction\>\>3%Ho et al\.\[[54](https://arxiv.org/html/2610.10648#bib.bib54)\]2026×\\times×\\times∘\\circ×\\times✓Ethnic BA gap \(pp\)6\.9→\\to1\.4Liang et al\.\[[55](https://arxiv.org/html/2610.10648#bib.bib55)\]2026∘\\circ×\\times∘\\circ×\\times∘\\circPopulation\-specific risk \(AUC\)0\.835–0\.903This StudyProposed framework2026✓✓✓✓✓3\-y MCI conv\. / sMCI\-pMCI/ dx \(AUC\)0\.884/0\.870/0\.935
M: multimodal; CT: continuous\-time; D\+P: diagnosis and prognosis; Com: explicit comorbidity modelling; R: robust optimisation or bias mitigation\.✓\\checkmark: addressed;∘\\circ: partially addressed;×\\times: not addressed in the cited work; –: not applicable\. Results differ in cohorts, endpoints, horizons, metrics and validation protocols and are not head\-to\-head comparisons\. n\.r\.: no single summary metric reported\. Values for the proposed model are three\-seed internal\-validation means\.

The proposed architecture combines causal clinical encoders, an adapted MRI encoder, task\-specific ODE\-GRU branches, auxiliary recorded\-comorbidity prediction and subgroup objectives\. The component and modality ablations assess this integrated design\.

## 3Methodology

Figure[1](https://arxiv.org/html/2610.10648#S3.F1)provides an overview of the proposed framework\. For each subjectiithe model processes a sequence of visits\{\(xi,kMRI,𝐱i,kClin,ti,k\)\}k=1Ti\\\{\(x^\{\\text\{MRI\}\}\_\{i,k\},\\mathbf\{x\}^\{\\text\{Clin\}\}\_\{i,k\},t\_\{i,k\}\)\\\}\_\{k=1\}^\{T\_\{i\}\}, whereti,kt\_\{i,k\}is the time in years since the first visit\. It outputs, at every visit, the current diagnosis, the probability of conversion to a more advanced stage within three years, current and next\-visit MMSE, next\-visit atrophy, and, at the last visit, a multi\-label comorbidity vector\. The visit\-level prediction paths are causal: the output at visitkkdepends only on visits1,…,k1,\\dots,k\.

![Refer to caption](https://arxiv.org/html/2610.10648v1/figures/figures/main_arch.jpg)Figure 1:SFCN–ODE\-GRU architecture\. A LoRA\-adapted SFCN MRI encoder and four task\-specific causal clinical Transformers are combined by shared causal fusion; each task branch evolves an ODE\-GRU state between irregular visits and feeds the diagnosis, progression, MMSE, atrophy and comorbidity heads\. Icons are illustrative\.### 3\.1Per\-Visit Encoding

#### 3\.1\.1MRI Encoder

The MRI encoder uses the Simple Fully Convolutional Network \(SFCN\) pretrained for brain\-age estimation on UK Biobank\[[56](https://arxiv.org/html/2610.10648#bib.bib56)\]\. The first four convolutional blocks are frozen \(including BatchNorm statistics\), and the 64\-dimensional pooled output is projected to a latent vector𝐡i,kM∈ℝ128\\mathbf\{h\}^\{M\}\_\{i,k\}\\in\\mathbb\{R\}^\{128\}\. The last two blocks \(conv4, conv5; 1\.79M parameters\) are fine\-tuned with a small learning rate, and their convolutions are additionally wrapped with LoRA adapters\[[57](https://arxiv.org/html/2610.10648#bib.bib57)\]:

Conv′​\(𝐱\)=Conv⁡\(𝐱\)\+αr​B​\(A⁡\(𝐱\)\),\\mathrm\{Conv\}^\{\\prime\}\(\\mathbf\{x\}\)=\\mathrm\{Conv\}\(\\mathbf\{x\}\)\+\\tfrac\{\\alpha\}\{r\}\\,B\\bigl\(A\(\\mathbf\{x\}\)\\bigr\),\(1\)whereAAis a rank\-rrconvolution with the original kernel size,BBis a1×1×11\\times 1\\times 1convolution initialised at zero \(so that training starts exactly at the pretrained encoder\),r=8r=8andα=16\\alpha=16\. Because brain\-age pretraining does not directly optimise the cohort\-defined diagnosis endpoint, an auxiliary cohort\-defined binary diagnosis head \(128→64→2128\\to 64\\to 2\) is applied to𝐡i,kM\\mathbf\{h\}^\{M\}\_\{i,k\}alone on visits with MRI and trained with cross\-entropy \(weightλimg=0\.2\\lambda\_\{\\text\{img\}\}=0\.2\), which encourages diagnosis\-relevant image representations\. Visits without MRI use a learnable embedding𝐞noMRI\\mathbf\{e\}\_\{\\text\{noMRI\}\}, and a learned availability embedding is added to every MRI latent\. During training the whole MRI stream of a subject is replaced by𝐞noMRI\\mathbf\{e\}\_\{\\text\{noMRI\}\}with probability 0\.05 \(modality dropout\), so that the model remains usable when imaging is absent\. SFCN is preferred to larger 3D CNNs because it is pretrained on brain MRI and has about 3\.0M parameters \(46\.6M for a 3D ResNet\-50\)\.

#### 3\.1\.2Task\-Specific Causal Clinical Encoders

Clinical inputs differ by task, because a feature that is legitimate for one task can be circular for another \(Section[2](https://arxiv.org/html/2610.10648#S4.F2)\)\. The architecture therefore includes four clinical encoders with identical architecture but different input sets:*diagnosis*,*forecast*,*MMSE*and*atrophy*\. Each concatenates its time\-varying features with temporal features \(time since first visit and visit order\), projects them to 128 dimensions and applies a 2\-layer, 4\-head Transformer encoder with a causal mask and a key\-padding mask:

𝐡i,1:TiC,b=Transformerb\(Wb\[𝐱i,1:Tib;𝐱i,1:Titime\]\),b∈ℬ,\\mathbf\{h\}^\{C,b\}\_\{i,1:T\_\{i\}\}=\\mathrm\{Transformer\}\_\{b\}\\bigl\(W\_\{b\}\[\\mathbf\{x\}^\{b\}\_\{i,1:T\_\{i\}\};\\mathbf\{x\}^\{\\text\{time\}\}\_\{i,1:T\_\{i\}\}\]\\bigr\),\\quad b\\in\\mathcal\{B\},\(2\)withℬ=\{diag,fc,mmse,atr\}\\mathcal\{B\}=\\\{\\text\{diag\},\\text\{fc\},\\text\{mmse\},\\text\{atr\}\\\}\. The causal mask guarantees that visitkknever attends to visitsk′\>kk^\{\\prime\}\>k\.

### 3\.2Shared Causal Attention Fusion

One fusion module, shared by all four task branches, combines the MRI latent with each task\-specific clinical stream\. The two representations are projected separately to 128 dimensions and concatenated to form a 256\-dimensional sequence:

𝐦i,k\\displaystyle\\mathbf\{m\}\_\{i,k\}=PM𝐡Mi,k,𝐜bi,k=PC𝐡C,bi,k,\\displaystyle=P\_\{M\}\\mathbf\{h\}^\{M\}\_\{i,k\},\\qquad\\mathbf\{c\}^\{b\}\_\{i,k\}=P\_\{C\}\\mathbf\{h\}^\{C,b\}\_\{i,k\},\(3\)𝐇bi,1:Ti\\displaystyle\\mathbf\{H\}^\{b\}\_\{i,1:T\_\{i\}\}=\[𝐦i,1:Ti;𝐜bi,1:Ti\]\.\\displaystyle=\[\\mathbf\{m\}\_\{i,1:T\_\{i\}\};\\mathbf\{c\}^\{b\}\_\{i,1:T\_\{i\}\}\]\.Four\-head self\-attention uses this same concatenated sequence for queries, keys and values, with causal and padding masks\. The attended output is projected to 128 dimensions, passed through dropout, and combined with a scaled modality residual:

𝐀bi,1:Ti\\displaystyle\\mathbf\{A\}^\{b\}\_\{i,1:T\_\{i\}\}=MHAcausal,pad\(𝐇bi,1:Ti,𝐇bi,1:Ti,𝐇bi,1:Ti\),\\displaystyle=\\mathrm\{MHA\}\_\{\\mathrm\{causal,pad\}\}\(\\mathbf\{H\}^\{b\}\_\{i,1:T\_\{i\}\},\\mathbf\{H\}^\{b\}\_\{i,1:T\_\{i\}\},\\mathbf\{H\}^\{b\}\_\{i,1:T\_\{i\}\}\),\(4\)𝐅bi,1:Ti\\displaystyle\\mathbf\{F\}^\{b\}\_\{i,1:T\_\{i\}\}=Dropout\(PO𝐀bi,1:Ti\)\\displaystyle=\\mathrm\{Dropout\}\(P\_\{O\}\\mathbf\{A\}^\{b\}\_\{i,1:T\_\{i\}\}\)\+λg\(wM𝐦i,1:Ti\+wC𝐜bi,1:Ti\),\\displaystyle\+\\lambda\_\{g\}\\left\(w\_\{M\}\\mathbf\{m\}\_\{i,1:T\_\{i\}\}\+w\_\{C\}\\mathbf\{c\}^\{b\}\_\{i,1:T\_\{i\}\}\\right\),𝐮i,kb\\displaystyle\\mathbf\{u\}^\{b\}\_\{i,k\}=Adaptb​\(𝐅i,kb\),\\displaystyle=\\mathrm\{Adapt\}\_\{b\}\(\\mathbf\{F\}^\{b\}\_\{i,k\}\),where\(wM,wC\)=softmax⁡\(𝐚\)\(w\_\{M\},w\_\{C\}\)=\\mathrm\{softmax\}\(\\mathbf\{a\}\)are learned global modality weights andλg=0\.01\\lambda\_\{g\}=0\.01\. These weights are shared across subjects, visits and tasks; they are not causal modality shares or patient\-specific explanations\. EachAdaptb\\mathrm\{Adapt\}\_\{b\}is a separate two\-layer MLP with a SiLU activation\. The shared fusion module is called separately for the diagnosis, progression, MMSE and atrophy clinical branches\.

### 3\.3Continuous\-Time Dynamics: ODE\-GRU

Each branch has its own ODE\-GRU cell\[[10](https://arxiv.org/html/2610.10648#bib.bib10),[38](https://arxiv.org/html/2610.10648#bib.bib38)\]with a 128\-dimensional hidden state𝐳b​\(t\)\\mathbf\{z\}^\{b\}\(t\)\.*Between visits*, the state evolves according to a learned vector field:

d​𝐳b​\(t\)d​t=fθb​\(𝐳b​\(t\),ϕ⁡\(t\)\)\+Sb​𝐳b​\(t\),\\frac\{d\\mathbf\{z\}^\{b\}\(t\)\}\{dt\}=f\_\{\\theta\_\{b\}\}\\bigl\(\\mathbf\{z\}^\{b\}\(t\),\\phi\(t\)\\bigr\)\+S\_\{b\}\\,\\mathbf\{z\}^\{b\}\(t\),\(5\)wherefθbf\_\{\\theta\_\{b\}\}is a 3\-layer MLP \(256 hidden units, SiLU\) whose last layer is initialised at zero,SbS\_\{b\}is a small orthogonally initialised linear skip, andϕ⁡\(t\)\\phi\(t\)is a sinusoidal time embedding \(8 frequencies from1/81/8to1616rad/year\)\. Equation \([5](https://arxiv.org/html/2610.10648#S3.E5)\) is integrated fromti,k−1t\_\{i,k\-1\}toti,kt\_\{i,k\}with the midpoint \(RK2\) method using two substeps per visit interval under the evaluated configuration\.*At a visit*, a GRU update mixes the propagated state with the fused observation𝐮i,kb\\mathbf\{u\}^\{b\}\_\{i,k\}, using update and reset gates that also receive the elapsed time:

𝐠k\\displaystyle\\mathbf\{g\}\_\{k\}=σ\(Wg\[𝐳b−k;𝐮bk;log\(1\+Δtk\)\]\),𝐠∈\{𝐳g,𝐫g\},\\displaystyle=\\sigma\\bigl\(W\_\{g\}\[\\mathbf\{z\}^\{b\-\}\_\{k\};\\mathbf\{u\}^\{b\}\_\{k\};\\log\(1\+\\Delta t\_\{k\}\)\]\\bigr\),\\quad\\mathbf\{g\}\\in\\\{\\mathbf\{z\}\_\{g\},\\mathbf\{r\}\_\{g\}\\\},\(6\)𝐧~k\\displaystyle\\tilde\{\\mathbf\{n\}\}\_\{k\}=tanh⁡\(Wn​\[𝐫g⊙𝐳kb−;𝐮kb\]\),\\displaystyle=\\tanh\\bigl\(W\_\{n\}\[\\mathbf\{r\}\_\{g\}\\odot\\mathbf\{z\}^\{b\-\}\_\{k\};\\mathbf\{u\}^\{b\}\_\{k\}\]\\bigr\),𝐳kb\\displaystyle\\mathbf\{z\}^\{b\}\_\{k\}=\(1−𝐳g\)⊙𝐳kb−\+𝐳g⊙𝐧~k,\\displaystyle=\(1\-\\mathbf\{z\}\_\{g\}\)\\odot\\mathbf\{z\}^\{b\-\}\_\{k\}\+\\mathbf\{z\}\_\{g\}\\odot\\tilde\{\\mathbf\{n\}\}\_\{k\},where𝐳kb−\\mathbf\{z\}^\{b\-\}\_\{k\}is the ODE\-propagated state just before visitkkandΔ​tk=ti,k−ti,k−1\\Delta t\_\{k\}=t\_\{i,k\}\-t\_\{i,k\-1\}\. The evolved sequence is refined by a causal temporal self\-attention layer with a residual connection and layer normalisation\. Compared with a discrete GRU, this formulation gives the latent disease state a well\-defined value*at any time between visits*\. An ODE\-drift penaltyℒdrift=𝔼​‖fθb​\(𝐳,ϕ⁡\(t\)\)\+Sb​𝐳‖2\\mathcal\{L\}\_\{\\text\{drift\}\}=\\mathbb\{E\}\\\|f\_\{\\theta\_\{b\}\}\(\\mathbf\{z\},\\phi\(t\)\)\+S\_\{b\}\\mathbf\{z\}\\\|^\{2\}\(λdrift=10−3\\lambda\_\{\\text\{drift\}\}=10^\{\-3\}\) discourages unnecessarily large vector fields\.

