acceptodds
Under review as a conference paper at ICLR 2027

PAID: Patient-Identity Contrastive Learning from Recurrent Admissions in Multimodal EHRs

Abstract

Generalizable representation learning from multimodal electronic health records (EHRs) is difficult because outcome labels are task-specific and each admission provides an incomplete view of patient state. Recurrent admissions offer outcome-independent supervision: they share patient-level context while differing in acute condition and modality availability. We propose Patient-Identity Contrastive Learning (PAID), which uses this relation in an asymmetric dual-space Transformer. A patient token (PAT) is aligned across admissions, while an admission token (ADM) retains a separate route for episode-specific evidence. Availability-masked attention, cross-modal reconstruction, and a structured matching score support learning from incomplete observations. On patient-disjoint MIMIC-IV and eICU cohorts, PAID achieves the highest AUROC on all 12 endpoints, with maximum gains of 0.0135 and 0.0247, respectively. Its prediction systems improve all 21 evaluated few-shot settings, and PAT achieves 51.4% Recall@5 for same-patient retrieval. These results support recurrent admissions as a useful source of supervision for reusable multimodal representations with task-dependent patient and admission roles.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
EHRs, Transformer, Contrastive Learning

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_rGJPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PAID, a multimodal EHR representation-learning framework that uses different admissions of the same patient as positive pairs. A shared Transformer produces a patient token (PAT) for cross-admission identity alignment and an admission token (ADM) intended to retain episode-specific information. Adversarial identity suppression, masked cross-modal reconstruction, and a missingness-aware matching score support this design. Experiments on MIMIC-IV and eICU evaluate retrospective clinical classification, few-shot learning, patient retrieval, and representation specialization.

Critical strengths

Strength 1: The paper makes good use of recurrent admissions from the same patient as positive pairs for contrastive learning. These admissions share patient-level information but may differ in clinical conditions and available modalities. This provides a natural source of supervision without requiring task-specific outcome labels.

Strength 2: The proposed dual-space design is intuitive. By aligning PAT representations across admissions while using adversarial learning to reduce patient identity information in ADM, the framework aims to capture both patient-level consistency and admission-specific differences. The use of availability-masked attention and cross-modal reconstruction also helps address missing modalities in EHR data.

Strength 3: The experimental results are encouraging. PAID achieves strong performance across multiple clinical prediction tasks on MIMIC-IV and eICU. The additional few-shot, patient retrieval, and ablation experiments provide further evidence of the usefulness of the learned representations.

Critical weaknesses

Weakness 1: The experimental comparisons are not fully consistent across methods. As mentioned in Section 4.3, different methods use different labeled training data and downstream predictors. In addition, some few-shot results combine PAID with XGBoost (Appendix A.3). This makes it difficult to determine whether the improvements come from the proposed method or the different experimental settings. A comparison under the same training data and downstream predictor would be more convincing.

Weakness 2: The experiments do not clearly demonstrate that PAT and ADM learn complementary information. In Table 5, PAT alone achieves an acute-task mean AUROC of 0.85961, while combining PAT and ADM only improves it to 0.85977. Although Table 3 shows that the dual-space design improves performance, it remains unclear how much additional admission-specific information ADM provides beyond PAT.

Questions for authors

Question 1: Could the authors compare PAID with a strong baseline using the same labeled training data and downstream classifier? This would help clarify whether the improvements come from the proposed representations or the different experimental settings.

Question 2: Table 5 shows very little improvement when ADM is added to PAT. Could the authors compare PAT and PAT+ADM using the same frozen encoder and downstream classifier to better demonstrate the additional contribution of ADM?

Question 3: In Section 4.4, the authors report a gain of 0.0135 for readmission over the strongest baseline. However, Table 1 shows that MUSE achieves 0.6172, which is higher than SimCLR (0.6067), so the actual improvement is only 0.0030. Could the authors clarify this discrepancy and check the reported performance gains?

2 comments
AuthorsOpenReview ↗

Reply 1/2 — Fair comparison and input-leakage audit (Weakness 1; Questions 1 and 2)

Comparison under the same labeled data and classifier. The original baselines were faithful reproductions of each method's published recipe (e.g., MUSE used Adam, lr 2e-4, batch 512, dropout 0.25, patient-level val fraction 0.1, final checkpoint), so small recipe differences are inherent to faithful reproduction; all methods shared the same pretraining admissions and feature pipeline. What differed was the downstream labeled pools and predictors (§4.3/A.3), so those tables are system-level comparisons. We agree they cannot isolate the representation's contribution, and re-ran the comparison under a fully unified protocol: the same 304,916 pretraining admissions; the same labeled patient pools and fit/validation/test manifests; the same label budgets (full/1%/5%); StandardScaler fit on training data plus the same logistic readout (same solver, regularization, class weights, tuning budget); validation-only model selection; fixed-final-epoch pretraining checkpoints with no test-based selection; fine-tuning with identical per-task heads, a 50-epoch cap and validation selection; three fixed seeds, 2,000-resample patient-clustered paired bootstrap, and all metrics recomputed from saved predictions. PAID uses the same training configuration as the baselines. Test AUROC (mean±SD):

TaskPAIDMUSESimCLR
AKI0.8432±0.00180.8342±0.00020.8270±0.0006
Readmission0.6362±0.00070.6226±0.00240.6253±0.0026

PAID leads both baselines on both tasks, and every PAID−MUSE seed-level paired CI is positive (AKI +0.0089±0.0020; readmission +0.0136±0.0021). The gain concentrates where patient history exists: stratifying AKI by the patient's admission count, the PAID−MUSE gap rises from +0.0036 (single admission) to +0.0066 (two admissions) to +0.0111 (three or more, all seed CIs positive).

