acceptodds
Under review as a conference paper at ICLR 2027

TheraWorld: A Medical World Model in Hyperbolic Space for Treatment-Driven Disease Trajectory Simulation

Abstract

Modeling treatment-driven disease trajectories is crucial for clinical decision support. However, existing large language models typically rely on Euclidean representation spaces that poorly capture the hierarchical structure of medical concepts and the rapidly expanding state space of severe patients. We introduce TheraWorld, a medical world model that predicts future lab values and patient-state trends from historical records and observed treatment sequences in hyperbolic space. TheraWorld adapts pretrained LLMs with LoRA in hyperbolic space, and further introduces hyperbolic consistency and severity-aware radial regularization to preserve clinical structure and align representation geometry with patient severity. We construct a treatment-round-based EHR dataset and evaluate TheraWorld across multiple model scales. TheraWorld achieves the best average score of 72.10, outperforming medical LLMs and large models such as MedGemma-27B (67.30) and GLM-5.1 (69.06). Further analysis shows that the learned hyperbolic radius reflects patient severity, demonstrating the interpretability of hyperbolic geometry for LLM-based medical world modeling.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
LLM, EHR, world model, hyperbolic space

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_CPezOpenReview ↗
Rating
3(+2)
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposed a medical world model built in hyperbolic space to adapt pretrained LLMs to predict future lab values from historical records and treatment sequences.

Critical strengths

This work is well-motivated. It outperforms models such as MedGemma-27B with a best average score of 72.10.

Critical weaknesses

In this paper, the central justification for hyperbolic geometry remains unconvincing. Moreover, the empirical evidence is insufficient to establish robustness and generalizability.

Questions for authors

Comments

  • The authors claim that existing EHR-based world and trajectory models are built upon Transformer representations. However, it does not show that all existing large language models rely on Euclidean representation spaces. It is overly broad.
  • The motivation for replacing Euclidean representations entirely with hyperbolic geometry is insufficient. Please discuss their complementary strengths and compare the proposed approach with hybrid Euclidean–hyperbolic representations.
  • Figure 1 appears to motivate hyperbolic geometry primarily through rare diseases. Please investigate whether its benefits extend to common diseases and identify the conditions under which each representation space is preferable.
  • Please specify the exact MIMIC-IV version used, together with the cohort selection and preprocessing procedures, to ensure reproducibility.
  • Please clarify the clinical settings and qualify the contributions accordingly.
  • Clinical inputs contain heterogeneous modalities and data types. Please clarify how they are encoded and integrated, and whether a shared hyperbolic representation is appropriate for all of them.
  • The construction of the SOFA-based severity levels is insufficiently explained. Please provide the thresholds, their rationale, and evidence supporting their clinical validity.
  • Eq. (7) introduces at least four hyperparameters, raising concerns about dependence on extensive tuning. Please report the search spaces, selection criteria, and tuning budgets.
  • The sensitivity analysis in Appendix Table 11 is restricted to Qwen3-8B and laboratory prediction. Please extend it to additional models, tasks, and datasets to establish the robustness.
  • Table 2 provides limited coverage of strong contemporary baselines, particularly larger models. Please broaden the comparisons and justify the selection of baseline architectures and model sizes.
  • Please analyze the computational complexity, report training time, inference latency, and token usage relative to the baselines.
4 comments
AuthorsOpenReview ↗

Author Response to Reviewer CPez: Part1

We thank the reviewer for their careful reading and constructive feedback. However, we respectfully disagree with the assessment that our motivation for hyperbolic geometry is unconvincing and that our empirical evidence lacks robustness.

On the Motivation of Hyperbolic Geometry: Our work is grounded in the intrinsic tree-like and hierarchical nature of multi-scale disease progression and clinical state transitions in Electronic Health Records (EHR) (Section 1 & 3). As mathematically and empirically demonstrated, Euclidean spaces fail to faithfully embed exponential growth and hierarchical clinical trajectories without severe distortion, whereas hyperbolic representations naturally capture these properties (Section 3.2).

On Empirical Sufficiency and Robustness: Our evaluation is rigorous and comprehensive, encompassing 16 contemporary baselines (ranging from 4B to 744B parameters, including proprietary frontier models, general open-source LLMs, and specialized medical LLMs) across 11,299 multi-round trajectories (Table 2). Furthermore, extensive sensitivity analyses (Appendix G.3), cross-scale zero-retuning transfer (4B/8B/14B), and multi-task validation collectively establish the stability and generalizability of our approach.

