Paper deep dive
PatientWM: Towards Generalizable Clinical World Models via Unified Patient States
Ziwen Kan, Jie Peng, Bo Yan, Davis Hobley, Zheng Zhang, Yu Tian, Tianlong Chen, Song Wang
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Clinical decisions depend on anticipating how a patient's condition will evolve as treatment is adjusted and new evidence becomes available. Existing medical world models capture patient-state evolution, but often define the evolving state around a particular modality or task, limiting reuse as inputs or prediction targets change. These models also retain history through observation sequences or recurrent or latent summaries rather than exposing preceding patient states and clinical actions as an ordered transition context. This distinction matters because similar current measurements may follow different recent changes, treatment exposures, and observed responses. To tackle these limitations, we propose PatientWM, a clinical world-model framework built around a unified patient state and sequential latent history. Availability-aware encoders integrate observations available at each decision time into the patient state, while typed decoders produce task-specific clinical readouts without defining the state around any single target. PatientWM further conditions each next-state prediction on the current state and supplied action together with a bounded, ordered history of preceding latent patient states and clinical actions, preserving recent treatment and response context. Experiments across MIMIC-III, MIMIC-IV, and three external clinical datasets show superior performance in clinical prediction, multi-step forecasting, and action-planning evaluation. Transfer experiments further demonstrate the generalizability of PatientWM's patient state. Together, these results establish its ability to forecast clinical evolution and support planning.
Tags
Links
- Source: https://arxiv.org/abs/2610.17059v1
- Canonical: https://arxiv.org/abs/2610.17059v1
Trouble viewing inline? Open PDF directly →
Full Text
96,161 characters extracted from source content.
Expand or collapse full text
PatientWM: Towards Generalizable Clinical World Models via Unified Patient States Ziwen Kan Jie Peng Affiliation: University of Science and Technology of China Bo Yan Affiliation: University of Central Florida Davis Hobley Affiliation: Princeton University Zheng Zhang Affiliation: Nanyang Technological University Yu Tian Affiliation: University of Central Florida Tianlong Chen Affiliation: University of North Carolina at Chapel Hill Song Wang Affiliation: University of Central Florida Abstract Clinical decisions depend on anticipating how a patient’s condition will evolve as treatment is adjusted and new evidence becomes available. Existing medical world models capture patient-state evolution, but often define the evolving state around a particular modality or task, limiting reuse as inputs or prediction targets change. These models also retain history through observation sequences or recurrent or latent summaries rather than exposing preceding patient states and clinical actions as an ordered transition context. This distinction matters because similar current measurements may follow different recent changes, treatment exposures, and observed responses. To tackle these limitations, we propose PatientWM, a clinical world-model framework built around a unified patient state and sequential latent history. Availability-aware encoders integrate observations available at each decision time into the patient state, while typed decoders produce task-specific clinical readouts without defining the state around any single target. PatientWM further conditions each next-state prediction on the current state and supplied action together with a bounded, ordered history of preceding latent patient states and clinical actions, preserving recent treatment and response context. Experiments across MIMIC-I, MIMIC-IV, and three external clinical datasets show superior performance in clinical prediction, multi-step forecasting, and action-planning evaluation. Transfer experiments further demonstrate the generalizability of PatientWM’s patient state. Together, these results establish its ability to forecast clinical evolution and support planning. 1 Introduction Clinical treatment often unfolds as a sequential decision process, in which clinicians repeatedly select or adjust treatments based on a patient’s current health status, prior treatment history, and observed responses over time (Yu et al., 2021). Each decision depends not only on the observations available at that moment, but also on how the patient’s condition has evolved under preceding treatments (Shickel et al., 2017). Conventional clinical AI systems commonly map longitudinal observations directly to predefined, task-specific outcomes, such as diagnoses or risk estimates (Choi et al., 2016; Rajkomar et al., 2018). As illustrated in Figure 1, such outputs answer a particular prediction question, but do not by themselves define a shared patient state that can be updated after treatment and carried forward to model subsequent clinical evolution. Figure 1: Conceptual Overview of PatientWM. Conventional clinical prediction focuses on learning task-specific predictions from heterogeneous observations. PatientWM instead learns a unified patient state from available observations and models how this state evolves under clinical actions and recent history. The learned state supports multiple clinical readouts and multi-step forecasting. Modeling this evolution requires representing the patient’s condition at the current decision time and predicting how that state changes as treatments are applied. World models are well suited to this setting because they represent a system’s current state and model its action-conditioned dynamics to predict future states. These models have been applied in autonomous driving, robotics, video generation, and social simulation to forecast future states and support planning or evaluation (Ha and Schmidhuber, 2018; Ding et al., 2025). In clinical settings, recent studies have begun to utilize this framework to model patient trajectories for specialized clinical modalities (Liu et al., 2026; Chen et al., 2026; Mu et al., 2026; Xu et al., 2026; Ding et al., 2026; Wang et al., 2026). Despite this progress, two challenges remain for the formulation considered here. (1) Limited-Scope State: Existing medical world models typically construct patient states for a fixed set of clinical inputs and prediction tasks. This makes the learned state harder to reuse when additional modalities or new prediction targets are introduced (Si et al., 2021; Hansen et al., 2024; Dou et al., 2025; Liu et al., 2026). (2) Implicit History: Existing methods generally retain history through raw observation sequences or compress it into recurrent or latent context, without making the ordered sequence of preceding patient states and clinical actions separately available to each transition (Wu et al., 2026; Xu et al., 2026; Ding et al., 2026; Wang et al., 2026). Consequently, it can be challenging for the transition model to learn how similar current states reached through different treatment-response trajectories may evolve differently. To address these limitations, we propose PatientWM, a clinical world model built around a unified patient state that represents the patient’s current condition, supports prediction of its future evolution, and serves as a shared basis for downstream clinical predictions. At each decision time, modality-specific encoders represent the available clinical observations, which are fused into the unified patient state, while typed decoders produce task-specific outputs on demand. PatientWM further represents recent history as sequential latent history: an ordered sequence of latent patient-state and clinical-action pairs. Each next-state prediction then conditions on this sequence together with the current state and supplied action. This explicit trajectory context helps the transition model distinguish similar current states reached through different treatment-response histories, which may imply different future clinical evolution. As illustrated in Figure 1, PatientWM models patient evolution directly in the unified state space, with sequential latent history and the supplied action conditioning each predicted state transition. Our main contributions are summarized below. • We propose PatientWM, a clinical world-model framework built around a unified patient state. It learns this state from available clinical observations, models patient evolution in the shared state space, and uses typed decoders to produce task-specific outputs on demand. • We represent patient history as an ordered sequence of latent patient-state and clinical-action pairs. Each transition conditions on this sequence together with the current state and supplied action, making trajectory context an explicit part of patient-state evolution. • Experiments across MIMIC-I, MIMIC-IV, and three external clinical datasets show superior performance in clinical prediction, multi-step forecasting, and action-planning evaluation. Transfer experiments further demonstrate the generalizability of PatientWM’s patient state. These results demonstrate its ability to forecast clinical evolution and support planning. 2 Related Work Clinical Prediction. Clinical prediction models use longitudinal EHRs and physiological signals (Shickel et al., 2017; Harutyunyan et al., 2019; Faust et al., 2018) to predict diagnoses, deterioration, clinical events, and outcomes across diverse clinical settings (Choi et al., 2016; Tomašev et al., 2019; Rajkomar et al., 2018; Steinberg et al., 2024; Hannun et al., 2019; Attia et al., 2019). These systems capture temporal and physiological patterns, but are typically organized around predefined diagnostic or prognostic targets. Such outputs do not by themselves represent how a patient’s condition evolves between observations and treatment updates, which is central to sequential clinical decision support (Liu et al., 2026; Chen et al., 2026). World Models. World models include latent state-space models such as World Models and Dreamer for imagined control (Ha and Schmidhuber, 2018; Hafner et al., 2019; Hafner et al., 2025), extensions such as DayDreamer for physical robot learning (Wu et al., 2023), and generative models such as Drive-WM and Genie for controllable environments (Wang et al., 2024; Bruce et al., 2024). Across these settings, the common principle is to connect an internal representation of the environment with action-conditioned future prediction, allowing possible trajectories to be simulated for planning, policy learning, or evaluation (Ding et al., 2025). This principle has recently been adapted to medicine by modeling patient states and their evolution under clinical interventions (Liu et al., 2026). Medical World Models. Medical world models and related trajectory models differ in what they treat as the evolving patient state. Some models define the evolving state directly in observable clinical space. EHRWorld and EHR2Path (Mu et al., 2026; Pellegrini et al., 2025) operate in record or pathway space, generating subsequent chart events or pathway updates from the preceding trajectory. SepsisAgent (Wu et al., 2026) predicts physiological evolution in structured observation space and remains tied to predefined clinical variables. Latent-state approaches include medDreamer (Xu et al., 2026), which supports action-conditioned rollouts for offline reinforcement learning, and ACWM (Fernandes et al., 2026), whose action labels denote physiological events rather than clinical treatments. Other medical world models target specialized imaging, postoperative, disease-specific, and physiological-signal settings (Ding et al., 2026; Yang et al., 2025; Chung et al., 2026; Wang et al., 2026; Luo et al., 2026). PatientWM instead integrates heterogeneous observations into a unified patient state with typed readouts and conditions each transition on an ordered history of latent patient states and supplied clinical actions. 