Label leakage audit. MIMIC inputs cannot directly reveal an endpoint: the 51 tabular features are demographics plus present-on-admission diagnosis chapter aggregates only (discharge codes, which define the sepsis label, are excluded by construction) plus ED lab/vital aggregates ≤24h; WARD/ICU labs, meds, transfers, procedures and orders are zeroed. The 64 time-series features are mean/std/min/max of ED vital channels only, with no lab channels, so the creatinine trajectory defining AKI is not an input. Readmission labels use only post-discharge admissions. Label code is shared and identical across methods. MIMIC: mortality = in-hospital death; LOS>3d/7d = discharge−admission timestamps; ICU utilization = ICU transfer during stay; readmission = next admission ≤30 days after discharge (eligible subset 12,853); AKI = creatinine +0.3 mg/dL within 48h or ≥1.5× admission minimum, during stay; sepsis = sepsis/septicemia ICD codes or titles. eICU: mortality = hospital expiry; LOS = unit discharge offset; sepsis = sepsis terms in stay diagnosis text/codes; Next ICU = exploratory subsequent-stay endpoint. Full eligibility filters and per-modality windows are in A.2. For eICU sepsis, whose legacy endpoint is a retrospective diagnosis classification, we compare results with diagnosis inputs removed (Reply 2/2).

AuthorsOpenReview ↗

Reply 2/2 — Complementarity, corrected claims, σ and ablation (Weakness 2; Question 3)

Complementarity, measured on one frozen encoder. Original Table 5 fine-tuned each readout (20+20 epochs, feature-mask refinement, best validation checkpoint, A.8); fine-tuning absorbs readout differences, so the small increments do not establish complementarity. New test: one shared frozen encoder per seed, identical logistic readout, splits and tuning budget, three fixed seeds, full labels, six-task macro AUROC:

ReadoutAUROC (mean±SD)
PAT0.8579±0.0033
ADM0.8593±0.0026
PAT+ADM0.8635±0.0008
PAT+ADM+σ0.8639±0.0006

PAT+ADM − PAT = +0.0056±0.0025, all three seed-level patient-clustered paired CIs positive; ADM alone matches PAT, not a degraded copy. The increment is larger on individual tasks (LOS>3d +0.0075, LOS>7d +0.0070, ICU utilization +0.0058, AKI +0.0049, sepsis +0.0055, all with three positive seed CIs) and grows when modalities are missing (LOS>7d +0.0171 without text, +0.0120 without ECG; LOS>3d +0.0133 without text; ICU +0.0142 without text), exactly the incomplete-coverage regime the dual space targets. σ adds a small further increment (+0.0004). The complementarity claim is rewritten as this consistent, measurable increment.

Corrected headline numbers. The strongest readmission baseline is MUSE (0.6172), not SimCLR; the improvement is 0.6202−0.6172 = +0.0030. Excluding exploratory Next ICU there are 11 clinical endpoints (7 MIMIC + 4 eICU), not 12. The mean MIMIC gain is 0.0040 (max 0.0074, ICU utilization); the PAID-only few-shot count is 18/21, with the three stacked cells decomposed as PAID-only / raw-XGBoost / ensemble. The abstract, §4.4 and conclusion have been revised. For eICU sepsis we compare results with diagnosis inputs removed: under the exact paper protocol (multi-task), removing diagnosis/ICD inputs and direct sepsis text gives 0.9966→0.9279 (downstream_test), 0.9959→0.9094 (held_out), 0.9963→0.9188 (all_unseen); the drop is expected because diagnosis fields partly constitute this retrospective endpoint, and the remaining 0.9188 reflects substantial non-diagnostic signal.

Table 3 and the controlled attribution. Table 3 ablations are real but jointly remove the associated sampling and auxiliary objectives (A.7), so they show system-level joint effects rather than per-loss attribution: removing the dual space (ADM+matcher) costs −0.0568 mean AUROC and removing the adversarial loss −0.0274, neither negligible. The new controlled matrix (below) gives the sharper attribution.

σ: three components and their use. σdisp\sigma_{\text{disp}} is the outer average over d coordinates of the availability-weighted per-coordinate standard deviation of visible modality tokens; with one visible modality it reduces to ε\sqrt{\varepsilon}, while context and missingness still contribute (A.4). σctx=softplus(fctx(hpat))\sigma_{\text{ctx}}=\mathrm{softplus}(f_{\text{ctx}}(h_{\text{pat}})) is a learned nonnegative readout of the patient-context token; σmiss=∑m(1−a~m)dm\sigma_{\text{miss}}=\sum_m (1-\tilde{a}_m)d_m with dm=softplus(bm)+softplus(fmiss(hpat)m)d_m=\mathrm{softplus}(b_m)+\mathrm{softplus}(f_{\text{miss}}(h_{\text{pat}})_m), so observed modalities contribute exactly zero. The squared sum (Eq. 6) aggregates these nonnegative sources so any one can down-weight matching and no term cancels another. Eq. 7 subtracts γ(σi+σk)\gamma(\sigma_i+\sigma_k) from the identity logit (γ=0.1\gamma=0.1), down-weighting identity matching when visible evidence is internally inconsistent, context is weak, or modalities are missing. Dispersion is computed before modality-type embeddings, so it reflects content rather than modality identity.