Below, we provide detailed, point-by-point responses (Points 1–11) addressing all specific concerns, backed by our comprehensive ablation data and efficiency profiles.

C1. Scope of Euclidean Representations in EHR World Models

To clarify, our claim specifically targets existing EHR-based world and trajectory models (e.g., Mu et al., 2026; Adam et al., 2026) rather than all generic LLMs. Because these current EHR models are built upon standard Transformer architectures, their hidden representations and operations natively reside in flat Euclidean spaces (Rn\mathbb{R}^n) unless explicit non-Euclidean geometric constraints are introduced. Thus, their reliance on Euclidean representations holds within this domain. To prevent any over-generalization, we will tighten the wording in the Abstract and Section 1 to explicitly reflect this scope (e.g., "standard Transformer-based EHR models operate in Euclidean spaces").

C2. Hybrid Architecture and Euclidean Control

MethodRank/αsMAPE ↓S@10 ↑S@15 ↑S@25 ↑
TheraWorld16/3230.6457.4066.1376.28
Euclidean control16/3239.7743.9850.5157.90
Standard LoRA16/3234.2756.4465.0474.72
SFT16/3232.1156.8165.6275.57

First, we clarify that TheraWorld does not entirely discard Euclidean representations, but inherently operates as a hybrid Euclidean–hyperbolic architecture: the frozen LLM backbone preserves its strong general linguistic and linear reasoning capabilities in Euclidean space, while HypLoRA injects hyperbolic geometric inductive bias into the adapter pathway via exponential and logarithmic maps to capture hierarchical medical concepts and exponentially expanding severe states. Second, to empirically verify the necessity of hyperbolic geometry over purely Euclidean spaces under identical structural supervision, we conducted a Euclidean control experiment (above table) by applying the exact same consistency and radial regularization losses to standard Euclidean LoRA. As shown below, forcing radial and consistency constraints in flat Euclidean space severely degrades performance across all metrics (e.g., S@10 drops from 57.40% to 43.98%), because polynomial volume growth (RdR^d) induces severe crowding of high-acuity patient states. We will explicitly highlight this hybrid synergy and include a discussion comparing pure, hybrid, and fully hyperbolic designs in the revised manuscript.

C3. Motivation: State Expansion and Long-Tail Distribution

We thank the reviewer for the question, but there is a misunderstanding of our core motivation: our geometric design is not tied to "rare versus common diseases," but rather to the intrinsic structural properties of EHR data at the vocabulary level (long-tailed frequency distribution) and the patient-state level (exponentially expanding organ-dysfunction states). As empirically demonstrated in our pilot experiments (Section 3.2), hyperbolic space is preferable when dealing with specialized low-frequency medical concepts and complex, high-acuity patient trajectories that require large representational capacity to prevent state crowding. Conversely, Euclidean space remains suitable for routine, low-complexity clinical records with sparse concept combinations.

AuthorsOpenReview ↗

Author Response to Reviewer CPez: Part2

C4. Cohort Construction and Preprocessing

We clarify that our cohort is extracted from MIMIC-IV v3.1 (accessed via PhysioNet under credentialed access). Furthermore, the detailed cohort selection, round partitioning, and preprocessing procedures are already documented in Appendix A (Sections A.1–A.4) of the manuscript. Specifically: (1) source table extraction (labevents, emar, procedures_icd, microbiologyevents, patients, admissions, discharge) is detailed in Appendix A.1; (2) timeline integration, deduplication (~7.03M duplicate events removed), length-of-stay (< P90) and event-count (P10–P90) filtering, and treatment-round partitioning are described in Appendix A.2; and (3) SOFA-lab3 baseline filtering (retaining 50,680 candidate rounds) and subject-level disjoint splitting (10,001 train / 302 val / 996 test; 11,299 total) are provided in Appendix A.4. In the revised manuscript, we will explicitly specify MIMIC-IV v3.1 and add a cross-reference to Appendix A in Section 3.1, and we will release the complete preprocessing scripts upon acceptance.

C5. Clinical Setting and Scope Boundaries

We thank the reviewer for the suggestion, but we would like to respectfully clarify that the clinical settings and boundaries of our work are already explicitly defined throughout the manuscript. Specifically, Section 3.3 (Task Definition) formally defines the scope as treatment-driven patient-state and laboratory prediction under EHR observations, and Section 1 clearly bounds the setting to ICU patient trajectories. Furthermore, our Ethics Statement explicitly qualifies that TheraWorld's outputs are probabilistic approximations of ICU trajectories from a single institution, and they should not be interpreted as causal intervention estimates, ground-truth forecasts, or direct clinical treatment recommendations.

