acceptodds
Under review as a conference paper at ICLR 2027

PatientWorld: A Physiology-Informed World Model for Learning Patient Dynamics

Abstract

Predicting how a patient's physiology evolves over time under treatment is fundamental to clinical decision making. Longitudinal electronic health records (EHRs) provide large-scale physiological and treatment trajectories from which patient dynamics can be learned. However, many clinical trajectory models either focus on direct future-state prediction without explicitly modeling treatment-conditioned transitions or encode heterogeneous physiological variables into shared representations, leaving physiological system organization and cross-system dynamics implicit. We introduce PatientWorld, a physiology-informed world model for learning reusable treatment-conditioned patient dynamics. PatientWorld organizes latent states by physiological systems, explicitly models generic treatments and cross-system effects, uses physiological priors to structure the effects of physiology-specific treatments, and models state-dependent local transition residuals. During pretraining on broad ICU trajectories, it learns cross-system interactions among physiological systems; downstream, the pretrained dynamics remain fixed while task-specific context adaptation uses task-relevant physiological relations and action-conditioned residuals to specialize transition contexts. Across MIMIC-III and MIMIC-IV, PatientWorld outperforms compared methods in nearly all multivariate trajectory forecasting settings, improves downstream adaptation over alternative reuse strategies, and achieves the highest mean on the primary decision metric among the compared world models on both tasks.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
EHR, World Model, Clinical Trajectory Prediction

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_77xeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PatientWorld is a latent world model for ICU trajectories. It factorizes the latent state into five physiological systems with global context. Each system’s transition combines three components: (1) GRU driven by a shared cross-system mixer and generic treatments, (2) physiology specific treatments routed to literature-defined target systems, and (3) gated local residual that limits rollout drift. The model is pretrained task-agnostically on MIMIC-III and MIMIC-IV (4-h bins, 48-h history, 12/24-h horizons). It is then adapted to sepsis treatment and RRT initiation timing on MIMIC-IV by freezing the backbone and learning residual corrections. It is evaluated on trajectory forecasting against 11 baselines, with the best error reported in 15 of 16 settings. There are also comparisons against other reuse strategies (from scratch, zero-shot, full fine-tuning, random-graph adapter).

Critical strengths
  • Well-motivated design with structural controls. Separating reusable, task-agnostic dynamics from task-specific context adaptation is a relevant direction for clinical world models. The physiological priors are tested with appropriate controls: shuffled treat breaks the treatment-to-system mapping, random graph adapt randomizes the task graph, and base WM is parameter-matched. These controls are what make the structural claims testable.

  • Broad, unified evaluation from forecasting to decisions. The paper evaluates diverse baselines under a shared protocol on two databases with three seeds, includes multiple OPE estimators and support/ESS diagnostics. I also appreciate the detailed appendices and code availability.

Critical weaknesses
  • The baseline comparison is not fully convincing for irregularly sampled EHR data. Although GRU and Transformer are included, common and relevant clinical irreuglar time-series models such as GRU-D, STraTS, DuETT, and mTAN are missing. Most comparisons rely on general-purpose forecasting models rather than EHR-specific approaches. A comparison with pretrained clinical time-series encoders using frozen backbones and forecasting heads would also help support the claim about reusing learned patient dynamics.

  • I am not fully convinced that the results demonstrate better treatment-conditioned dynamics. PatientWorld performs similarly to Transformer when future treatments are available, while most of its gains appear in blind forecasting. This raises the question of whether the improvement comes from better treatment modeling or simply greater robustness to missing future inputs. Also, the use of 4-hour bins may obscure the temporal relationship between treatments and physiological responses, particularly for fast-acting interventions. Fig 4 shows that predictions change under different actions, but does not establish whether these changes are physiologically plausible.

Questions for authors
  1. How are future treatment inputs handled in the blind setting? Were the same conditions applied to all baselines?
  2. What is the rationale for using 4-hour bins?
  3. How are test windows selected, and how are patients with less than 48 hours of history handled? An online-monitoring evaluation over the full ICU stay would also be interesting.
  4. The downstream results vary across OPE estimators, with other models outperforming PatientWorld on some metrics. How much of the improvement can be attributed to the world model itself rather than the shared planning components?
  5. Have you evaluated whether the predicted treatment responses are physiologically plausible, not just simply measuring how much predictions change under different actions?
5 comments
AuthorsOpenReview ↗

Response to Reviewer 77xe (Part 1/2)

We thank the reviewer for the constructive assessment and for recognizing the physiological design, structural controls, and common evaluation protocol. We added clinical EHR baselines and factual RRT treatment-response analyses, and clarify the relevant forecasting and downstream evaluation protocols below.

W1. Clinical EHR baselines

We added GRU-D, STraTS-supervised, and STraTS-SSL-frozen, covering missingness-aware modeling, supervised clinical sequence encoding, and a self-supervised pretrained clinical encoder kept frozen while training a forecasting head. All use the same MIMIC-III/MIMIC-IV patient splits, L12 history, and H3/H6 Conditional/Blind evaluation protocol as Table 1. Results are mean ± SD over three training seeds (lower is better).