Controlled protocol. The controlled arms disable all auxiliary identity sources simultaneously (wpred=wprior=wtemporal=0w_{\text{pred}}=w_{\text{prior}}=w_{\text{temporal}}=0: no CheXpert cross-admission target, no prototype, no temporal term), with reconstruction, VICReg, architecture, masks, data/steps/budget and readout fixed; only the pairing factor varies (true/random/synthetic/no-identity), three fixed seeds, paired CIs. The identity head reads a detached PAT via a separate projection, so identity gradients never enter the shared clinical encoder (wadv=0w_{\text{adv}}=0). Clinical gains are attributable to cross-modal content learning and the dual readout (+0.0056, CIs positive); identity alignment supplies cross-admission consistency and enables the longitudinal next-visit supervision that improves the macro by +0.0115/+0.0133/+0.0143 (full/5%/1%) over no-identity, all CIs positive.

Reviewer_46qtOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

PAID learns from incomplete multimodal health records by pairing different hospital admissions of the same patient. A shared Transformer produces two representations: PAT captures information shared across a patient’s admissions, while ADM is trained to retain less patient-identity information and focus on the current admission. Training also includes reconstructing hidden data, predicting CheXpert findings, and several extra losses to guide the representations and matching score.

The paper reports the best AUROC on its MIMIC-IV and eICU retrospective tasks and improvements in all 21 few-shot settings. It also retrieves another admission of the same patient among the top five results in 51.4% of queries. Using repeated admissions for training is a useful idea, but it is unclear how much of the clinical improvement comes from patient-identity learning or the two-representation design. The systems use different training data and classifiers, some few-shot results include an extra ensemble, and removing reconstruction or semantic supervision causes the largest performance drops. The controlled comparison also provides little evidence that ADM adds useful clinical information beyond PAT.

Critical strengths
  1. Repeated admissions provide useful training signals. Admissions from the same patient share background information but can differ in acute illness and available data. This offers a useful alternative to artificially changing inputs to create training pairs. On unseen test patients, PAT finds another admission of the same patient in the top five results for 51.4% of queries, without needing a patient-ID lookup during inference.

  2. The paper examines the results from several angles and openly explains important limits. It tests two datasets, multiple outcomes, limited-label settings, component removal, missing inputs, patient retrieval, and different ways of using the learned features. The appendix also makes clear that the main tables compare complete systems, training includes extra semantic labels, some few-shot results use ensembles, and the matching score’s link to missing data is partly built into its design. These details help readers judge what the results actually support.

Critical weaknesses
  1. The comparisons do not clearly show where the clinical gains come from. Methods use different labeled training data and classifiers, and some PAID few-shot results include an XGBoost ensemble. MUSE also uses different training and checkpoint-selection settings. Some reported claims need correction: PAID improves readmission AUROC over the strongest baseline by .0030, not .0135, and excluding the exploratory Next ICU task leaves 11 clinical endpoints rather than 12. Repeated runs are needed to judge small gains. The paper should also clearly define AKI/sepsis labels and input time windows, and check whether any input fields directly reveal the labels.

  2. The results do not clearly establish the value of identity learning or the two-representation design. Removing reconstruction or semantic supervision hurts performance much more than removing the adversarial loss. A fair comparison with random pairs or artificially changed views is needed while keeping other training signals the same. In Table 5, adding ADM and the matching score improves PAT’s acute-task AUROC by only .00022, which is too small to judge without uncertainty estimates. Strong patient retrieval shows identity consistency, but does not prove clinical benefit. The matching score’s time-order assumption also needs justification, and its link to missing data does not by itself show that it measures uncertainty well.

Questions for authors
  1. Could PAID and the strongest baselines be compared with the same patient training pools, label budget, validation selection, readout, and tuning budget? Please separate PAID-only predictors from PAID+XGBoost ensembles, report independent-run and patient-level paired uncertainty, and reconcile the readmission margin and endpoint count.

  2. What exactly are the rules, eligible cohorts, and modality-specific observation windows for AKI, sepsis, ICU utilization, 30-day readmission, and Next ICU? Do the 51 tabular or 64 time-series features include fields that directly define an endpoint? How does the .9963 eICU sepsis AUROC change when direct diagnostic information is removed?

  3. With reconstruction, semantic supervision, architecture, and training budget held fixed, what is the benefit of true patient pairing over random pairing or synthetic views? How are the identity signals in the prototype, cross-admission semantic target, and temporal term controlled, rather than only removing one contrastive loss?

2 comments
AuthorsOpenReview ↗

Reply 1/2 — Fair comparison and clinical reliability (Weakness 1; Questions 1 and 2)

PAID-only vs PAID+XGBoost. The three stacked few-shot cells, decomposed (test AUROC):

SettingPAID-onlyRaw XGBoostStacking
LOS>3d, 10%0.85980.86200.8682
ICU util., 5%0.91930.92360.9297
ICU util., 10%0.92660.92820.9339

PAID-only leads 18/21 settings; these three cells are system-level ensembles and are labeled as such, not used as representation-level evidence. With full labels and three fixed seeds, stacking matches or exceeds raw XGBoost on every task (largest gains mortality +0.0030, readmission +0.0021), so the representation adds complementary signal even for a strong tabular learner. The revision separates PAID-only, raw-XGBoost and ensemble columns throughout.

