Learning a patient world model for health system simulation in intensive care
Authors: Yingtai Li, Ruoyan Wang, Yu Li, Shaohua Kevin Zhou
Organizations: School of Biomedical Engineering, Division of Life Sciences and Medicine, University of Science and Technology of China (USTC), Hefei, Anhui, 230026, China · Medical Imaging, Robotics, Analytic Computing & Learning (MIRACLE) Lab, YRD-RIGHT, Suzhou Institute for Advanced Research, USTC, Suzhou, Jiangsu, 215123, China · School of Population Medicine and Public Health, Chinese Academy of Medical Sciences & Peking Union Medical College, Beijing, 100006, China · Innovation Center for Philosophy and Social Sciences in Higher Education, Innovation Center for Public Health Emergency Management, Beijing, 100006, China · State Key Laboratory of Respiratory Health and Multimorbidity, Beijing, 100006, China · Jiangsu Key Laboratory of Multimodal Digital Twin Technology, Suzhou, Jiangsu, 215123, China · State Key Laboratory of Precision and Intelligent Chemistry, USTC, Hefei, Anhui, 230026, China · Biomedical Basic Research Center (BBRC) of Jiangsu Province, Suzhou, Jiangsu, 215123, China
Designing healthcare strategies requires anticipating their clinical and operational consequences before implementation, a task that often relies on health system simulation. Such a simulation necessitates modeling how patient states evolve and generate care demands. However, conventional approaches typically tailor compact state transitions to a single policy question and omit history-dependent dynamics, thus requiring the redefinition of states and transitions when there is a new question that entails different clinical information. We propose to learn a patient world model from longitudinal EHRs using a neural network to provide a shared clinical representation of patient trajectories that can serve simulation questions across scenarios. Taking ICU operations as an example, we train the model on MIMIC-IV, freeze it, and couple its time-queryable predictions to shared ICU calendars, resources, and queues via scenario-specific mappings, so that one model renders admission demand, care tasks, and discharge proposals across simulations evaluating ICU bed capacity, care dispatch, and post-ICU transfer capacity. Experiments show that our model matches Markov and semi-Markov comparators in aggregate occupancy and directional responses to capacity and dispatch policies. It also reproduces policy gains from replayed observed trajectories. For patients entering the same support state from different prior trajectories, the learned representation captures a 24-h recovery difference (-6.4 predicted versus -6.7 observed percentage points), a contrast missed by compact state models, which propagates into distinct simulated care demand, queue burden, and bed use. These results indicate that a learned world model of patient dynamics has the potential to replace question-specific state models, allowing new care policies to be evaluated by reconfiguring the scenario without model rebuilding.
Figures & tables
Figure 1: A shared patient world model for different ICU simulation questions. a , Longitudinal EHRs are encoded as sparse state–event sequences over a shared clinical vocabulary, preserving variable values, events with categorical labels and irregular observation times to train a generative patient world model conditioned on history and forecast interval Δ . b , With learned parameters held fixed, querying the model at a requested future time returns state and event distributions, from which sampled updates are appended to history before the next query. Repeated rollouts from a single patient’s anchor history generate alternative individual futures; rollouts across patients construct a virtual cohort. c , Scenario-specific mappings connect outputs of the frozen patient model to admission demand, staff care tasks and ICU exit within a shared discrete-event simulator linking pre-ICU waiting, beds, staff queues and downstream service slots ( Discrete-event simulation and delivery feedback ). The schematic illustrates the full framework architecture; admission, care delivery and discharge are evaluated in separately configured scenarios. In subsequent figures, “Neural model” denotes this generative patient world model.
Figure 2: Clinical queries from a frozen patient model. a , AUROC across six clinical endpoints, each evaluated on 1,024 held-out test records from 956–981 patients. For ICU mortality, 0 and 24 h denote observed history after ICU admission. b , Brier score differences (model minus empirical-frequency reference); negative values favor the model. c , Uncalibrated and recalibrated slopes across 24 binary task–horizon queries; connecting lines pair estimates for the same query and the dashed line marks slope one. Whiskers in a – c show 95% patient-cluster bootstrap intervals from 1,000 resamples. All queries and calibration settings share the same frozen model parameters.
