acceptodds
Under review as a conference paper at ICLR 2027

Forecasting Patient Trajectories from the EHR with Action-Conditioned World Models

11 views

Abstract

We introduce EHR-WM, a predictive world model for electronic health records that represents patient histories as action-conditioned latent trajectories. Rather than modeling clinical events solely through autoregressive token prediction, EHR-WM separates patient states from observed clinical actions and learns their evolution through a state- and action-conditioned transition operator. A state-dependent linearization predictor propagates latent changes over time, supporting multi-step rollout and action-conditioned simulation. On MIMIC-IV, generated trajectories preserve task-relevant and temporal structure, and decoded free-running rollouts remain predictive of future clinical measurements against task-matched forecasting baselines. Across classification and time-to-event tasks, EHR-WM is competitive but task-dependent relative to substantially larger EHR foundation models, using a 4.39M-parameter model over a restricted high-frequency feature space. Perturbing supplied action sequences produces corresponding changes in generated trajectories, demonstrating action-sensitive learned dynamics. These results support world modeling as a complementary framework for explicit, time-indexed modeling of patient evolution.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
world models, electronic health records

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.50 over 2 reviews, above 99% of the others.
Reviewer_S43hOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Authors develop a world model that represents patient EHR using patient conditions, measurements, and medications. They create an algorithm to create patient future trajectories conditions on treatment. They also develop classfier for several conditions like sepsis, MIs, etc.

The main conribution is a new archtecture for medication conditions generations of future trajectories.

They use MIMIC-IV with patient disjoint train, val, and test sets. They evaluate future measurement accuracy, outcome prediction, similarity between generated and observed representations, and preservation of demographic associations. Results are mixed.

Generating patient trajectories is not itself new.

Changing medication inputs changes the generated trajectories, but the clinical correctness of those changes is not validated

Critical strengths

An interesting model achieving competative performance with other investiagtoed models capable of generating conditioned patient future trajectories

Critical weaknesses

The paper does not isolate whether generating future trajectory improves outcome prediction over directly using the same observed patient history.

Outcome definitions are doubioys considering that MIMIC-IV was used. Authors predict long term wheres MIMIC-IV is a reconst of ED and ICU followed by hospital admission. MIMIC is much better fro short-term inpatient forcasting (like sepsis).

Questions for authors

I am not really clear how medications inputs are selected. Are subsequent medication inputs the treatments actually administered later, or a plan known at prediction time? Where do the reported forecast windows begin relative to the two observed initialization periods, and do all baselines receive equivalent information?

You predict trajecroty based on latent representation and then predict outcomes based on trajecotry. What is vaue added of tarjecotry. What not to predict everything using latent info?

It needs to be explained how long term disease pancratatic cancer is established. In this dataset if patient have no ED or hospitalisatio that does not mean that it is negative for given outcome (pancreating cancer)

Figure 5 is not explaned in suffcient detail, what are the trajecotries what do they represnt. when medications for those patients where administered?

3 comments

The paper does not isolate whether generating future trajectories improves outcome prediction over directly using the same observed patient history.

Response: We thank the reviewer for this suggestion. We introduce a matched-information ablation comparing EHR-WM against direct prediction from initial latent states, with and without clinical actions. All models share the same initialization context, state encoder, and evaluation protocol. Importantly, STATES + ACTIONS receives the same recorded action sequence as EHR-WM, but without autoregressive trajectory generation, isolating the contribution of learned latent dynamics rather than additional clinical information.

As shown in Figure 6, EHR-WM improves AUROC over STATES + ACTIONS by 0.090, 0.079, and 0.029 for ICU admission, long LOS, and sepsis, respectively. These results demonstrate the additional predictive value of explicitly modeling patient-state evolution.

Trajectory-Based Prediction Ablation (AUROC, MIMIC-IV)

ModelICU AdmissionLong LOSSepsis
STATES0.7930.6560.705
STATES + ACTIONS0.8010.6890.757
EHR-WM0.8910.7680.786

Revision: Section 4, Figure 6.

Outcome definitions are dubious considering that MIMIC-IV was used. MIMIC is much better for short-term inpatient forecasting.

Response: We thank the reviewer for this observation. We revise the primary multi-site benchmark to emphasize inpatient outcomes, including ICU admission, sepsis, mortality, length of stay, and readmission where available. We remove long-horizon disease outcomes from the primary comparison (include them in the appendix) and clarify the limitations of disease-incidence labels derived from incomplete longitudinal follow-up.

Revision: Section 4 and Appendix D.