Fair comparison with the strongest baselines. Original MUSE numbers reproduced the MUSE paper's own recipe (Adam, lr 2e-4, batch 512, dropout 0.25, patient-level val fraction 0.1, final checkpoint), so configurations differed. All controlled comparisons now share the same 304,916 pretraining admissions, the same labeled patient pools and fit/validation/test manifests, the same label budgets (full/1%/5%), StandardScaler fit on training data plus the same logistic readout (solver, regularization, class weights, tuning budget), validation-only model selection, and fixed-final-epoch pretraining checkpoints with no test-based selection. Fine-tuning uses identical per-task heads, a 50-epoch cap, validation selection and three fixed seeds; PAID uses the same training configuration as the baselines. Test AUROC (mean±SD):

TaskPAIDMUSESimCLR
AKI0.8432±0.00180.8342±0.00020.8270±0.0006
Readmission0.6362±0.00070.6226±0.00240.6253±0.0026

PAID leads both baselines on both tasks. Every PAID−MUSE seed-level patient-clustered paired CI is positive (AKI +0.0089±0.0020; readmission +0.0136±0.0021). The corresponding text has been revised.

Readmission margin. Agreed: the strongest readmission baseline is MUSE (0.6172), not SimCLR (0.6067); the improvement is 0.6202−0.6172 = +0.0030, not +0.0135. The abstract, §4.4 and conclusion have been revised, including the "SimCLR strongest" statement.

Endpoint count. Agreed: excluding exploratory Next ICU, there are 11 clinical endpoints (7 MIMIC + 4 eICU), not 12. The abstract and conclusion have been revised to match A.10.

Repeats and uncertainty. All new experiments use three fixed pretraining seeds, report mean±sample SD, and give 2,000-resample patient-clustered paired bootstrap intervals per core comparison. All metrics are independently recomputed from saved per-admission predictions, and per-endpoint margins include intervals rather than single-run claims.

Label rules, windows, and diagnosis removal. Label code is shared and identical across methods. MIMIC: mortality = in-hospital death; LOS>3d/7d = discharge−admission timestamps; ICU utilization = ICU transfer during stay; readmission = next admission ≤30 days after discharge (eligible subset 12,853); AKI = creatinine +0.3 mg/dL within 48h or ≥1.5× admission minimum, during stay; sepsis = sepsis/septicemia ICD codes or titles. eICU: mortality = hospital expiry; LOS = unit discharge offset; sepsis = sepsis terms in stay diagnosis text/codes; Next ICU = exploratory subsequent-stay endpoint. Full eligibility filters and per-modality windows are in A.2.

MIMIC inputs cannot directly reveal an endpoint: 51 tabular features are demographics plus present-on-admission diagnosis chapter aggregates only (discharge codes, which define the sepsis label, are excluded by construction) plus ED lab/vital aggregates ≤24h; WARD/ICU labs, meds, transfers, procedures and orders are zeroed. The 64 time-series features are mean/std/min/max of ED vital channels only, with no lab channels, so the creatinine trajectory defining AKI is not an input. Readmission labels use only post-discharge admissions.

For eICU sepsis we compare results with diagnosis inputs removed. The legacy endpoint is a retrospective diagnosis classification (A.10) whose diagnosis inputs overlap the label; removing diagnosis/ICD inputs and direct sepsis text under the exact paper protocol (multi-task) gives:

CohortOriginalDiagnosis-free
downstream_test0.99660.9279
held_out0.99590.9094
all_unseen0.99630.9188

The drop is expected because diagnosis fields partly constitute this retrospective endpoint; the remaining 0.9188 reflects substantial non-diagnostic signal.

AuthorsOpenReview ↗

Reply 2/2 — Value of patient identity and the dual representation (Weakness 2; Question 3)

Table 5 conditions and new multi-seed results. Original Table 5 was a fine-tuning study: same pretrained encoder, 20+20-epoch fine-tuning with feature-mask refinement, each row at its best seven-task validation checkpoint (A.8). Fine-tuning plus refinement absorbs most readout differences, hence the small increments (+0.00016/+0.00022). The new frozen-encoder comparison uses the unified protocol (same features, labels, budgets, logistic readout, validation-only selection, three fixed seeds, six-task macro AUROC, full labels):

ReadoutAUROC (mean±SD)
PAT0.8579±0.0033
ADM0.8593±0.0026
PAT+ADM0.8635±0.0008
PAT+ADM+σ0.8639±0.0006

PAT+ADM − PAT = +0.0056±0.0025, all three seed-level patient-clustered paired CIs positive: ADM carries complementary clinical information. The increment is larger on individual tasks (LOS>3d +0.0075, LOS>7d +0.0070, ICU +0.0058, AKI +0.0049, sepsis +0.0055, all three seed CIs positive) and grows under missing modalities (LOS>7d +0.0171 without text, +0.0120 without ECG; LOS>3d +0.0133 without text; ICU +0.0142 without text). σ adds a small further increment (+0.0004).