MIMIC-IV

ModelMetricH3 Cond.H3 BlindH6 Cond.H6 Blind
GRU-DRMSE0.5807 ± 0.00020.6414 ± 0.01350.6175 ± 0.00150.6671 ± 0.0056
MAE0.3859 ± 0.00090.4455 ± 0.00910.4159 ± 0.00110.4749 ± 0.0052
STraTS-supervisedRMSE0.6166 ± 0.02650.6409 ± 0.03400.6402 ± 0.01450.6753 ± 0.0182
MAE0.4137 ± 0.02240.4542 ± 0.03310.4355 ± 0.01330.4891 ± 0.0166
STraTS-SSL-frozenRMSE0.5892 ± 0.00020.6041 ± 0.00070.6227 ± 0.00040.6395 ± 0.0017
MAE0.3941 ± 0.00070.4153 ± 0.00450.4217 ± 0.00100.4411 ± 0.0037

MIMIC-III

ModelMetricH3 Cond.H3 BlindH6 Cond.H6 Blind
GRU-DRMSE0.6328 ± 0.00050.7090 ± 0.01140.6653 ± 0.00110.7303 ± 0.0096
MAE0.4330 ± 0.00040.4993 ± 0.00890.4556 ± 0.00080.5133 ± 0.0077
STraTS-supervisedRMSE0.6856 ± 0.00710.6930 ± 0.00730.7102 ± 0.00730.7170 ± 0.0078
MAE0.4745 ± 0.00560.4782 ± 0.00490.5022 ± 0.00590.5049 ± 0.0058
STraTS-SSL-frozenRMSE0.6510 ± 0.00020.6709 ± 0.00590.6786 ± 0.00020.6978 ± 0.0061
MAE0.4471 ± 0.00040.4686 ± 0.00500.4758 ± 0.00040.4958 ± 0.0048

All three additional clinical baselines have higher mean RMSE and MAE than PatientWorld in the four forecasting settings on both datasets. We will add these comparisons to the revised paper.

Q1&W2. Blind treatment inputs

In the Conditional setting, models with a treatment-conditioning interface receive recorded future treatments. In the Blind setting, future treatment information is withheld, while historical treatment records remain available. All methods use the same patient-level splits, history windows, horizons, and metrics. Baselines that do not take treatments as inputs retain their respective forecasting interfaces. Thus, future treatment information is not withheld only from PatientWorld, nor supplied to competing methods in the Blind setting.

Q2&W2. Four-hour temporal discretization

We chose 4-hour bins to balance temporal resolution and observation density in sparse, irregular ICU data, especially laboratory measurements. This resolution is also used in established ICU decision-making work (e.g., AI Clinician; Komorowski et al., 2018) and enables a common evaluation protocol for both datasets and all baselines.

Observation masks and elapsed-time features represent missingness and sampling intervals (Appendix B.2). PatientWorld predicts next-step physiology from preceding-step treatment inputs. For downstream sepsis modeling, the state uses information available at the decision point, the action corresponds to treatment recorded over the following 4 hours, and subsequent physiology provides the prediction target (Appendix B.3).

Q3. Test windows, short-history patients, and online monitoring

Forecasting uses a 12-bin (48-hour) history to predict H3 or H6. A valid window must contain the full history and forecast range—at least 15 or 18 bins, respectively. Training samples at most eight eligible windows per ICU trajectory; testing evaluates eligible windows from the held-out patient-level test split.

Patients without enough history to form a complete L12 forecasting window are therefore excluded from this forecasting evaluation, and short-history padding is not used in this setting. Downstream adaptation permits shorter histories with appropriate padding and masks. Evaluating rolling or online monitoring across entire ICU stays, including early-admission short-history predictions, is an important direction for future work.

AuthorsOpenReview ↗

Response to Reviewer 77xe (Part 2/2)

Q4. World-model contribution under shared planning

All dynamics-based approaches use the same action space, planner, and OPE protocol for each downstream task; their main difference is the learned dynamics supplied to the planner. Thus, differences under this controlled comparison cannot be attributed to different planning algorithms and instead reflect the offline decision performance of different world models within the same planning pipeline. We also report model-free CQL, which does not rely on world-model-guided planning, as an independent decision-making baseline.

Q5. Physiological plausibility of treatment responses (also addressing W2)

We agree that variation across hypothetical actions alone does not establish agreement with observed physiological responses. We therefore added factual RRT treatment-response forecasting, restricted to test windows with observed RRT initiation (action initiate). Conditional forecasts use the recorded RRT treatment schedule and are compared with subsequent observed physiological changes; Blind forecasts use the same windows without future treatment information. This differs from the hypothetical-action sensitivity diagnostic in Figure 4.

We evaluate normalized continuous changes in BUN, creatinine, potassium, bicarbonate, and pH:

ModelConditional RMSE ↓Blind RMSE ↓Conditional gain ↑
PatientWorld0.63600.73290.0976
Transformer0.66260.73300.0705
MedDreamer0.76710.84980.0827

Conditional gain is defined as Blind RMSE − Conditional RMSE. PatientWorld's gain is 0.0969, with a patient-clustered bootstrap 95% CI of [0.0683, 0.1320]. Its Conditional RMSE is 0.0266 lower than Transformer's (paired patient-clustered 95% CI [0.0071, 0.0451]) and 0.1311 lower than MedDreamer's ([0.0878, 0.1773]). The patient-macro Spearman correlation between predicted and observed physiological changes is 0.5888.

These results show that, on observed RRT-initiation windows, PatientWorld more accurately predicts subsequent measured physiological changes, providing factual treatment-conditioned evidence beyond action sensitivity alone.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Dear Reviewer 77xe,

Thank you for your time and effort in reviewing our paper. We sincerely appreciate your constructive feedback.

