Multi-Disease Health System Models

• Multi-disease health system models are computational frameworks that jointly represent multiple diseases and health system responses, capturing multimorbidity and intervention effects.
• They integrate diverse data substrates including EHRs, imaging, and wearable sensor data to simulate disease trajectories and predict clinical outcomes.
• Advanced paradigms such as transformer-based sequence prediction, reinforcement learning, and modular digital twins enhance model accuracy and healthcare delivery simulation.

Multi-disease health system models (HSMs) are computational models that jointly represent multiple diseases and, in some formulations, “capture how healthcare delivery systems respond to population health needs” (Chalkley et al., 15 Aug 2025). In the recent literature, the term encompasses whole-population individual-based simulation, compartmental disease trajectory models, multi-state survival models, foundation models for disease trajectories, multimodal graph predictors, reinforcement-learning treatment policies, and modular health digital twins (Chalkley et al., 15 Aug 2025, Ledebur et al., 2024, Zhu et al., 14 May 2026, Wang et al., 9 Jun 2026). Across these variants, the unifying objective is to model multimorbidity, longitudinal disease evolution, cross-organ dependence, and system-level constraints more faithfully than single-disease or single-organ approaches.

1. Conceptual scope and formal objects

A central organizing concept is multimorbidity: chronic diseases “frequently co-occur in patterns that are unlikely to arise by chance,” and these patterns can be represented as longitudinal health states rather than isolated diagnoses (Einsiedler et al., 8 Oct 2025). In the compartmental disease trajectory model (CDTM), compartments correspond to multimorbidity patterns, and the model describes chronic disease trajectories across 132 distinct multimorbidity patterns (compartments) derived from 131 ICD-10 diagnostic groups (A00–N99) (Ledebur et al., 2024). Each individual occupies one compartment per year, and transitions are governed by empirically estimated age- and sex-specific probabilities $q_{g,a,k,j}$, yielding a multilayer directed network of disease-state evolution (Ledebur et al., 2024).

A related formalism is the multi-state model, in which disease histories are modeled as transitions between discrete states such as onset, progression, competing events, and death. Multi-state models are specifically described as suitable for “transitions between different disease stages in presence of competing risks,” but their estimation is complicated by dependent left-truncation, multiple time scales, index event bias, and interval-censoring (Wiegrebe et al., 24 Sep 2025). In that framework, the transition hazard for transition $k$ is written as

$$h_k(t \mid \mathbf{x}_{i,k}) := \lim_{\Delta t \to 0} \frac{1}{\Delta t} P\left(t \leq T < t + \Delta t, o \mid T \geq t, \ell, \mathbf{x}_{i,k}\right),$$

and piecewise exponential additive models (PAMs) are extended to estimate such hazards under the stated observational complexities (Wiegrebe et al., 24 Sep 2025).

The same conceptual space also includes whole-population HSMs. The Thanzi la Onse (TLO) model of Malawi is described as “the first of its kind” among multi-disease HSMs that build on individual-level epidemiological models of multiple diseases while also modeling healthcare delivery (Chalkley et al., 15 Aug 2025). By contrast, foundation models such as DT-Transformer treat a patient’s EHR as an ordered event sequence and learn a single model for broad next-event prediction across hundreds of disease categories (Zhu et al., 14 May 2026). These variants differ in granularity and purpose, but all treat disease histories as coupled trajectories rather than independent endpoints.

2. Data substrates and state representations

Recent HSMs are distinguished by the breadth of their data substrates. DT-Transformer was trained on 57.1M structured EHR entries over 1.7M patients from Mass General Brigham (MGB), spanning 11 hospitals and a broad network of outpatient clinics (Zhu et al., 14 May 2026). Its inputs include diagnoses coded as ICD-9/10 and truncated to 3-character codes, age in days, masked static covariates such as sex, smoking status, and alcohol status, intermittent “no-event” tokens, and death events (Zhu et al., 14 May 2026). This representation turns disease trajectories into a temporally ordered token stream.

Other HSMs are explicitly multimodal. DiffDT integrates multi-organ sensor data from brain, heart, liver, and kidney with tokenized healthcare events and digitalized SDoH proxies, specifically ICD-10 Chapters Z and V–Y (Wei et al., 10 May 2026). The UK Biobank data used in that study include 44,834 brain, 23,987 heart, 28,722 liver, and 32,155 kidney imaging samples, together with nearly 500k medical history sequences spanning ages 25–89 years (Wei et al., 10 May 2026). HGDC-Fuse uses EHR and chest X-ray data from MIMIC-IV and MIMIC-CXR, constructing a patient-centric heterogeneous graph with EHR nodes, one CXR node per image, cross-modal edges encoding relative acquisition time, and inter-patient EHR–EHR edges linking similar patients (Jiang et al., 19 Sep 2025).