Clinical benefit beyond retrieval. Agreed: retrieval alone shows identity consistency, not clinical benefit. Three results connect the design to clinical performance. First, the frozen dual-readout increment above (+0.0056, all seed CIs positive). Second, unified fine-tuning: PAID beats MUSE and SimCLR on AKI and readmission, with all PAID−MUSE paired CIs positive. Third, next-visit feature prediction on true pairs improves the frozen six-task macro over no-identity by +0.0115/+0.0133/+0.0143 (full/5%/1%, all seed CIs positive) and over identity-only true-pair training by +0.0096 (full, all CIs positive). The benefit concentrates where patient history exists: the PAID−MUSE AKI gap rises from +0.0036 (single admission) to +0.0111 (three or more, all seed CIs positive). Recurrent admissions contribute clinically usable signal, not just retrieval consistency.

Temporal-order assumption of the matching score. The temporal term is a one-sided soft hinge on the score only: Ltemp=E(e,l)[σl−σe]+\mathcal{L}_{\text{temp}}=\mathbb{E}_{(e,l)}[\sigma_l-\sigma_e]_+ (Eq. 14). It penalizes score increases from an earlier to a later admission and imposes nothing else; it shapes the scalar score head, not the clinical representation. Ordering uses real admission timestamps, and numeric identifier order is never a proxy for elapsed time (A.6). A.6 already states there is no clinical guarantee that later admissions are less uncertain; we treat the term as a regularizer, not a physiological claim, and removing it costs −0.0335 mean AUROC (Table 3). Controlled runs set wtemporal=0w_{\text{temporal}}=0, so no new claim depends on it.

True vs random/synthetic pairing. With reconstruction, VICReg, architecture, masks and training budget fixed, and all auxiliary identity sources off, the controlled matrix isolates the pairing factor over three fixed seeds. The distinctive clinical value of true pairing is what random and synthetic views cannot provide: a valid longitudinal target. Next-visit prediction, definable only on true same-patient pairs, yields the gains above across all budgets, and identity alignment supplies the cross-admission consistency behind tracking and retrieval (R@5 51.4% vs 0.16% random). Recurrent admissions therefore help clinically through longitudinal prediction, with the patient anchor organizing each visit against the patient's own history.

Controlling identity beyond one loss. A.6 already notes that removing one loss does not remove every identity dependence: the semantic predictor, patient prototype, temporal regularizer and paired sampler all carry longitudinal information. The controlled design disables all simultaneously (wpred=wprior=wtemporal=0w_{\text{pred}}=w_{\text{prior}}=w_{\text{temporal}}=0) and varies the sampler itself: true = two time-ordered admissions of one patient; random = different-patient admissions as pseudo-positive pairs, with explicit handling of accidental same-patient collisions and false negatives; synthetic = two independently MAE-masked views of the same admission; no_identity = the same synthetic views with contrastive and adversarial weights also zero. The identity head reads a detached PAT via a separate projection (stop-gradient), so identity gradients never enter the shared clinical encoder, and adversarial suppression is removed (wadv=0w_{\text{adv}}=0). All arms use three fixed seeds and identical data/steps/readout, isolating the pairing factor rather than one loss term.

Reviewer_aVh4OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

PAID uses recurrent admissions as positive pairs for multimodal EHR representation learning. A shared Transformer produces patient-level PAT and admission-level ADM representations, trained with identity alignment and adversarial identity suppression, respectively. Reconstruction and missingness-aware matching provide additional supervision.

Critical strengths
  1. Addresses the meaningful challenge of preserving stable patient context without suppressing admission-specific changes.
  2. Provides evaluation across two datasets, including clinical tasks, few-shot learning, retrieval, and component ablations.
Critical weaknesses
  1. Limited methodological novelty and evidence for complementarity. Patient-based contrastive learning already appears in PCLR and COMET. The main distinction is asymmetric PAT/ADM training, but its practical complementarity is weakly demonstrated: adding ADM and the scalar score increases the six-task AUROC from 0.85961 to only 0.85983 (Table 5).

  2. The comparisons do not isolate representation-learning benefits. Appendix A.3 specifies method-dependent labeled training pools and readouts, with several few-shot PAID results additionally using XGBoost ensembles. PAID also receives auxiliary CheXpert supervision. These differences complicate attribution of improvements specifically to patient-identity contrast.

Questions for authors

see in Critical Weaknesses

1 comment
AuthorsOpenReview ↗

Novelty and complementarity (Weakness 1); attribution (Weakness 2)

Novelty beyond PCLR/COMET. Patient-based contrast is not new: PCLR contrasts same-patient ECGs, and COMET adds patient-level relations to a hierarchy of time-series contrasts. PAID's contribution is what patient-based supervision becomes for recurrent, incomplete, multimodal admissions. Prior EHR pretraining models disease, population patterns from visits treated as i.i.d.; PAID models the patient. One individual's admissions are genuine multi-views that share a stable background but differ in acute state and modality coverage, not artificial augmentations. Aligning them extracts a persistent individual anchor, so each visit is predicted against the patient's own history and the same patient is trackable across visits without ID lookup (retrieval R@5 51.4% vs 0.16% random). Architecturally, one shared encoder produces two asymmetric readouts under availability-masked attention, PAT aligned across admissions and ADM retaining episode-specific evidence, trained with cross-modal reconstruction and an evidence-aware matching score (dispersion/context/missingness) that down-weights identity matching when evidence is weak. PCLR (single-modality ECG) and COMET (time-series hierarchy) do not address this dual-space, incomplete-multimodal setting.