As the author–reviewer discussion period is approaching its end, we would greatly appreciate it if you could take a moment to review our rebuttal and let us know whether your concerns have been adequately addressed.

Thank you again for your time and consideration.

Best regards,
The Authors

Reviewer_ksF5OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PatientWorld, a physiology-informed latent world model for predicting treatment-conditioned patient trajectories. It factorizes patient states into physiological systems, models cross-system interactions and treatment effects, and introduces state-dependent residuals for multi-step forecasting. A two-stage approach pretrains general patient dynamics and subsequently adapts task-specific transition contexts while freezing the pretrained backbone. Experiments on MIMIC-III and MIMIC-IV show improved forecasting performance, improved downstream adaptation, and promising offline decision-making results for sepsis treatment and renal replacement therapy (RRT) initiation.

Critical strengths
  1. Physiological system factorization and treatment-specific pathways introduce meaningful domain knowledge into patient dynamics modeling.

  2. The proposed method got the lowest mean forecasting error in 15 of 16 reported settings, with improvements under blind multi-step forecasting.

  3. Freezing pretrained dynamics while adapting task-specific contexts is a practical approach to reusing learned patient representations across clinical tasks.

Critical weaknesses
  1. Although physiological factorization and treatment-specific pathways improve forecasting, the paper does not directly validate whether learned cross-system interactions and treatment effects align with established physiological knowledge. It remains unclear whether the improvements reflect meaningful physiological modeling or architectural advantages.

  2. Decision-making experiments cover two MIMIC-IV tasks and rely entirely on offline evaluation. Different estimators and secondary metrics do not consistently favor PatientWorld, raising concerns about the robustness and generalizability of its reported decision-making improvements.

Questions for authors

Q1: How does PatientWorld distinguish actual treatment effects from confounding by indication in observational ICU trajectories? How reliable are predicted transitions under treatment actions that differ from those historically administered?

Q2: Have the learned cross-system interactions and treatment spillover pathways been compared against established physiological relationships? Can the authors demonstrate clinically meaningful interactions beyond forecasting improvements?

Q3: The local residual uses a learnable horizon embedding. How does the model generalize to rollout horizons beyond those used during training? Could horizon-specific corrections learn systematic dataset biases rather than genuine physiological dynamics?

Q4: How sensitive are the reported decision improvements to the choice of FQE estimator, behavior-policy support thresholds, and reward definitions? How should the disagreement between FQE and WIS results be interpreted?

Q5: Since both downstream tasks use MIMIC-IV and the models are pretrained separately on MIMIC-III and MIMIC-IV, how well does the pretrained dynamics backbone transfer across hospitals, datasets, or patient populations without retraining?

3 comments
AuthorsOpenReview ↗

Response to Reviewer ksF5 (Part 1/2)

We thank the reviewer for recognizing the physiology-informed design, multi-step forecasting results, and reuse of pretrained dynamics. Below we address treatment-conditioned prediction, physiological interactions, rollout horizons, offline evaluation, and cross-dataset transfer.

Q1. Treatment confounding and off-support actions

PatientWorld learns treatment-conditioned predictive dynamics from patient histories and recorded treatments; its objective is not to identify independent causal treatment effects. Accordingly, it does not attempt to separate treatment effects from confounding by indication in observational ICU data, and differences between action-conditioned predictions are not interpreted as deconfounded treatment effects.

To assess prediction under actual treatment, we evaluated test windows with observed RRT initiation (action initiate), using the recorded factual RRT schedule to predict subsequent physiological changes. PatientWorld achieves a Conditional RMSE of 0.6360, versus 0.6626 for Transformer and 0.7671 for MedDreamer, supporting its ability to use observed treatment information for forecasting. For target-policy actions differing from historical decisions, Table 3 reports behavior-policy support diagnostics—Mean Supp. and Low Supp. for sepsis and WIS ESS for RRT (Appendix G.5)—to characterize their support in the observational data.

Q2&W1. Physiological interactions and treatment spillover

We assess the physiological plausibility of cross-system interactions and treatment spillover through two complementary analyses.

(1) Physiological priors and structural controls. Appendix E.1 grounds the cardiac–renal (C↔K), renal–metabolic (K→M), and respiratory–renal (R↔K) relationships in clinical literature, with kidney (K) and metabolic/acid–base (M) systems designated as primary RRT targets. The shared cross-system mixer models inter-system dynamics, while direct/spillover pathways distinguish treatment effects on primary targets and other systems.

In Fig. 3, Shuffled Treat underperforms PatientWorld across both MIMIC datasets and Conditional/Blind settings. In Table 2, Random Graph Adapt performs worse on all eight metrics; H6 Global RMSE is 0.6284 vs. 0.6305 for Sepsis and 0.6698 vs. 0.6743 for RRT. These controls support the predictive value of clinically motivated relationships and treatment mappings over randomized alternatives.

(2) Factual physiological responses. Using recorded treatment schedules from actual RRT-initiation windows, we predict observed changes in BUN, creatinine, potassium, bicarbonate, and pH. PatientWorld achieves 0.6360/0.7329 Conditional/Blind RMSE, with a Conditional gain of 0.0969 (patient-clustered 95% CI [0.0683, 0.1320]) and patient-macro Spearman correlation of 0.5888 between predicted and observed changes. This provides evidence of clinically meaningful physiological response prediction under recorded treatment, beyond structural forecasting comparisons.