Are subsequent medication inputs the treatments actually administered later, or a plan known at prediction time? Where do forecast windows begin, and do all baselines receive equivalent information?

Response: We clarify that medication inputs are subsequently recorded administrations, not treatment plans known at prediction time. Actions are encoded as binary indicators across five treatment classes in six-hour intervals. EHR-WM initializes from two observed windows and generates future states autoregressively without future physiological observations, conditioned on recorded actions. Forecasting covers 6–48 hours after initialization, with all baselines receiving the same initial context and evaluated on identical variables, horizons, and observations.

Revision: Sections 3–4 and Appendix D.3.

What is the value added by trajectory generation? Why not predict everything using latent information?

Response: We address this directly through the matched-information ablation in Comment 1. Even when direct prediction receives the same initial states and action sequence, EHR-WM consistently performs better, demonstrating that autoregressive latent trajectory modeling contributes predictive value beyond directly using the available representations and actions.

Revision: Section 4, Figure 6.

How is pancreatic cancer established? The absence of hospitalization does not imply the absence of disease.

Response: We thank the reviewer for highlighting this limitation. Disease-incidence labels are defined using the first recorded ICD diagnosis within 365 days, excluding patients with prior diagnoses. We clarify that absence of a recorded diagnosis does not establish absence of disease, since care outside the observed health system is unavailable. We therefore prioritize inpatient outcomes in the primary evaluation and explicitly discuss this limitation and move the remainder of the tasks to the appendix.

Revision: Appendix D.1 and Section 5.

Figure 5 is not explained in sufficient detail. What do the trajectories represent, and when were medications administered?

Response: We thank the reviewer for requesting clarification. We have expanded the intervention analysis to specify that trajectories represent simulated latent patient-state evolution under recorded or modified medication sequences as suggested by reviewer edcc. The matched-cohort experiments initialize trajectories using the first 24 hours of patient history, introduce interventions between 24 and 54 hours, and evaluate physiological responses at 54–60 hours against observed treated–control differences. We have revised the figure captions to clarify intervention timing and interpretation.

Revision: Section 4, Figure 4.


Thank you again for the constructive feedback! Please let us know if these additional experiments and analysis provide sufficient evidence to warrant a score raise.

I appreciate the authors’ efforts to address my comments on such short notice. I find their responses and the corresponding revisions satisfactory.

I would, however, like to highlight one omission that I overlooked in my original review and subsequently raised in my comment to the Area Chair. ETHOS is particularly relevant prior work: it was trained and evaluated on MIMIC-IV, has a publicly available implementation, and generates patient-specific future trajectories conditioned on longitudinal medical histories. It should therefore be cited and discussed directly, including how EHR-WM’s explicit separation of patient states and clinical actions differs from this earlier approach.

Reference: Renc et al. “Zero shot health trajectory prediction using transformer.” npj Digital Medicine 7, 256 (2024). https://doi.org/10.1038/s41746-024-01235-0.

AuthorsOpenReview ↗

Response S43h

Thank you again for your constructive feedback and for sharing the relevant literature. We have provided a detailed response to your original comments in the Senior AC thread above, including an explanation of the corresponding revisions to the manuscript. We hope these revisions adequately address the remainder of your concerns and would appreciate your consideration of an updated score if it addresses all your concerns.

Updated: Thank you again for making our submission stronger with your feedback and follow up experiments :)

Reviewer_edccOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper models patient trajectories as obtained from the MIMIC-IV datasets, i.e. a discrete time-stamped sequence of observations (say a blood pressure reading) and clinical interventions (say the administration of a drug) with a JEPA-style world model, by encoding these explicit observation in a latent space and then learning to predict the evolution of that patient's latent state. The paper proposes some adaptations of an existing architecture to make it work for the patient trajectories. The method is evaluated for the fidelity and the utility of its learned representations, as well as for its controllability of the state evolution when keeping the observations but altering action sequences. The approach produces superior or comparable representations compared to their selected baselines while enabling controlability.

Critical strengths

I think EHR patient trajectories provide a suitable environment to formulate it as a problem for world models, thus the research is novel and timely. The results suggest that the learned representation are indeed useful and en par with other state-of-the-art patient trajectory representation learning methods. The paper is clearly written and easy to follow, the narrative makes sense, all the experiments I would expect to see are there and the setup seems generally sound. This is good work.