3 Method 3.1 Problem Formulation PatientWM models clinical evolution at discrete time steps. Let OtO_t denote the heterogeneous clinical observations available at step t, together with their availability indicators. Because these records are shaped by measurement and documentation processes, OtO_t provides evidence about the patient but is not the patient state itself. PatientWM infers a learned predictive patient state tz_t from OtO_t. This state supports prediction of subsequent clinical evolution; it is neither an observed modality nor a claim about the patient’s true physiological state. Let ta_t denote the clinical action supplied for the transition from step t to t+1t+1, and let HtH_t denote the recent state–action history. The factual transition target is p∗(t+1∣t,t,Ht),p^*\! (z_t+1 _t,a_t,H_t ), (1) where p∗p^* denotes the conditional distribution induced by the observed clinical process, not an interventional treatment-response law. The modeled object is the next predictive state in the shared latent space, while typed clinical observations ground and evaluate this state. The formulation specifies a conditional target, whereas the implemented transition produces a deterministic prediction toward a stopped-gradient momentum target. Figure 2: PatientWM encodes available observations and the supplied action into a unified patient state ztz_t and action representation αt _t. Conditioned on ztz_t, αt _t, and sequential latent history HtH_t, the transition predicts z^t+1 z_t+1. The query decoder produces clinical event and outcome predictions and query-specific readouts, including laboratory values. Black arrows show the training-time forward pass with observed history; blue arrows show recursive multi-step rollout, where predicted states update the history for subsequent transitions. 3.2 State Representation Availability-Aware Observation Encoding. Let ℳtM_t be the modalities available at step t, with m indexing one modality, otmo_t^m its observation, and htmh_t^m its encoded representation. The modality-specific encoders EmE_m are fused by EOE_O through ffusionf_fusion to obtain the predictive patient state: htm=Em(otm),t=EO(Ot)=ffusion(htm:m∈ℳt).h_t^m=E_m(o_t^m), _t=E_O(O_t)=f_fusion (\h_t^m:m _t\ ). (2) The fusion uses only modalities available at that step. The resulting tz_t is the unified predictive patient state used by the transition and typed readouts. Typed Clinical Readouts. For target types q∈q , let Yt+1(q)Y_t+1^(q) and Y^t+1(q) Y_t+1^(q) denote the observed and predicted targets. A type-specific decoder Dψ(q)D_ψ^(q) maps the predicted state to the corresponding target: Y^t+1(q)=Dψ(q)(^t+1). Y_t+1^(q)=D_ψ^(q)( z_t+1). (3) These decoders ground the shared patient state in observable clinical quantities without defining that state around any single prediction target. Here q indexes clinical event, outcome, and continuous laboratory targets, each with its corresponding readout and loss. Sharing tz_t makes it a common predictive substrate whose dynamics support multiple clinical queries, rather than an endpoint-specific representation. Equation 3 describes the shared readout interface. Continuous heads may additionally use the transition representation or an output-specific causal prior, which remain decoder-side inputs rather than additional shared patient states. Persistence-Aware Continuous Readouts. For a continuous target, rt+1r_t+1 and r^t+1 r_t+1 denote the observed and predicted values. PatientWM predicts a residual Δtr _t^r around the persistence prior rtpriorr_t^prior, controlled by a learned gate gtrg_t^r: r^t+1=rtprior+gtr⊙Δtr. r_t+1=r_t^prior+g_t^r _t^r. (4) Here ⊙ denotes element-wise multiplication. The gate controls how much of the predicted residual is applied to the prior. The waveform prior uses the most recent quality-qualified embedding with a learned fallback when none exists. The laboratory prior uses the last observed normalized value and train-mean zero for a never-observed item. These priors remain within their typed readouts and do not alter the recursively reused shared-state slot. 3.3 Action-Conditioned Latent Dynamics Sequential Latent History. The supplied clinical action is encoded separately, and PatientWM retains the preceding C−1C-1 latent-state and action representations: t=EA(t),Ht=((s,s))s=t−C+1t−1. α_t=E_A(a_t), H_t= ((z_s, α_s) )_s=t-C+1^t-1. (5) Together with (t,t)(z_t, α_t), this provides each transition with an ordered context of at most C steps. Encoding the action separately keeps treatment-side information explicit without assuming independence from the observation-derived state. The ordered history distinguishes transitions that reach similar current states through different preceding state–action paths, rather than requiring tz_t to compress the complete trajectory. Residual Transition. PatientWM predicts the next patient state from the current state, supplied action, and sequential latent history: ^t+1=Fθ(t,t,Ht)=t+Δθ(t,t,Ht). z_t+1=F_θ(z_t, α_t,H_t)=z_t+ _θ(z_t, α_t,H_t). (6) The residual form parameterizes state changes relative to the current predictive patient state, with Δθ _θ representing the complete action- and history-conditioned update. The formulation is architecture agnostic; our implementation uses a causally masked sequence model and learned gate to retain supported information while applying predicted changes. Further details are provided in the appendix. Training Objective. To provide a stable target patient state for learning the latent transition, we maintain a momentum-updated copy E¯O E_O of the observation encoder, which encodes the next-step observations as follows: t+1tar=stopgrad(E¯O(Ot+1)).z_t+1^tar=stopgrad\! ( E_O(O_t+1) ). (7) The momentum encoder provides a slowly evolving shared-state target, while the typed objectives ground it in clinically observable quantities. The one-step objective is ℒdirect=ℒz(^t+1,t+1tar)+∑q∈mt+1(q)λqℒq(Y^t+1(q),Yt+1(q)).L_direct=L_z( z_t+1,z_t+1^tar)+ _q m_t+1^(q) _qL_q\! ( Y_t+1^(q),Y_t+1^(q) ). (8) Here mt+1(q)m_t+1^(q) indicates whether target q is observed and included in the loss, and λq _q weights its type-specific objective. The latent loss aligns the predicted and target patient states, while the typed losses ground the transition in observable clinical quantities. 3.4 Factual Rollout Recursive State Evolution. To model evolution beyond the next step, PatientWM recursively reuses its predicted patient states. Starting from an observed state ^t0=t0 z_t_0=z_t_0, it computes ^t0+k+1=Fθ(^t0+k,t0+k,H^t0+k),k=0,…,K−1. z_t_0+k+1=F_θ\! ( z_t_0+k, α_t_0+k, H_t_0+k ), k=0,…,K-1. (9) where H^t0+k H_t_0+k contains the retained observed or previously predicted state–action pairs. The history is truncated according to Eq. 5, and the action at each step remains externally supplied. After initialization, each predicted state is appended to the retained history and becomes the state input to the next transition. Future observations are not re-encoded during rollout, distinguishing recursive state evolution from repeated one-step prediction with observed history. Recursive Training. Under the recursive training configuration, the direct objective predicts the next state from observed history, while an additional rollout term feeds generated states into subsequent transitions to supervise multi-step prediction: ℒrollout=ℒdirect+λr(t)∑k=1Hℒdirect(t+k∣t),L_rollout=L_direct+ _r(t) _k=1^HL_direct^(t+k t), (10) where H is the training rollout horizon and λr(t) _r(t) controls the recursive loss contribution. Decoder-side priors are carried across rollout steps but do not constitute complete write-back of generated clinical observations; further details are provided in the appendix. 4 Experiments 4.1 Experimental Setup Datasets and Transition Unit. We evaluate PatientWM on MIMIC-I and MIMIC-IV clinical trajectories, including time-aligned waveform inputs where available (Johnson et al., 2016; Moody et al., 2020; Johnson et al., 2023). Both datasets are organized into contiguous six-hour event threads: observations and source-available treatment records at step t form the model input, while the following window provides factual state and readout targets. Dataset construction, cohort statistics, action definitions, and information boundaries are detailed in Appendix A. Comparisons and Evaluation Roles. Our experiments evaluate factual prediction through typed clinical readouts, dependence on the proposed model components, recursive reuse of generated patient states, action-planning evaluation over bounded candidates, and generalization through frozen-state reuse, label-efficient cross-dataset adaptation, and external cohorts. We compare SepsisAgent, medDreamer, ACWM, and SMB-Structure through the same clinical prediction interface. Dataset-facing input layers allow each method to consume the available laboratory and clinical-fact inputs together with embeddings from the same frozen waveform encoder, while preserving its defining dynamics or representation core. All methods use the same visible inputs, action timing where applicable, target definitions, eligible support, and evaluator. The comparisons therefore assess the retained method structures under matched multimodal information. Implementation details are provided in Appendix C. Metrics. We report AUPRC (Saito and Rehmsmeier, 2015), AUROC (Davis and Goadrich, 2006), and Brier score (Glenn and others, 1950) for clinical events and outcomes, normalized MAE (nMAE) for observed laboratory values, and coverage and top-k agreement for the action-planning evaluation. Metrics remain target-specific rather than being combined into a heterogeneous aggregate. Full protocols and metric definitions are provided in Appendix D.1. 