Q3. Generalization beyond trained rollout horizons

We evaluated H3-trained checkpoints at H4–H6 without retraining. Inference at H1–H3 is unchanged; beyond the training horizon, we compare Clamp-to-H3 with No-horizon-embedding:

StepConditional: ClampConditional: No embeddingBlind: ClampBlind: No embedding
H40.6340 ± 0.00090.6341 ± 0.00090.6508 ± 0.00350.6507 ± 0.0034
H50.6729 ± 0.00170.6730 ± 0.00170.6920 ± 0.00490.6915 ± 0.0044
H60.6900 ± 0.00300.6899 ± 0.00330.7151 ± 0.00530.7131 ± 0.0036

Values are per-step RMSE (mean ± SD). The model continues to roll out beyond its training horizon, with similar errors under the two embedding rules. Furthermore, removing the local residual in the MIMIC-IV L12H6 ablation changes Blind RMSE from 0.6223 to 0.6239, indicating that multi-step performance does not primarily depend on horizon-specific residual correction.

AuthorsOpenReview ↗

Response to Reviewer ksF5 (Part 2/2)

Q4&W2. OPE estimators, behavior support, and reward sensitivity

(1) Estimator disagreement and behavior support. FQE estimates state–action values through Bellman-based fitting, whereas WIS reweights observed outcomes using target-to-behavior policy ratios. For deterministic policies, mismatched trajectories receive zero weight, making WIS sensitive to historical policy support.

MedDreamer's WIS advantage may be affected by finite-sample importance-weighting bias. Compared with PatientWorld, it has lower WIS ESS (61.06% vs. 75.42%) and higher across-seed WIS SD (0.695 vs. 0.070). However, FQE, DR, and WDR favor PatientWorld, so the WIS advantage is not consistent across estimators.

(2) Evaluator sensitivity. We tested seven RRT configurations covering FQE ridge regularization (10⁻⁴, 10⁻³, 10⁻²), propensity floors (10⁻⁴, 10⁻³, 10⁻²), and importance-weight clipping (20, 100, 500). With policies and the test cohort fixed, maximum absolute deviations from default OPE estimates were only 4.61 × 10⁻⁵ for PatientWorld and 5.19 × 10⁻⁶ for MedDreamer, with rankings unchanged within each estimator. The 0.01 Low Supp. threshold is diagnostic only and does not affect policy selection or OPE estimates.

(3) Reward sensitivity. Removing the lactate reward component while keeping the PatientWorld sepsis policy fixed changed mean Reward-FQE only from 6.613 to 6.581, indicating limited sensitivity to this reward variant.

Q5. Cross-dataset zero-shot transfer

We evaluated MIMIC-III ↔ MIMIC-IV zero-shot forecasting by directly applying a source-trained H6 model to the target dataset's test set, without target-dataset parameter updates:

TransferConditional RMSEBlind RMSEConditional MAEBlind MAE
MIMIC-III → IV0.9206 ± 0.06030.9229 ± 0.05800.6612 ± 0.03860.6590 ± 0.0356
MIMIC-IV → III0.6862 ± 0.01900.6853 ± 0.00990.4896 ± 0.01150.4923 ± 0.0054

The pretrained dynamics can generate cross-dataset forecasts without parameter updates, although performance differs substantially by transfer direction.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Dear Reviewer ksF5,

Thank you for your time and effort in reviewing our paper. We sincerely appreciate your constructive feedback.

As the author–reviewer discussion period is approaching its end, we would greatly appreciate it if you could take a moment to review our rebuttal and let us know whether your concerns have been adequately addressed.

Thank you again for your time and consideration.

Best regards,
The Authors

Reviewer_Jiq3OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces PatientWorld, a physiology-informed clinical world model designed to predict ICU patient dynamics under treatments and support clinical decision-making. The method first pretrains general dynamics by organizing latent states by organ systems, separating generic and targeted treatment pathways, and adding a local transition residual; it then freezes the pretrained backbone for downstream tasks and adapts only the transition context, demonstrating strong performance on MIMIC-III and MIMIC-IV.

Critical strengths

First, the problem is clearly motivated. Organizing latent states by physiological systems and separating generic from targeted treatment pathways aligns very well with ICU clinical intuition and practical clinical challenges.

Second, the experimental evaluation on two large-scale datasets, MIMIC-III and MIMIC-IV, is thorough and represents a complete, substantial amount of work. The results are detailed and credible, well-supported by extensive appendices and code for reproducibility.

Critical weaknesses

While the claimed contributions are appealing, I have several concerns regarding the model's actual working mechanisms and architectural novelty:

  1. My biggest concern lies in the multi-step rollout mechanism. The authors claim that the model performs better multi-step rollouts, yet it relies on introducing a local residual term rtr_t. The main function of this residual term is to prevent multi-step predictions from drifting too far. However, this not only restricts the model to a fixed rollout window (H<=6H <= 6), but also raises doubts about whether the model truly learns reliable physiological state-transition rules. See Question 1 for details.

  2. It is questionable whether the model truly achieves organ-specific separation and unchanged physiological dynamics in practice. On one hand, the dense mixer (fmixf_{mix}) at every step allows information from all organs to mix freely, which cannot guarantee that the latent variables are genuinely disentangled by organ. On the other hand, although the transition network weights are frozen during fine-tuning, modifying its inputs (and even adding an extra input pathway for the RRT action) still changes the effective transition rules.