Table 5: new frozen-encoder comparison. The original +0.00022 came from a fine-tuning study (20+20 epochs, feature-mask refinement, best validation checkpoint per readout, A.8); fine-tuning absorbs readout differences, so it cannot establish complementarity. New test: one shared frozen encoder per seed, identical logistic readout, splits and tuning budget, three fixed seeds, full labels, six-task macro AUROC, same training configuration as the baselines:

ReadoutAUROC (mean±SD)
PAT0.8579±0.0033
ADM0.8593±0.0026
PAT+ADM0.8635±0.0008
PAT+ADM+σ0.8639±0.0006

PAT+ADM − PAT = +0.0056±0.0025, all three seed-level patient-clustered paired CIs positive; ADM alone matches PAT, not a degraded copy. The increment is larger on individual tasks (LOS>3d +0.0075, LOS>7d +0.0070, ICU utilization +0.0058, AKI +0.0049, sepsis +0.0055, all with three positive seed CIs) and grows when modalities are missing (e.g. LOS>7d +0.0171 without text, +0.0120 without ECG), exactly the incomplete-coverage regime the dual space targets. σ adds a small further increment (+0.0004).

Labeled pools, readouts, ensembles. Original tables are system-level comparisons (A.3 disclosed the differences). The new unified protocol shares the same 304,916 pretraining admissions, labeled patient pools and fit/val/test manifests, label budgets (full/1%/5%), train-fit StandardScaler and identical logistic readout (solver, regularization, class weights, tuning budget), validation-only selection, fixed-final-epoch checkpoints, identical fine-tune heads with a 50-epoch cap, and three fixed seeds with patient-clustered paired bootstrap. Fine-tuning test AUROC:

TaskPAIDMUSESimCLR
AKI0.8432±0.00180.8342±0.00020.8270±0.0006
Readmission0.6362±0.00070.6226±0.00240.6253±0.0026

PAID leads both baselines; every PAID−MUSE seed-level paired CI is positive (AKI +0.0089±0.0020, readmission +0.0136±0.0021). For XGBoost, 3 of 21 few-shot cells used validation-selected stacking (LOS>3d@10%, ICU@5%, ICU@10%): PAID-only 0.8598/0.9193/0.9266 vs raw XGBoost 0.8620/0.9236/0.9282 vs stacking 0.8682/0.9297/0.9339. PAID-only leads 18/21 settings; ensemble cells are labeled system-level. With full labels and three seeds, stacking matches or exceeds raw XGBoost on every task (largest gains mortality +0.0030, readmission +0.0021).

CheXpert and attribution. Controlled arms disable all auxiliary identity sources simultaneously (wpred=wprior=wtemporal=0w_{\text{pred}}=w_{\text{prior}}=w_{\text{temporal}}=0: no CheXpert cross-admission target, no prototype, no temporal term), with reconstruction, VICReg, architecture, masks, data/steps and readout fixed; only the pairing factor varies (true/random/synthetic/no-identity), three fixed seeds, paired CIs. Clinical gains are attributable to cross-modal content learning and the dual readout (+0.0056, CIs positive). The main role of identity alignment is cross-admission consistency (retrieval) and enabling longitudinal supervision: next-visit feature prediction, definable only on true pairs, improves the macro by +0.0115/+0.0133/+0.0143 (full/5%/1%) over no-identity, all CIs positive.

Summary. The revision centers the contribution on recurrent admissions as outcome-free, patient-indexed supervision, with clinical gains from content learning, the dual readout and fusion, and identity alignment as the patient-level anchor for tracking and longitudinal prediction. The numerical corrections (readmission +0.0030 vs MUSE; 11 clinical endpoints; mean MIMIC gain 0.0040; 18/21 PAID-only) have been applied throughout.

Reviewer_dxQqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a representation learning method for electronic health records called PAID, a contrastive learning method that aligns separate admissions of the same patient together. The goal is to learn an encoder that maps observed modalities at a given admission into a patient-level and admission-level representation. The visit-level representations are aligned after adjusting for missing observations. Experimental validation on two ICU datasets show that the approach outperforms various baselines on classification tasks.

Critical strengths
  • The concept of aligning visit-level representations within the same patient is potentially a useful additional inductive bias specifically for representation learning approaches that embed each visit independently. The exploration and empirical evaluation of this concept offer useful insights for learning robust patient representations.

  • The paper provides experimental validation across two real-world datasets and over diverse clinical classification tasks, as well as ablations for each component of the proposed framework. The experimental validation is also comprehensive as it evaluates practical real-world challenges such as missing modalities and sparse label scenarios.

Critical weaknesses
  • The motivation for using a contrastive objective (Section 3.4) to align visit-level representations of the same patient requires further justification, either through a theoretical rationale or illustrative examples showing when it is beneficial. Specifically, it does not seem necessary when modeling P(Y∣History)P(Y \mid \text{History}) for EHR inference tasks, where patient identity is largely orthogonal to extracting representations predictive of the target variable Y. This is further supported by Table 3, where removing the adversarial loss and "dual space" (presumably omitting hPATh_{PAT} ) does not significantly degrade performance. In fact, enforcing distinct representations across different patients can be a counterproductive inductive bias, given that clinically similar patients should yield similar predictions.

  • The experimental setup is difficult to establish that these particular design choices in the loss provide significant performance gain over baseline methods. Aside from the lack of confidence intervals, a more convincing argument would be to demonstrate that the patient-identity contrastive loss offers unique supervisory benefits beyond alternative standard modern pretraining objectives for EHR like next-token prediction or time-to-event modeling (ex. MOTOR, ORA), which already likely establish strong within-patient supervision and learn representations that naturally encode patient history. This would be best demonstrated via targeted loss ablations that isolate the objective while maintaining a fixed backbone/architecture.