Figure 3: Longitudinal generation of virtual ICU cohorts. Sixteen recalibrated 14-day rollouts are generated from each of 670 held-out test patients at ICU admission (10,720 trajectories). a , Day-14 support-transition and terminal-event incidence; lines pair observed and model estimates. b , Fraction of patients remaining in the ICU over 14 days. c , Patient-level heterogeneity: generated-to-observed variance log-ratios for per-patient support initiations ( 0 indicates equal variance), dual-support occupancy as a fraction of follow-up patient-hours, and mean duration of respiratory episodes ending with an explicit stop (unadjusted for terminal censoring). Whiskers and bands show 95% patient-bootstrap intervals from 1,000 resamples.
Figure 4: Operational responses across admission, care-delivery and discharge simulations. Headers above each column indicate the mapped patient-model output and the varied operational control across separately configured scenarios, comparing the frozen neural model with task-fitted Markov and semi-Markov baselines. a , ICU admission fraction and restricted pre-ICU waiting across staffed-bed capacities, sharing pre-ICU admission requests from the calibrated neural model (bands, repeat-bootstrap 95% intervals over eight replications). b , Acute-first-minus-FIFO differences in acute and non-acute queue burden (cumulative task-waiting hours per initial patient) across staffing levels using model-specific trajectories with delivery feedback disabled ( 128 test patients, 128 beds, 168-h horizon; whiskers, paired Student- t 95% intervals over 16 replications). c , Mean time from live-exit proposal to bed release (among completed transfers) and ICU admission fraction across one to four concurrent post-ICU transfer slots with a 4-h transfer duration (bands, Student- t 95% intervals over eight replications).
Figure 5: Bed-and-staff capacity interactions and observed-trajectory operational replay. The upper schematic illustrates how expanding bed access shifts demand into staff-constrained care queues when staffing is fixed. a , Restricted bed-request waiting (hours per request) and started non-acute task waiting (hours per started task) across bed capacities ( 16 versus 64 beds) and staffing levels at two arrival intensities ( 64 arrivals over 72 h or 24 h ). b , Paired capacity-expansion effects on bed waiting and on acute and non-acute queue burden (cumulative task-waiting hours per initial patient) relative to the 16-bed , low-staffing baseline. c , FIFO-minus-acute-first reductions in acute queue burden driven by observed versus generated trajectories for the same 128 held-out patients ( 128 beds, 168-h horizon). Points in a , b and bars in c show means; whiskers show Student- t 95% intervals across 16 replications (conditional on the fixed cohort in c ).
Figure 6: Retained clinical history within a shared support state and its operational consequences. a , Matched entry paths in 404 held-out test pairs sharing the same single-organ support state St and hours since ICU admission at state entry ( Dt=0 ), entered either from no support (upward entry) or from dual support (downward entry). Schematics compare inputs of the Markov ( St and ICU hours), semi-Markov (adding state dwell time Dt , clock) and neural models (multivariable history Ht and query interval Δ ). HR, heart rate; MAP, mean arterial pressure. b , Upward-minus-downward 24-h differences in recovery (reaching no-support status within 24 h ) and terminal exit (death or live ICU discharge). Thin capped whiskers show patient-sampling 95% bootstrap intervals; thick segments show Monte Carlo 95% intervals across 16 paired replications. c , Upward-minus-downward differences in support-related task requests under unconstrained staffing (left), and in acute queue burden (middle) and occupied bed time (right) per patient under shared bed and staff constraints across staffing levels (observed whiskers, patient-paired 95% bootstrap intervals; model whiskers, paired Student- t95% intervals across 16 replications). pp, percentage points.