  3. Finally, regarding architectural novelty and training tricks, this work looks more like a combination of a standard RSSM and "frozen backbone + adapter tuning" from transfer learning. Moreover, model training involves too many manually designed rules and fine-tuning tricks. For instance, the pretraining loss function alone has multiple loss components and manually set weights. This raises concerns about whether the model's performance relies heavily on fine-grained hyperparameter tuning, and whether it can remain reproducible and generalizable.

Questions for authors
  1. Why does PatientWorld perform almost identically to a basic Transformer in the Conditional setting (where true future treatments are provided), while its advantage is mainly concentrated in the Blind setting where future treatments are hidden (Table 1)? Beyond the H3 ablation in Fig. 3, could the authors clarify how much of this blind-forecasting gain is simply driven by rtr_t pulling predictions back from drifting?

  2. Pretraining uses multiple loss components along with manually set rules such as percentile thresholds and spillover weights. How much would the model's performance be affected if these auxiliary losses were simplified or if the model were allowed to learn some of these components automatically?

  3. What motivated the choice of 4-hour bins rather than a finer temporal resolution, such as 1 hour? Since relevant clinical interventions often take effect within minutes, does aggregating over 4-hour windows obscure short-term physiological responses and confound pre-treatment deterioration with post-treatment effects within the same bin?

6 comments
AuthorsOpenReview ↗

Response to Reviewer Jiq3 (Part 1/2)

We thank the reviewer for the careful assessment and for recognizing the physiological organization, treatment pathways, and evaluation on two datasets. We address the multi-step forecasting mechanism, auxiliary objectives, temporal discretization, and the distinction between physiological factorization and frozen-backbone adaptation.

Q1&W1. Blind forecasting and the local residual

In Conditional forecasting, both PatientWorld and the Transformer receive future treatments, narrowing their performance gap. In Blind forecasting, future treatments are hidden, making learned state-transition dynamics more important.

The local residual rtr_t is a gated latent correction, not an explicit pull toward historical states. Our Fig. 3 already includes a No-local ablation at H3. To quantify its longer-horizon contribution, we additionally tested MIMIC-IV L12H6 (single seed):

VariantConditional RMSE ↓Blind RMSE ↓
Full0.60460.6223
No-local0.60580.6239

Removing rtr_t increases Blind RMSE by only 0.0016 (0.26%), indicating that the local residual contributes modestly rather than driving the Blind advantage.

To address the fixed-horizon concern, we extended H3-trained checkpoints to H6 without retraining. At H6, Clamp-to-H3 and No-horizon-embedding achieve similar Blind RMSEs (0.7151/0.7131). Thus, rollout is not strictly limited to the trained horizon, supporting the role of learned transition dynamics beyond local residual correction.

Q2 & W3. Auxiliary losses and manually specified components

To assess dependence on auxiliary supervision, we removed all four auxiliary losses while retaining the architecture and core training objective. The single-training-seed results on MIMIC-IV L12H6 are:

ConfigurationConditional RMSE ↓Blind RMSE ↓
Full0.60460.6223
No auxiliary losses0.60440.6240

Without auxiliary losses, Conditional RMSE improves slightly and Blind RMSE increases by only 0.0017. Thus, the main forecasting performance is largely retained without auxiliary supervision.

The percentile thresholds are used only to construct auxiliary event labels for the local residual; these labels are no longer used when the auxiliary losses are removed. The spillover scale is a fixed coefficient (default 0.1) controlling secondary-effect strength, whereas state-dependent treatment responses are modeled by learned treatment gates, projections, and transition networks. Appendix F.6 further shows limited sensitivity to the rollout-loss weight over λ_rollout ∈ {0.25, 0.5, 1.0}. These results indicate that the main predictive performance does not rely heavily on the auxiliary-loss design.

Q3. Four-hour temporal discretization

We use 4-hour bins because ICU measurements, particularly laboratory tests, are sparse and irregular. Relative to 1-hour bins, aggregation reduces empty time steps from low-frequency measurements and balances temporal resolution against observation density. This resolution is also consistent with established ICU decision-making studies, including AI Clinician (Komorowski et al., 2018), and provides a common protocol across MIMIC-III, MIMIC-IV, and all baselines. Observation masks and elapsed-time features encode missingness and time since the last observation (Appendix B.2).

For temporal alignment, PatientWorld uses the preceding treatment input, at−1a_{t−1}, to predict physiology at the next step. In downstream sepsis modeling, the patient history available at the decision point defines the current state, treatment recorded over the following 4 hours defines the action, and the subsequent physiological observation is the prediction target (Appendix B.3).

We therefore view 4-hour discretization as a resolution–density trade-off: it reduces sparsity but limits the representation of within-bin short-term dynamics and immediate treatment responses. Exploring finer-grained or adaptive temporal modeling is a direction for future work.

W2. Physiological factorization and frozen-backbone adaptation

Two distinctions clarify our design. First, physiological factorization does not aim to learn statistically independent or causally disentangled organ states. It organizes observations and system-specific transition modules by physiological system, while a shared cross-system mixer models their interactions. Factorization therefore preserves system-specific representations without assuming independence.

Second, frozen pretrained dynamics means that pretrained backbone parameters remain fixed during downstream adaptation; it does not imply that the effective transitions remain identical after incorporating task-specific context. Relational and action-conditioned adapters introduce task-relevant interactions and treatment conditioning while reusing the pretrained backbone.