4.2 Factual Prediction And Clinical Readout Clinical Targets. We evaluate clinical prediction in three panels: dataset-specific MIMIC-I Core-7 outcomes and a shared Common Event-7 interface on MIMIC-I and MIMIC-IV (Table 1). The shared panel combines six acute next-window events present in both datasets with dynamic ICU mortality at its native 24-hour horizon. PatientWM achieves the strongest overall aggregate performance across all three panels. Figure 3 and the corresponding AUPRC and Brier analyses in Appendix D.2 further characterize its performance across individual clinical events and outcomes. Together, these results demonstrate the effectiveness of PatientWM for clinical prediction across heterogeneous targets and datasets. Table 1: Clinical event and outcome prediction on the MIMIC-I and MIMIC-IV test sets. The three panels report MIMIC-I Core-7 outcomes and Common Event-7 readouts on both datasets. Bold indicates the best result in each column. MIMIC-I: Core-7 MIMIC-I: Common Event-7 MIMIC-IV: Common Event-7 Method AUPRC↑ AUROC↑ Brier↓ AUPRC↑ AUROC↑ Brier↓ AUPRC↑ AUROC↑ Brier↓ medDreamer 0.0579 0.5130 0.2192 0.0278 0.5049 0.2631 0.0341 0.5050 0.2547 SMB-Structure 0.0638 0.5412 0.2165 0.0522 0.6000 0.2278 0.0604 0.6375 0.2572 SepsisAgent 0.0772 0.6191 0.2181 0.0517 0.6631 0.2457 0.1742 0.7971 0.1875 ACWM 0.0762 0.5945 0.2235 0.0444 0.5879 0.2444 0.1768 0.7959 0.1846 PatientWM (Ours) 0.2227 0.7974 0.0761 0.0598 0.6837 0.1964 0.1784 0.7988 0.1704 Figure 3: Per-target AUROC profiles for (a) MIMIC-I Core-7 outcomes, (b) MIMIC-I Common Event-7 readouts, and (c) MIMIC-IV Common Event-7 readouts. Corresponding per-target AUPRC and Brier results are reported in Appendix D.2. Laboratory Values. PatientWM outperforms all baselines in laboratory-value prediction at every evaluated horizon on both MIMIC-I and MIMIC-IV, as shown in Table 2. These consistent gains from 6 to 96 hours demonstrate that PatientWM more accurately reconstructs future observed laboratory values from the evolving patient state. Table 2: Laboratory-value prediction on the MIMIC-I and MIMIC-IV test cohorts across 6–96-hour horizons. Bold marks the best result within each panel. MIMIC-I Observed Laboratory Values Method 6 h nMAE ↓ 24 h nMAE ↓ 48 h nMAE ↓ 96 h nMAE ↓ Mean nMAE ↓ medDreamer 0.4296 0.4791 0.5387 0.6079 0.5138 SMB-Structure 0.6257 0.6236 0.6300 0.6286 0.6270 SepsisAgent 0.4295 0.4792 0.5388 0.6086 0.5140 ACWM 0.4296 0.4789 0.5386 0.6081 0.5138 PatientWM (Ours) 0.3248 0.3888 0.4430 0.4966 0.4133 MIMIC-IV Observed Laboratory Values Method 6 h nMAE ↓ 24 h nMAE ↓ 48 h nMAE ↓ 96 h nMAE ↓ Mean nMAE ↓ medDreamer 0.4243 0.4527 0.5093 0.5728 0.4898 SMB-Structure 0.5879 0.5862 0.5926 0.6022 0.5922 SepsisAgent 0.4191 0.4520 0.5085 0.5690 0.4871 ACWM 0.4240 0.4519 0.5084 0.5729 0.4893 PatientWM (Ours) 0.4093 0.4161 0.4556 0.4958 0.4442 4.3 Ablations and Mechanism Analysis Table 3: MIMIC-I Core-7 component and modality ablations. MIMIC-I Core-7 Method AUPRC ↑ AUROC ↑ Brier ↓ PatientWM (Ours) 0.2227 0.7974 0.0761 Without action 0.0861 0.6067 0.2024 Without history 0.0840 0.5507 0.2523 Laboratory only 0.0800 0.5874 0.2171 Clinical facts only 0.0777 0.5888 0.2137 Waveform only 0.0772 0.5853 0.2198 We conduct an ablation study to evaluate the effectiveness of PatientWM’s key transition components and modalities. Table 3 compares five ablations on the MIMIC-I Core-7 outcomes. The first two remove action conditioning or ordered history. PatientWM outperforms both, supporting the contribution of these mechanisms to learning clinically informative patient-state dynamics. The larger degradation without ordered history further shows the importance of preserving recent treatment-response trajectories. The remaining three variants restrict state construction to laboratory values, clinical facts, or waveforms alone. Their consistently lower performance shows that no single modality is sufficient and supports combining complementary clinical modalities in the unified patient state. Figure 4: MIMIC-IV multi-step rollout comparison. Panel (a) reports recursive Event-7 macro-AUPRC, and panel (b) reports observed-laboratory nMAE from the typed laboratory decoder. Detailed results are reported in Appendix D.3. 4.4 Multi-Step Patient-State Rollout Multi-step prediction requires models to propagate their own predicted states across successive transitions. From the same observed initial states, we recursively roll out PatientWM and baselines using recorded actions, while reserving future clinical events and laboratory values solely for evaluation. Although laboratory error increases and event ranking becomes more variable with horizon, PatientWM achieves the highest Event-7 macro-AUPRC and the lowest observed-laboratory nMAE at every horizon. These results support more accurate continued forecasting under recorded actions. Figure 4 summarizes the trend, with exact values and support audits reported in Appendix D.3. 4.5 Action-Planning Evaluation Table 4: Recorded-action agreement in the action-planning evaluation. Method Top-1 covered Top-2 covered medDreamer 0.4713 0.7997 SMB-Structure 0.3616 0.7925 SepsisAgent 0.6920 0.9081 ACWM 0.3576 0.7833 PatientWM (Ours) 0.7477 0.9126 To compare action-planning capability, we evaluate whether predicted future consequences can support ranking a bounded set of candidate clinical actions. For every candidate, each method predicts the subsequent clinical-event and laboratory burdens without access to the observed future; the predicted burdens are then combined by the same fixed scoring rule. For decisions in which the clinician’s recorded action is represented in the candidate set, covered top-k agreement measures whether that action is ranked among the top one or two candidates. As shown in Table 4, PatientWM achieves the highest covered top-1 and top-2 agreement, demonstrating the strongest action-planning performance under this bounded protocol. Candidate construction, scoring, and interpretation details are reported in Appendix D.3. Figure 5: Frozen-state transfer on MIMIC-I Common Event-7. Panels (a)–(c) report AUPRC, AUROC, and Brier score, respectively, with full PatientWM and the four Table 1 baselines shown on the same test support. Higher is better for AUPRC and AUROC; lower is better for Brier. 4.6 Frozen-State Reuse We next evaluate the generalizability of the learned PatientWM state to downstream clinical targets that are unavailable during world-model training. To simulate this setting, we train the MIMIC-I world model without Common Event-7 supervision, freeze all world-model parameters, and subsequently train only a new typed decoder using the Common Event-7 labels, the learned ordered 16-state history, and decision-boundary treatment context. This protocol tests whether a patient state learned independently of the downstream target can be reused when a new clinical prediction task is introduced. Figure 5 compares the resulting frozen-state transfer model with full PatientWM and the four MIMIC-I Common Event-7 baselines from Table 1, which are trained with the downstream supervision, under identical target definitions, per-target support, and evaluation. Without using Common Event-7 supervision during world-model training, the frozen state ranks third in AUPRC and second in both AUROC and Brier relative to the four baselines. These results indicate that the unified patient state retains clinically relevant information beyond its original training objectives and can serve as a reusable basis for new typed prediction interfaces. This target-level reuse supports the central goal of PatientWM: moving toward a clinical world model whose state is not tied to one predefined output task. Detailed results are reported in Appendix D.4. 4.7 Label-Efficient Cross-Dataset Adaptation Table 5: MIMIC-I to MIMIC-IV adaptation with 1% target training data. Method AUPRC↑ AUROC↑ Brier↓ medDreamer 0.0323 0.4985 0.2204 SMB-Structure 0.0450 0.5944 0.2061 SepsisAgent 0.0515 0.5904 0.2035 ACWM 0.0386 0.5532 0.2163 PatientWM (Transfer) 0.0517 0.6418 0.2023 We next evaluate whether the patient state learned on MIMIC-I supports adaptation to the MIMIC-IV Common Event-7 targets when target supervision is limited. For each method, we freeze the source predictor and train the same lightweight residual adapter using only 1% of the MIMIC-IV training data. All methods use the same target examples, outcome definitions, and evaluator. As shown in Table 5, PatientWM achieves the strongest aggregate performance across AUPRC, AUROC, and Brier score. These results show that the frozen PatientWM state provides a transferable basis for label-efficient adaptation to a new clinical cohort. Protocol details are provided in Appendix D.5. 4.8 External Framework Evaluation To test whether PatientWM remains effective beyond the MIMIC cohorts, we evaluate it and the baselines on three external datasets with different clinical data objects: ALOTT (Lawrence et al., 2025), VitalDB (Lee et al., 2022), and MC-MED (Kansal et al., 2025), whose provenance and interfaces are detailed in Appendix A. ALOTT and VitalDB evaluate sequential next-state prediction, whereas MC-MED provides a pointwise prediction setting. Within each dataset, all methods use the same visible inputs, targets, data splits, training budget, and evaluator. As shown in Figure 6, PatientWM achieves the highest AUROC on all three datasets. These results support the applicability of the PatientWM framework across distinct clinical data interfaces; detailed AUPRC, Brier, and AUROC results are reported in Appendix D.2. Figure 6: External AUROC comparison. Panels (a), (b), and (c) report results on ALOTT, VitalDB, and MC-MED, respectively. Higher is better in all panels. 5 Conclusion To address the limited state scope and insufficient trajectory context of existing clinical world models, we proposed PatientWM, a framework built around a unified latent patient state and ordered state–action history. It learns the patient state from available clinical observations, models its evolution under clinical actions, and exposes typed clinical readouts. Experiments on MIMIC-I, MIMIC-IV, and external datasets show strong factual prediction and recursive rollout performance, while the action-planning evaluation shows higher recorded-action agreement within fixed candidate sets. Together, these results show that PatientWM learns a reusable patient state for clinical prediction, maintains more accurate forecasts across transitions, and provides stronger predictive support for bounded action planning. AI use statement In this work, we used generative AI tools to aid and polish writing; retrieve, discover, and synthesize related literature; support research ideation and execution, including experiment planning, code development, result organization and interpretation, and figure and table preparation; and draft sections of the manuscript. The authors reviewed AI-assisted outputs against cited sources, executable configurations, experiment logs, and aggregate results and take responsibility for the final text, claims, code, and artifacts. Ethics statement This retrospective study uses deidentified, access-controlled clinical datasets under their respective data-use terms. It does not provide treatment recommendations or estimate causal treatment effects. Row-level records remain within approved environments; released artifacts are limited to code, schemas, hashes, and aggregate statistics. Appendix A details data governance and intended use. Reproducibility statement The main text reports the evaluation roles, comparison protocols, and primary results. The appendix documents dataset construction and information boundaries, model implementation, comparison controls, supplementary evaluations, model parameters, and data splits. References Attia et al. (2019) Z. I. Attia, S. Kapa, F. Lopez-Jimenez, P. M. McKie, D. J. Ladewig, G. Satam, P. A. Pellikka, M. Enriquez-Sarano, P. A. Noseworthy, T. M. Munger, et al. Screening for cardiac contractile dysfunction using an artificial intelligence–enabled electrocardiogram. Nature medicine 25 (1), p. 70–74. Cited by: §2. Bruce et al. (2024) J. Bruce, M. D. Dennis, A. Edwards, J. Parker-Holder, Y. Shi, E. Hughes, M. Lai, A. Mavalankar, R. Steigerwald, C. Apps, et al. Genie: generative interactive environments. In Forty-first international conference on machine learning, Cited by: §2. Chen et al. (2026) Z. Chen, Z. Cong, Z. Jin, W. Fan, D. Zhou, Q. Ai, H. Gong, C. Liao, X. Liu, and C. Wang Medical world models in healthcare: foundations, applications, and challenges for trustworthy clinical translation. arXiv preprint arXiv:2607.25242. Cited by: §1, §2. Choi et al. (2016) E. Choi, M. T. Bahadori, J. Sun, J. Kulas, A. Schuetz, and W. Stewart Retain: an interpretable predictive model for healthcare using reverse time attention mechanism. Advances in neural information processing systems 29. Cited by: §1, §2. Chung et