Critical weaknesses

There are no critical weaknesses as I see them. A few minor points, which are quite usual for such papers:

  • only MIMIC-IV dataset was used for evaluation (EHRShot could be another one to check - it would also help to see model transfer)

  • I think the main advantage of world models, controlability is a bit undersold here. Generally, the way I understand it, World models are strongest where they learn the underlying "world" (patient state in this case) from observations alone, i.e. without the inclusion of explicit actions, by modelling them latently. Here, the actions are given "for free", so that is not necessary. Furthermore, that aspect is also the weakest in the evaluation, the paper basically just shows that changing actions changes the outputs (so actions are indeed conditioned upon), not necessarily whether it's "good" or "correct".

  • This is not a weakness of this paper, but rather of such work in general: While technically, this is a sound paper, clinical utility is not touched upon. I am not saying it should (in this paper) - ICLR is a technical venue after all. This is more of an encouragement to continue validating the technology in more clinically relevant use cases.

Questions for authors

Regarding evaluating actions: I understand quality of the action conditioning here is hard to evaluate beyond "it changes something about the state", because there are no one-to-one counterfactually aligned patient trajectories with different outcomes based on including/withholding an intervention. One could, however, use a statistical measure here, similar to RCT setups - two otherwise indistinguishable cohorts that differ in a single intervention (i.e action). Applying the intervention to members of the cohort where it was not observed in the data should decode final states more similar to that of (or a linear probe should predict the outcome more in line with) the group which had the intervention applied in data. A full experiment might be out of scope for the rebuttal, but a simple conceptual setup and demonstration would greatly demonstrate the power of the proposed approach here.

2 comments

Only MIMIC-IV dataset was used for evaluation (EHRShot could be another one to check - it would also help to see model transfer).

Response: We thank the reviewer for this suggestion. We extend our evaluation to two independent inpatient icu datasets, eICU-CRD and NWICU, using harmonized clinical features and consistent evaluation protocols. EHR-WM is trained and evaluated independently on each dataset. We report the mean predictive performance across classification and TTE tasks below.

Classification Performance (Mean AUROC)

ModelMIMIC-IVeICU-CRDNWICUOverall
GBM0.72030.70030.69450.7055
CLMBR0.77100.74430.72730.7478
MOTOR0.78280.75970.72030.7537
EHR-WM (Rollout)0.76680.75270.73500.7514
EHR-WM (Oracle)0.76230.76930.74980.7596

Time-to-Event Performance (Mean C-index)

ModelMIMIC-IVeICU-CRDNWICUOverall
Cox PH0.66050.76370.73300.7150
DeepHit0.67680.74770.71250.7091
DeepSurv0.61150.77870.74530.7057
EHR-WM (Rollout)0.70080.76530.73650.7314
EHR-WM (Oracle)0.69880.76900.74500.7347

EHR-WM achieves overall scores of 0.7514 AUROC for classification and 0.7314 C-index for time-to-event prediction, remaining competitive with substantially larger EHR foundation models and achieving the highest average time-to-event performance among the evaluated models. These experiments demonstrate robustness across independently trained inpatient cohorts, rather than direct cross-dataset model transfer. We clarify this distinction and identify cross-site transfer evaluation and EHRShot (outpatient ehr) as future directions in the Limitations.

Revision: Section 4, Table 1, and Sections 5 and D.

I think the main advantage of world models, controllability is a bit undersold here.

Response: We thank the reviewer for this observation. To better emphasize controllability, we have revised our formulation to explicitly distinguish physiological states from observed clinical actions and introduced additive action forcing, which enables interventions to induce state-dependent latent changes even when the preceding physiological trajectory is stable. Furthermore, we have added a matched-cohort intervention evaluation demonstrating that simulated physiological responses align more strongly with observed outcomes for matching interventions than for nonmatching interventions. These revisions strengthen our evaluation of controllability by demonstrating intervention-specific physiological responses rather than simply showing that changing actions alters model predictions.

Revision: Sections 3.1 and 4, Figures 4–5.

Clinical utility is not touched upon. This is more of an encouragement to continue validating the technology in more clinically relevant use cases.

Response: We thank the reviewer for this encouragement. Our work establishes the technical feasibility of learning controllable patient dynamics rather than immediate clinical utility. We expand the Limitations to identify prospective clinical validation, causal evaluation, and clinically grounded applications as important future directions.

Revision: Section 5, Limitations and Future Work.

One could use a statistical measure here, similar to RCT setups, with two otherwise indistinguishable cohorts that differ in a single intervention.

Response: We thank the reviewer for this valuable suggestion. We implement a matched-cohort intervention analysis on MIMIC-IV across five treatment classes: antibiotics, intravenous fluids, vasopressors, analgesics, and sedatives. Treated and untreated patients are matched on pre-intervention physiological states, demographic characteristics, and concomitant treatments. We then introduce interventions into untreated trajectories and compare simulated physiological responses with observed treated–control differences.