AuthorsOpenReview ↗

Response to Reviewer Jiq3 (Part 2/2)

W3. Architectural contribution

PatientWorld builds on RSSM and the frozen-backbone adaptation paradigm. Our architectural contribution is the integration of system-specific physiological transitions, cross-system interactions, and treatment-specific direct/spillover pathways into a structured patient dynamics model, together with relational and action-conditioned adapters for task-specific reuse. We do not claim RSSM or generic adapter tuning as new.

The paper includes parameter-matched Base WM, Shuffled Treat, and Random Graph Adapt controls. The new results further show lower Conditional/Blind RMSE for physiological versus random grouping (0.6046/0.6223 vs. 0.6114/0.6349), and lower RRT H6 RMSE with both adapters than with neither (0.6697 vs. 0.6921). These comparisons provide direct evidence for the structured design and adaptation components.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Dear Reviewer Jiq3,

Thank you for your time and effort in reviewing our paper. We sincerely appreciate your constructive feedback.

As the author–reviewer discussion period is approaching its end, we would greatly appreciate it if you could take a moment to review our rebuttal and let us know whether your concerns have been adequately addressed.

Thank you again for your time and consideration.

Best regards,
The Authors

Reviewer_M1wXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PatientWorld is a physiology-informed latent world model for treatment-conditioned patient trajectory forecasting and downstream decision support. It partitions clinical variables into physiological groups, maintains group-specific latent states and recurrent transitions, and combines cross-system context with treatment-specific direct and spillover pathways. A gated local-history residual provides additional corrections during open-loop forecasting. For downstream tasks, the pretrained backbone is frozen while task-specific relational and action-conditioned context residuals are trained. The paper evaluates forecasting on MIMIC-III and MIMIC-IV and downstream adaptation and offline policy utility on sepsis treatment and renal replacement therapy initiation tasks from MIMIC-IV. PatientWorld reports the lowest mean error in 15 of 16 aggregate forecasting comparisons, with its clearest gains under blind forecasting. It also reports the highest mean primary FQE metric on both decision tasks, although rankings differ under complementary estimators.

Critical strengths
  1. The architecture provides a concrete way to incorporate physiological organization into treatment-conditioned dynamics. Group-specific transitions and clinically motivated treatment routing make the proposed inductive biases explicit and testable. This is a useful design direction for clinical world models, particularly when compared with a single undifferentiated latent state.

  2. The evaluation provides encouraging evidence across forecasting and downstream reuse. The study uses patient-level splits, training-only normalization, multiple seeds, a capacity-matched monolithic control, and shuffled treatment-routing controls. The forecasting results are generally strong, and the frozen-backbone adapter performs competitively with full fine-tuning. Sharing planners across dynamics-based methods and reporting multiple offline estimators are also valuable experimental choices.

Critical weaknesses
  1. PatientWorld is selected using validation RMSE, while standard forecasting baselines use a combined value-and-mask loss; this difference could affect rankings, particularly where the reported improvements are small.
  2. Existing controls do not establish whether the gains come from clinically meaningful organization rather than generic variable grouping or residual adaptation, and the relational and action-conditioned adapters are not evaluated separately.
  3. Rankings differ across FQE, WIS, and WDR, while the reported three-seed standard deviations do not quantify patient-level uncertainty or sensitivity to evaluator choices.
  4. Section 4.1 excludes primary target systems and G from spillover effects, whereas Appendix C.2 aggregates spillover outputs without this exclusion, leaving the architecture used in the experiments unclear.
Questions for authors

See weaknesses

2 comments

We thank the reviewer for the constructive assessment and for recognizing the value of PatientWorld's physiology-informed modeling, treatment pathways, controlled experiments, and downstream reuse. We address the four concerns with checkpoint-selection checks, targeted ablations, and additional offline policy evaluation (OPE) analyses.

W1. Checkpoint selection

We checked whether selecting checkpoints by validation loss rather than validation RMSE changes the reported comparisons. Both criteria select the same checkpoint for GRU, Transformer, DLinear, PatchTST, iTransformer, MOMENT, Moirai 2.0, and TFT-multi. For Phys-JEPA, the selected checkpoint changes in only one of three training seeds (9903): Conditional RMSE changes from 0.6166 to 0.6104, while Blind RMSE changes from 0.9505 to 1.0341. Neither change affects the corresponding model ranking.

MedDreamer already uses validation RMSE for checkpoint selection. TDSTF uses validation denoising loss, given its diffusion training objective and the computational cost of sampling-based validation RMSE. We will clarify the checkpoint-selection protocol for each method in the revision.

W2. Contributions of physiological grouping, local residuals, and task adapters

We conducted three targeted ablations with a fixed training seed.

(1) Physiological vs. random grouping. We kept the number of groups and the number of variables per group unchanged and compared the physiological partition with three randomly generated partitions (MIMIC-IV, L12H6):

GroupingConditional RMSE ↓Blind RMSE ↓
Physiological0.60460.6223
Random grouping (3 grouping seeds)0.6114 ± 0.00040.6349 ± 0.0006

The physiological partition outperforms the tested random partitions under both settings. For random grouping, the reported mean ± SD is over three grouping seeds, with the training seed held fixed.

(2) Local residual and horizon correction. We compared the full model with variants removing the local residual or sharing the horizon correction:

VariantConditional RMSE ↓Blind RMSE ↓
Full0.60460.6223
No-local0.60580.6239
Shared-horizon0.60500.6228

The differences are small, indicating that PatientWorld's blind-forecasting performance does not primarily arise from local residual drift correction.