C6. Structured Serialization and Token-Level Integration

As detailed in Appendix A.3 and Appendix H (Tables 15–16), all heterogeneous clinical inputs (including laboratory values, microbiology results, medications, procedures, and timestamps) are serialized into a unified structured JSON-like text schema, encoded as token sequences via the LLM tokenizer, and integrated through standard Transformer attention. Crucially, there are no separate modality encoders or modality-specific hyperbolic projections: our HypLoRA adapters operate locally on individual token-level activation vectors within the attention and feed-forward projection layers, leaving the backbone's cross-modal token interactions intact. Furthermore, this heterogeneity is precisely why a shared hyperbolic space is advantageous—its negative curvature and exponential capacity naturally accommodate multi-granular, hierarchical, and heterogeneous clinical concepts within a single unified manifold without geometric distortion. We will add a brief summary of this serialization and integration mechanism in Section 4.1 for greater clarity.

AuthorsOpenReview ↗

Author Response to Reviewer CPez: Part3

C7. Clinical Validation of SOFA-lab3

We thank the reviewer for this comment. As detailed in Appendix A.4 and Table 6, our severity levels are directly constructed from the standard Sequential Organ Failure Assessment (SOFA) criteria using three objectively measured laboratory-based organ components, each scored from 00 to 44 (aggregate score 0-120\text{-}12). The rationale for selecting these three laboratory-grounded components is to avoid subjective clinical charting noise (e.g., Glasgow Coma Scale) and missing bedside ventilator/vasopressor charting while maintaining high temporal precision across ICU treatment rounds. Clinically, SOFA is the gold-standard ICU severity metric endorsed by the Sepsis-3 consensus guidelines[1] and extensively validated as a robust predictor of ICU morbidity and mortality[2,3].

To further validate the clinical fidelity of our SOFA-lab3 construction and predictions, we provide two additional quantitative analyses across the test set:

  1. Clinical Fidelity of Predicted SOFA-lab3 (996 test samples): From predicted next-round laboratory values, we derive SOFA-lab3 and compare it against ground truth. Both models achieve ρ>0.91\rho > 0.91, substantially outperforming the previous-round-as-prediction baseline (ρ≈0.84\rho \approx 0.84). Crucially, when the predicted SOFA-lab3 changes from the previous round, TheraWorld is accurate 67.3% of the time (vs. 17.3% for the previous-round baseline), confirming genuine clinical trajectory prediction rather than historical copying.
MetricQwen3-8BQwen3-14B
SOFA exact match80.6%79.9%
SOFA within ±199.2%99.0%
Spearman ρ (predicted vs. actual SOFA)0.9150.910
Spearman ρ (previous round as baseline)0.8430.839
  1. Monotonic Geometric Alignment Across Scales: Across all three model scales, the learned hyperbolic radius increases strictly monotonically with clinical severity:
SOFA groupQwen3-4B radiusQwen3-8B radiusQwen3-14B radius
00.8660.8800.823
11.0841.0360.993
21.1941.1391.114
≥31.2911.3301.304

We will integrate these clinical validation references and supplementary quantitative results into Section 3.1 and Appendix G.1 of the revised manuscript.

C8. Hyperparameter Sensitivity and Search Budget

We would like to point out that a comprehensive sensitivity analysis across all four hyperparameters in Eq. (7) is already provided in Appendix G.3 and Table 11 (spanning λS∈{0.00,0.01,0.05,0.10}\lambda_S \in \{0.00, 0.01, 0.05, 0.10\}, λP∈{0,0.002,0.004}\lambda_P \in \{0, 0.002, 0.004\}, λB∈{0,0.0001,0.0005}\lambda_B \in \{0, 0.0001, 0.0005\}, and λT∈{0,0.01,0.02,0.04}\lambda_T \in \{0, 0.01, 0.02, 0.04\}, alongside LoRA rank r∈{8,16,32,64}r \in \{8, 16, 32, 64\}).

As demonstrated in both Table 11, model performance remains remarkably stable across the entire grid, varying any single radial weight (λP,λB,λT\lambda_P, \lambda_B, \lambda_T). This confirms that these regularization terms shape the latent geometry smoothly without requiring extensive hyperparameter tuning.

[1]Singer, M., Deutschman, C. S., Seymour, C. W., et al. (2016). The Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3). JAMA, 315(8), 801–810.

[2]Ferreira, F. L., Bota, D. P., Bross, A., Mélot, C., & Vincent, J. L. (2001). Serial evaluation of the SOFA score to predict outcome in critically ill patients. JAMA, 286(14), 1754–1758.