Figure S2 : Hourly five-state occupancy in generated virtual cohorts. a–e , Hourly cohort proportions in no support, single-organ support, dual support, live ICU discharge and death over 14 days, comparing observed clinical courses with raw and recalibrated trajectories across 670 test patients (16 rollouts per patient per variant, averaged within each patient and weighted equally across patients). Live discharge and death are absorbing states; vertical axis ranges differ across panels.
Figure S3 : Generation depth, finite-rollout precision and trajectory diversity. a , Event-frequency total variation and transition Jensen–Shannon divergence across prefix caps of 1 to 256 events in the 14-day trajectories of 670 test patients (16 rollouts per patient per variant; after recalibration, drift increased by 0.0379 [95% CI, 0.0325–0.0421] and 0.0238 [0.0197–0.0275] per cap doubling, respectively). Shaded bands show 95% patient-bootstrap intervals. b , Estimated Monte Carlo standard errors (MCSEs) for raw and recalibrated cohort means at budgets of R=1,2,4,8 and 16 rollouts per patient, computed as MCSE(R)=∑i=1Nsi2/(N2R) , where si2 is the sample variance across the 16 generated rollouts for patient i ( N=670 fixed patients; at R=16 , recalibrated MCSEs equal 33–54% of the corresponding patient-bootstrap standard errors). c , Within-patient trajectory diversity across the 16 rollouts per patient over 1 to 14 days, measured by the proportion of unique trajectory signatures, the modal-signature fraction and the mean pairwise Jaccard distance, averaged equally across patients. Signatures comprise ordered event types, sampled event-label bins and 6-h time bins (with order and multiplicity ignored for Jaccard distance); when all 16 trajectory signatures for a patient are distinct, the modal fraction reaches its minimum of 1/16 (6.25%, dotted line). Shaded bands show 95% patient-bootstrap intervals from 1,000 resamples. pp, percentage points.
Figure S4 : Markov and semi-Markov comparator benchmarks and support-onset timing. a , Cohort-level benchmark across 256 test ICU stays with matched admission-state initialization over 14 days, comparing hourly five-state occupancy mean absolute error (MAE), restricted mean ICU length of stay, and cumulative live-discharge and death incidence MAE. Whiskers show 95% patient-bootstrap intervals from 1,000 resamples. b , First- start and all- start 1-Wasserstein distances (in hours) over 7 days for 244 eligible test patients without baseline support. All start s include recurrences; marker shapes identify endpoints, lines connect estimates within each model, and whiskers show 95% patient-bootstrap intervals.
Figure S5 : Resource requirements and execution sensitivity. a , Role-specific task requirements and triangular service-time distributions (minimum/mode/maximum in minutes: admission—nurse 25/35/45, attending physician 15/20/25 and resident 10/15/20; hourly monitoring—nurse 6/8/10; 12-h rounding—resident 10/12.5/15; acute care—nurse 5/7.5/10, attending physician 5/10/15 and resident 5/10/15; simulation workflow in Discrete-event simulation and delivery feedback ). b , Paired 30- and 120-min clinical update-interval contrasts relative to the 60-min reference across 128 test patients over 168 h (whiskers, paired Student- t 95% intervals across four replications). c , FIFO-minus-acute-first reductions in acute queue burden per initial patient with delivery feedback disabled or enabled ( 128 test patients, 128 beds, 168-h horizon; bars and whiskers, means and paired Student- t 95% intervals across 16 replications). At lower staffing, enabling feedback reduces the benefit from 13.41 to 9.02 h per patient (paired attenuation 4.39 h , 95% interval [3.79,5.00] ).