al. (2026) Y. Chung, Y. Liu, A. F. Hassan, H. Feng, M. M. Maleckar, N. Marrouche, and J. Hamm Intervention-aware clinical world model for post-op outcome forecasting in cardiology. arXiv preprint arXiv:2608.13518. Cited by: §2. Davis and Goadrich (2006) J. Davis and M. Goadrich The relationship between precision-recall and roc curves. In Proceedings of the 23rd international conference on Machine learning, p. 233–240. Cited by: §D.1, §4.1. Ding et al. (2025) J. Ding, Y. Zhang, Y. Shang, Y. Zhang, Z. Zong, J. Feng, Y. Yuan, H. Su, N. Li, N. Sukiennik, F. Xu, and Y. Li Understanding world or predicting future? a comprehensive survey of world models. ACM Computing Surveys 58 (3), p. 1–38. Cited by: §1, §2. Ding et al. (2026) T. Ding, Y. Zou, C. Chen, M. Shah, and Y. Tian CLARITY: medical world model for guiding treatment decisions by simulating context-aware disease trajectories. In European Conference on Computer Vision, p. 202–221. Cited by: §D.3, §1, §1, §2. Dou et al. (2025) Y. Dou, Y. Liu, H. Zou, W. Zeng, K. Xu, and S. Peng A survey on electronic health record driven multimodal representation learning. Information Fusion, p. 103810. Cited by: §1. Faust et al. (2018) O. Faust, Y. Hagiwara, T. J. Hong, O. S. Lih, and U. R. Acharya Deep learning for healthcare applications based on physiological signals: a review. Computer methods and programs in biomedicine 161, p. 1–13. Cited by: §2. Fernandes et al. (2026) J. G. Fernandes, L. Facury, P. R. Dutenhefner, and W. Meira Jr Beyond patient invariance: learning cardiac dynamics via action-conditioned jepas. arXiv preprint arXiv:2604.22618. Cited by: §2. Glenn et al. (1950) W. B. Glenn et al. Verification of forecasts expressed in terms of probability. Monthly weather review 78 (1), p. 1–3. Cited by: §D.1, §4.1. Ha and Schmidhuber (2018) D. Ha and J. Schmidhuber World models. arXiv preprint arXiv:1803.10122. Cited by: §1, §2. Hafner et al. (2019) D. Hafner, T. Lillicrap, J. Ba, and M. Norouzi Dream to control: learning behaviors by latent imagination. arXiv preprint arXiv:1912.01603. Cited by: §2. Hafner et al. (2025) D. Hafner, J. Pasukonis, J. Ba, and T. Lillicrap Mastering diverse control tasks through world models. Nature 640 (8059), p. 647–653. Cited by: §2. Hannun et al. (2019) A. Y. Hannun, P. Rajpurkar, M. Haghpanahi, G. H. Tison, C. Bourn, M. P. Turakhia, and A. Y. Ng Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network. Nature medicine 25 (1), p. 65–69. Cited by: §2. Hansen et al. (2024) E. R. Hansen, T. Sagi, and K. Hose Multimodal representation learning for medical analytics—a systematic literature review. Health Informatics Journal 30 (4), p. 1–20. External Links: Document Cited by: §1. Harutyunyan et al. (2019) H. Harutyunyan, H. Khachatrian, D. C. Kale, G. Ver Steeg, and A. Galstyan Multitask learning and benchmarking with clinical time series data. Scientific data 6 (1), p. 96. Cited by: §D.1, §2. Hyndman and Koehler (2006) R. J. Hyndman and A. B. Koehler Another look at measures of forecast accuracy. International journal of forecasting 22 (4), p. 679–688. Cited by: §D.1. Johnson et al. (2023) A. E. Johnson, L. Bulgarelli, L. Shen, A. Gayles, A. Shammout, S. Horng, T. J. Pollard, S. Hao, B. Moody, B. Gow, et al. MIMIC-iv, a freely accessible electronic health record dataset. Scientific data 10 (1), p. 1. Cited by: §A.6, §4.1. Johnson et al. (2016) A. E. Johnson, T. J. Pollard, L. Shen, L. H. Lehman, M. Feng, M. Ghassemi, B. Moody, P. Szolovits, L. Anthony Celi, and R. G. Mark MIMIC-i, a freely accessible critical care database. Scientific data 3 (1), p. 160035. Cited by: §A.1, §A.6, §4.1. Kansal et al. (2025) A. Kansal, E. Chen, T. Jin, P. Rajpurkar, and D. Kim Multimodal Clinical Monitoring in the Emergency Department (MC-MED). PhysioNet. Note: Version 1.0.1 External Links: Document, Link Cited by: §A.5, §A.6, §4.8. Lawrence et al. (2025) J. Lawrence, M. Rayo, and T. Huerta ALarms, Outcomes Telemetry with Timing (ALOTT): A Bedside-EMR Database. PhysioNet. Note: Version 1.0.0 External Links: Document, Link Cited by: §A.5, §A.6, §4.8. Lee et al. (2022) H. Lee, Y. Park, S. B. Yoon, S. M. Yang, D. Park, and C. Jung VitalDB, a high-fidelity multi-parameter vital signs database in surgical patients. Scientific Data 9 (1), p. 279. Cited by: §A.5, §A.6, §4.8. Liu et al. (2026) K. Liu, M. Li, Y. Bao, T. Zhang, C. Chu, J. Bu, and H. Wang Medical world models: representing medical states, modelling clinical dynamics and guiding intervention policies. arXiv preprint arXiv:2606.16721. Cited by: §1, §1, §2, §2. Luo et al. (2026) Y. Luo, X. Chen, Y. Chen, L. Zhang, M. M. Islam, S. Zhao, P. Kotanko, S. Dasgupta, A. Campbell, R. Malhotra, et al. Toward world modeling of physiological signals with chaos-theoretic balancing and latent dynamics. arXiv preprint arXiv:2605.15465. Cited by: §2. Moody et al. (2020) B. Moody, G. Moody, M. Villarroel, G. D. Clifford, and I. Silva MIMIC-I Waveform Database. PhysioNet. Note: Version 1.0 External Links: Document, Link Cited by: §A.1, §A.6, §4.1. Mu et al. (2026) L. Mu, Z. Huang, Y. Gu, S. Qin, S. Zhang, and X. Zhang EHRWorld: a patient-centric medical world model for long-horizon clinical trajectories. arXiv preprint arXiv:2602.03569. Cited by: §1, §2. Pellegrini et al. (2025) C. Pellegrini, E. Özsoy, D. Bani-Harouni, M. Keicher, and N. Navab From ehrs to patient pathways: scalable modeling of longitudinal health trajectories with llms. arXiv preprint arXiv:2506.04831. Cited by: §2. Rajkomar et al. (2018) A. Rajkomar, E. Oren, K. Chen, A. M. Dai, N. Hajaj, M. Hardt, P. J. Liu, X. Liu, J. Marcus, M. Sun, et al. Scalable and accurate deep learning with electronic health records. NPJ digital medicine 1 (1), p. 18. Cited by: §1, §2. Saito and Rehmsmeier (2015) T. Saito and M. Rehmsmeier The precision-recall plot is more informative than the roc plot when evaluating binary classifiers on imbalanced datasets. PloS one 10 (3), p. e0118432. Cited by: §D.1, §4.1. Shickel et al. (2017) B. Shickel, P. J. Tighe, A. Bihorac, and P. Rashidi Deep ehr: a survey of recent advances in deep learning techniques for electronic health record (ehr) analysis. IEEE journal of biomedical and health informatics 22 (5), p. 1589–1604. Cited by: §1, §2. Si et al. (2021) Y. Si, J. Du, Z. Li, X. Jiang, T. Miller, F. Wang, W. J. Zheng, and K. Roberts Deep representation learning of patient data from electronic health records (ehr): a systematic review. Journal of biomedical informatics 115, p. 103671. Cited by: §1. Steinberg et al. (2024) E. Steinberg, J. Fries, Y. Xu, and N. Shah MOTOR: a time-to-event foundation model for structured medical records. In International Conference on Learning Representations, Vol. 2024, p. 41038–41077. Cited by: §2. Tomašev et al. (2019) N. Tomašev, X. Glorot, J. W. Rae, M. Zielinski, H. Askham, A. Saraiva, A. Mottram, C. Meyer, S. Ravuri, I. Protsyuk, et al. A clinically applicable approach to continuous prediction of future acute kidney injury. Nature 572 (7767), p. 116–119. Cited by: §2. Vaswani et al. (2017) A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin Attention is all you need. Advances in neural information processing systems 30. Cited by: §B.1. Wang et al. (2026) J. Wang, X. Chen, J. He, X. Xu, S. Xie, and F. Han ChronoMedicalWorld: a medical world model for learning patient trajectories from longitudinal care data. arXiv preprint arXiv:2605.21963. Cited by: §1, §1, §2. Wang et al. (2024) Y. Wang, J. He, L. Fan, H. Li, Y. Chen, and Z. Zhang Driving into the future: multiview visual forecasting and planning with world model for autonomous driving. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 14749–14759. Cited by: §2. Wu et al. (2026) M. Wu, Y. Yan, Z. Cai, K. Ji, C. Fang, Z. Sheng, X. Wang, R. Wang, H. Zhang, S. Li, et al. Agentifying patient dynamics within llms through interacting with clinical world model. arXiv preprint arXiv:2605.14723. Cited by: §D.3, §1, §2. Wu et al. (2023) P. Wu, A. Escontrela, D. Hafner, P. Abbeel, and K. Goldberg Daydreamer: world models for physical robot learning. In Conference on robot learning, p. 2226–2240. Cited by: §2. Xu et al. (2026) Q. Xu, G. Habib, F. Wu, D. Perera, and M. Feng Meddreamer: model-based reinforcement learning with latent imagination on complex ehrs for clinical decision support. In Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1, p. 1693–1704. Cited by: §1, §1, §2. Yang et al. (2025) Y. Yang, Z. Wang, Q. Liu, S. Sun, K. Wang, R. Chellappa, Z. Zhou, A. Yuille, L. Zhu, Y. Zhang, et al. Medical world model. In 2025 IEEE/CVF International Conference on Computer Vision (ICCV), p. 8319–8329. Cited by: §2. Yu et al. (2021) C. Yu, J. Liu, S. Nemati, and G. Yin Reinforcement learning in healthcare: a survey. ACM Computing Surveys (CSUR) 55 (1), p. 1–36. Cited by: §1. Appendix A Datasets And Information Boundary A.1 MIMIC-I State, Outcomes, And Splits We align MIMIC-I clinical and waveform resources (Johnson et al., 2016; Moody et al., 2020) to contiguous six-hour ICU windows; empty windows remain in the timeline. The observed record is Ot=(ℒt,tL,t,ty,t,mtw),O_t=(L_t,m^L_t,y_t,m^y_t,w_t,m^w_t), (11) where ℒtL_t is a sparse set of normalized laboratory tokens, ty_t contains current-window clinical events, and t∈ℝ512w_t ^512 is a frozen waveform embedding. Every component has an availability mask. A transition-derived label for t→t+1t→ t+1 is stored in window t+1t+1, never in the source window. The resulting cohort contains 53,329 ICU episodes, 911,164 windows, and 857,835 adjacent transitions. Subject-level splitting produces 42,669/5,323/5,337train/validation/test episodes from 30,835/3,848/3,827subjects; no subject crosses splits. We report seven core binary outcomes with non-treatment-derived target definitions: hypercapnic acidemia, severe acidemia, creatinine- and urine-defined KDIGO stage 2/3 AKI, GCS drop, incident CAM positivity, and mortality within 24 hours. A.2 Structured Action Ledger Action truth comes from INPUTEVENTS_CV, INPUTEVENTS_MV, and PROCEDUREEVENTS_MV. Each event retains treatment identity, execution time, chart-availability time, amount, rate, unit, route when available, source table, and execution evidence. Notes may later supply decision-time context or candidate mentions, but do not override structured execution records. We distinguish the active treatment state known at the decision boundary from treatment records that become available later. Only boundary-known, chart-available treatment information enters the current action port, and it remains separate from the patient observation state. Later treatment records are excluded from the current input and may enter only as externally supplied logged actions at later factual-replay steps; they are not interpreted as causal intervention assignments. Table 6 reports the dataset-wide coverage of the structured event ledger. A transition is counted as covered when its corresponding event window contains at least one executed event; this descriptive statistic is distinct from the availability of the boundary-known action input. Table 6: Structured event-ledger coverage. Coverage is the fraction of transitions whose corresponding event window contains at least one executed event of any supported type. This dataset statistic does not measure boundary-known action availability or prediction performance. Split Transitions Event windows Coverage Events Episodes Train 685,561 620,559 90.52% 11,870,214 41,172 Validation 85,295 77,490 90.85% 1,460,107 5,141 Test 86,979 78,673 90.45% 1,531,140 5,146 A.3 MIMIC-IV Port We use MIMIC-IV for two principal native-cohort evaluation roles under the same six-hour transition