[3]Raith, E. P., Udy, A. A., Bailey, M., et al. (2017). Prognostic accuracy of the SOFA score, SIRS criteria, and qSOFA score for in-hospital mortality among adults with suspected infection admitted to the intensive care unit. JAMA, 317(3), 290–300.

AuthorsOpenReview ↗

Author Response to Reviewer CPez: Part4

C9. Cross-Scale Generalization and Robustness

We thank the reviewer for this suggestion. While we appreciate the goal of establishing broader robustness, we respectfully argue that exhaustive hyperparameter grid searches across additional backbones, tasks, and datasets are unnecessary to support our core conclusions for two reasons:

Zero-Retuning Cross-Scale Generalization: Rather than tuning hyperparameters separately for each model scale, we applied the exact default configuration selected on Qwen3-8B directly to both Qwen3-4B and Qwen3-14B without any re-tuning. As shown in Table 2, Table 3, and Figure 6, TheraWorld consistently outperforms Euclidean baselines and achieves monotonic severity-radius alignment across all three scales (p<10−17p < 10^{-17}), directly confirming that the hyperparameters transfer seamlessly across model sizes.

Rich Clinical Diversity and Computational Practicality: Derived from MIMIC-IV, our dataset already spans 387 distinct lab labels, 1,212 medications, 53 organisms, and 11,299 multi-round trajectories across diverse ICU populations. Given that our sensitivity analyses (Table 11) already demonstrate minimal variance across parameter ranges, repeating full multi-parameter sweeps across additional backbones and external datasets would incur prohibitive GPU costs (34–62 GPU-hours per run) without altering our central finding that hyperbolic geometry drives the observed gains.

C10. Baseline Coverage and Selection Rationale

We thank the reviewer for this comment, though we would like to respectfully clarify that Table 2 already provides extensive coverage of 16 contemporary baselines across three distinct categories and a wide parameter spectrum (4B to 744B): (1) proprietary frontier models (GPT-5-Mini and Claude-Haiku-4.5); (2) 9 open-source general-purpose models spanning small-to-massive scales, including Qwen3 (4B, 8B, 14B, 235B-A22B), Qwen3.5-122B-A10B, MiniMax-M2.7 (229B), DeepSeek-V4-Flash (284B), Llama-4-Maverick (400B), and GLM-5.1 (744B); and (3) 5 specialized open-source medical LLMs (MedGemma-4B-IT, MedGemma-27B-IT, HuatuoGPT-01-8B, FineMedLM-01, and Baichuan-M2-32B). Notably, our largest evaluated baselines (GLM-5.1 at 744B and Llama-4-Maverick at 400B) are over 28× to 53× larger than our largest model (TheraWorld-14B), yet TheraWorld-14B still outperforms all of them (72.10 vs. 69.06 and 69.15 average score). Our selection rationale was specifically designed to: (i) ensure strict same-backbone fairness across 4B/8B/14B scales (against Base, LoRA, and SFT in Table 3), (ii) isolate domain pretraining effects via specialized medical LLMs, and (iii) test whether geometric inductive bias can surpass massive general-purpose scaling.

C11. Computational Complexity and Training Efficiency

We thank the reviewer for this suggestion. We note that our distributed training setup and baseline computational budgets were initially outlined in Appendix F.2. To provide a complete complexity and efficiency breakdown relative to baselines, we report the exact parameter counts, training wall-clock time, inference latency, and token usage below:

MethodTrainable paramsTotal paramsTraining wall-clock
LoRA-8B-r1683M8.2B~18h
SFT-8B8.2B8.2B~34h
TheraWorld-8B-r1685M8.2B~20h

Complexity & Training Overhead: HypLoRA introduces only lightweight rank-rr adapter updates plus exponential/logarithmic map operations of O(r⋅d)\mathcal{O}(r \cdot d) complexity per adapted layer, along with a 2M-parameter state projection head. The auxiliary self-supervised state consistency loss (LS\mathcal{L}_S, Section 4.3) requires a teacher forward pass, resulting in only a 7% per-step overhead (11% total training overhead over standard LoRA) while remaining substantially more efficient than full SFT.

Zero Inference Overhead: At inference time, the teacher branch and auxiliary regularization heads are completely discarded (Section 4.3), and HypLoRA adapters are merged directly into the frozen backbone weights. Consequently, TheraWorld incurs zero additional inference latency or memory overhead compared to standard LoRA or SFT, averaging ~13 seconds per sample with input ~2,837 tokens and output ~571 tokens on a single A100-80GB.