(3) Relational and action-conditioned adapters. We independently enabled the two adapters in RRT factual forecasting:

ConfigurationGlobal RMSE@H6 ↓
Neither0.6921
Relational only0.6813
Action only0.6908
Both0.6697

Each adapter improves on the Neither configuration, while their combination achieves the lowest error. Together, these results support the contributions of physiological grouping and task-specific adapters.

W3. Robustness of offline policy evaluation

FQE is our primary decision-evaluation metric, with WIS, DR, and WDR reported as complementary estimators. Their rankings need not coincide because they rely on different estimation mechanisms.

To quantify patient-level uncertainty, we performed a paired patient-clustered bootstrap for RRT using 2,955 patients, 3,711 episodes, and 2,000 patient-level resamples. The trained policies and OPE nuisance models were held fixed. The differences below are MedDreamer − PatientWorld (higher HFD60 is better):

EstimatorDifferencePatient-clustered 95% CI
FQE−0.5751[−0.6656, −0.3476]
WIS+1.7134[+0.7007, +1.8945]
WDR−0.8810[−2.2500, −0.5033]

FQE and WDR favor PatientWorld, whereas WIS favors MedDreamer. These patient-resampling intervals complement the three-training-seed mean ± SD reported in the paper.

We also tested seven RRT evaluator configurations, including the default and alternatives for FQE ridge regularization (10⁻⁴, 10⁻³, 10⁻²), propensity floor (10⁻⁴, 10⁻³, 10⁻²), and importance-weight clipping (20, 100, 500). With fixed policies and the same test cohort, the relative ranking of PatientWorld and MedDreamer remained unchanged within each estimator across the tested configurations.

W4. Spillover formulation and implementation

We thank the reviewer for identifying the inconsistency between Section 4.1 and Appendix C.2. In the implementation, the spillover pathway excludes the treatment's primary target systems and the clinical-context group G. Primary targets receive treatment inputs through the direct pathway, while the global context X retains a weak spillover contribution. Thus, the experiments follow the masking rule stated in Section 4.1; the aggregation formula in Appendix C.2 omitted this condition. We will correct the formula.

AuthorsOpenReview ↗

Follow-up on Our Rebuttal

Dear Reviewer M1wX,

Thank you for your time and effort in reviewing our paper. We sincerely appreciate your constructive feedback.

As the author–reviewer discussion period is approaching its end, we would greatly appreciate it if you could take a moment to review our rebuttal and let us know whether your concerns have been adequately addressed.

Thank you again for your time and consideration.

Best regards,
The Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to All Reviewers (Part I)

We thank all reviewers for their insightful and constructive feedback. We are encouraged by their recognition of PatientWorld's physiology-informed modeling design (M1wX, Jiq3, ksF5, 77xe), strong multi-step forecasting performance (M1wX, Jiq3, ksF5), and controlled experimental evaluation and downstream reuse (M1wX, ksF5, 77xe).

We are especially grateful to Reviewer Jiq3 for the constructive follow-up discussion and for positively updating their assessment in light of our additional evidence and clarifications.

We have updated the manuscript, with all revisions highlighted in blue.

Main Edits in the Manuscript

  1. Expanded Table 1 with GRU-D, STraTS-supervised, and STraTS-SSL-frozen, including their complete MIMIC-III/MIMIC-IV H3/H6 Conditional/Blind forecasting results.
  2. Added implementation details and references for the new clinical baselines in Appendix F.2.
  3. Corrected the spillover aggregation formula in Appendix C.2 to match Section 4.1 and the actual implementation.
  4. Updated the affected forecasting comparisons and rankings after incorporating the additional baselines.

Beyond these manuscript revisions, our rebuttal provides targeted architectural ablations, checkpoint-selection checks, factual treatment-response analyses, offline evaluation robustness tests, and generalization experiments. We believe these additional results and detailed clarifications have sufficiently addressed the major concerns raised by Reviewers M1wX and ksF5, although we have not received further feedback from them.

Below, we summarize our responses to the major common questions, with further experimental details provided in our individual reviewer responses.

Common Question 1: Do Physiological Structure and Task-Specific Adapters Provide Meaningful Contributions?

(Addressed to Reviewers M1wX, Jiq3, ksF5)

We conducted targeted ablations to isolate the contributions of physiology-informed state organization and the two task-specific adapters.

Physiological versus random grouping. We compared the physiological partition with three random partitions while preserving the number of groups and variables per group on MIMIC-IV L12H6.

Table R1: Physiological grouping ablation (RMSE ↓)

GroupingConditionalBlind
Physiological0.60460.6223
Random grouping (3 grouping seeds)0.6114 ± 0.00040.6349 ± 0.0006

The training seed was held fixed, and the random-grouping mean ± SD was computed across three grouping seeds. The physiological partition achieves lower forecasting errors under both settings. Together with the Base WM and Shuffled Treat controls in Figure 3, these results support the contribution of physiology-informed state factorization and treatment alignment.

Independent contributions of task-specific adapters. We separately enabled the relational and action-conditioned adapters for RRT factual forecasting.

Table R2: Task-specific adapter ablation (Global RMSE@H6 ↓)

ConfigurationRMSE
Neither0.6921
Relational only0.6813
Action only0.6908
Both0.6697

Each adapter improves performance over using neither, while their combination achieves the lowest error. These experiments provide direct evidence for the contributions of physiological grouping and the two complementary adaptation mechanisms, addressing the architectural concerns raised by M1wX and ksF5.