Figure 4 shows stronger alignment for matching than nonmatching interventions (mean Pearson correlation 0.40 versus 0.25, permutation p=0.017p=0.017). These findings provide evidence of intervention-specific predictive fidelity beyond generic action sensitivity.

Revision: Section 4, Figure 4, and supplementary intervention analyses.

Thanks for the response and the additional experiments, the results further underline the quality of your work. I think this is good work and I'm happy to recommend acceptance.

Please continue engaging with the other reviewer(s) as necessary to satisfy their concerns.

Comments on the paper

5 comments
Senior_Area_ChairsOpenReview ↗

Open Review

This paper develops a method to build predictive world models from electronic health records (EHRs) and evaluates it extensively.

Strengths

  • Excellent review of prior related work in world models and foundation models for EHRs

  • Excellent novel combination of various existing tools and techniques from the above literature as well as curriculum learning to compose the final system, EHR-WM

  • Good discussion of the many practical data preprocessing steps that were required and were completed

  • Extensive experimental evaluation methodology and promising results, including forecasting and ablation studies

  • Discussion of limitations was included, though quite brief

Weaknesses

  • For inpatient data, much of which in MIMIC IV is virtually continuously collected, 6 hours is an awfully large time step at which to aggregate, potentially losing much information, especially in ICU

  • Using only top k most frequent clinical variables also seems to yield major information loss

  • It feels like Epic's Curiosity is a very related patient trajectory tool. Some mention of that seems useful.

  • More broadly, the work seems very related to synthetic data simulators for EHRs, such as for example in Epic's CoMET and the 2026 NeurIPS work of Rishi Kamalaswaran and Pan Xu.

  • Much of the innovation seems to be around connecting existing tools in a novel way. It would be helpful to see a discussion of which individual parts of the system are truly innovative, if any.

We thank the Senior AC for these constructive comments, which have helped us substantially strengthen the manuscript.

For inpatient data, much of which in MIMIC IV is virtually continuously collected, 6 hours is an awfully large time step at which to aggregate, potentially losing much information, especially in ICU.

Response: We agree that six-hour aggregation may obscure high-frequency physiological changes. However, EHR measurements vary substantially in sampling frequency, making temporal discretization a trade-off between resolution and sparsity. However in the original submission, we tested this ablation in Appendix G.3. Table 20 temporal resolution ablation across 3, 6, 9, and 12-hour intervals while holding the architecture and training protocol fixed.

As shown in Appendix G.3, Table 20, six-hour intervals consistently achieve the highest AUROC across all evaluated tasks, supporting their use as a balance between temporal fidelity and statistical efficiency.

IntervalICU AdmissionLong LOSSepsis
3 hours0.8820.7600.779
6 hours0.8910.7680.786
9 hours0.8840.7610.781
12 hours0.8760.7540.774

We additionally acknowledge that six-hour aggregation may miss rapid physiological fluctuations.

Revision: Section 3 and Appendix G.3, Table 20.

Using only top k most frequent clinical variables also seems to yield major information loss.

Response: We agree that frequency-based selection alone may exclude clinically informative variables. We have revised our feature schema to retain 45 relevant variables, including physiological measurements and demographic features shared across our three development and evaluation datasets. We additionally group medications into five clinically meaningful intervention classes, enabling consistent action conditioning across institutions while retaining broader medication families.

Revision: Section 3 and Appendix Tables 8–9.

It feels like Epic's Curiosity is a very related patient trajectory tool. Some mention of that seems useful.

Response: We thank the Senior AC for highlighting this work. We have expanded the related-work discussion to include Epic's Curiosity and CoMET, clarifying their relationship to EHR-WM. While these models generate future clinical events autoregressively, EHR-WM explicitly separates patient states from clinical actions and models latent physiological evolution under specified intervention sequences.

Revision: Section 2.

More broadly, the work seems very related to synthetic data simulators for EHRs, such as for example in Epic's CoMET and the 2026 NeurIPS work of Rishi Kamalaswaran and Pan Xu.

Response: We thank the Senior AC for these references. We have expanded our discussion of synthetic EHR generation and simulation approaches, including the work of Kamaleswaran and Xu. Unlike conventional synthetic data generation, which emphasizes producing plausible patient records, EHR-WM models the conditional evolution of an individual patient's latent physiological state under specified clinical actions, enabling controlled trajectory simulation rather than population-level record synthesis.

Revision: Section 2 and Appendix B.

Much of the innovation seems to be around connecting existing tools in a novel way. It would be helpful to see a discussion of which individual parts of the system are truly innovative, if any.