Questions for authors
  • How should Lidentity\mathcal L_{identity} be interpreted and justified in the context of downstream EHR classification tasks, where the goal is to extract a representation that captures patient characteristics relevant to the classification task? When and why would Lidentity\mathcal L_{identity} be beneficial, and does this provide distinct advantages over alternative pretraining objectives like next-token prediction and time-to-event?

  • To help readers better parse the proposed methods and key insights, a few aspects of the text could benefit from improved organization and streamlining. For example, the methodology section is currently difficult to navigate because core design choices are not clearly prioritized over minor technical implementation details. For instance, Section 3.2 introduces the framework before defining key terminology (e.g., adversarial suppression or matching score). Furthermore, non-essential equations and notation (ex. type embedding eme_m, Equation (3)) are included in the main text despite not being central to the proposed method or referenced later. Moving such secondary details to the Appendix would help keep the main text focused on core concepts. Finally, the core contributions lack sufficient detail and intuition, such as equation 5 and 6. For example, what is the exact purpose of capturing dispersion in Equation (5), and what does the outer summation over r=1…dr=1…d represent? In Equation (6), how are fmissf_{miss} and fctxf_{ctx} defined? Do they operate on different subcomponents of h, and what is the underlying rationale for summing all σ\sigma terms?

3 comments
AuthorsOpenReview ↗

Reply 1/3 — Motivation and inductive bias (Weakness 1); interpretation and when it helps (Question 1)

Why align visits of the same patient. Most EHR pretraining objectives, such as masked modeling, next-token prediction and SimCLR-style contrast, model disease: they learn population-level clinical patterns from admissions treated as independent samples, and the representation answers "what clinical state does this visit resemble?" The same disease presents differently across patients, however, and a patient's visits are correlated through stable individual factors such as chronic baseline, physiology, comorbidity structure and care-seeking behavior. A disease-level model must re-estimate this individual context from each visit's limited observations.

PAID adds patient-level modeling. Aligning visits of the same patient forces the encoder to extract what persists across that patient's admissions, an individual anchor, while the admission route (ADM) retains visit-specific state. Each visit can then be predicted against the patient's own longitudinal context rather than the population average: who is this patient, and how does this visit deviate from their baseline? This is the information a clinician uses when they know the patient, and it is unavailable to objectives that treat visits as i.i.d.

The benefit is largest exactly where patient history exists. Under a unified protocol (same data, labeled pools, readout and three fixed seeds), PAID leads both MUSE and SimCLR on AKI (+0.0089 over MUSE) and readmission (+0.0136 over MUSE). Stratifying by the patient's number of admissions, the PAID−MUSE gap on AKI grows from +0.0036 for single-admission patients to +0.0066 for two admissions and +0.0111 for three or more, with all seed-level paired CIs positive in the recurrent group. Longitudinal next-visit supervision built on true patient pairs improves the six-task macro by +0.0115/+0.0133/+0.0143 (full/5%/1% labels) over no-identity training, largest at 1%. The anchor is patient-specific: retrieval R@5 51.4% vs 0.16% random, with no ID lookup at inference.

Whether separating patients harms clinically similar cases. We agree that clinically similar patients should yield similar predictions, and the design contains this risk rather than ignoring it. Identity alignment acts on the patient anchor PAT, while the clinical admission representation is trained by masked cross-modal reconstruction and downstream supervision, which pull clinically similar visits together regardless of identity. The identity objective is a class-balanced pairwise sigmoid loss (Eq. 8), not row-normalized InfoNCE: each pair is classified directly, so a clinically similar negative contributes a bounded per-pair term instead of distorting a batch-wide softmax, and each batch draws distinct patients with one ordered pair per patient. The identity head reads a detached PAT through its own projection, so separation pressure, including any erroneous pressure from same-disease negatives, never reaches the shared clinical encoder. Clinical performance is preserved while the same patient remains trackable across visits.

One correction to the reading of Table 3: removing the dual space (ADM+matcher) costs −0.0568 mean AUROC and removing the adversarial loss −0.0274, neither negligible. These are joint effects under the full recipe (A.7), which is why we also report the fixed-backbone isolation in Reply 2.

How to read the representation downstream, and when it helps. PAT is a patient-level context anchor, not a label-bearing feature. The identity objective organizes the patient space during pretraining; downstream classifiers read ADM or PAT+ADM, whose clinical content comes from reconstruction and supervised fine-tuning. The anchor lets each visit be interpreted against the patient's own history, the individual baseline a population-level disease model lacks. It helps most for recurrent-admission patients, for tasks sensitive to individual baseline (readmission, and AKI whose label is defined relative to the patient's own creatinine), and under scarce labels, where longitudinal structure supplies outcome-free supervision. The role of identity alignment is to provide this patient-level consistency and to enable the longitudinal supervision that produces the clinical gains above.