### 3\.4Prediction Heads

Classification and next\-visit residual heads use evolved branch states; current MMSE and volume heads instead use the task\-adapted fused representations𝐮i,kb\\mathbf\{u\}^\{b\}\_\{i,k\}:

- 1\.Current diagnosis\(cohort\-defined binary endpoint\):y^i,kdx=softmax⁡\(MLPdx​\(𝐳i,kdiag\)\)\\hat\{y\}^\{\\text\{dx\}\}\_\{i,k\}=\\mathrm\{softmax\}\(\\mathrm\{MLP\}\_\{\\text\{dx\}\}\(\\mathbf\{z\}^\{\\text\{diag\}\}\_\{i,k\}\)\)\. An auxiliary 3\-way stage head \(0/1/2; cohort\-specific mapping\) on the same state uses all visits including MCI \(weight 0\.01\)\.
- 2\.3\-year conversion:y^i,kconv=softmax⁡\(MLPconv​\(Wm​𝐳i,kfc\+Ws​𝐬i\)\)\\hat\{y\}^\{\\text\{conv\}\}\_\{i,k\}=\\mathrm\{softmax\}\\bigl\(\\mathrm\{MLP\}\_\{\\text\{conv\}\}\(W\_\{m\}\\mathbf\{z\}^\{\\text\{fc\}\}\_\{i,k\}\+W\_\{s\}\\mathbf\{s\}\_\{i\}\)\\bigr\), where𝐬i\\mathbf\{s\}\_\{i\}denotes the covariate vector used by the direct head pathway at visitkk\(demographics, baseline imaging and fluid biomarkers, creatinine and family history, with family history treated as a distinct predictor from the participant’s own medical history\) entering through a linear projection\. Each variable enters with its causally available value at visitkk\(last observation carried forward; variables recorded only at baseline enter as recorded\)\. The vector is a subset of the diagnosis\-branch inputs and therefore obeys the same prohibited\-input rules\.
- 3\.Current MMSE and atrophy:m^i,k=MLPmmse​\(𝐮i,kmmse\)\\hat\{m\}\_\{i,k\}=\\mathrm\{MLP\}\_\{\\text\{mmse\}\}\(\\mathbf\{u\}^\{\\text\{mmse\}\}\_\{i,k\}\),a^i,k=MLPatr​\(𝐮i,katr\)\\hat\{a\}\_\{i,k\}=\\mathrm\{MLP\}\_\{\\text\{atr\}\}\(\\mathbf\{u\}^\{\\text\{atr\}\}\_\{i,k\}\)\. The atrophy target is normalised whole\-brain volume \(nWBV\) in OASIS\-2 and intracranial\-volume\-normalised hippocampal volume in ADNI, standardised per cohort on the training fold\.
- 4\.Next\-visit MMSE and atrophy\(residual forecasts\):m^i,knext=mi,kobs\+Δ​m^i,k\\hat\{m\}^\{\\text\{next\}\}\_\{i,k\}=m^\{\\text\{obs\}\}\_\{i,k\}\+\\Delta\\hat\{m\}\_\{i,k\}, wheremi,kobsm^\{\\text\{obs\}\}\_\{i,k\}is the current observed value \(or the model’s current estimate when unobserved\) andΔ​m^i,k=MLPΔ​\(𝐳i,kmmse\)\\Delta\\hat\{m\}\_\{i,k\}=\\mathrm\{MLP\}\_\{\\Delta\}\(\\mathbf\{z\}^\{\\text\{mmse\}\}\_\{i,k\}\); atrophy is handled identically\. The residual head learns a change relative to the current value\. The observed current MMSE \(or volume\) is an explicit forecast anchor, even though that variable is excluded from the corresponding current\-estimation encoder\. The head is not conditioned on the interval to the next visit; next\-visit results therefore describe prediction at the next observed visit, whose interval varies, rather than at a fixed horizon\.
- 5\.Comorbidities\(multi\-label, 9 outputs\): 𝐜^i=σ⁡\(MLPcom​\(𝐳i,Tidiag\)\)\\hat\{\\mathbf\{c\}\}\_\{i\}=\\sigma\(\\mathrm\{MLP\}\_\{\\text\{com\}\}\(\\mathbf\{z\}^\{\\text\{diag\}\}\_\{i,T\_\{i\}\}\)\)read from the diagnosis\-branch state at the last visit, so that comorbidity information shapes the representation used for diagnosis\.

### 3\.5Robust Optimisation Objective

The training objective combines task balancing, subgroup robustness and trajectory regularisation\.

#### 3\.5\.1Base Task Losses

Each taskτ\\tauuses a masked loss

ℒτ=∑i∑kmi,kτ​ℓτ​\(i,k\)∑i∑kmi,kτ\+ϵ,\\mathcal\{L\}\_\{\\tau\}=\\frac\{\\sum\_\{i\}\\sum\_\{k\}m^\{\\tau\}\_\{i,k\}\\,\\ell\_\{\\tau\}\(i,k\)\}\{\\sum\_\{i\}\\sum\_\{k\}m^\{\\tau\}\_\{i,k\}\+\\epsilon\},\(7\)wheremi,kτm^\{\\tau\}\_\{i,k\}indicates that the label is*actually observed*at that visit \(imputed values are never used as targets\)\.ℓτ\\ell\_\{\\tau\}is cross\-entropy with label smoothing \(0\.1\) for diagnosis, focal loss for conversion \(from epoch 5\), Huber loss \(β=2\\beta=2\) for MMSE and atrophy, and binary cross\-entropy with capped positive\-class weights for comorbidities\. The task set is

𝒯=\{\\displaystyle\\mathcal\{T\}=\\\{dx,dx3,conv,MMSEcur\+next,\\displaystyle\\text\{dx\},\\ \\text\{dx\}\_\{3\},\\ \\text\{conv\},\\ \\text\{MMSE\}\_\{\\text\{cur\+next\}\},atrophycur\+next,com\}\.\\displaystyle\\text\{atrophy\}\_\{\\text\{cur\+next\}\},\\ \\text\{com\}\\\}\.

#### 3\.5\.2Comorbidity Mask

Rule\-derived recorded\-comorbidity information is available only for ADNI participants\. For OASIS\-2 and MIRIAD, comorbidities are treated asunknown, not as zeros\. A subject\-level availability mask distinguishes observed records from unavailable information:

micom=\{1ADNI medical history available,0otherwise\.m\_\{i\}^\{\\text\{com\}\}=\\begin\{cases\}1&\\text\{ADNI medical history available\},\\\\ 0&\\text\{otherwise\}\.\\end\{cases\}The comorbidity loss, the comorbidity part of the Group\-CVaR penalty and all comorbidity subgroup analyses use only subjects withmicom=1m\_\{i\}^\{\\text\{com\}\}=1\. This prevents conflating “no comorbidity” with “missing data”\.

#### 3\.5\.3Task\-DRO: Scale\-Free Dynamic Task Balancing

Task losses lie on very different scales \(focal vs\. Huber vs\. BCE\), so a softmax over raw losses would be dominated by whichever loss is numerically largest\. The entropy\-regularised DRO update\[[18](https://arxiv.org/html/2610.10648#bib.bib18)\]is therefore applied to*relative*losses\. Each task keeps an exponential moving averageℒ¯τ\\bar\{\\mathcal\{L\}\}\_\{\\tau\}\(decay 0\.98\), and

qτ=CappedSoftmaxqmax​\(log⁡\(1\+ℒτ/ℒ¯τ\)γ\),q\_\{\\tau\}=\\mathrm\{CappedSoftmax\}\_\{q\_\{\\max\}\}\\\!\\left\(\\frac\{\\log\\bigl\(1\+\\mathcal\{L\}\_\{\\tau\}/\\bar\{\\mathcal\{L\}\}\_\{\\tau\}\\bigr\)\}\{\\gamma\}\\right\),\(8\)withγ=0\.5\\gamma=0\.5\. The capped softmax iteratively clips any weight aboveqmax=0\.3q\_\{\\max\}=0\.3and redistributes the remaining mass, so that no single task can dominate\. The primary loss isℒTaskDRO=\|𝒯\|​∑τqτ​wτ​ℒτ\\mathcal\{L\}\_\{\\text\{TaskDRO\}\}=\|\\mathcal\{T\}\|\\sum\_\{\\tau\}q\_\{\\tau\}w\_\{\\tau\}\\mathcal\{L\}\_\{\\tau\}, wherewτw\_\{\\tau\}are base weights ramped in by a smooth curriculum \(diagnosis from epoch 1, MMSE/atrophy and comorbidity ramping in during the first epochs\)\. The base weights arewdx=0\.5w\_\{\\text\{dx\}\}=0\.5,wdx3=0\.01w\_\{\\text\{dx3\}\}=0\.01,wconv=1\.5w\_\{\\text\{conv\}\}=1\.5,wMMSE=0\.1w\_\{\\text\{MMSE\}\}=0\.1,watr=0\.05w\_\{\\text\{atr\}\}=0\.05andwcom=0\.2w\_\{\\text\{com\}\}=0\.2\. A task with a larger loss relative to its recent average receives a larger adaptive weight\.

#### 3\.5\.4Group\-CVaR over Cohorts and Comorbidity Strata

To target groups with higher losses, the objective includes a Conditional Value\-at\-Risk penalty\[[18](https://arxiv.org/html/2610.10648#bib.bib18)\]on per\-subject classification losses \(diagnosis and conversion\)\. For a set of groups𝒢\\mathcal\{G\}present in a mini\-batch, with mean lossℒg\\mathcal\{L\}\_\{g\},

ℛ𝒢=η\+1\(1−α\)​\|𝒢\|​∑g∈𝒢max⁡\(0,ℒg−η\),\\mathcal\{R\}\_\{\\mathcal\{G\}\}=\\eta\+\\frac\{1\}\{\(1\-\\alpha\)\|\\mathcal\{G\}\|\}\\sum\_\{g\\in\\mathcal\{G\}\}\\max\(0,\\mathcal\{L\}\_\{g\}\-\\eta\),\(9\)whereη\\etais theα\\alpha\-quantile \(linear interpolation\) of the group mean losses,α=0\.7\\alpha=0\.7\. The penalty is applied twice: over cohorts \(ADNI, OASIS\-2, MIRIAD\) and over coarse comorbidity strata \{none, cardiovascular \(hypertension, CAD, stroke\), metabolic \(diabetes, hyperlipidaemia\), depression, other \(arthritis, cancer, COPD\)\} for subjects withmicom=1m^\{\\text\{com\}\}\_\{i\}=1\. The strata are mutually exclusive: a subject with several conditions is assigned by priority depression\>\>cardiovascular\>\>metabolic\>\>other\. The penalty is skipped when fewer than two groups or fewer than two subjects per group are present\. The penalty is computed from unreduced per\-subject losses, so that group means genuinely differ\.ℛCVaR=ℛcohort\+ℛcom\\mathcal\{R\}\_\{\\text\{CVaR\}\}=\\mathcal\{R\}\_\{\\text\{cohort\}\}\+\\mathcal\{R\}\_\{\\text\{com\}\}, withλcvar=0\.1\\lambda\_\{\\text\{cvar\}\}=0\.1\.

#### 3\.5\.5Monotonicity Constraints

For subjects with a dementia label at any visit, a small hinge penaltyℒmono\\mathcal\{L\}\_\{\\text\{mono\}\}\(λmono=0\.01\\lambda\_\{\\text\{mono\}\}=0\.01\) discourages increases in predicted MMSE and atrophy\-target values between consecutive visits, averaged over visit pairs\. Here the atrophy targets are volume measures \(nWBV or normalised hippocampal volume\), so the discouraged direction is an increase in volume, rather than an increase in tissue loss\. The soft penalty applies to the current\-value estimates of MMSE and the atrophy target\. It encodes a preference for non\-increasing trajectories rather than a hard biological constraint: upward predictions remain possible, and actual measurements may fluctuate\. Its purpose is regularisation, not a claim that every individual follows a strictly monotonic course\.