Summary

We once again thank the reviewer for their time, careful reading, and constructive engagement with our work. We kindly invite the reviewer to review the specific architectural designs, ablation results, and extensive baseline comparisons detailed throughout our manuscript and this response. Our work represents a rigorous, principled effort to introduce non-Euclidean geometric inductive bias into clinical EHR modeling, backed by solid theoretical motivation, extensive empirical evaluations across 16 state-of-the-art baselines, and thorough clinical validation. We hope this comprehensive evidence addresses all concerns and demonstrates the substantial contributions of TheraWorld.

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

The paper proposes TheraWorld, a way to adapt pretrained LLMs for predicting a patient’s next-round lab values after an observed treatment sequence. It argues that Euclidean representations fit clinical data poorly for two reasons: medical vocabulary is long-tailed, and severe patients occupy a much more diverse state space. The authors build a MIMIC-IV dataset split into treatment rounds, with a SOFA-subset severity label from platelets, bilirubin, and creatinine.

Critical strengths
  • A well-motivated task. The treatment-round MIMIC-IV construction targets intervention-conditioned prediction rather than fixed time windows.

  • Table 2 covers 18 baselines across closed, open general, and medical LLMs. Table 3 compares with LoRA and SFT on the same backbones and shows TheraWorld ahead of both at 8B and 14B on every reported metric.

  • The paper states that it does not estimate causal effects, the radius is a learned proxy rather than a validated score, and that results come from a single institution. This is consequential for avoiding misuse of the world model label.

Critical weaknesses
  • Table 4 shows HypLoRA alone at 8B (S@10 55.07, S@25 72.94, sMAPE 39.07). Table 3 shows plain Euclidean LoRA on the same backbone at 56.44, 74.72, and 34.27. So HypLoRA without the auxiliary losses is worse than Euclidean LoRA on all four metrics.

  • Hyperparameter tuning for LoRA/SFT is not described beyond matched rank and scale. LoRA at 14B is worse than at 8B, which hints at under-tuned baselines.

  • The text attributes TheraWorld’s advantage to targeted hyperbolic adaptation, but the fine-tuning effect and the geometry effect are confounded there.

Questions for authors
  • HypLoRA alone is below plain LoRA at 8B on all four metrics. How do you reconcile this with the claim that hyperbolic adaptation is the driver?

  • What learning rates and other hyperparameters were searched for LoRA and SFT, on what validation data, and how large was the search?

  • How many test patients and samples fall in each SOFA group in Fig. 4, and on which backbone?

2 comments
AuthorsOpenReview ↗

Author Response to Reviewer diCy: Part 1

We thank the reviewer for the focused and precise critique. We address each question directly below with new experiments that resolve the central concern.

W1&Q1. HypLoRA and Geometry Attribution

We thank the reviewer for this critical question. The lower performance of HypLoRA-only (without auxiliary losses) does not invalidate our claim; rather, it reveals a fundamental synergy between hyperbolic geometry and structured supervision:

  1. Geometry and Supervision are Jointly Necessary. HypLoRA alone lacks task-specific geometric alignment, while standard Euclidean space cannot accommodate hierarchical severity objectives.

  2. The Decisive Euclidean Control Experiment. To isolate the geometric contribution, we trained a standard Euclidean LoRA with the exact same auxiliary losses (state consistency, pairwise ranking, and boundary regularization). As shown below, forcing these objectives in flat Euclidean space actively degrades performance (sMAPE drops from 34.23 to 39.77), because polynomial volume growth (RdR^d) causes severe state crowding.

MethodRank/αsMAPE ↓S@10 ↑S@15 ↑S@25 ↑
TheraWorld16/3230.6457.4066.1376.28
Euclidean control16/3239.7743.9850.5157.90
Standard LoRA16/3234.2756.4465.0474.72
SFT16/3232.1156.8165.6275.57
  1. Reconciling the Results.
    • HypLoRA-only < plain LoRA: adapters alone need supervision.
    • Euclidean + auxiliary losses < plain LoRA: auxiliary losses harm Euclidean spaces.
    • TheraWorld (HypLoRA + auxiliary losses) ≫\gg all baselines: hyperbolic geometry provides the structural substrate that makes the auxiliary losses beneficial rather than harmful.

We will revise Section 6.3 to explicitly clarify this joint design.