unit and decision-time information boundary. The first is the cross-dataset Common Event-7 clinical readout reported in Table 1, which shares its target definitions with the corresponding MIMIC-I panel. The second trains a MIMIC-IV transition model used for the multi-step rollout in Figure 4 and the action-planning evaluation in Table 4. The frozen transition also supports the MIMIC-IV typed laboratory decoder reported in Table 2. These roles share the dataset construction, patient-disjoint split, availability-mask and leakage contracts. Separately, the label-efficient adaptation experiment uses MIMIC-IV as the target cohort for evaluating transfer from a frozen MIMIC-I source predictor. A.4 Common Event-7 Panel To support a direct cross-dataset event comparison, we derive a separate Common Event-7 outcome sidecar from the native query matrices in both datasets. The shared heads are any acute deterioration, new invasive ventilation, new pressor, new circulation pressor, respiratory IMV escalation, Sepsis-3 onset, and dynamic ICU mortality. The first six heads are next-window events; dynamic ICU mortality retains its six-hour gap and 24-hour target horizon. Each row keeps the observed mask and native target timing. The sidecar is used for task-level readout evaluation and is not inserted into the six-dimensional recursive state label matrix. The MIMIC-I Core-7 and MIMIC-IV cardiac-arrest extensions remain dataset-specific transition diagnostics. A.5 External Cohorts The supplementary question is whether the same unified-patient-state framework applies across different clinical data objects. ALOTT (Lawrence et al., 2025) and VitalDB (Lee et al., 2022) have verified adjacent states; MC-MED (Kansal et al., 2025) has one prediction anchor and serves as a pointwise control. Before task eligibility filtering, their train/test counts are 974,025/215,239, 881,421/111,162, and 227,886/33,119, respectively. VitalDB aggregates 22 forecasting tasks and excludes three concurrent-state controls. One MC-MED task with only three positive test cases after waveform QC is excluded from the aggregate. Every method receives pre-anchor clinical context and a known-empty executed-action port. A.6 Data Governance And Intended Use MIMIC-I is a deidentified, credentialed PhysioNet resource governed by its data-use agreement (Johnson et al., 2016); the waveform resource is separately released through PhysioNet (Moody et al., 2020). MIMIC-IV is separately governed and deidentified (Johnson et al., 2023). VitalDB is used under the terms of the versioned release (Lee et al., 2022). ALOTT and MC-MED are used from their versioned PhysioNet releases (Lawrence et al., 2025; Kansal et al., 2025). Public artifacts contain code, schemas, hashes, and aggregate statistics only; governed row-level trajectories remain inside their approved environments. PatientWMis intended for retrospective representation and factual-forecasting research, not bedside monitoring, treatment selection, or autonomous simulation. Clinical use would require external and prospective validation, subgroup and calibration audits, and a causal design appropriate to the intended decision. Appendix B PatientWM Method And Implementation B.1 Architecture Observation Encoder. The main MIMIC-I checkpoint instantiates EOE_O with three active observation branches: sparse laboratory tokens, masked clinical facts, and frozen waveform-derived embeddings. Each branch emits a 256-dimensional modality representation, and availability masks determine which branch outputs are eligible for fusion. Sparse laboratory observations are represented as at most 128 valid item tokens per window from a 693-item lab vocabulary. For item j, the encoder combines a learned identity embedding ejL∈ℝ256e_j^L ^256 with an 11-dimensional numeric feature vector xt,jLx_t,j^L containing normalized value summaries, observation counts, abnormality summaries, timing or recency, and missingness-related indicators. The token representation is ht,jL=LayerNorm(ejL+WLxt,jL+bL).h_t,j^L=LayerNorm\! (e_j^L+W_Lx_t,j^L+b_L ). (12) A learned lab query attends over valid lab tokens using the lab key-padding mask. If no valid lab token exists in the window, the branch returns a learned empty-lab embedding and marks the lab modality absent for fusion. The clinical-fact branch consumes six binary current-window facts with their availability masks. Unavailable fact values are masked before projection, and the concatenated values and masks are mapped to a 256-dimensional fact representation: hty=LayerNorm(Wy[t⊙ty;ty]+by).h_t^y=LayerNorm (W_y[y_t _t^y;m_t^y]+b_y ). (13) The fact modality is marked present when at least one fact is observed. The world model does not ingest raw waveform samples. A frozen pretrained waveform encoder, without downstream fine-tuning, converts valid waveform frames into embeddings, which are mean-pooled into one 512-dimensional representation per window. When this representation is present, a linear projection maps it to the common 256-dimensional branch width; when absent, a learned empty-waveform embedding is used and the waveform modality is marked absent. For modality fusion, each branch representation receives a learned modality type embedding. A learned observation query attends over the stack of available branch states: t=EO(Ot)=LayerNorm(MHA(qO,Gtmod,Gtmod,tmod)),z_t=E_O(O_t)=LayerNorm (MHA(q_O,G_t^mod,G_t^mod;m_t^mod) ), (14) where Gtmod∈ℝ3×256G_t^mod ^3× 256 stacks the lab, fact, and waveform branch states, and tmodm_t^mod records branch availability. If all branches are absent, the encoder returns a learned empty-state embedding. Online And Target Encoders. Training uses an online observation encoder EOE_O and a momentum target encoder E¯O E_O. The target encoder supplies the stopped-gradient next-window latent target, t+1tar=stopgrad(E¯O(Ot+1)),z_t+1^tar=stopgrad\! ( E_O(O_t+1) ), (15) and its parameters follow the online encoder after each optimizer update: ϕ¯←0.99ϕ¯+0.01ϕ. φ← 0.99\, φ+0.01\,φ. (16) This target is a joint-embedding predictive supervision signal. It is not interpreted as recovered true latent physiology. Action Encoder. The main checkpoint encodes the source-boundary active-treatment record from the structured action ledger. A treatment item is included in ta_t only when it started before the end of window t, remains active at that boundary, and is already chart-available. Target-window duration, future delivered amount, future start or stop, end-of-window activity, and future status changes are zeroed or excluded. The active-treatment set contains at most 64 tokens from a 308-entry treatment vocabulary. Each token has a 25-dimensional numeric feature vector covering available lifecycle, exposure, amount/rate/unit/source, execution-evidence, and timing features. Let ejA∈ℝ256e_j^A ^256 be the treatment identity embedding and xt,jA∈ℝ25x_t,j^A ^25 its numeric feature vector: ht,jA=LayerNorm(ejA+WAxt,jA+bA).h_t,j^A=LayerNorm (e_j^A+W_Ax_t,j^A+b_A ). (17) A learned action query attends over the valid action tokens: t=EA(t)=LayerNorm(MHA(qA,GtA,GtA,tA))∈ℝ256, α_t=E_A(a_t)=LayerNorm (MHA(q_A,G_t^A,G_t^A;m_t^A) ) ^256, (18) where GtAG_t^A stacks token states and tAm_t^A masks invalid token slots. A known-empty action ledger returns a learned empty-action embedding and an explicit presence indicator; it is not treated as missing data. Transition Model. The main transition first projects the encoded patient state and encoded action state into a transition token: xttr=LayerNorm(GELU(Wtr[t;t]+btr))∈ℝ256.x_t^tr=LayerNorm (GELU (W_tr[z_t; α_t]+b_tr ) ) ^256. (19) The primary dynamics operator is a pre-norm causal Transformer (Vaswani et al., 2017). It uses strict causal masking and a transition-valid padding mask so that position t attends only to legal positions through t. The architecture and training settings are collected in Table 9. For final hidden state ht∈ℝ256h_t ^256, ht=CausalTransformer(xt−C+1:ttr).h_t=CausalTransformer (x_t-C+1:t^tr ). (20) The latent transition is decoded as a gated residual around the current latent state: Δtz=WΔzht+bΔz,gtz=σ(Wgzht+bgz),^t+1=t+gtz⊙Δtz. _t^z=W_ ^zh_t+b_ ^z, g_t^z=σ(W_g^zh_t+b_g^z), z_t+1=z_t+g_t^z _t^z. (21) The residual projection is zero-initialized, and the gate bias initializes the gate probability near 0.1. The initialized model is therefore close to latent copy-forward and learns departures only when supported by training targets. The main MIMIC-I configuration optimizes 5,685,029 parameters. The momentum target encoder follows the online encoder by exponential moving average and is not optimized directly. Typed Readout Interface. Following Equation 3, each type-specific decoder maps the predicted patient state ^t+1 z_t+1 to its corresponding clinical target. Continuous heads may additionally use the transition representation or an output-specific causal prior. These remain decoder-side inputs rather than additional shared patient states, and the recursively reused patient-state slot remains ^t+1 z_t+1. B.2 Typed Decoders And Recursive State Persistence-Aware Continuous Readouts. As described in Equation 4, continuous readouts predict a gated residual around a causal persistence prior: r^t+1=rtprior+gtr⊙Δtr. r_t+1=r_t^prior+g_t^r _t^r. (22) For waveform embeddings, the prior is the most recent quality-qualified embedding, with a learned fallback when none is available. For laboratory values, the prior is the last observed normalized value for a previously measured item and the training-set mean for a never-observed item. These priors use only source-visible information and remain within their corresponding decoders during rollout. Decoder Roles. Each query type selects a decoder with a target-specific output space. The transition checkpoint also includes auxiliary heads that ground the learned state in next-window observations. Table 7 summarizes these roles. Table 7: Typed decoder outputs and their roles in training and evaluation. Decoder Output Role clinical event/outcome target-specific probability clinical readout waveform 512-D embedding waveform prediction and cache update clinical fact six probabilities transition supervision lab availability 693 probabilities observation-process supervision joint lab value 693 normalized values transition supervision and cache update queried lab value one normalized value observed-laboratory prediction Clinical And Waveform Readouts. The query type selects the corresponding decoder. Binary clinical-event and outcome decoders emit probabilities for their requested targets according to Equation 3. These task-specific readouts share the predicted patient state but use separate output parameters and losses. The transition checkpoint additionally predicts six next-window clinical facts with a linear logit head, ^t+1y=σ(Wydecht+bydec). p_t+1^y=σ(W_y^dech_t+b_y^dec). (23) These auxiliary factual targets ground transition training and are distinct from the Core-7 and Common Event-7 clinical readout panels used in the main evaluation. The waveform decoder predicts a residual in the frozen embedding space: ^t+1=tprior+gtw⊙Δtw,Δtw=WΔwht+bΔw,gtw=σ(Wgwht+bgw). w_t+1=w_t^prior+g_t^w _t^w, _t^w=W_ ^wh_t+b_ ^w, g_t^w=σ(W_g^wh_t+b_g^w). (24) If the waveform prior is unavailable, a separate linear fallback maps hth_t to the 512-dimensional embedding. The decoded waveform is a model-generated embedding-space representation, not raw waveform generation. Joint Laboratory Heads. The laboratory-availability decoder emits one next-window observation probability per item in the 693-item vocabulary: ^t+1L,present=σ(WmLht+bmL)∈[0,1]693. p_t+1^L,present=σ(W_m^Lh_t+b_m^L)∈[0,1]^693. (25) The corresponding dense value head predicts normalized laboratory values around the causal prior: ^t+1joint=tprior+tL⊙ΔtL. l_t+1^joint=l_t^prior+g_t^L _t^L. (26) This head contributes loss only for numerically observed target item-value pairs. During recursive rollout, its predictions update the cache for items that have already been observed; the availability probabilities remain typed outputs and do not require thresholding to create new synthetic measurements. Accumulated Laboratory Decoder. For the MIMIC-I observed-laboratory evaluation, the trained transition is frozen and only a query-specific laboratory decoder is optimized. The decoder receives the rolled patient state or transition context, the queried item identity, its last source-visible normalized value and age, whether it has been observed previously, and causal summary features computed from the available laboratory history. Its GapShrink prior is l^j,t+hgap=βj,aj+hlj,tlast,βj,g∈[0,1], l_j,t+h^gap= _j,a_j+hl_j,t^last, _j,g∈[0,1], (27) where aja_j is the age of the last observed value at rollout start. The coefficients are estimated from training data only, and gaps beyond the modeled range fall back to persistence. The learned query branch combines the 256-dimensional rolled state, a 128-dimensional item embedding, and the source-visible laboratory ledger into a 128-dimensional query representation qj,hq_j,h. Cross-laboratory attention allows queried item j to use measurements of other previously observed items: Ajk=softmaxk((WQqj,h)⊤(WKsk)128),cjcross=∑k:seenAjkWVsk.A_jk=softmax_k ( (W_Qq_j,h) (W_Ks_k) 128 ), c_j^cross= _k:seenA_jkW_Vs_k. (28) The attended context is projected back into the query representation, q~j,h=qj,h+WOcjcross q_j,h=q_j,h+W_Oc_j^cross. The output projection WOW_O is zero-initialized so that training begins from the item-specific query model and adds cross-item corrections only when supported by the data. The final prediction is a gated residual around GapShrink: δj,h=Wδq~j,h+bδ,gj,h=σ(Wgq~j,h+bg),l^j,t+h=l^j,t+hgap+gj,hδj,h. _j,h=W_δ q_j,h+b_δ, g_j,h=σ(W_g q_j,h+b_g), l_j,t+h= l_j,t+h^gap+g_j,h _j,h. (29) The gate bias initializes the update probability near 0.1, and the decoder is optimized with pointwise L1L_1 loss on observed item–target pairs. For each item, the accumulated source-only features comprise the running mean, running standard deviation, deviation of the latest value from the running mean, fast and slow exponentially weighted moving averages, their difference, and positive and negative CUSUM statistics. The fast and slow coefficients are 0.5 and 0.1, respectively; CUSUM uses retention 0.85 and drift 0.05 with gap-aware decay on the six-hour grid. These statistics are computed only from source-visible observations and are never refreshed from the future target ledger. Recursive Implementation State. Recursive evaluation reuses the predicted patient state together with the causal continuous-readout caches: Rt=(^t,^t,tL,tL,tL),R_t= ( z_t, w_t,c_t^L,s_t^L,a_t^L ), (30) where ^t z_t is the predicted latent patient state, ^t w_t is the generated waveform-embedding cache, tLc_t^L is the dense causal last-value lab cache, tLs_t^L records whether each lab item has ever been seen, and tLa_t^L records lab-value age in windows. The additional fields support typed persistence and query decoding; tz_t remains the unified patient state used by the transition model. Clinical-event, outcome, and laboratory-availability probabilities are decoded from the evolving state when requested, whereas the continuous caches above are carried to provide causal priors for subsequent predictions. B.3 Training Details Direct Transition Objective. For a valid transition t→t+1t→ t+1, PatientWM jointly optimizes the latent transition and the available typed targets: ℒdirect=λzℒz+λwℒw+λyℒy+λmLℒmL+λLℒL.L_direct= _zL_z+ _wL_w+ _yL_y+ _m^LL_m^L+ _LL_L. (31) The five terms correspond to the next patient state, waveform embedding, clinical facts, laboratory availability, and observed laboratory values. In the reported transition configurations, every enabled top-level term has weight 1.0; unavailable targets are removed by their corresponding masks. The latent objective aligns the predicted patient state with the stopped-gradient target from the momentum encoder: ℒz=1−cos(^t+1,t+1tar).L_z=1- ( z_t+1,z_t+1^tar). (32) The waveform objective is evaluated only when a quality-qualified next-window embedding is available: ℒw=[mt+1w=1](1−cos(^t+1,t+1)).L_w=1[m_t+1^w=1] (1- ( w_t+1,w_t+1) ). (33) Clinical-fact and laboratory-availability heads use masked binary cross-entropy. Their positive weights are estimated from the training split as negative-to-positive ratios and clipped to [1,100][1,100]. The joint laboratory value head uses masked Smooth-L1L_1 loss only for item–target pairs with an observed numeric value. Residual-Update Regularization. The magnitudes of gated latent, waveform, and laboratory updates are regularized by ℒresid=13(‖gz⊙Δz‖22+‖gw⊙Δw‖22+‖gL⊙ΔL‖22).L_resid= 13 (\|g^z ^z\|_2^2+\|g^w ^w\|_2^2+\|g^L ^L\|_2^2 ). (34) This term discourages unsupported departures from the causal persistence priors while allowing the learned gates to apply larger updates when supported by the training targets. Recursive Training Objective. Transition models trained for multi-step use combine the teacher-forced direct objective with a generated-history branch: ℒ=ℒdirect+λr(t)∑k=1Hℒdirect(t+k∣t)+λΔℒresid+ℒaux.L=L_direct+ _r(t) _k=1^HL_direct^(t+k t)+ _ L_resid+L_aux. (35) The generated-history branch receives the observed initial state once and then advances the predicted latent state, waveform embedding, and existing-item laboratory cache. Logged source-boundary actions remain externally supplied at each transition. Future observations, values, and availability masks enter only as supervised targets. The rollout coefficient λr(t) _r(t) is warmed up from zero to its final value in the MIMIC-I transition training protocol, avoiding long generated prefixes at the beginning of optimization. Table 8: Recursive training settings used by the reported transition experiments. Experiment H Final λr _r Additional objective MIMIC-I transition rollout 16 0.25 residual and waveform terms MIMIC-IV primary rollout 16 0.50 residual term One-step/readout training – 0 none For recursive transition training, λΔ=0.001 _ =0.001. Modality-specific auxiliary losses are included only when the corresponding target is part of the stated experiment. Teacher-Forced And Free-Running Modes. Teacher-forced evaluation constructs OtO_t and ta_t from the legal observed record and uses the observed prefix to form the ordered state–action history. The next record Ot+1O_t+1 is joined only to provide targets. This mode therefore measures factual next-window prediction without testing error accumulation under generated history. Logged-action free-running evaluation initializes once from observed Ot0O_t_0. At each subsequent step, it encodes the externally supplied logged action, applies the transition to the retained predicted state–action prefix, updates the latent, waveform, and existing-item laboratory caches, and decodes the requested outputs. The observed future record is joined only after the prediction for scoring. The principal reported horizons of 1, 4, 8, and 16 transitions correspond to 6, 24, 48, and 96 hours, respectively. B.4 Hyperparameters Table 9: Implementation settings for the main MIMIC-I causal-Transformer transition and laboratory readout. Transition Architecture Setting Value Width 256 Layers 4 Attention heads 8 Feed-forward hidden width 1,024 Activation GELU Dropout 0.1 Maximum context 16 six-hour windows Positional representation learned absolute position embedding Causal mask strict upper-triangular attention mask Transition Optimization Setting Value Optimizer AdamW Learning rate 3×10−43× 10^-4 Weight decay 0.01 Micro-batch 8 Gradient accumulation 4 Effective batch 32 Optimizer steps 1,024 Balanced shard contribution 32 optimizer-step equivalents per shard Sampled trajectories 32,768 Checkpoint criterion direct validation loss plus 0.25 closed-loop loss Query-Only Laboratory Readout Setting Value Complete frozen-core-plus-decoder parameters 9,132,456 Trainable decoder parameters 191,107 Item-query dimension 128 Training mode recursive rollout Point loss micro-target masked L1L_1 NLL weight 0 Batch size 8 Gradient accumulation 4 Learning rate 3×10−43× 10^-4 Weight decay 0.01 Optimizer steps 2,048 Examples seen 65,536 across all 32 shards B.5 Tensor Contract The following contract describes the main MIMIC-I transition implementation. The MIMIC-IV implementations preserve the same temporal alignment, masking, and source–target separation while using their dataset-specific observation and target schemas. Let B be the batch size, C≤16C≤ 16 the number of transitions, NL=128N_L=128 the maximum number of laboratory tokens per window, NA=64N_A=64 the maximum number of active-treatment tokens, VL=693V_L=693 the laboratory vocabulary size, Ky=6K_y=6 the auxiliary clinical-fact dimension, DW=512D_W=512 the waveform-embedding dimension, and D=256D=256 the patient-state dimension. Table 10: Tensor contract for the main MIMIC-I transition model. The state axis contains C+1C+1 windows because C transitions connect consecutive source and target states. Tensor role Shape Meaning laboratory item identities [B,C+1,NL][B,C+1,N_L] sparse item indices for each state window laboratory numeric features [B,C+1,NL,11][B,C+1,N_L,11] value summaries, counts, timing, and availability features laboratory token mask [B,C+1,NL][B,C+1,N_L] valid, non-padding laboratory positions clinical-fact values [B,C+1,Ky][B,C+1,K_y] current-window binary clinical facts clinical-fact mask [B,C+1,Ky][B,C+1,K_y] availability of each clinical fact waveform embedding [B,C+1,DW][B,C+1,D_W] frozen waveform-derived representation waveform mask [B,C+1][B,C+1] availability of the waveform representation treatment item identities [B,C,NA][B,C,N_A] source-boundary active-treatment indices treatment numeric features [B,C,NA,25][B,C,N_A,25] lifecycle, exposure, amount, evidence, and timing features treatment token mask [B,C,NA][B,C,N_A] valid, non-padding treatment positions transition-valid mask [B,C][B,C] legal, non-padding source–target transitions encoded patient states z [B,C+1,D][B,C+1,D] observation-derived state sequence encoded actions α [B,C,D][B,C,D] treatment-set representations transition hidden states h [B,C,D][B,C,D] causal history-conditioned representations predicted patient states z [B,C,D][B,C,D] predictions aligned with state positions 2:C+12:C+1 predicted waveform embeddings [B,C,DW][B,C,D_W] next-window waveform representations clinical-fact logits [B,C,Ky][B,C,K_y] auxiliary next-window fact predictions laboratory-availability logits [B,C,VL][B,C,V_L] next-window item-observation predictions laboratory-value predictions [B,C,VL][B,C,V_L] dense normalized next-window values For transition index c, the inputs at state position c and action position c predict targets at state position c+1c+1. Padding masks are applied both within each sparse set and across the transition sequence; unavailable values are masked before encoding and do not contribute to their typed losses. Longer histories retain the most recent C valid transitions. During recursive evaluation, the observed tensors at future state positions remain scoring targets: after the initial window, the transition consumes the predicted patient state and causal continuous caches together with the externally supplied action