AuthorsOpenReview ↗

Reply 2/3 — Unique benefit and targeted ablation (Weakness 2); comparison with next-token and time-to-event (Question 1)

Benefit beyond standard pretraining objectives. We implemented two longitudinal adaptations on the same backbone, true pairs, MAE/VICReg, masks, batch 1024, 100 epochs, fixed-final checkpoint, three fixed seeds and frozen readout: next_visit predicts the later admission's raw multimodal features from the earlier (the analog of next-token for pre-extracted admission features), and time_to_event predicts log⁡(1+days to next admission)\log(1+\text{days to next admission}) from real admission timestamps (the analog of time-to-event). These are adaptations to our feature setting, not full MOTOR/ORA implementations, and we label them as such.

On the six-task macro (full labels), next_visit exceeds identity-only true-pair PAT by +0.0096±0.0014 (all seed CIs positive) and no-identity training by +0.0115; the gap over no-identity widens to +0.0133 at 5% and +0.0143 at 1% labels. time_to_event alone matches identity-level clinical performance but adds no further signal (+0.0007).

These objectives model different things and are complementary rather than substitutes. Identity alignment models who, a persistent patient anchor; next-visit prediction models what changes, longitudinal content dynamics; time-to-event models when. Content dynamics add clinical signal on top of identity training, so identity alone does not capture all longitudinal information. Conversely, neither alternative explicitly optimizes cross-admission identity consistency, the capability behind patient-level tracking and retrieval (R@5 51.4% vs 0.16%), and neither defines a valid longitudinal target on its own, since a random pair has no "next visit". The supervisory benefit of patient identity is therefore structural: it turns recurrent admissions into a patient-indexed longitudinal resource that is outcome-free, uses no future information at inference, and makes no token-sequence or event-time assumptions. The clinical gains of that resource are realized through longitudinal content prediction, with identity gradient-isolated so the two roles coexist without interference.

Targeted loss ablation with a fixed backbone. The ablation you asked for is our controlled matrix: same backbone and architecture, same reconstruction and VICReg, same masks, same data/steps/budget, same readout, with all auxiliary identity sources disabled simultaneously (wpred=wprior=wtemporal=0w_{\text{pred}}=w_{\text{prior}}=w_{\text{temporal}}=0, so no CheXpert cross-admission target, no patient prototype, no temporal term). Only the pairing factor varies across true, random, synthetic (two MAE-masked views of one admission) and no-identity, over three fixed seeds with 2,000-resample patient-clustered paired bootstrap; the identity head acts on detached PAT (weight 0.03, wadv=0w_{\text{adv}}=0). The clinical gains are attributable to cross-modal content learning and the dual readout (PAT+ADM−PAT = +0.0056±0.0025, all CIs positive) and to the longitudinal supervision that true pairs enable (+0.0096 over identity-only). Identity alignment contributes cross-admission consistency and makes that longitudinal supervision possible. This is a sharper attribution than Table 3, which removes one component per run under the full recipe.

AuthorsOpenReview ↗

Reply 3/3 — Presentation and equations (Question 2)

Ordering of §3.2. The section was written in forward-pass order (inputs, shared encoder, PAT/ADM, objectives), so the framework overview preceded the loss-specific terminology such as adversarial suppression and matching score, defined at first use in §3.3–3.4. We agree this forced readers to hold undefined terms in mind. The revised version places a short terminology-and-notation block (PAT, ADM, identity alignment, adversarial suppression, matching score) before the framework and prioritizes the core design choices over implementation details.

Type embedding and Eq. 3. These are standard Transformer machinery rather than method-specific components. The revised version moves them to the appendix; the main text retains only the availability-masked attention rule, which is central to learning under incomplete coverage.

Eq. 5 (dispersion). For each coordinate r, the inner term is the availability-weighted variance of the visible modality tokens around their weighted mean, and the square root makes it a per-coordinate standard deviation. The outer average over the d coordinates collapses these into a single scalar per admission: how much the visible modalities disagree about the patient's state, i.e. evidence heterogeneity. The identity logit subtracts γ(σi+σk)\gamma(\sigma_i+\sigma_k) (Eq. 7, γ=0.1\gamma=0.1), so when visible evidence is internally inconsistent, identity matching is down-weighted rather than trusted. With one visible modality the dispersion term reduces to ε\sqrt{\varepsilon} while context and missingness continue to contribute (A.4). Dispersion is computed before modality-type embeddings are added, so it reflects content, not modality identity.

Eq. 6 (components and rationale). The three terms act on different subcomponents: σdisp\sigma_{\text{disp}} on the visible modality tokens; σctx=softplus(fctx(hpat))\sigma_{\text{ctx}}=\mathrm{softplus}(f_{\text{ctx}}(h_{\text{pat}})), a learned nonnegative readout of the patient-context token; and σmiss=∑m(1−a~m)dm\sigma_{\text{miss}}=\sum_m (1-\tilde{a}_m)d_m with dm=softplus(bm)+softplus(fmiss(hpat)m)d_m=\mathrm{softplus}(b_m)+\mathrm{softplus}(f_{\text{miss}}(h_{\text{pat}})_m), learned per-modality penalties summed only over absent modalities (observed modalities contribute exactly zero), gated by the availability mask. Each term is a distinct, nonnegative source of evidence incompleteness: internal disagreement, weak patient context, and missing modalities. The squared sum σ2=σdisp2+σctx2+σmiss2+ε\sigma^2=\sigma_{\text{disp}}^2+\sigma_{\text{ctx}}^2+\sigma_{\text{miss}}^2+\varepsilon aggregates them into one nonnegative, differentiable scalar, so any single source can down-weight matching and no term can cancel another. A.4 derives the score's properties and gradients.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.