W2&Q2. Baseline Tuning and Scaling Behavior

First, the observation that full SFT or larger backbones do not always outperform 8B LoRA is likely an inherent characteristic of structured EHR numeric forecasting under ~10k trajectories, where full-parameter tuning is not inherently superior for this task and tends to overfit on fine-grained numeric outputs. Second, we clarify that we did not perform hyperparameter tuning or search alternative configurations for either the LoRA or SFT baselines, instead applying standard default settings (r=16,α=32r=16, \alpha=32 for LoRA) on the 10,001-sample training set (validated on the 302-sample validation split). Third, to verify whether the LoRA baseline was limited by its learning rate, we conducted a supplementary sweep over LR∈{1×10−4,2×10−4,5×10−4}\text{LR} \in \{1\times 10^{-4}, 2\times 10^{-4}, 5\times 10^{-4}\} on Qwen3-8B under the identical training recipe

LRVal lossS@10 ↑S@15 ↑S@25 ↑sMAPE ↓
1e-4 (paper config)0.116756.4465.0474.7234.27
2e-40.112352.0060.0468.6548.42
5e-40.109852.6060.7069.4646.96

As shown above, our default learning rate (1×10−41\times 10^{-4}) already achieves the best downstream performance for LoRA across all metrics, whereas higher learning rates lower validation token loss at the expense of severe numerical overfitting. Finally, we did not search additional hyperparameters for the baselines because doing so is unlikely to alter our main experimental conclusions while incurring substantial computational overhead. We will explicitly document these baseline configurations in Appendix F.

AuthorsOpenReview ↗

Author Response to Reviewer diCy: Part 2

W3. Geometric vs. Fine-Tuning Confound

We thank the reviewer for raising this important methodological point. We decouple the fine-tuning effect from the hyperbolic geometry effect through a two-stage controlled comparison on the exact same training data and parameter budget.

First, Both Standard LoRA and full SFT are fine-tuned on the exact same MIMIC-IV training split as TheraWorld. While domain fine-tuning improves over the zero-shot base model, TheraWorld consistently outperforms both Euclidean fine-tuning baselines across all scales (Table 3), ruling out simple domain adaptation as the sole driver.

Second, To ensure that the performance gain is not merely confounded by the auxiliary losses (LS,LP,LB,LT\mathcal{L}_S, \mathcal{L}_P, \mathcal{L}_B, \mathcal{L}_T), our Euclidean control holds everything constant, varying only the underlying manifold (Euclidean LoRA vs. hyperbolic HypLoRA). As shown in our ablation table above, replacing hyperbolic geometry with Euclidean geometry under identical supervision causes sMAPE to degrade from 30.64 to 39.77 and S@10 to drop from 57.40% to 43.98%.

This single-variable contrast cleanly isolates the geometric effect, confirming that hyperbolic representation space is indispensable for realizing the gains of TheraWorld. We will clarify this disentanglement explicitly in Section 6.3.

Q3. Figure 4 Statistics and Backbone Specification

Figure 4 is evaluated on the Qwen3-14B backbone across the 996-sample test set (comprising 967 parseable rounds with valid baseline SOFA-lab3 scores). The exact patient and round counts across each severity group are summarized below:

SOFA-lab3 groupPatientsRounds
0 (no organ dysfunction)244474
1 (mild)117201
2 (moderate)78122
≥3 (severe)72170

We will explicitly add the backbone specification and these exact patient/sample counts to the caption of Figure 4 in the revised manuscript.

Reviewer_mTd8OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This article introduced TheraWorld, and the authors considered it a medical world model that predicts future lab values and patient-state trends from historical records and observed treatment sequences in hyperbolic space. TheraWorld adapts pretrained LLMs with LoRA in hyperbolic space. This article further introduces hyperbolic consistency and severity-aware radial regularization. Authors claimed that these could preserve clinical structure and align representation geometry with patient severity.