#### 3\.5\.6Longitudinal Contrastive Regularisation

To encourage subject\-specific and temporally consistent diagnosis representations, a small MLP projects the evolved states at the last two visits of each subject,\(𝐳i,Ti−1diag,𝐳i,Tidiag\)\(\\mathbf\{z\}^\{\\text\{diag\}\}\_\{i,T\_\{i\}\-1\},\\mathbf\{z\}^\{\\text\{diag\}\}\_\{i,T\_\{i\}\}\), and a symmetric NT\-Xent loss is applied across the mini\-batch:

ℒcon=−12​B∑i\[log⁡esi​i/κ∑jesi​j/κ\+logesi​i/κ∑jesj​i/κ\],\\begin\{split\}\\mathcal\{L\}\_\{\\text\{con\}\}=\-\\tfrac\{1\}\{2B\}\\sum\_\{i\}\\Bigl\[&\\log\\tfrac\{e^\{s\_\{ii\}/\\kappa\}\}\{\\sum\_\{j\}e^\{s\_\{ij\}/\\kappa\}\}\\\\ &\+\\log\\tfrac\{e^\{s\_\{ii\}/\\kappa\}\}\{\\sum\_\{j\}e^\{s\_\{ji\}/\\kappa\}\}\\Bigr\],\\end\{split\}\(10\)wheresi​js\_\{ij\}is the cosine similarity between the projected states of subjectsiiandjj,κ=0\.1\\kappa=0\.1andλcon=0\.05\\lambda\_\{\\text\{con\}\}=0\.05\.

#### 3\.5\.7Components Deliberately Disabled

Uncertainty weighting\[[13](https://arxiv.org/html/2610.10648#bib.bib13)\]is not used, to avoid introducing a second task\-weighting mechanism alongside Task\-DRO\.

#### 3\.5\.8Final Objective

ℒtotal=ℒTaskDRO\+λcvar​ℛCVaR\+λimg​ℒimg\+λmono​ℒmono\+λcon​ℒcon\+λdrift​ℒdrift\+λΔ​‖Δ​𝐲^‖2,\\boxed\{\\begin\{aligned\} \\mathcal\{L\}\_\{\\text\{total\}\}&=\\mathcal\{L\}\_\{\\text\{TaskDRO\}\}\+\\lambda\_\{\\text\{cvar\}\}\\mathcal\{R\}\_\{\\text\{CVaR\}\}\+\\lambda\_\{\\text\{img\}\}\\mathcal\{L\}\_\{\\text\{img\}\}\+\\lambda\_\{\\text\{mono\}\}\\mathcal\{L\}\_\{\\text\{mono\}\}\\\\ &\\quad\+\\lambda\_\{\\text\{con\}\}\\mathcal\{L\}\_\{\\text\{con\}\}\+\\lambda\_\{\\text\{drift\}\}\\mathcal\{L\}\_\{\\text\{drift\}\}\+\\lambda\_\{\\Delta\}\\\|\\Delta\\hat\{\\mathbf\{y\}\}\\\|^\{2\},\\end\{aligned\}\}\(11\)whereℒimg\\mathcal\{L\}\_\{\\text\{img\}\}is the image\-level auxiliary loss and the last term \(λΔ=0\.05\\lambda\_\{\\Delta\}=0\.05\) penalises residual forecast magnitude after the first five anchor epochs\.

### 3\.6Network–Tabular Hybrid and Calibration

Gradient\-boosted trees remain very strong on tabular clinical data\. The complementary tabular stack is fitted on the training fold only and comprises the mean rank of a class\-balanced logistic regression and a histogram gradient\-boosting classifier using the same causal features as the corresponding network branch \(no MRI\)\. The hybrid score is a rank mixturew​rank​\(p^net\)\+\(1−w\)​rank​\(p^tab\)w\\,\\mathrm\{rank\}\(\\hat\{p\}\_\{\\text\{net\}\}\)\+\(1\-w\)\\,\\mathrm\{rank\}\(\\hat\{p\}\_\{\\text\{tab\}\}\), with one weightw∈\{0,0\.1,…,1\}w\\in\\\{0,0\.1,\\dots,1\\\}per cohort selected on validation predictions\. Probabilities are calibrated by temperature scaling\[[58](https://arxiv.org/html/2610.10648#bib.bib58)\]\. Within each fold, the temperature is fitted on that fold’s validation predictions, and internal calibration metrics are computed on the same predictions; internal calibration is therefore in\-sample with respect to the temperature\. For external evaluation, the temperatures fitted on the internal folds are applied unchanged to OASIS\-3, so external calibration is out\-of\-sample\. Unless stated otherwise, all claims about the model refer to the*network alone*; the hybrid is reported as a complementary result\.

### 3\.7Training Algorithm

Algorithm[1](https://arxiv.org/html/2610.10648#alg1)summarises the training procedure\.

Algorithm 1Training with Task\-DRO, Group\-CVaR and Auxiliary Terms1:Input:model

θ\\theta, tasks

𝒯\\mathcal\{T\}, cohorts and comorbidity strata

𝒢\\mathcal\{G\},

γ\\gamma,

qmaxq\_\{\\max\},

α\\alpha,

λcvar\\lambda\_\{\\text\{cvar\}\},

λmono\\lambda\_\{\\text\{mono\}\},

λcon\\lambda\_\{\\text\{con\}\},

λdrift\\lambda\_\{\\text\{drift\}\}\.

2:foreach epoch

eedo

3:foreach mini\-batch of subjectsdo

4:Encode MRI \(SFCN \+ LoRA\) and the four clinical streams \(causal Transformers\)

5:Fuse each stream with MRI \(Eq\.[4](https://arxiv.org/html/2610.10648#S3.E4)\); evolve with ODE\-GRU \(Eqs\.[5](https://arxiv.org/html/2610.10648#S3.E5)–[6](https://arxiv.org/html/2610.10648#S3.E6)\)

6:Compute masked task losses

ℒτ\\mathcal\{L\}\_\{\\tau\}on observed labels only \(Eq\.[7](https://arxiv.org/html/2610.10648#S3.E7)\)

7:Update EMA

ℒ¯τ\\bar\{\\mathcal\{L\}\}\_\{\\tau\}; compute capped Task\-DRO weights

qτq\_\{\\tau\}\(Eq\.[8](https://arxiv.org/html/2610.10648#S3.E8)\)

8:

ℒTaskDRO←\|𝒯\|​∑τqτ​wτ​\(e\)​ℒτ\\mathcal\{L\}\_\{\\text\{TaskDRO\}\}\\leftarrow\|\\mathcal\{T\}\|\\sum\_\{\\tau\}q\_\{\\tau\}w\_\{\\tau\}\(e\)\\mathcal\{L\}\_\{\\tau\}
9:// Group\-CVaR on unreduced per\-subject losses

10:

ℛCVaR←ℛcohort\+ℛcom\\mathcal\{R\}\_\{\\text\{CVaR\}\}\\leftarrow\\mathcal\{R\}\_\{\\text\{cohort\}\}\+\\mathcal\{R\}\_\{\\text\{com\}\}\(Eq\.[9](https://arxiv.org/html/2610.10648#S3.E9); groups with

<2<2subjects skipped;

ℛcom\\mathcal\{R\}\_\{\\text\{com\}\}only on

mcom=1m^\{\\text\{com\}\}=1\)

11:Compute

ℒimg\\mathcal\{L\}\_\{\\text\{img\}\}on visits with MRI,

ℒmono\\mathcal\{L\}\_\{\\text\{mono\}\}\(Sec\.[3\.5\.5](https://arxiv.org/html/2610.10648#S3.SS5.SSS5)\),

ℒcon\\mathcal\{L\}\_\{\\text\{con\}\}\(Eq\.[10](https://arxiv.org/html/2610.10648#S3.E10)\), and

ℒdrift\\mathcal\{L\}\_\{\\text\{drift\}\}
12:Compute

ℒtotal\\mathcal\{L\}\_\{\\text\{total\}\}\(Eq\.[11](https://arxiv.org/html/2610.10648#S3.E11)\); accumulate gradients \(2 steps\), clip, AdamW update

13:endfor

14:Evaluate on the validation fold; keep the checkpoint with the best selection score; early stopping

15:endfor

16:Fit temperature and conversion threshold on validation predictions; fit the tabular stack on the training fold

## 4Experimental Setup

### 4\.1Datasets

The development dataset combines three publicly available longitudinal cohorts with heterogeneous populations, imaging protocols and comorbidity profiles: OASIS\-2\[[59](https://arxiv.org/html/2610.10648#bib.bib59)\], MIRIAD\[[60](https://arxiv.org/html/2610.10648#bib.bib60)\]and ADNI\[[61](https://arxiv.org/html/2610.10648#bib.bib61)\]\. The assembled ADNI subset contains 2,430 subjects\. The aggregated development set comprises2,649 subjects and 17,317 visits, with1,081 indexed MRI acquisitions across 896 MRI\-available visits from 219 subjects\(Table[2](https://arxiv.org/html/2610.10648#S4.T2)\)\. OASIS\-3 is held out entirely for external testing\.

Table 2:Dataset statistics per source\. MIRIAD provides several acquisitions per visit \(708 acquisitions over 523 visits\)\.SourceSubjectsVisitsMRI acquisitionsOASIS\-2150373373MIRIAD69523708ADNI2,43016,4210Total2,64917,3171,081Diagnosis label mapping\.The primary task is a*cohort\-defined binary diagnosis endpoint*, not a uniform CN\-versus\-dementia definition\. ADNI excludes MCI visits from this task, MIRIAD contrasts controls and AD, and OASIS\-2 contrasts CDR<1<1with CDR≥1\\geq 1\. In particular, OASIS\-2 CDR 0\.5 visits belong to the negative class and are not cognitively normal controls\. The class counts are 4,526 negative and 2,839 positive visits\. Table[3](https://arxiv.org/html/2610.10648#S4.T3)gives the mapping; stage labels are harmonisation codes rather than interchangeable clinical diagnoses\.

Table 3:Diagnosis label mapping per cohort\. Stage codes: 0 = CN/control/CDR 0, 1 = ADNI MCI or OASIS\-2 CDR 0\.5, 2 = dementia/AD/CDR≥1\\geq 1\. Stage 1 is not clinically equivalent across cohorts\.CohortSource labelsBinary label \(stage\)OASIS\-2CDR 0 / 0\.5 /≥\\geq10 = CDR<<1 \(0, 1\)1 = CDR≥\\geq1 \(2\)MIRIADControl, AD0 = control \(0\)1 = AD \(2\)ADNICN, MCI, Dementia0 = CN \(0\)1 = Dementia \(2\)MCI: excluded \(1\)Nominal 3\-year progression label\.For visits with stage code 0 or 1, a positive label is assigned when the first later observed visit with a higher stage occurs within 3 years \(tolerance 0\.15 years, i\.e\.≤3\.15\\leq 3\.15years\); a negative label requires no observed increase and at least 2\.85 years of follow\-up\. Shorter follow\-up without progression is treated as an unknown outcome\. Stage decreases are not events: conversion is defined by the first observed stage increase\. There are 5,375 labelled at\-risk visits, including 1,443 positives \(ADNI 27%, OASIS\-2 23%\)\. MIRIAD has no stage\-1 at\-risk visits\. Evaluation covers three views: \(i\) MCI visits, \(ii\) first\-MCI\-visit sMCI versus pMCI, and \(iii\) all stage\-0/1 visits\. For ADNI, MCI\-to\-dementia is a clinical conversion endpoint; OASIS\-2 CDR\-stage progression is different\. The MCI views contain stage\-1 visits from ADNI \(clinical MCI\) and OASIS\-2 \(CDR 0\.5\): 3,005 visits and 850 subjects\.

#### 4\.1\.1OASIS\-2

OASIS\-2 provides MRI, age, sex, education, SES, eTIV, nWBV, ASF and MMSE; it lacks PET, CSF, regional volumes, other cognitive tests and medical history\. CDR defines the label and is never an input\.

#### 4\.1\.2MIRIAD

MIRIAD provides repeated T1 scans, age, sex and MMSE; it lacks education, SES, brain volumes, PET, CSF and medical history\. The label is the study group \(control or AD\); CDR is not used as an input\.

#### 4\.1\.3ADNI

ADNI provides regional MRI volumes\[[62](https://arxiv.org/html/2610.10648#bib.bib62)\], PET, CSF, APOE4, cognitive and functional tests, creatinine and family history; raw ADNI MRI is not used\. Medications, the participant’s own medical\-history items and condition\-defining laboratory values \(glucose, HbA1c, cholesterol, triglycerides\) are excluded from the inputs because they define the comorbidity labels; family history is a separate predictor and is not used to construct those labels\.

#### 4\.1\.4Comorbidity Representation

Nine binary comorbidities \(hypertension, diabetes, depression, hyperlipidaemia, arthritis, cancer, coronary artery disease \(CAD\), stroke and COPD\) are derived from ADNI medical histories by keyword rules that ignore negated and family\-history mentions\. They are available for 1,739 ADNI subjects \(micom=1m\_\{i\}^\{\\text\{com\}\}=1\), from 835 hypertension to 54 COPD cases; for OASIS\-2 and MIRIAD they are unknown, not zero\.

### 4\.2Preprocessing and Leakage Control

Figure[2](https://arxiv.org/html/2610.10648#S4.F2)summarises cohort alignment, subject\-level folds, fold\-wise transformation fitting and model\-ready sequences\.

![Refer to caption](https://arxiv.org/html/2610.10648v1/figures/figures/data_preprocessing.jpg)Figure 2:Data preparation and internal validation\. Cohorts are aligned, labels constructed and features filled causally; imputation and scaling are fitted on training subjects only, MRI is processed per volume, and all visits of a participant remain within one of five subject\-grouped folds\.#### 4\.2\.1Imaging

T1 volumes undergo foreground masking using Otsu thresholding and the largest connected component, are resized to1283128^\{3\}and intensity\-normalised with the statistics of the brain mask of the*same volume*\. No dataset\-level statistic is involved\. No spatial registration or voxel\-spacing resampling is applied: each volume is resized as a whole to the fixed grid, and the encoder operates in native orientation\.