sequence. Appendix C Comparison Methods And Controls Why Adaptation Is Necessary. The source methods were developed for different clinical data interfaces and prediction objectives. We adapt them to a common clinical prediction interface to compare how their retained dynamics or representation structures model patient evolution from matched information and supervision. Adaptation Principle And Fairness. For each comparison, we retain the defining computational core of the source method and replace the dataset-facing input and output interfaces. Methods are matched on patient splits, temporal windows, source-visible inputs, action timing, targets, missingness masks, eligible support, typed readouts, losses, optimization budget, validation-based selection, and test evaluation. Parameter counts and internal recurrence remain properties of the retained method structures. MIMIC-IV Comparisons. Table 11 summarizes the MIMIC-IV comparison contract. SepsisAgent, ACWM, and medDreamer are retrained on the common MIMIC-IV tensors after their native input and output layers are replaced by shared laboratory and clinical-fact encoders, the same frozen waveform encoder, the common action port, and typed readouts. Their recurrent or latent dynamics blocks retain the source-inspired update structure. SMB-Structure receives the same source-visible modality representations; its frozen predictive core is followed by an action-aware transition and the shared typed readouts for the rollout and action-planning experiments. Table 11: Adaptation and fairness contract for the MIMIC-IV comparisons. Each method retains its defining source-inspired core while using the stated common benchmark interface. Method Retained core Interface adaptation Fairness controls medDreamer Dreamer-style recurrent state-space dynamics Replace task-specific interfaces with shared laboratory and clinical-fact encoders, frozen waveform embeddings, action inputs, and typed readouts Same split, windows, visible inputs, action timing, targets, masks, training budget, selection rule, support, and evaluator SMB-Structure Frozen predictive representation Map the same laboratory, clinical-fact, and frozen-waveform representations into the frozen core, then fit an action-aware transition and shared typed readouts Same split, visible inputs, action timing, clinical targets, masks, support, and evaluator SepsisAgent Recurrent history-to-next-state update Replace native feature and output layers with shared laboratory and clinical-fact encoders, frozen waveform embeddings, action inputs, and typed readouts Same split, windows, visible inputs, action timing, targets, masks, training budget, selection rule, support, and evaluator ACWM Latent state-transition update Replace native signal/label interface with shared laboratory and clinical-fact encoders, frozen waveform embeddings, action inputs, and typed readouts Same split, windows, visible inputs, action timing, targets, masks, training budget, selection rule, support, and evaluator MIMIC-I Target-Specific Controls. MIMIC-I uses references appropriate to each predicted object rather than one pooled baseline. Waveform-embedding prediction is compared with copy-forward and state-only, aligned-action, and shuffled-action controls. Observed laboratory values are compared with last-value carry-forward and the training-only GapShrink prior. Clinical outcomes use train prevalence and recurrent clinical-prediction baselines evaluated with identical target masks. Laboratory availability uses mask-copy or training prevalence where applicable. Each comparison therefore preserves the target definition, eligible support, and metric used by PatientWM. Component And Modality Ablations. The five Core-7 ablations in Table 3 use the same MIMIC-I split, outcome definitions, readout architecture, optimization budget, checkpoint selection rule, and evaluator as the full configuration. The first group removes action conditioning or ordered state–action history. The second group restricts the observation encoder to laboratory values, clinical facts, or waveform embeddings. These controls isolate transition mechanisms and modality composition from changes in the downstream prediction task. External-Cohort Adaptations. The same adaptation rule is applied independently within ALOTT, VitalDB, and MC-MED. Dataset-specific encoders map the available variables into a common state width, after which each method applies its retained dynamics core and the same target-specific output heads. Within a dataset, methods share training and validation splits, visible inputs, target definitions, losses, optimization budget, checkpoint rule, and evaluator. ALOTT and VitalDB provide sequential examples. MC-MED has no verified adjacent-state sequence, so all temporal methods operate with context length one. None of these external interfaces provides a verified executed-action update; the action port is therefore explicitly known-empty for every method. Appendix D Experimental Protocol And Supplementary Results D.1 Evaluation Protocol Evaluation Roles. The experiments separate four roles that answer different questions. Factual transition evaluation measures next-state prediction along recorded trajectories. Typed readout evaluation measures requested clinical events, outcomes, and continuous values from the learned state. Recursive rollout evaluation repeatedly advances generated states without consuming future observations. The action-planning evaluation passes a bounded set of externally supplied candidates through the same transition model and measures agreement with the recorded clinician action. Temporal Evaluation Modes. Teacher-forced evaluation encodes each legal observed source state and scores the immediately following target. Logged-action free-running evaluation observes the initial prefix once, then reuses generated patient states and causal caches while supplying the recorded source-boundary action externally at every transition. The principal rollout tables report 1, 4, 8, and 16 six-hour transitions, corresponding to 6, 24, 48, and 96 hours. Future observations, values, and availability masks are joined after prediction for scoring. Generated states and causal-cache entries remain model outputs and are not relabeled as observed clinical records. Aligned, removed, and patient-shuffled action controls change only the action input and probe factual action dependence. Target-Specific Metrics. Average precision (AP) is the empirical precision–recall area used here as AUPRC. For target q, after sorting its NqN_q predictions by decreasing score, we compute APq=∑r=1Nq(Rq,r−Rq,r−1)Pq,r,MacroAUPRC=1|eval|∑q∈evalAPq,AP_q= _r=1^N_q (R_q,r-R_q,r-1 )P_q,r, = 1|Q_eval| _q _evalAP_q, (36) where Pq,rP_q,r and Rq,rR_q,r are precision and recall at rank r, and evalQ_eval contains heads with valid evaluation support. Macro-AUROC uses the same unweighted aggregation over per-target AUROC. AUPRC is emphasized for imbalanced binary targets, while AUROC measures case–control ranking (Davis and Goadrich, 2006; Saito and Rehmsmeier, 2015; Harutyunyan et al., 2019). For predicted probability pi,qp_i,q and binary target yi,qy_i,q, probability accuracy is measured by Brierq=1Nq∑i=1Nq(pi,q−yi,q)2,Brier_q= 1N_q _i=1^N_q (p_i,q-y_i,q )^2, (37) with lower values indicating better probabilistic accuracy (Glenn and others, 1950). The Common Event-7 interface contains six shared acute next-window events and dynamic ICU mortality: acute deterioration, new invasive ventilation, acute pressor initiation, circulatory pressor initiation, respiratory IMV escalation, Sepsis-3 onset, and mortality. Mortality retains its native six-hour gap and 24-hour prediction horizon. Dataset-specific transition heads are not included in this cross-dataset macro. The per-target metrics use each head’s observed support before the seven values are macro-averaged. Observed-laboratory nMAE is absolute error after item-wise normalization using training statistics and is evaluated only for item–target pairs measured in the future. This avoids aggregating errors across incompatible physical scales (Hyndman and Koehler, 2006). Let μj _j and σj _j be the training-set mean and scale for laboratory item j. Over the observed future item–target set Ωh _h at horizon h, nMAEh=1|Ωh|∑(i,j)∈Ωh|x^i,j,h−μjσj−xi,j,h−μjσj|.nMAE_h= 1| _h| _(i,j)∈ _h | x_i,j,h- _j _j- x_i,j,h- _j _j |. (38) Laboratory-availability AUPRC and Brier score the observation mask rather than an unmeasured numeric value. The bounded candidate-action probe reports tie-aware top-k recovery of the recorded action. Supplementary diagnostics additionally report risk-rule and dose-token agreement where those fields are defined. These quantities measure behavioral agreement within the supplied candidate set. Let ℐcovI_cov contain cases for which the recorded action is represented in the bounded candidate set. If ai⋆a_i is the recorded action, si(a)s_i(a) is the candidate burden score, and si,(k)s_i,(k) is the kkth-smallest score for case i, then Agree@k=1|ℐcov|∑i∈ℐcov[si(ai⋆)≤si,(k)].Agree@k= 1|I_cov| _i _cov1\! [s_i(a_i )≤ s_i,(k) ]. (39) The non-strict inequality includes all candidates tied at the kkth-lowest score. Metrics from different target types are never summed or averaged into a single system score because their semantics and scales differ. D.2 Factual Readouts And Prediction Controls Clinical Outcomes, Readout Design, And Probability Accuracy. Tables 12–14 provide the per-target results underlying the aggregate clinical comparison in Table 1 and Figure 3. Each target is scored on its observed test support, and the macro metrics are unweighted averages across the seven targets in the corresponding panel. The panels retain their dataset-specific target horizons, so results are summarized within each panel. Table 12: Per-target MIMIC-I Core-7 test results at the native target horizons. Macro scores are unweighted averages across the seven targets; the AUROC rows correspond to Figure 3(a). AUPRC ↑ Method Hypercapnic acidemia Severe acidemia Creatinine AKI Urine-output AKI GCS decrease Incident CAM+ 24-hour mortality Macro AUPRC medDreamer 0.0581 0.0262 0.0226 0.2236 0.0102 0.0401 0.0244 0.0579 SMB-Structure 0.0637 0.0312 0.0155 0.2484 0.0115 0.0597 0.0170 0.0638 SepsisAgent 0.0660 0.0404 0.0270 0.2518 0.0198 0.1003 0.0354 0.0772 ACWM 0.0647 0.0389 0.0227 0.2637 0.0194 0.0982 0.0261 0.0762 PatientWM (Ours) 0.2556 0.1747 0.1192 0.5179 0.0386 0.1092 0.3440 0.2227 AUROC ↑ Method Hypercapnic acidemia Severe acidemia Creatinine AKI Urine-output AKI GCS decrease Incident CAM+ 24-hour mortality Macro AUROC medDreamer 0.5012 0.4821 0.6079 0.5859 0.3527 0.4419 0.6192 0.5130 SMB-Structure 0.5405 0.5408 0.5124 0.6359 0.4710 0.5500 0.5380 0.5412 SepsisAgent 0.5790 0.6139 0.6632 0.6462 0.5671 0.6166 0.6479 0.6191 ACWM 0.5757 0.6136 0.6431 0.6452 0.4620 0.6251 0.5971 0.5945 PatientWM (Ours) 0.8143 0.8095 0.8278 0.8204 0.7239 0.7006 0.8850 0.7974 BRIER ↓ Method Hypercapnic acidemia Severe acidemia Creatinine AKI Urine-output AKI GCS decrease Incident CAM+ 24-hour mortality Macro BRIER medDreamer 0.2910 0.2654 0.2169 0.2753 0.0621 0.1851 0.2387 0.2192 SMB-Structure 0.2571 0.2306 0.2334 0.2267 0.0809 0.2863 0.2005 0.2165 SepsisAgent 0.2679 0.2379 0.2143 0.2555 0.0848 0.2277 0.2390 0.2181 ACWM 0.2828 0.2428 0.1940 0.2767 0.0958 0.2166 0.2559 0.2235 PatientWM (Ours) 0.1062 0.0668 0.0387 0.1711 0.0120 0.0741 0.0641 0.0761 Table 13: Per-target MIMIC-I Common Event-7 test results. Macro scores are unweighted averages across the seven targets; the AUROC rows correspond to Figure 3(b). Method Any