Critical strengths
  1. This article is well-written and easy to follow.
  2. The theoretical and empirical justification is relatively clear and convincing, especially in section 3.2.
  3. The article provides helpful details in the appendix, which largely facilitated understanding and can be used to guide reproduction.
  4. Comprehensive testing across multiple models.
Critical weaknesses
  1. Tuning an LLM using LoRA based on MIMIC-IV is hard to define as building a world model. MMIC-IV is a limited-scope dataset focusing on EHR records in ER and IC, recorded in one healthcare institute. Saying that a "world model" based on this dataset is overclaiming. This is good work itself that does not need to be over-decorated using a popular buzzword like the "world model".
  2. The dataset construction should not exclude data points beyond 10% - 90%, especially those beyond 90% event counts. The longer the event sequence, the better the sample is for conducting a "stress test" on TheraWorld.
  3. SOFA is a limited method to evaluate severity. The data labeling/ranking lacks medical experts' guidance.
  4. It was unknown what tools were used to process the MIMIC-IV dataset and convert it to the training/validation/test dataset for TheraWorld.
  5. The results are technical without sufficient medical interpretation. For instance, Creatinine level of 4.5mg/dL vs. 4.7mg/dL is numerically different, but clinically in the same range of "significantly elevated". Thus, how does the prediction accuracy difference link to medical interpretation difference was not well discussed.
  6. The results reporting often misses units for medical measurements. Please ask a clinical individual to help address the problem.
Questions for authors
  1. Could you please provide a trajectory example beyond a single event.
5 comments
AuthorsOpenReview ↗

Author Response to Reviewer mTd8: Part 1

We thank the reviewer for the careful reading and constructive feedback. We address each concern below with new experiments and clarifications.

W1. Terminology Scope and Clinical World Model Definition

We thank the reviewer for raising this thoughtful point and for the encouraging assessment of our technical work. First, following the functional definition in recent medical world-model literature [1], world models are "internal models that learn how an environment evolves and how actions change future states." TheraWorld directly fulfills this definition: our hyperbolic representation serves as the structured internal states, while our treatment-conditioned forecasting and counterfactual evaluations (Table 12) model clinical dynamics and intervention-driven state transitions. Second, framing domain-specific simulators as medical world models follows established field conventions, such as EHRWorld[2] (single-institution inpatient EHR), MeWM[3] (post-TACE liver tumor cohorts), and CLARITY[4] (targeted glioma/breast cancer cohorts). Finally, we agree on the importance of avoiding over-claims and we will pursue multi-center, multi-task generalization in future work.

W2. Long-Sequence Stress Testing Beyond the 90th Percentile

We thank the reviewer for this insightful suggestion. We agree that long sequences provide a valuable "stress test" for evaluating trajectory modeling capacity.

Clarification on Original Filtering: In our initial pipeline (Appendix A.2), records with event counts beyond the 90th percentile (>376 events) were filtered solely to ensure stable context-window management and training batch efficiency during fine-tuning, rather than due to data quality concerns.

New Stress-Test Evaluation: To directly address the reviewer's concern, we completed the requested stress test by constructing a rigorous, patient-isolated test subset of 500 admissions with event counts strictly above P90 (>376 events), drawn entirely from subjects excluded from the training and validation splits. This subset represents extreme clinical complexity, with a median input length of ~2,316 tokens (P90: 4,150 tokens; max: 6,085 tokens).

MethodS@10 ↑S@15 ↑S@25 ↑sMAPE ↓Stat-F1 ↑
TheraWorld 8B60.7369.4879.9722.7587.84
LoRA 8B52.2159.0568.3548.9681.73

As shown in the table above, while standard LoRA suffers a severe performance drop on these ultra-long, high-acuity trajectories (e.g., sMAPE jumping to 48.96), TheraWorld maintains remarkable stability and superior performance across all metrics (e.g., S@10 of 60.73 and sMAPE of 22.75).

W3. Clinical Validity of SOFA Metrics and Expert Evaluation Integration

We thank the reviewer for this comment and completely agree on the importance of clinical grounding. First, SOFA is not an arbitrary metric but the gold-standard ICU severity framework endorsed by Sepsis-3[5] and extensively validated as a robust predictor of mortality[6,7]. By utilizing objective laboratory-based organ components (platelets, bilirubin, and creatinine), we avoid subjective charting noise while ensuring high temporal fidelity. Second, acknowledging that proxy metrics cannot entirely replace expert judgment, we are actively expanding our evaluation protocols with licensed physician annotations across multiple clinical departments.

[1] Liu, K., et al. (2026). Medical world models: representing medical states, modelling clinical dynamics and guiding intervention policies. arXiv:2606.16721.

[2] Mu, L., et al. (2026). EHRWorld: A Patient-Centric Medical World Model for Long-Horizon Clinical Trajectories. arXiv:2602.03569.

[3] Yang, Y., et al. (2025). Medical World Model: Generative Simulation of Tumor Evolution for Treatment Planning. arXiv:2506.02327.

[4] Ding, T., et al. (2026). CLARITY: Medical World Model for Guiding Treatment Decisions by Simulating Context-Aware Disease Trajectories. ECCV 2026 / arXiv:2512.08029.

[5] Singer, M., Deutschman, C. S., Seymour, C. W., et al. (2016). The Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3). JAMA, 315(8), 801–810.