#### 4\.2\.2Causal Feature Construction

Time is measured in years since each subject’s first visit\. Sparse longitudinal measures \(MMSE, volumes, PET, CSF, cognitive tests\) are filled with*last observation carried forward only*; the first visit is never back\-filled, and each filled value is accompanied by an availability flag and the time since its last observation\. Rates of change are computed from the two most recent observations*at or before*the current visit \(minimum gap 0\.25 years\)\. All remaining medians, scalers, target standardisers and the comorbidity class weights are fitted on the training fold and applied unchanged to the validation fold\. Inputs are clipped to±5\\pm 5standard deviations\.

#### 4\.2\.3Branch\-Specific Prohibited Inputs

Several variables are part of how the target itself is defined and are therefore removed from the corresponding branch \(Table[4](https://arxiv.org/html/2610.10648#S4.T4)\)\. In particular, the*diagnosis*branch never sees MMSE, any cognitive or functional test, CDR or the current clinical stage, because these assessments can contribute to clinical diagnosis or encode closely related target information\. The diagnosis branch therefore uses biological measures, demographics and history\. The forecast branch may use cognitive tests and, for ADNI, the current stage \(MCI or not\), because these are known at the time of the forecast and are not the future outcome\. The MMSE and atrophy branches start from the forecast input set and remove their own targets: the MMSE branch excludes MMSE, its rates and composite scores that contain MMSE, and the atrophy branch excludes nWBV and hippocampal volume and their rates\. Under the evaluated configuration, both therefore also receive current clinical\-stage information for ADNI \(MCI or not\), which is part of the clinical diagnostic assessment, together with the remaining cognitive tests and regional volumes\. The stage indicator is set only for cohorts with a clinical diagnosis and is never derived from CDR, so OASIS\-2 stage is not an input\. CDR is excluded everywhere because it defines the OASIS\-2 labels\. Prohibited inputs are removed when each branch’s input list is built, and an assertion checks each branch against its prohibited list\.

Table 4:Branch\-specific prohibited inputs \(removed when input lists are built\)\.BranchProhibited inputsDiagnosisCDR, MMSE, all cognitive/functional tests, current stageForecastCDR \(and its sum of boxes\)MMSECDR, MMSE and its rates, composite scores that contain MMSEAtrophyCDR, nWBV, hippocampal volume and their rates
#### 4\.2\.4Cross\-Validation

Internal evaluation uses 5\-foldStratifiedGroupKFoldgrouped by subject \(split seed 42\): every visit and every MRI file of a subject belongs to exactly one fold\. In each fold the model is trained on four\-fifths of the subjects and evaluated on the held\-out fifth, which also drives early stopping, the conversion threshold, the temperature and the hybrid weights\. The resulting metrics are therefore*internal\-validation*estimates\. The subject\-bootstrap intervals condition on the saved predictions\. Mini\-batches are balanced across cohorts\.

#### 4\.2\.5Empirical Leakage Checks

Three checks are run for every fold model and written to a fold\-level report:\(P1\) partition: zero subjects and zero MRI files are shared between training and validation, and no label\-like column or prohibited input reaches any branch;\(P2\) feature causality: for 60 random subjects, every derived feature at visitkkis recomputed after deleting all visits afterkkand must be identical;\(P3\) model causality: for 40 random subjects, every model output at visit≤k\\leq kmust be unchanged \(tolerance10−310^\{\-3\}\) when visits\>k\>kare deleted\.

### 4\.3Training Details

Training uses AdamW \(weight decay5×10−45\\times 10^\{\-4\}\) with differentiated learning rates: Transformers and fusion5×10−55\\times 10^\{\-5\}, classification heads1×10−41\\times 10^\{\-4\}, regression heads3×10−43\\times 10^\{\-4\}, MMSE head3×10−53\\times 10^\{\-5\}, ODE\-GRU1×10−41\\times 10^\{\-4\}, LoRA adapters1×10−41\\times 10^\{\-4\}and the fine\-tuned SFCN blocks1×10−51\\times 10^\{\-5\}\. The batch size is 8 subjects with gradient accumulation over 2 steps, automatic mixed precision \(ODE integration in float32\), gradient\-norm clipping at 0\.5 and a ReduceLROnPlateau scheduler\. Training runs for at most 25 epochs with early stopping \(patience 5, minimum 7 epochs\) on a validation selection score0\.35​AUCdx\+0\.45​AUCconv\+0\.10​SMMSE\+0\.10​AUCcom0\.35\\,\\text\{AUC\}\_\{\\text\{dx\}\}\+0\.45\\,\\text\{AUC\}\_\{\\text\{conv\}\}\+0\.10\\,S\_\{\\text\{MMSE\}\}\+0\.10\\,\\text\{AUC\}\_\{\\text\{com\}\}, whereSMMSE=clip⁡\(1−MAEnext/MAEpersist,0,1\)S\_\{\\text\{MMSE\}\}=\\mathrm\{clip\}\(1\-\\text\{MAE\}\_\{\\text\{next\}\}/\\text\{MAE\}\_\{\\text\{persist\}\},0,1\)compares the next\-visit MMSE error with that of carrying the current value forward\. The latent dimension is 128 throughout\.

The remaining hyperparameters areγ=0\.5\\gamma=0\.5,qmax=0\.3q\_\{\\max\}=0\.3,α=0\.7\\alpha=0\.7,λcvar=0\.1\\lambda\_\{\\text\{cvar\}\}=0\.1,λmono=0\.01\\lambda\_\{\\text\{mono\}\}=0\.01,λcon=0\.05\\lambda\_\{\\text\{con\}\}=0\.05\(κ=0\.1\\kappa=0\.1\),λdrift=10−3\\lambda\_\{\\text\{drift\}\}=10^\{\-3\},λimg=0\.2\\lambda\_\{\\text\{img\}\}=0\.2, LoRAr=8r=8,αLoRA=16\\alpha\_\{\\text\{LoRA\}\}=16\.

The full model is trained with three seeds \(42, 123, 2026\) on the same folds\. Training uses one NVIDIA A40 GPU per run\.

### 4\.4Evaluation Metrics

Primary metrics use pooled out\-of\-fold predictions: the corresponding network was not fitted on the subject, but its fold was reused for model selection and post\-processing\. 95% confidence intervals use a subject\-level bootstrap \(1,000 resamples\)\. NR: not reported\. Unless stated otherwise,±\\pmdenotes the standard deviation across the three training seeds; fold\-level standard deviations are labelled as such\. Arrows indicate the optimisation direction\.

- 1\.Diagnosis:AUROC↑\\uparrow, average precision↑\\uparrow, balanced accuracy↑\\uparrow, expected calibration error \(ECE\)↓\\downarrowbefore and after temperature scaling, with 10 equal\-width bins; bins with fewer than five predictions are omitted\.
- 2\.3\-year conversion:AUROC↑\\uparrowon the three views defined above; conversion ECE uses all eligible CN\+MCI visits rather than MCI visits alone\.
- 3\.Cognition and atrophy:current and next\-visit MMSE MAE↓\\downarrowandR2R^\{2\}↑\\uparrow; next\-visit atrophyR2R^\{2\}↑\\uparrow\. Regression metrics use only*truly observed*targets, never imputed values\.
- 4\.Comorbidities:per\-condition AUROC↑\\uparrowand mean AUROC over the nine conditions \(ADNI subjects withmcom=1m^\{\\text\{com\}\}=1\)\.
- 5\.Subgroup robustness:diagnosis and conversion AUROC within each of the nine comorbidity strata \(subjects*with*the condition\); each task is summarised by the worst stratum \(the lowest value, never the best\), the mean of the three worst, and the disparity \(best minus worst\)\.
- 6\.Paired comparisons:Δ\\DeltaAUROC between two models on the same subjects, with a 95% CI from a paired subject\-level bootstrap \(1,000 resamples\); a CI excluding zero is described as nominal evidence of a difference\. No multiple\-comparison adjustment is applied\.

### 4\.5Ablation Protocol

The ablations reuse the same five folds and seed 42, so comparisons can be paired at the subject level\. Most remove one component; the comorbidity variant jointly removes its prediction task and its CVaR strata:

1. 1\.w/o Task\-DRO:fixed curriculum weightswτw\_\{\\tau\}\.
2. 2\.w/o Group\-CVaR:λcvar=0\\lambda\_\{\\text\{cvar\}\}=0\.
3. 3\.w/o contrastive:λcon=0\\lambda\_\{\\text\{con\}\}=0\.
4. 4\.w/o comorbidity head:no comorbidity task and no comorbidity strata in CVaR\.
5. 5\.w/o LoRA:LoRA adapters removed \(conv4–5 still fine\-tuned\)\.
6. 6\.Clinical\-only:MRI replaced by the no\-MRI embedding for all subjects\.
7. 7\.Imaging\-focused:clinical encoder inputs removed and temporal features retained\. The diagnosis head then sees only the image embedding and visit\-time features, whereas the conversion head still receives the static\-covariate projection, so its conversion results are not image\-only\.
8. 8\.GRU \(no ODE\):identical cell and time\-aware gates \(Eq\.[6](https://arxiv.org/html/2610.10648#S3.E6)\), but the state is held constant between visits instead of being integrated\.

## 5Results

### 5\.1Main Results

Table[5](https://arxiv.org/html/2610.10648#S5.T5)summarises the full model on pooled out\-of\-fold predictions\. The network reaches a diagnosis AUROC of0\.936\[0\.925, 0\.948\] \(0\.935±\\pm0\.002 over three seeds\) and a3\-year stage\-1 progression AUROC of 0\.881\[0\.858, 0\.902\] \(0\.884±\\pm0\.003\)\. At the subject level, the first stage\-1 visit separates stable from progressive cases with an AUROC of 0\.872 \(0\.870±\\pm0\.005\)\. Combining the network with the tabular stack gives0\.951,0\.909and0\.896\(three\-seed means\)\. The tabular stack alone reaches 0\.948, 0\.908 and 0\.894 \(seed 42; Figure[3](https://arxiv.org/html/2610.10648#S5.F3)\), so the hybrid improves only slightly on it\. Current MMSE is estimated with an MAE of 1\.58 points and next\-visit MMSE with an MAE of 1\.62 points \(R2=0\.692R^\{2\}=0\.692\)\. Temperature scaling lowers the diagnosis ECE from 0\.135 to 0\.031\. Across comorbidity strata, diagnosis stays at or above0\.819\(worst: COPD\) and conversion at or above0\.765\(worst: stroke\)\. The auxiliary comorbidity head achieves a mean AUROC of 0\.565 from the permitted inputs\. Seed standard deviations are below 0\.03 for all listed metrics\. Figure[3](https://arxiv.org/html/2610.10648#S5.F3)shows the ROC curves\.

Table 5:Overall performance of the full model \(SFCN, internal 5\-fold validation\)\. Value: seed 42, pooled out\-of\-fold, with 95% subject\-bootstrap CI; last column: mean±\\pmSD over seeds 42/123/2026\. Hybrid = network \+ tabular stack\.MetricValue \[95% CI\]3 seedsDiagnosis \(cohort\-defined binary endpoint\)AUROC, network↑\\uparrow0\.936 \[0\.925, 0\.948\]0\.935±\\pm0\.002AUROC, hybrid↑\\uparrow0\.951 \[0\.941, 0\.961\]0\.951±\\pm0\.001Average precision, network / hybrid↑\\uparrow0\.917 / 0\.9360\.915 / 0\.936Balanced accuracy↑\\uparrow0\.8700\.865±\\pm0\.008ECE, raw→\\tocalibrated↓\\downarrow0\.135→\\to0\.0310\.134→\\to0\.0253\-year conversionMCI visits, network↑\\uparrow0\.881 \[0\.858, 0\.902\]0\.884±\\pm0\.003MCI visits, hybrid↑\\uparrow0\.9100\.909±\\pm0\.001sMCI vs\. pMCI, network↑\\uparrow0\.872 \[0\.849, 0\.894\]0\.870±\\pm0\.005sMCI vs\. pMCI, hybrid↑\\uparrow0\.8970\.896±\\pm0\.001CN\+MCI visits, network / hybrid↑\\uparrow0\.858 / 0\.8880\.858 / 0\.888ECE, raw→\\tocalibrated↓\\downarrow0\.167→\\to0\.0570\.178→\\to0\.059Cognition and atrophyCurrent MMSE MAE↓\\downarrow1\.5771\.592±\\pm0\.021Next\-visit MMSE MAE↓\\downarrow1\.6151\.611±\\pm0\.010Next\-visit MMSER2R^\{2\}↑\\uparrow0\.6920\.698±\\pm0\.006Next\-visit atrophyR2R^\{2\}↑\\uparrow0\.8280\.826±\\pm0\.002Comorbidity and subgroups \(ADNI,mcom=1m^\{\\text\{com\}\}=1\)Comorbidity mean AUROC↑\\uparrow0\.5650\.571±\\pm0\.010Worst\-stratum diagnosis↑\\uparrow0\.819 \(COPD\)0\.827±\\pm0\.008 \(COPD\)Worst\-stratum conversion↑\\uparrow0\.765 \(stroke\)0\.774±\\pm0\.029Hardest comorbidity↑\\uparrow0\.512 \(diabetes\)0\.517±\\pm0\.006

Entries without brackets: no CI computed\. Seed SD is distinct from a bootstrap CI\. Worst\-stratum conversion is the lowest stratum in each seed \(stroke, diabetes or COPD\)\.