acute deterioration New invasive ventilation New pressor Circulation pressor Respiratory IMV Sepsis-3 onset Dynamic ICU mortality Macro MIMIC-I Common Event-7 Test Panel: AUPRC ↑ medDreamer 0.0423 0.0335 0.0185 0.0234 0.0276 0.0317 0.0178 0.0278 SMB-Structure 0.0733 0.0725 0.0302 0.0495 0.0518 0.0694 0.0187 0.0522 SepsisAgent 0.0656 0.0530 0.0477 0.0507 0.0416 0.0786 0.0250 0.0517 ACWM 0.0517 0.0324 0.0404 0.0501 0.0437 0.0668 0.0256 0.0444 PatientWM (Ours) 0.0691 0.0620 0.0409 0.0718 0.0483 0.0758 0.0505 0.0598 MIMIC-I Common Event-7 Test Panel: AUROC ↑ medDreamer 0.4873 0.4981 0.4679 0.5058 0.5029 0.5276 0.5449 0.5049 SMB-Structure 0.5431 0.5707 0.5910 0.6387 0.5849 0.7157 0.5557 0.6000 SepsisAgent 0.6352 0.6554 0.6841 0.7040 0.6469 0.7135 0.6028 0.6631 ACWM 0.4869 0.4889 0.5822 0.6682 0.6001 0.6802 0.6090 0.5879 PatientWM (Ours) 0.6340 0.6984 0.6971 0.7493 0.6597 0.7044 0.6428 0.6837 MIMIC-I Common Event-7 Test Panel: Brier ↓ medDreamer 0.2721 0.2604 0.3267 0.2306 0.2329 0.2884 0.2304 0.2631 SMB-Structure 0.2303 0.2576 0.2068 0.1818 0.3104 0.1280 0.2797 0.2278 SepsisAgent 0.2455 0.3527 0.1867 0.2775 0.2492 0.2184 0.1901 0.2457 ACWM 0.2483 0.2612 0.3150 0.1509 0.2351 0.2689 0.2313 0.2444 PatientWM (Ours) 0.2069 0.2197 0.1937 0.1369 0.2351 0.2151 0.1673 0.1964 Table 14: Per-target MIMIC-IV Common Event-7 test results. Macro scores are unweighted averages across the seven targets; the AUROC rows correspond to Figure 3(c). Method Any acute deterioration New invasive ventilation New pressor Circulation pressor Respiratory IMV Sepsis-3 onset Dynamic ICU mortality Macro MIMIC-IV Common Event-7 Test Panel: AUPRC ↑ medDreamer 0.0653 0.0428 0.0253 0.0257 0.0346 0.0272 0.0178 0.0341 SMB-Structure 0.0901 0.0589 0.0548 0.0514 0.0528 0.0494 0.0657 0.0604 SepsisAgent 0.2733 0.1193 0.2093 0.2044 0.1228 0.0819 0.2087 0.1742 ACWM 0.2725 0.1202 0.2107 0.2042 0.1259 0.0827 0.2214 0.1768 PatientWM (Ours) 0.2816 0.1253 0.2066 0.1970 0.1265 0.0843 0.2278 0.1784 MIMIC-IV Common Event-7 Test Panel: AUROC ↑ medDreamer 0.5347 0.5111 0.4734 0.4891 0.5293 0.5094 0.4877 0.5050 SMB-Structure 0.5994 0.5933 0.6661 0.6635 0.6044 0.6460 0.6898 0.6375 SepsisAgent 0.8105 0.7341 0.8416 0.8389 0.7535 0.7478 0.8531 0.7971 ACWM 0.8089 0.7339 0.8404 0.8378 0.7556 0.7446 0.8504 0.7959 PatientWM (Ours) 0.8124 0.7375 0.8428 0.8394 0.7563 0.7477 0.8552 0.7988 MIMIC-IV Common Event-7 Test Panel: Brier ↓ medDreamer 0.2817 0.2631 0.2595 0.2519 0.2555 0.2181 0.2528 0.2547 SMB-Structure 0.2889 0.2566 0.2182 0.2314 0.3242 0.2821 0.1987 0.2572 SepsisAgent 0.1933 0.2746 0.1404 0.1408 0.2360 0.2004 0.1266 0.1875 ACWM 0.1976 0.2632 0.1436 0.1465 0.2257 0.1944 0.1212 0.1846 PatientWM (Ours) 0.1699 0.2216 0.1387 0.1414 0.1845 0.2007 0.1358 0.1704 External Point Estimates. Figure 6 in the main text summarizes the external AUROC comparison. Table 15 reports the exact test AUPRC, AUROC, and Brier point estimates. PatientWM leads the listed adapted clinical world-model baselines on all three datasets. Table 15: External test evaluation of PatientWM and available clinical-model baselines. All rows use the same per-target observed test support within each dataset. Bold marks the best value within each dataset and metric. Dataset Method Macro AUPRC ↑ Macro AUROC ↑ Brier ↓ ALOTT medDreamer 0.0675 0.5659 0.0474 SMB-Structure 0.0651 0.5672 0.0525 SepsisAgent 0.2007 0.7864 0.0451 ACWM 0.1838 0.7757 0.0451 PatientWM (Ours) 0.2268 0.7915 0.0437 VitalDB medDreamer 0.0538 0.5082 0.0441 SMB-Structure 0.0540 0.5160 0.0545 SepsisAgent 0.0859 0.6114 0.0436 ACWM 0.0868 0.6264 0.0435 PatientWM (Ours) 0.1000 0.6725 0.0426 MC-MED medDreamer 0.2049 0.5039 0.1485 SMB-Structure 0.2493 0.5726 0.1963 SepsisAgent 0.3239 0.6600 0.1391 ACWM 0.3232 0.6565 0.1384 PatientWM (Ours) 0.3423 0.6764 0.1352 D.3 Rollout And Action-Planning Details Multi-Method Recursive Rollout. Figure 4 in the main text summarizes the multi-method rollout comparison. Table 16 reports the exact Event-7 and laboratory values below. All methods are evaluated on the same trajectories, horizons, recorded actions, targets, and evaluator. Table 16: MIMIC-IV multi-step rollout on the same test trajectories and targets. We report recursive Event-7 macro-AUPRC and observed-laboratory nMAE from the typed laboratory decoder. Bold marks the best value in each column. Event-7 macro-AUPRC ↑ Observed-laboratory nMAE ↓ Method 6 h 24 h 48 h 96 h 6 h 24 h 48 h 96 h medDreamer 0.0474 0.0733 0.1038 0.1052 0.4243 0.4527 0.5093 0.5728 SMB-Structure 0.1141 0.0546 0.0493 0.0425 0.6528 0.7082 0.8018 0.9533 SepsisAgent 0.1637 0.0844 0.0927 0.0939 0.4191 0.4520 0.5085 0.5690 ACWM 0.1378 0.0665 0.0761 0.0714 0.4240 0.4519 0.5084 0.5729 PatientWM (Ours) 0.1755 0.1005 0.1090 0.1308 0.4093 0.4161 0.4556 0.4958 Action-Planning Evaluation. Table 4 summarizes the main-text result. The action-planning evaluation includes 15,951 test trajectories for each method. Its candidate provenance hash is identical across methods. Candidate actions come only from a prespecified empty update or the same patient’s earlier recorded history. The current action, actions from other patients, and future outcomes are not used to create or score candidates. All methods use the source-boundary boundary_known action view. The primary denominator is 13,143 test decisions with non-empty recorded treatment updates. Item-level candidate coverage is 0.5573 (95% patient-cluster interval 0.5458–0.5672); coverage across all decisions, including empty updates, is 0.6352. Ties are retained. The event score averages predicted sepsis-onset and cardiac-arrest probabilities. The laboratory score is the predicted absolute standardized value, weighted by predicted measurement probability. The combined score first scales both burdens within each candidate set and then adds a fixed uncertainty penalty shared by all methods. The main comparison uses shared host inputs, heads, and evaluator for all methods. SMB-Structure passes its frozen predictive state through an action-aware transition before scoring each candidate. Under the fixed score, PatientWM obtains the highest covered top-1 and top-2 agreement. Surrogate Boundary. This diagnostic follows a common limitation of retrospective treatment-planning evaluations: recorded clinician regimens, expert preference, guideline adherence, and off-policy value estimates can support behavioral plausibility, but they do not directly reveal the true outcome of an unexecuted action for the same patient. The observed record contains the outcome under historical care, while alternative candidate outcomes remain unobserved. We therefore interpret recorded-action agreement as a clinician-behavior surrogate, not as evidence of treatment optimality, counterfactual correctness, or patient-level action quality. Action-conditioned systems such as SepsisAgent and CLARITY (Wu et al., 2026; Ding et al., 2026) can query alternative actions and report policy or regimen surrogates, but without additional causal identification, support diagnostics, or randomized-like validation, model-predicted risk or agreement with historical or expert choices does not by itself establish treatment benefit. D.4 Frozen-State Reuse Table 17 reports the exact values summarized in Figure 5. The frozen PatientWM representation is trained without clinical-event label loss and evaluated with a newly trained Common Event-7 readout. All rows use the same seven targets, 14,044 test examples, per-target observed and positive counts, and evaluator. The full PatientWM row provides an in-family reference for the cost of removing event supervision during source training; the remaining rows reproduce the baseline results from Table 1. Table 17: MIMIC-I Common Event-7 frozen-state transfer on the Table 1 test support. Bold marks the best result in each column. Method Macro AUPRC ↑ Macro AUROC ↑ Macro Brier ↓ medDreamer 0.0278 0.5049 0.2631 SMB-Structure 0.0522 0.6000 0.2278 SepsisAgent 0.0517 0.6631 0.2457 ACWM 0.0444 0.5879 0.2444 PatientWM (Ours) 0.0598 0.6837 0.1964 PatientWM (Frozen transfer) 0.0500 0.6122 0.2401 D.5 Label-Efficient Cross-Dataset Adaptation The cross-dataset experiment transfers the Common Event-7 prediction interface from MIMIC-I to MIMIC-IV while keeping each source predictor frozen. For every method, the only target-domain component learned is the same residual logit adapter with 56 trainable parameters. The adapter is trained on 328 examples, corresponding to 1% of a fixed 32,768-example MIMIC-IV training support. All methods use the same target examples, seven outcome definitions, and evaluator. This protocol isolates label-efficient cross-cohort reuse: the source representation and prediction interface remain fixed, while the small target-domain adapter accounts for cohort-level shifts in the output space. The aggregate MIMIC-IV test results are reported in Table 5. Appendix E Reproducibility And Artifacts We report the data splits and model parameters used for the principal experiments. The MIMIC-I transition model uses all 32 training shards with equal contribution and is evaluated on all 32 test shards. The test split contains 3,827 subjects, 5,337 ICU episodes, and 14,317 transition trajectories; paired waveform support is 2,291/1,911/1,401/877 targets at 6/24/48/96 hours. The model uses micro-batches of eight with four-step gradient accumulation, AdamW learning rate 3×10−43× 10^-4, weight decay 0.01, and a closed-loop objective that advances up to 16 six-hour windows from one observed initial state. The MIMIC-IV transition model for recursive rollout and action-planning evaluation uses 15,995 validation trajectories and 15,951 common test trajectories across all 32 shards. It uses a two-layer width-128 GRU, dropout 0.2, context length 16, ordered state–action history, concatenated transition fusion, and gated residual latent update. It trains with batch size eight, AdamW learning rate 3×10−43× 10^-4, weight decay 0.01, and a 16-window closed-loop term weighted by 0.50. For the MIMIC-IV laboratory results in Table 2, the typed laboratory decoder trains with the transition and GapShrink prior frozen. It has 36,483 trainable parameters, a 32-dimensional item query, dual item–cross-laboratory attention, and source-only trend, history, count, window-summary, and laboratory-ledger features. The decoder is selected by validation all-target laboratory nMAE. Exact common test support is 246,487/149,654/105,953/58,858 observed item-targets at 6/24/48/96 hours. For the MIMIC-I Core-7 results in Table 1, the clinical-outcome readout uses 2,283 source features and appends 228 source-history features for 2,511 inputs. The appended source-history features are available for 106,753 of 107,859 training windows and 14,033 of 14,122 test windows; the remaining rows retain the availability indicator off. Each outcome uses 120 histogram-gradient-boosting iterations, 31 leaves, learning rate 0.05, minimum leaf size 20, L2=1L_2=1, and balanced class weights. For the external-dataset evaluations in Figure 6, the models use context length eight, shared host width 192, batch size 64, and AdamW learning rate 2×10−42× 10^-4. PatientWM uses three layers and six attention heads. SepsisAgent, medDreamer, and ACWM receive the same host state and known-empty external-action port, while SMB-Structure maps the same host representation through its frozen predictive core. Every learned row shares a balanced binary outcome loss plus 0.10-weighted next-EHR and next-waveform SmoothL1 losses where adjacent targets exist. Dataset-specific patient-group training, validation, and test splits are used for all external evaluations.