Response: We thank the Senior AC for recognizing the integration of existing techniques as a strength and for requesting clarification of our methodological contributions. EHR-WM introduces an explicitly controllable latent dynamics formulation that enables alternative-action rollouts from the same initial patient state. We further introduce additive action forcing, allowing interventions to induce state-dependent latent changes independently of preceding physiological dynamics. Our new matched-cohort evaluation across five treatment classes shows stronger alignment between simulated and observed responses for matching versus nonmatching interventions (Pearson correlation 0.40 versus 0.25, permutation p = 0.017). We also extend evaluation to MIMIC-IV, eICU-CRD, and NWICU, including multi-site forecasting and downstream prediction. These revisions clarify our methodological contributions and provide empirical evidence of intervention-specific dynamics beyond generic action sensitivity across multiple public datasets.

Revision: Sections 2–4, Table 1, Figures 3–4, and Appendix D.

AuthorsOpenReview ↗

Thank you to all the reviewers and senior area chairs for their constructive comments

Summary of Major Manuscript Revisions

We thank the reviewers and Senior AC for their constructive feedback. We have substantially revised the manuscript to address the concerns regarding generalizability, controllability, trajectory generation, and clinical relevance. The major updates are summarized below.

  • Revised state and action representations (Senior AC, edcc). We replace frequency-based feature selection with a harmonized schema of 45 clinically relevant variables, spanning vital signs, laboratory measurements, and demographics present across three brand new evaluation sites. We also introduce five medication action classes, antibiotics, intravenous fluids, vasopressors, analgesics, and sedatives, to support consistent action conditioning across datasets. These tables are included in Appendix Section C in Appendix Tables 8 & 9

  • Improved controllable latent dynamics (Senior AC, edcc). We introduce additive action forcing, enabling interventions to induce state-dependent latent changes independently of preceding physiological displacement. This extends the original transition formulation and allows interventions to affect patient trajectories even when prior physiological dynamics are stable.

  • Expanded evaluation across three clinical datasets (edcc, Senior AC). We extend evaluation beyond MIMIC-IV to eICU-CRD and NWICU, using harmonized clinical features and consistent evaluation protocols. The expanded experiments include classification, time-to-event prediction, and physiological forecasting across independent inpatient cohorts.

  • New matched-cohort intervention analysis (edcc, Senior AC). We evaluate action controllability across five intervention classes by introducing treatments into matched untreated patient trajectories. Simulated responses show stronger alignment with observed treated–control differences for matching versus nonmatching interventions (Pearson correlation 0.40 versus 0.25, permutation p = 0.017), providing evidence of intervention-specific physiological responses beyond generic action sensitivity.

  • New trajectory-generation ablation (S43h). We compare EHR-WM against direct prediction from STATES and STATES + ACTIONS under matched input conditions. Importantly, STATES + ACTIONS receives the same recorded action sequence without autoregressive trajectory generation. EHR-WM improves AUROC by 0.090, 0.079, and 0.029 for ICU admission, long LOS, and sepsis, respectively, isolating the predictive contribution of learned trajectory dynamics.

  • Expanded physiological forecasting evaluation (S43h, edcc). We evaluate free-running rollouts over 6–48-hour horizons against observed clinical measurements, using common initialization contexts and task-matched baselines, including GRU-D, Latent ODE, and Transformer forecasting models. We clarify that future medication inputs correspond to recorded administrations rather than prospective treatment plans.

  • Temporal resolution and preprocessing (Senior AC). We clarify the motivation for six-hour intervals and highlight our temporal discretization ablation across 3, 6, 9, and 12 hours, where six-hour intervals achieve the highest AUROC across all evaluated tasks. We also acknowledge the potential loss of high-frequency physiological information.

  • Revised clinical outcomes and limitations (S43h, edcc). We prioritize inpatient outcomes in the primary benchmark and clarify the limitations of long-horizon disease labels derived from incomplete clinical follow-up. We further distinguish multi-site evaluation from direct cross-site model transfer and identify causal evaluation, EHRShot, and prospective clinical validation as future directions.

  • Expanded related work and methodological positioning (Senior AC). We incorporate discussion of Curiosity, CoMET, and synthetic EHR simulators, clarifying that EHR-WM focuses on explicitly controlled latent patient-state evolution rather than generating synthetic patient records. We also distinguish our methodological extensions from the established world-model components on which they build.

These revisions strengthen the empirical evaluation of EHR-WM and clarify its primary contribution as an explicitly controllable framework for learning and simulating action-conditioned patient trajectories.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.