Figure 3:ROC curves on pooled out\-of\-fold predictions \(seed 42\) for diagnosis and two conversion views \(all MCI visits; first MCI visit per subject\), for the network, the tabular stack and the hybrid\. Legends give the AUROC with its 95% subject\-bootstrap CI\. The two progression panels use different visit\-selection rules and denominators; they are complementary evaluation views, not independent replication\.
### 5\.2External Validation on OASIS\-3

OASIS\-3 was excluded from model development and reserved for external validation after the model, fold preprocessors, MMSE and atrophy scalers, tabular baselines and temperature scalers were frozen on the internal folds\. The five fold models trained on ADNI, OASIS\-2 and MIRIAD were applied unchanged to every OASIS\-3 visit; no OASIS\-3 label, image or statistic entered fitting, threshold selection, calibration or hybrid weighting\. An attribute\-based overlap screen was applied to the OASIS\-3 candidate pool before model evaluation to reduce potential participant overlap with OASIS\-2\. A subject was flagged when sex, education and handedness matched an OASIS\-2 participant, the OASIS\-3 baseline age exceeded the OASIS\-2 baseline age by 3–20 years, the absolute difference in estimated total intracranial volume was at most 30,000 mm3, and the OASIS\-3 baseline age was at least 60\. The screen excluded 135 participants; the frozen exclusion list is reused by the build script, which asserts that no excluded subject appears in the prepared cohort\. One T1 was indexed per MRI session, and each session was linked to the nearest UDS visit within±270\\pm 270days, a matching window that permits imaging and clinical assessments to occur months apart\. The prepared cohort contains 1,152 subjects and 2,407 visits \(1,995 diagnosis eligible, 707 progression eligible, 980 comorbidity eligible\); every prepared visit carries a T1, so the MRI encoder was evaluated on all records\. Within each seed, every fold model is applied with its own training\-fold preprocessing, tabular stack and hybrid weight; each fold’s saved temperature is applied to its own predictions, and raw and calibrated predictions are then averaged per subject–visit across the five fold models, with the decision threshold set to the mean of the fold thresholds\. The primary external result is the per\-visit average of the three seeds’ predictions, computed before any metric is evaluated\. External progression is evaluated at the next observed visit, not at a fixed 3\-year horizon\. The external analysis evaluates next\-visit transfer separately from the internal 3\-year endpoint\.

Table[6](https://arxiv.org/html/2610.10648#S5.T6)summarises the external results\. The diagnosis AUROC is 0\.763 for the network and 0\.767 for the hybrid\. For next\-observed\-visit progression, the network achieves 0\.733 and the hybrid 0\.764, compared with 0\.710 for logistic regression\. The tabular stack achieves a progression AUROC of 0\.768, while logistic regression achieves the highest diagnosis point estimate \(0\.787\)\. Transferred temperature scaling reduces diagnosis ECE from 0\.197 to 0\.052; progression ECE changes from 0\.401 to 0\.303\. Next\-visit MMSE prediction achieves an MAE of 0\.86 andR2=0\.24R^\{2\}=0\.24, and current MMSE estimation achieves an MAE of 1\.80 andR2=−0\.03R^\{2\}=\-0\.03\. Comorbidity prediction was evaluated externally for eight of the nine conditions \(hypertension, diabetes, depression, hyperlipidaemia, arthritis, cancer, coronary artery disease and stroke\); COPD was not evaluable because none of the 980 OASIS\-3 subjects with comorbidity records was COPD\-positive\. The mean comorbidity AUROC across these eight conditions is 0\.565\. These results extend the evaluation to a separate cohort across diagnostic, progression, cognitive and auxiliary outcomes\. External atrophy prediction is outside the present evaluation\.

Table 6:External validation on OASIS\-3, three\-seed per\-visit ensemble\. Counts denote eligible records; diagnosis and progression are AUROC; MMSE rows report MAE /R2R^\{2\}\. Progression uses the next observed visit rather than the internal 3\-year horizon\. No paired confidence intervals are available for these contrasts\.Task / metricNetworkTabularHybridLogisticDiagnosis AUROC \(n=1,995n=1\{,\}995\)0\.7630\.7620\.7670\.787Progression AUROC \(n=707n=707\)0\.7330\.7680\.7640\.710Current MMSE MAE /R2R^\{2\}\(n=2,407n=2\{,\}407\)1\.80 /−0\.03\-0\.03–––Next\-visit MMSE MAE /R2R^\{2\}\(n=1,255n=1\{,\}255\)0\.86 / 0\.24–––Diagnosis ECE, raw→\\tocalibrated0\.197→\\to0\.052–––Progression ECE, raw→\\tocalibrated0\.401→\\to0\.303–––
Mean comorbidity AUROC: 0\.565 \(n=980n=980\)\. The evaluation includes fewer conditions than the nine\-output training head\.

### 5\.3Context with Published Results

The studies in Table[1](https://arxiv.org/html/2610.10648#S2.T1)use different cohorts, label definitions, horizons and validation designs, so their scores provide context rather than a ranking\.

### 5\.4Ablation Studies

Table[7](https://arxiv.org/html/2610.10648#S5.T7)reports all ablations together with paired subject\-bootstrap tests against the full model on the same subjects\. For the component ablations, the observed differences in pooled diagnosis, MCI\-visit conversion and sMCI/pMCI AUROC lie between−0\.0038\-0\.0038and\+0\.0036\+0\.0036, and the paired intervals include zero except for MCI\-visit conversion without Group\-CVaR \(\+0\.0036\+0\.0036\[0\.0001, 0\.0069\], nominal and unadjusted\)\. Key observations:

- 1\.Modalities\.Removing the clinical encoder inputs \(imaging\-focused\) lowers diagnosis by 0\.272 \[0\.252, 0\.293\], MCI\-visit conversion by 0\.051 \[0\.031, 0\.072\] and sMCI/pMCI by 0\.056 \[0\.033, 0\.078\], and raises the MMSE error from 1\.58 to 2\.83\. Removing MRI \(clinical\-only\) gives differences whose intervals include zero \(diagnosis−0\.002\-0\.002\[−0\.007\-0\.007, 0\.004\]\); in this configuration the pooled signal is carried mainly by clinical inputs and derived volumes\.
- 2\.Longitudinal contrastive loss\.Without it, the mean comorbidity AUROC falls from 0\.565 to 0\.540, with lower values for eight of the nine conditions\.
- 3\.LoRA\.Without LoRA \(conv4–5 still fine\-tuned\), the worst\-stratum diagnosis AUROC is 0\.891, compared with 0\.819 for the full model; the worst strata are small and these are single\-seed estimates\.
- 4\.Task\-DRO\.Without Task\-DRO, the MMSE errors are 1\.554 \(current\) and 1\.599 \(next visit\), compared with 1\.577 and 1\.615\.
- 5\.Comorbidity head\.Removing the auxiliary task gives paired differences of\+0\.001\+0\.001\[−0\.003\-0\.003, 0\.006\] and−0\.002\-0\.002\[−0\.008\-0\.008, 0\.005\] for diagnosis and MCI\-visit conversion, respectively\.
- 6\.Group\-CVaR\.Its subgroup effect is analysed in Section[5\.9](https://arxiv.org/html/2610.10648#S5.SS9)\.

Table 7:Ablation study \(network, seed 42, same five folds\)\. \* = paired subject\-bootstrap 95% CI of the AUROC difference to the full model excludes 0 \(the three AUROC columns\)\. The last row gives the seed standard deviation of the full model \(seeds 42/123/2026\) to describe variation over three runs, not as a significance threshold\. Asterisks denote nominal, unadjusted contrasts\.ExperimentDiagnosis3\-year conversion AUROC↑\\uparrowCognition / atrophyComorbidity \(ADNI\)AUROC↑\\uparrowMCI visitssMCI/pMCIMMSE MAE↓\\downarrowNext MMSE MAE↓\\downarrowNext atrophyR2R^\{2\}↑\\uparrowMean AUROC↑\\uparrowWorst\-stratum dx↑\\uparrowFull model0\.9360\.8810\.8721\.5771\.6150\.8280\.5650\.819Component ablationsw/o Task\-DRO0\.9370\.8800\.8691\.5541\.5990\.8330\.5770\.834w/o Group\-CVaR0\.9370\.885\*0\.8741\.5971\.6190\.8340\.5650\.828w/o contrastive0\.9380\.8840\.8711\.5571\.5940\.8390\.5400\.837w/o comorbidity head0\.9370\.8800\.8701\.5741\.6100\.834NRNRw/o LoRA0\.9380\.8800\.8751\.5941\.6170\.8350\.5650\.891Modality ablationsClinical\-only0\.9350\.8870\.8701\.5751\.6110\.8290\.5740\.787Imaging\-focused0\.664\*0\.830\*0\.817\*2\.8252\.2330\.582NRNRTemporal modelGRU, no ODE \(time\-aware gates\)0\.9370\.8840\.8691\.5761\.6030\.8370\.5700\.851Full model, seed SD±\\pm0\.002±\\pm0\.003±\\pm0\.005±\\pm0\.021±\\pm0\.010±\\pm0\.002±\\pm0\.010±\\pm0\.008

\* Paired subject\-bootstrap 95% CI of the difference to the full model excludes 0 \(nominal, unadjusted\); markers apply to the three AUROC columns only\. NR: comorbidity prediction metrics are unavailable where the auxiliary head is absent; the saved outputs of these runs contain no comorbidity labels, so worst\-stratum diagnosis was not computed for them\. Seed SD: standard deviation of the full model over seeds 42/123/2026\. Paired differences quoted in the text are computed from unrounded AUROCs and can differ by 0\.001 from differences of the rounded table values\.

### 5\.5MRI Encoder Comparison

To assess the choice of MRI encoder, the SFCN encoder was replaced by a MedicalNet\-pretrained 3D ResNet\-50 with the same adaptation \(last\-block fine\-tuning and LoRA, rank 8\) and all other settings unchanged \(seed 42\)\. Table[8](https://arxiv.org/html/2610.10648#S5.T8)reports accuracy and cost\. The three encoder configurations reach comparable AUROC, with differences of at most 0\.006 and overlapping 95% subject\-bootstrap CIs \(ResNet\-50: diagnosis 0\.932 \[0\.920, 0\.945\], MCI conversion 0\.884 \[0\.862, 0\.904\]\), while SFCN uses about one seventh of the total parameters of ResNet\-50 and requires less training time\. SFCN with LoRA is therefore retained as the MRI encoder\.

Table 8:MRI encoder comparison \(seed 42, internal 5\-fold validation\)\. Dx: diagnosis AUROC; MCI: MCI\-conversion AUROC; sMCI: sMCI/pMCI AUROC\. Parameters are for the full model; s/epoch is mean training time per epoch; 5\-fold time is total wall\-clock time over the five folds\. The two SFCN rows repeat the corresponding AUROCs of Table[7](https://arxiv.org/html/2610.10648#S5.T7); this table adds the ResNet\-50 comparison and computational cost\.EncoderDxMCIsMCIParams \(total/trainable\)s/epoch5\-fold timeSFCN \+ LoRA \(main\)0\.9360\.8810\.8726\.97M / 4\.02M1875\.33 hSFCN w/o LoRA0\.9380\.8800\.8756\.91M / 3\.96M2175\.32 hResNet\-50 \+ LoRA0\.9320\.8840\.87150\.50M / 4\.34M2086\.18 h

### 5\.6Continuous\-Time Dynamics

The ODE\-GRU and the time\-aware GRU share the same cell and gates and differ only in between\-visit integration\. The paired differences \(GRU minus ODE\) are\+0\.001\+0\.001\[−0\.003\-0\.003, 0\.004\] for diagnosis,\+0\.003\+0\.003\[−0\.003\-0\.003, 0\.009\] for MCI\-visit conversion and−0\.003\-0\.003\[−0\.010\-0\.010, 0\.004\] for sMCI/pMCI; no statistically detectable AUROC difference was observed\. The ODE formulation additionally provides a latent state defined between visits\.

### 5\.7Performance per Cohort

Table[9](https://arxiv.org/html/2610.10648#S5.T9)and Figure[4](https://arxiv.org/html/2610.10648#S5.F4)break down diagnosis by cohort\. ADNI, with rich biomarkers, reaches 0\.947 \(hybrid 0\.961\), and OASIS\-2 reaches 0\.629 \(hybrid 0\.803\)\. On MIRIAD, the diagnosis AUROC of the fused model is 0\.424, which is no better than chance: there, the diagnosis branch sees only demographics and the MRI embedding, because MMSE is prohibited, and this cohort has fewer permitted clinical inputs than ADNI\. When the clinical encoder inputs are removed \(imaging\-focused setting\), the image pathway reaches 0\.597 on MIRIAD and 0\.621 on OASIS\-2\. The imaging\-focused comparison complements the fused\-model evaluation by characterising performance under reduced clinical input\.

Table 9:Diagnosis AUROC per cohort \(seed 42, pooled out\-of\-fold\)\.CohortVisitsNetworkHybridADNI6,4690\.9470\.961OASIS\-23730\.6290\.803MIRIAD5230\.4240\.456Figure 4:Cohort\-specific internal diagnosis AUROC for the network and hybrid, seed 42\. Intervals use 1,000 subject\-cluster bootstrap resamples retaining all visits\. Cohort\-specific endpoints are heterogeneous\. Cohort differences reflect distinct populations, inputs and label mappings; pooled performance should not be interpreted as uniformly strong discrimination in every cohort\.
### 5\.8Comorbidity Prediction

Table[10](https://arxiv.org/html/2610.10648#S5.T10)reports the per\-condition AUROC\. Using inputs that exclude medication, medical\-history and condition\-defining laboratory records, the auxiliary head achieves a mean AUROC of 0\.565: coronary artery disease \(0\.634\) and depression \(0\.605\) are the best detected, and diabetes \(0\.512\) and hyperlipidaemia \(0\.519\) are close to chance\. Group occlusion of the comorbidity head \(not shown\) indicates that the remaining signal comes mainly from demographics \(up to 0\.08 AUROC\) and, for COPD, from structural volumes \(0\.05\)\. Without Group\-CVaR, the mean comorbidity AUROC is also 0\.565, and without the longitudinal contrastive loss it is 0\.540 \(Table[7](https://arxiv.org/html/2610.10648#S5.T7)\)\.