Common Question 2: What Drives Multi-Step Forecasting Performance?

(Addressed to Reviewers M1wX, Jiq3, ksF5)

PatientWorld's multi-step forecasting performance is primarily supported by physiology-factorized state transitions and physiology-aligned treatment pathways, as evidenced by the ablation results in Figure 3.

We further evaluated the local residual, horizon-specific correction, and auxiliary objectives on MIMIC-IV L12H6. Removing the local residual, sharing the horizon correction, or removing all four auxiliary losses produces only small changes in forecasting performance.

These findings clarify the roles of the different components. The local residual is designed as a bounded correction to mitigate accumulated rollout drift, rather than serving as the primary dynamics mechanism. Similarly, the four auxiliary losses provide additional physiological supervision and local-residual regularization. The main forecasting performance is largely retained without these complementary components, supporting the central contribution of the learned treatment-conditioned state transitions.

These additional analyses directly address the questions about the mechanisms underlying multi-step forecasting and the respective contributions of the local residual and auxiliary objectives.

AuthorsOpenReview ↗

General Response to All Reviewers (Part II)

Common Question 3: Are the Forecasting Comparisons Fair and Clinically Relevant?

(Addressed to Reviewers M1wX, 77xe)

We added GRU-D, STraTS-supervised, and STraTS-SSL-frozen, covering missingness-aware, supervised, and self-supervised pretrained clinical sequence modeling.

All three use the same patient-level splits, history windows, horizons, and metrics. PatientWorld achieves lower mean RMSE and MAE than all three across both datasets and four settings (updated Table 1).

All Stage-I baselines, including GRU-D and both STraTS variants, receive historical treatment information. Conditional and Blind differ in future-treatment availability, not access to treatment history. Thus, PatientWorld's Blind advantage cannot be attributed to exclusive access to past treatments.

We compared validation-loss and validation-RMSE checkpoint selection: both agree for eight baselines, while a Phys-JEPA difference does not change rankings. We also clarified selection procedures for MedDreamer and TDSTF.

These additional baselines and implementation checks strengthen the fairness and clinical relevance of the forecasting evaluation.

Common Question 4: Do the Learned Dynamics Capture Physiologically Meaningful Treatment Responses?

(Addressed to Reviewers ksF5, 77xe)

PatientWorld learns treatment-conditioned predictive dynamics from observed patient histories and recorded treatments. To evaluate whether these predictions agree with actual physiological changes, we conducted an additional factual RRT treatment-response analysis.

We selected held-out test windows with recorded RRT initiation and evaluated predictions of subsequent changes in BUN, creatinine, potassium, bicarbonate, and pH under the observed treatment schedules.

PatientWorld achieves a Conditional RMSE of 0.6360, compared with 0.6626 for Transformer and 0.7671 for MedDreamer. The patient-macro Spearman correlation between predicted and observed physiological changes is 0.5888.

The analysis also provides patient-clustered uncertainty estimates, with a 95% confidence interval of [0.0683, 0.1320] for PatientWorld's treatment-conditioning gain.

Together with the physiological structural controls in Figure 3 and Table 2, these results provide direct evidence that PatientWorld's treatment-conditioned predictions agree with observed physiological responses, beyond sensitivity to hypothetical actions.

Common Question 5: Are the Downstream Evaluations Reliable, and Can the Dynamics Generalize Beyond Their Training Settings?

(Addressed to Reviewers M1wX, Jiq3, ksF5, 77xe)

Offline policy evaluation robustness. We conducted additional patient-level uncertainty and evaluator-sensitivity analyses for RRT. FQE is our primary decision-evaluation metric, while WIS, DR, and WDR provide complementary estimates.

We performed a paired patient-clustered bootstrap using 2,955 patients, 3,711 episodes, and 2,000 patient-level resamples, holding the trained policies and OPE nuisance models fixed.

Table R3: Paired patient-clustered RRT OPE differences

EstimatorMedDreamer − PatientWorldPatient-clustered 95% CI
FQE−0.5751[−0.6656, −0.3476]
WIS+1.7134[+0.7007, +1.8945]
WDR−0.8810[−2.2500, −0.5033]

Higher HFD60 is better. FQE and WDR favor PatientWorld, whereas WIS favors MedDreamer. We further tested seven evaluator configurations varying FQE ridge regularization, propensity floors, and importance-weight clipping. The relative rankings remained unchanged within each estimator across these configurations.

We also clarified that all dynamics-based approaches share the same downstream action space, planner, and OPE protocol, enabling comparisons of different learned dynamics under a common planning pipeline.

Generalization beyond training horizons. We evaluated H3-trained checkpoints at H4–H6 without retraining. The models continue recursive forecasting beyond the trained horizon, with similar performance when the horizon embedding is clamped or removed. These results demonstrate that rollout is not strictly limited to its original training horizon.

Cross-dataset transfer. We additionally evaluated MIMIC-III ↔ MIMIC-IV zero-shot forecasting using source-trained H6 checkpoints without target-dataset parameter updates. These experiments demonstrate that the learned dynamics can generate predictions across datasets.

Temporal discretization. We clarified the use of 4-hour bins to balance temporal resolution and observation density in sparse ICU trajectories, together with observation masks and elapsed-time features for representing irregular measurements.

These complementary experiments and analyses strengthen the evidence for reliable downstream evaluation and reusable patient dynamics, directly addressing the robustness and generalization concerns raised during review.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.