[6] Ferreira, F. L., Bota, D. P., Bross, A., Mélot, C., & Vincent, J. L. (2001). Serial evaluation of the SOFA score to predict outcome in critically ill patients. JAMA, 286(14), 1754–1758.

[7] Raith, E. P., Udy, A. A., Bailey, M., et al. (2017). Prognostic accuracy of the SOFA score, SIRS criteria, and qSOFA score for in-hospital mortality among adults with suspected infection admitted to the intensive care unit. JAMA, 317(3), 290–300.

AuthorsOpenReview ↗

Author Response to Reviewer mTd8: Part 2

W4. Data processing tools and MIMIC-IV version

We thank the reviewer for pointing this out. While our preprocessing pipeline, round partitioning, and patient-level splitting criteria are already detailed in Appendix A (Sections A.1–A.4), we did not explicitly specify the software stack and dataset version.

Specifically, the MIMIC-IV (v3.1) dataset (accessed via PhysioNet) was processed using Python and pandas for table merging, deduplication, and time alignment. Following the workflow in Appendix A, data proceeds from source table extraction and event filtering (App. A.1–A.2) to treatment-round partitioning (App. A.2) and patient-level disjoint splitting into 10,001 train, 302 validation, and 996 test samples (App. A.4). We will explicitly add the MIMIC-IV version, software stack details, and pipeline specifications to Appendix A, and release the complete preprocessing scripts upon acceptance.

W5. Numerical Precision vs. Clinical Relevance

We thank the reviewer for this insightful comment and fully agree that minor numerical deviations (e.g., 4.54.5 vs. 4.7 mg/dL4.7\text{ mg/dL} for creatinine) often fall within the same clinical category without altering medical interpretation. In our evaluation framework, we explicitly decouple technical regression precision from clinical significance: while sMAPE and S@10/15/25 strictly benchmark fine-grained numerical modeling capability, Stat-F1 directly captures clinical interpretation by mapping continuous predictions into clinically established normal versus abnormal reference ranges—treating minor variations within the same tier identically while penalizing errors that cross critical clinical thresholds. We will add a dedicated discussion in Section 5 to clarify this distinction and better link our Stat-F1 and SOFA-lab3 improvements to practical clinical decision-making.

W6. Missing units in results reporting

We thank the reviewer for catching this important detail and fully agree that standardized measurement units are essential for clinical clarity and accuracy.

To clarify the current manuscript: standard clinical units are already included in our SOFA-lab3 severity criteria (Appendix A.4 and Table 6) as well as within our structured input/output serialization schemas (Appendix A.3 and Appendix H, Tables 15–16), which preserve the original MIMIC-IV unit fields (valueuom) for every laboratory measurement. However, we acknowledge that units were inadvertently omitted when reporting specific numeric examples and case studies in the main text (Section 5) and parts of Appendix G.

Following the reviewer's recommendation, we will consult with our collaborating clinicians to perform a thorough, manuscript-wide audit, ensuring that standardized medical units are explicitly and consistently attached to every laboratory value and clinical measurement across all text, tables, and figures in the revised version.

Q1. Trajectory example beyond a single event

We thank the reviewer for this suggestion. To illustrate TheraWorld's multi-round trajectory forecasting capability beyond single-event predictions, we present a concrete longitudinal case study from our test set (Subject 13973768) spanning three consecutive clinical horizons (+6h, +24h, and +48h), complete with standardized clinical units:

Lab (unit)+6h Pred+6h True+24h Pred+24h True+48h Pred+48h True
Glucose (mg/dL)127124127119105108
Potassium (mEq/L)4.64.54.54.44.34.2
Sodium (mEq/L)137136137138137137
Creatinine (mg/dL)1.61.51.61.61.41.3
Urea Nitrogen (mg/dL)686568646158

We will include the full table with all matched labs and units in the revision.

Thanks for answering my questions. The > 90% results are good to be added. The SOFA is for tracking organ dysfunction and sepsis severity, which does not fit all the cases/goals of the dataset.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to Reviewer CPez

We sincerely thank the reviewer for their constructive feedback and for acknowledging our clarifications by raising their score. We also greatly appreciate the helpful suggestion regarding the organization of the paper.

As suggested, we have now updated the main text, specifically adding the core cohort construction and preprocessing details into Section 5.1 (Dataset Construction Details), to make the core methodology and data pipeline clearer.

We will ensure these additions are fully integrated into the camera-ready version.

Thank you once again for your valuable time and expertise in helping us improve our manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.