Table 10:Per\-condition comorbidity AUROC, ADNI subjects withmcom=1m^\{\\text\{com\}\}=1\(n=1,739n=1\{,\}739\)\. Full: seed 42; w/o CVaR: ablation, seed 42\.ConditionCasesFullw/o CVaR3\-seed meanDepression5180\.6050\.6150\.598Diabetes1620\.5120\.5400\.551Hypertension8350\.5470\.5440\.547Coronary artery disease2230\.6340\.6370\.636Stroke740\.5700\.5460\.588Hyperlipidaemia8010\.5190\.5230\.521COPD540\.5540\.5290\.544Arthritis7080\.5580\.5690\.564Cancer5120\.5850\.5860\.592Mean0\.5650\.5650\.571
### 5\.9Subgroup Robustness

Table[11](https://arxiv.org/html/2610.10648#S5.T11)and Figure[5](https://arxiv.org/html/2610.10648#S5.F5)evaluate diagnosis and conversion within each of the nine comorbidity strata \(ADNI only, since comorbidity status is unknown in OASIS\-2 and MIRIAD\)\. Diagnosis ranges from 0\.819 \(COPD\) to 0\.954 \(cancer\); conversion is more sensitive, with its worst stratum in stroke \(0\.765; 24 subjects with at least one positive conversion label across 74 positive MCI visits\), about 0\.12 below the pooled value, identifying stroke as a priority subgroup for further evaluation\. The COPD stratum contains only 5 subjects with at least one positive conversion label across 15 positive MCI visits \(conversion AUROC 0\.785\), and its estimates are correspondingly uncertain\. Without Group\-CVaR, the worst diagnosis stratum is 0\.828 \(COPD\) and the worst conversion stratum 0\.768 \(stroke\), against 0\.819 and 0\.765 with it; the mean of the three worst strata and the best–worst disparity are also similar\.

Table 11:Subgroup robustness over the nine comorbidity strata \(ADNI, seed 42\)\. The stratum is given in parentheses\. Seed SD: full model over three seeds\.Worst\-stratumdiagnosisWorst\-stratumconversionHardestcomorbidityComorbiditymeanFull model0\.819 \(COPD\)0\.765 \(stroke\)0\.512 \(diab\.\)0\.565w/o CVaR0\.828 \(COPD\)0\.768 \(stroke\)0\.523 \(hyperl\.\)0\.565Seed SD±\\pm0\.008±\\pm0\.029±\\pm0\.006±\\pm0\.010Mean of 3 worst strata: diagnosis 0\.883, conversion 0\.785 \(full\); 0\.889, 0\.799 \(no CVaR\)Disparity \(best−\-worst\): diagnosis 0\.135, conversion 0\.145 \(full\); 0\.123, 0\.144 \(no CVaR\)

Figure 5:Network AUROC within recorded\-comorbidity strata, seed 42, for binary diagnosis and MCI\-visit conversion\. Intervals use 1,000 subject\-cluster bootstrap resamples\. Numbers at the right of each panel \(n\+\) give the subjects with at least one positive label for that task\. Dashed lines show the corresponding overall ADNI AUROC\. Strata may overlap, and small subgroups have substantial uncertainty\. Point estimates and intervals characterise subgroup performance and its uncertainty\.
### 5\.10Interpretability

Network input dependence is assessed with Integrated Gradients \(IG\)\[[63](https://arxiv.org/html/2610.10648#bib.bib63)\]and group occlusion\. Network explanations use the first batches of each validation loader, capped at 250 subjects \(the batch boundary can extend this to 256\)\. This subset comprises ADNI participants without raw MRI\. The network explanations therefore describe a sampled ADNI tabular population rather than all three cohorts\. MRI embeddings are cached and held fixed during IG, which is computed for diagnosis and forecast input tensors relative to a zero\-input baseline\. The scalar target is the sum of eligible\-output logit contrasts; the attribution describes aggregate model dependence and is not a causal effect or an explanation of one selected visit\. Progression explanations include eligible progression\-labelled visits and are not restricted to the MCI population used for the primary conversion metric\.

#### 5\.10\.1Group Occlusion

Occlusion measures how much AUROC is lost when a whole input group is removed \(set to “never measured”\), which examines joint dependence on a feature group but can introduce out\-of\-distribution missingness and remains sensitive to correlated groups \(Figure[6](https://arxiv.org/html/2610.10648#S5.F6)\)\. Fordiagnosis, removing the structural MRI volumes costs 0\.099±\\pm0\.027 AUROC \(mean±\\pmSD over folds\) and removing PET/CSF biomarkers 0\.022±\\pm0\.011\. Forconversion, removing cognitive tests costs 0\.062±\\pm0\.018, PET/CSF 0\.020±\\pm0\.011 and structural volumes 0\.016±\\pm0\.014\. The raw\-MRI perturbation is not informative in this subset because these subjects lack raw MRI; it is omitted from the figure\. The two methods assess different forms of input dependence\. Network occlusion replaces a group by its missing\-input representation, whereas tabular permutation disrupts its correspondence with the outcome\. A negative bar, such as family history in the network progression panel, means that AUROC increased after perturbation\. It does not imply that family history protects against disease\. This pattern can reflect correlated inputs, sampling variability or perturbation sensitivity\. Fold SD describes variability, not a confidence interval or a significance test\.

Figure 6:Input\-group perturbation for diagnosis \(top\) and eligible\-visit progression \(bottom\): network occlusion \(left\) and tabular\-stack permutation \(right\), seed 42\. Bars and error bars show the fold mean and sample SD\. Positive values mean that perturbation lowers AUROC; negative values mean that it increases AUROC\. Network explanations use a capped ADNI subset; the tabular panels can use different eligible rows, so magnitudes should not be compared directly\. Progression explanations are not MCI\-only\. Structural MRI volumes are tabular measurements; raw MRI is not evaluated in this subset\.
#### 5\.10\.2Top Features

The top\-ranked individual features include structural, metabolic and fluid measures associated with AD \(Figure[7](https://arxiv.org/html/2610.10648#S5.F7)\)\. Fordiagnosis, the ranking is led by hippocampal and entorhinal volume and FDG\-PET, followed by ventricular and middle\-temporal volume, amyloid PET \(AV45\), whole\-brain volume, CSF Aβ42\\beta\_\{42\}, p\-tau and t\-tau, and fusiform volume\. These are the established medial\-temporal, metabolic and amyloid/tau markers of AD\. Forconversion, the ranking is led by the Functional Activities Questionnaire \(FAQ\), followed by hippocampal volume, FDG\-PET, ADAS\-13, AV45, middle\-temporal volume, CSF Aβ42\\beta\_\{42\}, MMSE and RAVLT immediate recall, i\.e\. early loss of instrumental activities of daily living and episodic memory, supported by neurodegeneration and amyloid markers\. Attribution magnitudes identify the inputs the model relies on, not causal effects\.

Figure 7:Tabular features ranked by mean absolute IG for diagnosis \(left\) and eligible\-visit progression \(right\); bars and error bars show the equally weighted fold mean and sample SD\. The IG target sums eligible\-output logit contrasts, with raw MRI held fixed\. Related value, availability and history features are grouped by base feature\. Absolute IG shows the magnitude of dependence, not whether a feature raises or lowers predicted risk; related inputs can share attribution\.

## 6Discussion

### 6\.1Main Findings

A single leakage\-controlled model jointly performs diagnosis, conversion forecasting, cognitive and atrophy prediction, and comorbidity prediction, with diagnosis and stage\-1 progression AUROCs of 0\.935 and 0\.884 across three seeds and low seed variation\. The sampled ADNI tabular explanations identify dependence on features associated with AD: diagnosis relies on hippocampal and entorhinal volume, FDG\-PET and amyloid and tau markers, whereas conversion relies on functional decline \(FAQ\) and episodic memory, supported by the same biomarkers\.

Part of this division of evidence is imposed by the branch\-specific input rules, and the attributions describe model associations rather than causal mechanisms\.

The contribution of this work is the implemented multimodal multitask framework and its evaluation: leakage\-controlled integration of heterogeneous cohorts, tasks and modalities, comorbidity prediction without condition\-defining records, stratified subgroup analysis and external validation\. The paired ablations characterise the sensitivity of this configuration to its components; observed differences in pooled AUROC were small \(at most 0\.004 in absolute value\) for the selected endpoints and seed, and the longitudinal contrastive loss was associated with higher comorbidity discrimination\.

### 6\.2Continuous\-Time Modelling

With elapsed time entering the GRU gates in both temporal models, no statistically detectable AUROC difference was observed between ODE integration and a time\-aware GRU in this configuration\. The ODE additionally provides a latent state defined at any time between visits, which supports future work on forecasting at chosen horizons\.

### 6\.3Comorbidity and Modality Contributions

Auxiliary comorbidity modelling and subgroup evaluation provide complementary views of the clinical setting\. The head predicts nine recorded conditions without receiving their defining medical\-history, medication or laboratory records, achieving a mean AUROC of 0\.571 over three seeds\. Recorded conditions also define training strata for Group\-CVaR and support condition\-specific diagnosis and progression analyses\. This design integrates auxiliary supervision and subgroup assessment without using recorded conditions as direct predictive inputs\. With and without Group\-CVaR, the worst diagnosis and conversion strata were similar \(0\.819 vs\. 0\.828; 0\.765 vs\. 0\.768\), and conversion in patients with a history of stroke was the lowest\-performing stratum\.

The sampled ADNI explanations identify dependence on derived imaging volumes, PET/CSF biomarkers and permitted clinical features\. The imaging\-focused comparator achieves diagnosis AUROCs of 0\.597 on MIRIAD and 0\.621 on OASIS\-2, compared with 0\.424 and 0\.629 for the fused network\. Together, the cohort and modality analyses characterise how performance varies with input availability and population\. The tabular attribution analysis covers ADNI subjects without raw MRI; individual raw\-image localisation is outside its scope\.

### 6\.4Limitations and Future Work

The validation folds also guide model selection, threshold and temperature fitting and hybrid weighting, so internal estimates may be optimistic; external evaluation on OASIS\-3 provides an independent assessment\. The binary diagnosis and progression endpoints are cohort\-defined, and external progression uses the next observed visit\. Comorbidity labels are rule\-derived from medical histories without independent clinical adjudication\. On the internal folds, the tabular stack alone exceeds the network AUROC\. Raw MRI is not available for ADNI, and volumes are processed without anatomical registration\. Ablations use a single seed\. Future work will strengthen the image pathway, harmonise comorbidity labels against clinical records and extend external evaluation to cohorts with matched endpoints\.

## 7Conclusion

This study presents an integrated framework for cohort\-defined diagnosis, longitudinal progression, cognitive and volume prediction, and auxiliary comorbidity modelling under missing modalities and irregular observation times\. Its contribution combines an adapted SFCN encoder, causal clinical streams, task\-aware fusion, ODE\-GRU dynamics, task balancing, subgroup objectives and explicit input controls in a common pipeline\. Three\-seed internal validation achieves mean diagnosis and stage\-1 progression AUROCs of 0\.935 and 0\.884, increasing to 0\.951 and 0\.909 with the tabular hybrid\. Paired ablations characterise the sensitivity of performance to the individual components\. Task\-specific explanations and comorbidity\-stratified evaluation connect the predictive results with their input and population context\. External validation on OASIS\-3 extends the assessment to another cohort, with hybrid diagnosis and next\-visit progression AUROCs of 0\.767 and 0\.764, improved diagnosis calibration and a next\-visit MMSE MAE of 0\.86\. Together, these experiments provide a structured evaluation of multimodal multitask prediction across heterogeneous Alzheimer’s disease research cohorts\.

## CRediT authorship contribution statement

Akeem Temitope Otapo:Conceptualization, Methodology, Investigation, Formal analysis, Data curation and Writing – original draft\.Alice Othmani:Supervision, Writing – review & editing and Project administration\.Ghazaleh Khodabandelou \(corresponding author\):Conceptualization, Validation and Writing – review & editing\.Zuheng Ming:Conceptualization, Validation and Writing – review & editing\.

## Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper\.

## Funding

This work was supported by the Tertiary Education Trust Fund \(TETFund\), Nigeria \[grant number TETF/ES/COE/YABA/TSAS/2022\]\.

## Data availability

The ADNI, OASIS\-2, OASIS\-3 and MIRIAD data used in this study are available from their respective repositories to registered users under each cohort’s data use agreement\. The code is available upon request\.

## Acknowledgements

The authors thank the OASIS, MIRIAD and ADNI investigators for making their data publicly available\.