Figure S6 : Organ-support triggers underlying matched-history care-demand differences. Values are acute trigger groups per initial patient over 24 h across 404 matched test pairs (neural estimates use 16 paired replications under unconstrained staffing). Simultaneous support signals for a patient were grouped into a single trigger group, each generating one task request for each of the three staff roles; 14 observed groups contained multiple support-event types, so the stop and change categories overlap and deduplicated summaries count each group once. a , Upward-minus-downward differences in observed and neural-generated demand by trigger group, showing that lower total acute demand in upward-entry patients was driven primarily by fewer stop - and change -related groups. Whiskers show 95% intervals from 2,000 pair-level bootstrap resamples for observations and paired Student- t Monte Carlo (MC) intervals across the 16 replications for neural estimates; the dashed line marks zero. † Actual- start values are point estimates without confidence intervals; counting all proposed neural start events instead gives 1.9692 versus 1.8773 per patient (contrast +0.0919 , MC 95% CI [0.0339,0.1499] ). b , Absolute mean demand for deduplicated stop -or- change and all trigger groups in upward-entry (filled circles) and downward-entry (open circles) patients.
Medical diagnosis and treatment are dynamic processes in which patient states evolve over time and clinical interventions alter future outcomes. Although current medical AI can detect disease, estimate risk and generate reports, many systems still return static labels or scores, offering limited insight into how illness may progress or how alternative interventions may reshape its trajectory. Medical world models adapt the world-model idea from artificial intelligence to healthcare by learning internal simulators of patient-state dynamics. Their long-term goal is to help clinicians anticipate deterioration, compare treatment-conditioned futures and tailor care to individual patients. Yet relevant work remains scattered across foundation models, longitudinal modelling, disease simulation, treatment-effect estimation, reinforcement learning and digital twins. To bridge this gap, this review outlines a roadmap for advancing medical AI from isolated diagnosis and prediction toward medical world models that simulate disease evolution and support intervention decisions. This roadmap is organized around three coupled capabilities: patient-state construction, clinical dynamics modelling and intervention decision support. Across representative systems, the comparison highlights what each capability contributes and how partial components can be integrated into more mature perception--dynamics--planning systems. Finally, we identify the challenges involved in turning plausible rollouts into clinically useful simulators. Related literature is available at https://github.com/1999kevin/awesome_medical_world_models.
Ke Liu, Mengxuan Li, Yanyi Bao +4
College of Computer Science, Zhejiang University, Hangzhou, China. · School of Medicine, Zhejiang University, China. · Department of Biomedical Informatics, Harvard University, USA.
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.
Ziwen Kan, Jie Peng, Bo Yan +5
University of Science and Technology of China · University of Central Florida · Princeton University +2
Long-horizon clinical simulation -- predicting how a patient's physiology evolves over years under specified interventions -- is central to chronic-disease care, yet existing electronic health record (EHR) models are predominantly discriminative, and general-purpose large language models drift under repeated interventions. We propose the ChronoMedicalWorld Model (CMWM), an action-conditioned latent world-model framework for learning patient trajectories from longitudinal care data. CMWM couples a joint-embedding state encoder with a wide action encoder that admits both structured intervention indicators and free-text communication embeddings, and trains a recurrent latent transition module under a six-term objective: next-observation supervision, next-latent prediction, SIGReg latent regularisation, and three physiology-aware shape priors (slope, continuity, large-jump penalty). A closed-loop rollout-prefix protocol matches training to deployment, so the model is optimised against the same multi-step error it exhibits at inference. As a concrete case study, we instantiate CMWM for annual estimated glomerular filtration rate (eGFR) trajectory forecasting in chronic kidney disease (CKD). On a 2{,}232-patient nephrology cohort, the CKD instantiation achieves a dynamic-50% history rollout test mean absolute error (MAE) of 7.384 and root-mean-square error (RMSE) of 10.256, against 7.964 and 11.069 for a tuned GPT-5.5 structured-prompting baseline (−7.28% MAE, −7.35% RMSE), with the gain dominated by the dialogue portion of patient--health-coach communication. The framework is not CKD-specific: its architecture, loss design, and training protocol apply to any chronic condition that can be cast as periodic clinical state interleaved with structured and conversational interventions.
Jiangyuan Wang, Xuyong Chen, Junwei He +3
Beijing KidneyTec Medical Technology Co., Ltd., Beijing, China