## References

- \[1\]E\. Nichols, J\. D\. Steinmetz, S\. E\. Vollset et al\., “Estimation of the global prevalence of dementia in 2019 and forecasted prevalence in 2050: an analysis for the Global Burden of Disease Study 2019,” The Lancet Public Health, vol\. 7, no\. 2, pp\. e105–e125, 2022\. \[Online\]\. Available:[https://doi\.org/10\.1016/S2468\-2667\(21\)00249\-8](https://doi.org/10.1016/S2468-2667(21)00249-8)
- \[2\]A\. Wimo, K\. Seeher, R\. Cataldi et al\., “The worldwide costs of dementia in 2019,” Alzheimer’s & Dementia, vol\. 19, no\. 7, pp\. 2865–2873, 2023\. \[Online\]\. Available:[https://doi\.org/10\.1002/alz\.12901](https://doi.org/10.1002/alz.12901)
- \[3\]M\. Lamar, D\. J\. Libon, A\. V\. Ashley, J\. J\. Lah, A\. I\. Levey, and F\. C\. Goldstein, “The impact of vascular comorbidities on qualitative error analysis of executive impairment in Alzheimer’s disease,” Journal of the International Neuropsychological Society, vol\. 16, no\. 1, pp\. 77–83, 2010\. \[Online\]\. Available:[https://doi\.org/10\.1017/S1355617709990981](https://doi.org/10.1017/S1355617709990981)
- \[4\]B\. Rajamaki, B\. Braithwaite, S\. Hartikainen, and A\.\-M\. Tolppanen, “Identifying Comorbidity Patterns in People with and without Alzheimer’s Disease Using Latent Dirichlet Allocation,” Journal of Alzheimer’s Disease, vol\. 101, no\. 4, pp\. 1393–1403, 2024\. \[Online\]\. Available:[https://doi\.org/10\.3233/JAD\-240490](https://doi.org/10.3233/JAD-240490)
- \[5\]R\. Duara and W\. Barker, “Heterogeneity in Alzheimer’s Disease Diagnosis and Progression Rates: Implications for Therapeutic Trials,” Neurotherapeutics, vol\. 19, no\. 1, pp\. 8–25, 2022\. \[Online\]\. Available:[https://doi\.org/10\.1007/s13311\-022\-01185\-z](https://doi.org/10.1007/s13311-022-01185-z)
- \[6\]K\. Maciejewska, K\. Czarnecka, and P\. Szyman’ski, “A review of the mechanisms underlying selected comorbidities in Alzheimer’s disease,” Pharmacological Reports, vol\. 73, no\. 6, pp\. 1565–1581, 2021\. \[Online\]\. Available:[https://doi\.org/10\.1007/s43440\-021\-00293\-5](https://doi.org/10.1007/s43440-021-00293-5)
- \[7\]Y\. Li, H\. Li, Z\. Xu, and S\. K\. Zhou, “CLIS: Causality\-inspired Longitudinal Image Synthesis and its application to Alzheimer’s disease characterization,” Medical Image Analysis, vol\. 112, p\. 104126, 2026\. \[Online\]\. Available:[https://doi\.org/10\.1016/j\.media\.2026\.104126](https://doi.org/10.1016/j.media.2026.104126)
- \[8\]M\. Nguyen, T\. He, L\. An, D\. C\. Alexander, J\. Feng, and B\. T\. Yeo, “Predicting Alzheimer’s disease progression using deep recurrent neural networks,” NeuroImage, vol\. 222, p\. 117203, 2020\. \[Online\]\. Available:[https://doi\.org/10\.1016/j\.neuroimage\.2020\.117203](https://doi.org/10.1016/j.neuroimage.2020.117203)
- \[9\]R\. T\. Q\. Chen, Y\. Rubanova, J\. Bettencourt, and D\. K\. Duvenaud, “Neural ordinary differential equations,” in Advances in Neural Information Processing Systems, vol\. 31\. Curran Associates, Inc\., 2018\. \[Online\]\. Available:[https://proceedings\.neurips\.cc/paper/2018/hash/69386f6bb1dfed68692a24c8686939b9\-Abstract\.html](https://proceedings.neurips.cc/paper/2018/hash/69386f6bb1dfed68692a24c8686939b9-Abstract.html)
- \[10\]Y\. Rubanova, R\. T\. Q\. Chen, and D\. K\. Duvenaud, “Latent Ordinary Differential Equations for Irregularly\-Sampled Time Series,” in Advances in Neural Information Processing Systems, vol\. 32\. Curran Associates, Inc\., 2019\. \[Online\]\. Available:[https://proceedings\.neurips\.cc/paper/2019/hash/42a6845a557bef704ad8ac9cb4461d43\-Abstract\.html](https://proceedings.neurips.cc/paper/2019/hash/42a6845a557bef704ad8ac9cb4461d43-Abstract.html)
- \[11\]P\. Kidger, J\. Morrill, J\. Foster, and T\. Lyons, “Neural controlled differential equations for irregular time series,” in Advances in Neural Information Processing Systems, vol\. 33\. Curran Associates, Inc\., 2020, pp\. 6696–6707\. \[Online\]\. Available:[https://proceedings\.neurips\.cc/paper/2020/file/4a5876b450b45371f6cfe5047ac8cd45\-Paper\.pdf](https://proceedings.neurips.cc/paper/2020/file/4a5876b450b45371f6cfe5047ac8cd45-Paper.pdf)
- \[12\]Y\. Fang and S\. Jin, “Neural Ordinary Differential Equations with Differentiable Hidden State for Irregular Time Series,” in 2025 5th International Conference on Neural Networks, Information and Communication Engineering \(NNICE\), 2025, pp\. 1949–1952\. \[Online\]\. Available:[https://doi\.org/10\.1109/NNICE64954\.2025\.11064674](https://doi.org/10.1109/NNICE64954.2025.11064674)
- \[13\]A\. Kendall, Y\. Gal, and R\. Cipolla, “Multi\-task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics,” in 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018, pp\. 7482–7491\. \[Online\]\. Available:[https://doi\.org/10\.1109/CVPR\.2018\.00781](https://doi.org/10.1109/CVPR.2018.00781)
- \[14\]Z\. Chen, V\. Badrinarayanan, C\.\-Y\. Lee, and A\. Rabinovich, “GradNorm: Gradient Normalization for Adaptive Loss Balancing in Deep Multitask Networks,” in Proceedings of the 35th International Conference on Machine Learning, ser\. Proceedings of Machine Learning Research, vol\. 80\. PMLR, 2018, pp\. 794–803\. \[Online\]\. Available:[https://proceedings\.mlr\.press/v80/chen18a\.html](https://proceedings.mlr.press/v80/chen18a.html)
- \[15\]J\. Zhou, L\. Yuan, J\. Liu, and J\. Ye, “A multi\-task learning formulation for predicting disease progression,” in Proceedings of the 17th ACM SIGKDD international conference on Knowledge discovery and data mining, 2011, pp\. 814–822\. \[Online\]\. Available:[https://doi\.org/10\.1145/2020408\.2020549](https://doi.org/10.1145/2020408.2020549)
- \[16\]H\. Harutyunyan, H\. Khachatrian, D\. C\. Kale, G\. Ver Steeg, and A\. Galstyan, “Multitask learning and benchmarking with clinical time series data,” Scientific Data, vol\. 6, no\. 1, p\. 96, 2019\. \[Online\]\. Available:[https://doi\.org/10\.1038/s41597\-019\-0103\-9](https://doi.org/10.1038/s41597-019-0103-9)
- \[17\]F\. Abuhantash, R\. Welsch, S\. Finkelstein, and A\. AlShehhi, “Alzheimer’s disease risk prediction using machine learning for survival analysis with a comorbidity\-based approach,” Scientific Reports, vol\. 15, no\. 1, p\. 28723, 2025\. \[Online\]\. Available:[https://doi\.org/10\.1038/s41598\-025\-14406\-0](https://doi.org/10.1038/s41598-025-14406-0)
- \[18\]S\. Sagawa, P\. W\. Koh, T\. B\. Hashimoto, and P\. Liang, “Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst\-Case Generalization,” in International Conference on Learning Representations, 2020\. \[Online\]\. Available:[https://openreview\.net/forum?id=ryxGuJrFvS](https://openreview.net/forum?id=ryxGuJrFvS)
- \[19\]S\. R\. Pfohl, H\. Zhang, Y\. Xu, A\. Foryciarz, M\. Ghassemi, and N\. H\. Shah, “A comparison of approaches to improve worst\-case predictive model performance over patient subpopulations,” Scientific Reports, vol\. 12, no\. 1, p\. 3254, 2022\. \[Online\]\. Available:[https://doi\.org/10\.1038/s41598\-022\-07167\-7](https://doi.org/10.1038/s41598-022-07167-7)
- \[20\]Y\. Wang, R\. Bone, W\. Fleisher et al\., “Using Under\-Represented Subgroup Fine Tuning to Improve Fairness for Disease Prediction,” in Proceedings of the 18th International Joint Conference on Biomedical Engineering Systems and Technologies \(BIOSTEC\), Volume 2: HEALTHINF, vol\. 2, 2025, pp\. 240–253\. \[Online\]\. Available:[https://doi\.org/10\.5220/0013318600003911](https://doi.org/10.5220/0013318600003911)
- \[21\]J\. Wen, E\. Thibeau\-Sutre, M\. Diaz\-Melo et al\., “Convolutional neural networks for classification of Alzheimer’s disease: Overview and reproducible evaluation,” Medical Image Analysis, vol\. 63, p\. 101694, 2020\. \[Online\]\. Available:[https://doi\.org/10\.1016/j\.media\.2020\.101694](https://doi.org/10.1016/j.media.2020.101694)
- \[22\]S\. Grueso and R\. Viejo\-Sobera, “Machine learning methods for predicting progression from mild cognitive impairment to Alzheimer’s disease dementia: a systematic review,” Alzheimer’s Research & Therapy, vol\. 13, no\. 1, p\. 162, 2021\. \[Online\]\. Available:[https://doi\.org/10\.1186/s13195\-021\-00900\-w](https://doi.org/10.1186/s13195-021-00900-w)
- \[23\]I\. Afifi, M\. Elgendy, M\. Abdelfatah, and S\. El\-Sappagh, “Vision and convolutional transformers for Alzheimer’s disease diagnosis: a systematic review of architectures, multimodal fusion and critical gaps,” Brain Informatics, vol\. 13, no\. 1, p\. 1, 2026\. \[Online\]\. Available:[https://doi\.org/10\.1186/s40708\-025\-00286\-7](https://doi.org/10.1186/s40708-025-00286-7)
- \[24\]C\. Sun, M\. Song, D\. Cai, B\. Zhang, H\. Li, and S\. Hong, “A Review of Deep Learning Methods for Irregularly Sampled Medical Time Series Data,” Health Data Science, vol\. 6, p\. 0456, 2026\. \[Online\]\. Available:[https://doi\.org/10\.34133/hds\.0456](https://doi.org/10.34133/hds.0456)
- \[25\]A\. Abrol, M\. Bhattarai, A\. Fedorov, Y\. Du, S\. Plis, and V\. Calhoun, “Deep residual learning for neuroimaging: An application to predict progression to Alzheimer’s disease,” Journal of Neuroscience Methods, vol\. 339, p\. 108701, 2020\. \[Online\]\. Available:[https://doi\.org/10\.1016/j\.jneumeth\.2020\.108701](https://doi.org/10.1016/j.jneumeth.2020.108701)
- \[26\]D\. Zhang, D\. Shen, and Alzheimer’s Disease Neuroimaging Initiative, “Predicting Future Clinical Changes of MCI Patients Using Longitudinal and Multimodal Biomarkers,” PLoS ONE, vol\. 7, no\. 3, p\. e33182, 2012\. \[Online\]\. Available:[https://doi\.org/10\.1371/journal\.pone\.0033182](https://doi.org/10.1371/journal.pone.0033182)
- \[27\]G\. Lee, K\. Nho, B\. Kang et al\., “Predicting Alzheimer’s disease progression using multi\-modal deep learning approach,” Scientific Reports, vol\. 9, no\. 1, p\. 1952, 2019\. \[Online\]\. Available:[https://doi\.org/10\.1038/s41598\-018\-37769\-z](https://doi.org/10.1038/s41598-018-37769-z)
- \[28\]T\. Zhou, M\. Liu, K\.\-H\. Thung, and D\. Shen, “Latent Representation Learning for Alzheimer’s Disease Diagnosis With Incomplete Multi\- Modality Neuroimaging and Genetic Data,” IEEE Transactions on Medical Imaging, vol\. 38, no\. 10, pp\. 2411–2422, 2019\. \[Online\]\. Available:[https://doi\.org/10\.1109/TMI\.2019\.2913158](https://doi.org/10.1109/TMI.2019.2913158)
- \[29\]W\. Huang, “Multimodal Contrastive Learning and Tabular Attention for Automated Alzheimer’s Disease Prediction,” in 2023 IEEE/CVF International Conference on Computer Vision Workshops \(ICCVW\), 2023, pp\. 2465–2474\. \[Online\]\. Available:[https://doi\.org/10\.1109/ICCVW60793\.2023\.00261](https://doi.org/10.1109/ICCVW60793.2023.00261)
- \[30\]D\. Agostinho, M\. Simões, and M\. Castelo\-Branco, “Predicting conversion from mild cognitive impairment to Alzheimer’s disease: a multimodal approach,” Brain Communications, vol\. 6, no\. 4, p\. fcae208, 2024\. \[Online\]\. Available:[https://doi\.org/10\.1093/braincomms/fcae208](https://doi.org/10.1093/braincomms/fcae208)
- \[31\]B\. Simkhada, T\. Y\. Liang, X\. Cui et al\., “Multimodal prediction of MCI\-to\-dementia conversion over a two\-year window,” GeroScience, 2026, published online 22 September 2026; volume and final pagination not assigned in the retrieved record\. \[Online\]\. Available:[https://doi\.org/10\.1007/s11357\-026\-02551\-x](https://doi.org/10.1007/s11357-026-02551-x)
- \[32\]R\. Zhang, D\. Chen, Y\. Qin, H\. Han, and H\. Yu, “Longitudinal multi\-modal data prediction model for mild cognitive impairment by deep survival analysis,” BMC Medical Informatics and Decision Making, vol\. 26, no\. 1, p\. 81, 2026\. \[Online\]\. Available:[https://doi\.org/10\.1186/s12911\-026\-03387\-3](https://doi.org/10.1186/s12911-026-03387-3)
- \[33\]Y\. Lu, S\. K\. Hammonds, and A\. Fernandez\-Quilez, “Clinical pathways matter for multimodal deep learning in early Alzheimer’s disease detection,” Scientific Reports, vol\. 16, no\. 1, p\. 28381, 2026\. \[Online\]\. Available:[https://doi\.org/10\.1038/s41598\-026\-57861\-z](https://doi.org/10.1038/s41598-026-57861-z)
- \[34\]Y\. Geng and H\. Zhang, “Machine\-learning prediction and risk stratification of 12\-month cognitive decline in Alzheimer’s disease using routine clinical and MRI data,” Scientific Reports, vol\. 16, no\. 1, p\. 12227, 2026\. \[Online\]\. Available:[https://doi\.org/10\.1038/s41598\-026\-43321\-1](https://doi.org/10.1038/s41598-026-43321-1)
- \[35\]I\. Kiani, E\. Hajiasadi, Z\. Tahamtani Torkamani et al\., “Deep learning models predicting MCI\-to\-Alzheimer’s conversion via structural MRI: A systematic review and meta\-analysis,” European Journal of Radiology Artificial Intelligence, vol\. 7, p\. 100094, 2026\. \[Online\]\. Available:[https://doi\.org/10\.1016/j\.ejrai\.2026\.100094](https://doi.org/10.1016/j.ejrai.2026.100094)
- \[36\]E\. Bélanger, N\. Rosendaal, R\. Gutman et al\., “Identifying Medicare beneficiaries with Alzheimer’s disease and related dementia using home health OASIS assessments,” Journal of the American Geriatrics Society, vol\. 71, no\. 10, pp\. 3229–3236, 2023\. \[Online\]\. Available:[https://doi\.org/10\.1111/jgs\.18487](https://doi.org/10.1111/jgs.18487)
- \[37\]R\. V\. Marinescu, N\. P\. Oxtoby, A\. L\. Young et al\., “The Alzheimer’s Disease Prediction Of Longitudinal Evolution \(TADPOLE\) Challenge: Results after 1 Year Follow\-up,” Machine Learning for Biomedical Imaging, vol\. 2021, no\. December 2021, pp\. 1–60, 2021\. \[Online\]\. Available:[https://www\.melba\-journal\.org/papers/2021:019\.html](https://www.melba-journal.org/papers/2021:019.html)
- \[38\]E\. De Brouwer, J\. Simm, A\. Arany, and Y\. Moreau, “GRU\-ODE\-Bayes: Continuous Modeling of Sporadically\-Observed Time Series,” in Advances in Neural Information Processing Systems, vol\. 32\. Curran Associates, Inc\., 2019\. \[Online\]\. Available:[https://proceedings\.neurips\.cc/paper/2019/hash/455cb2657aaa59e32fad80cb0b65b9dc\-Abstract\.html](https://proceedings.neurips.cc/paper/2019/hash/455cb2657aaa59e32fad80cb0b65b9dc-Abstract.html)
- \[39\]D\. Lachinov, A\. Chakravarty, C\. Grechenig, U\. Schmidt\-Erfurth, and H\. Bogunovic’, “Learning Spatio\-Temporal Model of Disease Progression With NeuralODEs From Longitudinal Volumetric Data,” IEEE Transactions on Medical Imaging, vol\. 43, no\. 3, pp\. 1165–1179, 2024\. \[Online\]\. Available:[https://doi\.org/10\.1109/TMI\.2023\.3330576](https://doi.org/10.1109/TMI.2023.3330576)
- \[40\]M\. N\. Bossa and H\. Sahli, “A multidimensional ODE\-based model of Alzheimer’s disease progression,” Scientific Reports, vol\. 13, no\. 1, p\. 3162, 2023\. \[Online\]\. Available:[https://doi\.org/10\.1038/s41598\-023\-29383\-5](https://doi.org/10.1038/s41598-023-29383-5)
- \[41\]R\. Sharma, V\. Kukreja, S\. Hariharan, and S\.\-Y\. Chen, “cLDM\-ODE: A Multimodal Generative Framework for Uncertainty\-Aware Forecasting of Alzheimer’s Disease Progression,” IRBM, vol\. 47, no\. 1, p\. 100926, 2026\. \[Online\]\. Available:[https://doi\.org/10\.1016/j\.irbm\.2025\.100926](https://doi.org/10.1016/j.irbm.2025.100926)
- \[42\]J\. Wang, Y\. Mao, X\. Liu, W\. Hao, and for the Alzheimer’s Disease Neuroimaging Initiative, “Learning patient\-specific spatial biomarker dynamics via operator learning for Alzheimer’s disease progression,” npj Systems Biology and Applications, vol\. 12, no\. 1, p\. 104, 2026\. \[Online\]\. Available:[https://doi\.org/10\.1038/s41540\-026\-00719\-x](https://doi.org/10.1038/s41540-026-00719-x)
- \[43\]Y\. Zhao, T\. Che, X\. Wang, and S\. Li, “Forecasting individualized progression of Alzheimer’s disease using structural MRI and population spatiotemporal priors,” Frontiers in Aging Neuroscience, vol\. 18, p\. 1691084, 2026\. \[Online\]\. Available:[https://doi\.org/10\.3389/fnagi\.2026\.1691084](https://doi.org/10.3389/fnagi.2026.1691084)
- \[44\]T\. He, K\. Jiang, A\. Zhao et al\., “A Stage\-Aware Mixture of Experts Framework for Neurodegenerative Disease Progression Modelling,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol\. 40, no\. 26, 2026, pp\. 21 663–21 671\. \[Online\]\. Available:[https://doi\.org/10\.1609/aaai\.v40i26\.39316](https://doi.org/10.1609/aaai.v40i26.39316)
- \[45\]X\. Fan, M\. Zhou, J\. Qi, Y\. Yang, and P\. Yang, “A multi\-target multi\-task approach based on correlated multiple cognitive scores for AD progression prediction,” in 2024 International Joint Conference on Neural Networks \(IJCNN\), 2024, pp\. 1–8\. \[Online\]\. Available:[https://doi\.org/10\.1109/IJCNN60899\.2024\.10650319](https://doi.org/10.1109/IJCNN60899.2024.10650319)
- \[46\]M\. Wang, D\. Zhang, D\. Shen, and M\. Liu, “Multi\-task exclusive relationship learning for alzheimer’s disease progression prediction with longitudinal data,” Medical Image Analysis, vol\. 53, pp\. 111–122, 2019\. \[Online\]\. Available:[https://doi\.org/10\.1016/j\.media\.2019\.01\.007](https://doi.org/10.1016/j.media.2019.01.007)
- \[47\]W\. Liang, K\. Zhang, P\. Cao, X\. Liu, J\. Yang, and O\. R\. Zaiane, “Exploiting task relationships for Alzheimer’s disease cognitive score prediction via multi\-task learning,” Computers in Biology and Medicine, vol\. 152, p\. 106367, 2023\. \[Online\]\. Available:[https://doi\.org/10\.1016/j\.compbiomed\.2022\.106367](https://doi.org/10.1016/j.compbiomed.2022.106367)
- \[48\]D\. Ma, C\. Pabalan, A\. Rajagopal et al\., “Predicting categorical and continuous Alzheimer’s disease outcomes from a single MRI scan,” Nature Aging, vol\. 6, no\. 5, pp\. 1121–1137, 2026\. \[Online\]\. Available:[https://doi\.org/10\.1038/s43587\-026\-01121\-2](https://doi.org/10.1038/s43587-026-01121-2)
- \[49\]W\. Zhou, Z\. Xue, J\. Liang, J\. Cui, X\. Guo, and Y\. Wen, “Modeling and application of alzheimer’s disease complex trait prediction based on multi\-task learning,” Scientific Reports, vol\. 16, no\. 1, p\. 7749, 2026\. \[Online\]\. Available:[https://doi\.org/10\.1038/s41598\-026\-37820\-4](https://doi.org/10.1038/s41598-026-37820-4)
- \[50\]S\. Jahani, G\. Roshanaei, L\. Tapak, and for the Alzheimer’s Disease Neuroimaging Initiative, “Assessing the accuracy of survival machine learning and traditional statistical models for Alzheimer’s disease prediction over time: a study on the ADNI cohort,” BMC Medical Research Methodology, vol\. 25, no\. 1, p\. 250, 2025\. \[Online\]\. Available:[https://doi\.org/10\.1186/s12874\-025\-02689\-w](https://doi.org/10.1186/s12874-025-02689-w)
- \[51\]H\. Ren, F\. Jing, Y\. Zhou et al\., “Utilizing multimodal models to forecast Alzheimer’s disease progression and clinical subtypes,” Health Information Science and Systems, vol\. 14, no\. 1, p\. 10, 2026\. \[Online\]\. Available:[https://doi\.org/10\.1007/s13755\-025\-00407\-w](https://doi.org/10.1007/s13755-025-00407-w)
- \[52\]B\. K\. Karaman, M\. Nguyen, H\. Kim, and M\. R\. Sabuncu, “A Deep Survival Model for Predicting Alzheimer’s Diagnosis Based on Multi\-Modal Longitudinal Data,” Big Data Mining and Analytics, vol\. 9, no\. 2, pp\. 465–480, 2026\. \[Online\]\. Available:[https://doi\.org/10\.26599/BDMA\.2025\.9020064](https://doi.org/10.26599/BDMA.2025.9020064)
- \[53\]C\. Wachinger, A\. Rieckmann, and S\. Pölsterl, “Detect and correct bias in multi\-site neuroimaging datasets,” Medical Image Analysis, vol\. 67, p\. 101879, 2021\. \[Online\]\. Available:[https://doi\.org/10\.1016/j\.media\.2020\.101879](https://doi.org/10.1016/j.media.2020.101879)
- \[54\]N\.\-H\. Ho, S\. Charisis, N\. Honnorat et al\., “Advancing fair and explainable machine learning for neuroimaging dementia pattern classification in multi\-racial and multi\-ethnic populations,” Nature Communications, vol\. 17, no\. 1, p\. 8026, 2026\. \[Online\]\. Available:[https://doi\.org/10\.1038/s41467\-026\-74515\-w](https://doi.org/10.1038/s41467-026-74515-w)
- \[55\]M\. Liang, Z\. Ye, G\. Velma et al\., “Population\-specific dementia risk prediction using deep transfer learning in diverse populations,” npj Digital Medicine, 2026\. \[Online\]\. Available:[https://doi\.org/10\.1038/s41746\-026\-03065\-8](https://doi.org/10.1038/s41746-026-03065-8)
- \[56\]H\. Peng, W\. Gong, C\. F\. Beckmann, A\. Vedaldi, and S\. M\. Smith, “Accurate brain age prediction with lightweight deep neural networks,” Medical Image Analysis, vol\. 68, p\. 101871, 2021\. \[Online\]\. Available:[https://doi\.org/10\.1016/j\.media\.2020\.101871](https://doi.org/10.1016/j.media.2020.101871)
- \[57\]E\. J\. Hu, Y\. Shen, P\. Wallis et al\., “LoRA: Low\-Rank Adaptation of Large Language Models,” in International Conference on Learning Representations, 2022\. \[Online\]\. Available:[https://openreview\.net/forum?id=nZeVKeeFYf9](https://openreview.net/forum?id=nZeVKeeFYf9)
- \[58\]C\. Guo, G\. Pleiss, Y\. Sun, and K\. Q\. Weinberger, “On Calibration of Modern Neural Networks,” in Proceedings of the 34th International Conference on Machine Learning, ser\. Proceedings of Machine Learning Research, vol\. 70\. PMLR, 2017, pp\. 1321–1330\. \[Online\]\. Available:[https://proceedings\.mlr\.press/v70/guo17a\.html](https://proceedings.mlr.press/v70/guo17a.html)
- \[59\]Open Access Series of Imaging Studies, “Open Access Series of Imaging Studies \(OASIS\) Portal,” accessed: 2 October 2026\. \[Online\]\. Available:[https://sites\.wustl\.edu/oasisbrains/](https://sites.wustl.edu/oasisbrains/)
- \[60\]University College London, “MIRIAD Data,” accessed: 2 October 2026\. \[Online\]\. Available:[https://www\.ucl\.ac\.uk/drc/research/miriad\-data](https://www.ucl.ac.uk/drc/research/miriad-data)
- \[61\]Alzheimer’s Disease Neuroimaging Initiative, “Alzheimer’s Disease Neuroimaging Initiative \(ADNI\),” accessed: 2 October 2026\. \[Online\]\. Available:[https://adni\.loni\.usc\.edu/](https://adni.loni.usc.edu/)
- \[62\]B\. Fischl, “FreeSurfer,” NeuroImage, vol\. 62, no\. 2, pp\. 774–781, 2012\. \[Online\]\. Available:[https://doi\.org/10\.1016/j\.neuroimage\.2012\.01\.021](https://doi.org/10.1016/j.neuroimage.2012.01.021)
- \[63\]M\. Sundararajan, A\. Taly, and Q\. Yan, “Axiomatic Attribution for Deep Networks,” in Proceedings of the 34th International Conference on Machine Learning, ser\. Proceedings of Machine Learning Research, vol\. 70\. PMLR, 2017, pp\. 3319–3328\. \[Online\]\. Available:[https://proceedings\.mlr\.press/v70/sundararajan17a\.html](https://proceedings.mlr.press/v70/sundararajan17a.html)