Population-level HSMs rely on registry-scale aggregation. CDTM uses Austria-wide data from approximately 45 million hospital stays spanning 17 years, while the comparative Austria–Denmark study defines yearly health states as binary diagnosis vectors over 131 disease blocks and derives 132 distinct, interpretable clusters per country via divisive hierarchical clustering (Ledebur et al., 2024, Einsiedler et al., 8 Oct 2025). TLO, by contrast, simulates individuals in a synthetic Malawi population and represents demand through disease and symptom status, health-seeking behavior, and health system interaction events that consume workforce time, consumables, equipment, and beds (Chalkley et al., 15 Aug 2025).

Wearable and movement sensing extends HSMs beyond EHR and imaging. ADH-MTL uses wearable sensor data and patient profile information to jointly assess diabetes, cardiovascular disease, high cholesterol, and depression in a multi-task setting (Chai et al., 20 Nov 2025). The gait foundation model uses 3D skeletal motion from a single depth camera during five standardized motor/postural tasks, generating embeddings from 26 3D joints over 900 frames (30 s) and testing them against 3,210 phenotypic targets spanning 18 body systems (Gabet et al., 26 Mar 2026). This broadening of input space suggests that HSM state representations are increasingly multi-scale, combining clinical events, sensor-derived phenotypes, and contextual covariates.

Paradigm	State representation	Data scope
CDTM (Ledebur et al., 2024)	Annual multimorbidity compartments and transition network	Austria-wide hospital stays
DT-Transformer (Zhu et al., 14 May 2026)	Temporal token sequence of diagnoses, ages, static covariates, no-event and death tokens	MGB health system EHR
DiffDT (Wei et al., 10 May 2026)	Tokenized event history plus multi-organ digital twins and SDoH proxies	UK Biobank
HGDC-Fuse (Jiang et al., 19 Sep 2025)	Patient-centric multimodal heterogeneous graph	MIMIC-IV and MIMIC-CXR
TLO (Chalkley et al., 15 Aug 2025)	Individual life-course simulation with health system interaction events	Synthetic Malawi population
OmniBioTwin (Wang et al., 9 Jun 2026)	Modular digital twins coupled in a multi-layer network	Multiscale HDT framework

3. Principal computational paradigms

A dominant line of work models disease histories as sequence prediction. DT-Transformer adapts the GPT architecture for disease trajectory prediction, using 12 transformer layers, 12 attention heads, an embedding dimension of 120, and approximately 2.2M parameters (Zhu et al., 14 May 2026). It predicts both the next event and time-to-event, combining cross-entropy for event prediction with negative log-likelihood for an exponential waiting-time model (Zhu et al., 14 May 2026). DiffDT also uses an autoregressive transformer for tokenized ICD sequences, with an adaptive tokenizer and embeddings

$$y_t = \mathrm{Embed}_\text{ICD}(s_t) + \mathrm{Embed}_\text{age}(\tau_t),$$

and an autoregressive loss

$$\mathcal{L}_{\mathrm{AR}} = -\sum_t \log p(s_{t+1} \mid s_0, \ldots, s_t; \phi)$$

to encode history before conditioning a generative sensor twin (Wei et al., 10 May 2026).

A second line uses generative mediation. DiffDT formalizes disease reasoning through

$$P(\text{future disease} \mid \text{history}) = \int P(\text{future disease} \mid \text{biomarker}) \; P(\text{biomarker} \mid \text{history}) \; d\,\text{biomarker},$$

thereby coupling event histories to multi-organ biomarkers (Wei et al., 10 May 2026). For tabular organ traits it uses conditional DDPMs with classifier-free guidance, and for topological brain connectivity it introduces SPD-VQVAE, which encodes symmetric positive definite matrices through the Cholesky factorization $M = LL^\top$ so that generated connectomes remain valid SPD objects while avoiding $O(N^3)$ spectral operations (Wei et al., 10 May 2026). This is a distinct shift from purely event-level disease predictors toward explicit physiological mediation.

A third family uses heterogeneous graph learning and disease-aware fusion. HGDC-Fuse represents each patient as a graph $\mathcal{G}_s = (\mathcal{V}_s, \mathcal{E}_s)$ with modality-specific node types and time-aware edge attributes (Jiang et al., 19 Sep 2025). It aggregates messages separately from same-type neighbors and learns a disease correlation matrix

$$A_{ij} = \frac{\text{co-occur}(i, j)}{\text{count}(i)}, \quad i \neq j,$$

which is thresholded and processed by a GCN to produce disease prototypes that guide disease-specific attention over EHR features, similar-patient messages, and temporally weighted CXR messages (Jiang et al., 19 Sep 2025). The defining feature is that fusion weights are disease-specific rather than globally shared.

A fourth line uses multi-task heterogeneity modeling. ADH-MTL treats each disease assessment as a task and models both disease heterogeneity and patient heterogeneity through group-level models $k$0, a decomposition

$k$1

and a Bayesian network with variational inference over relationship parameters and model parameters (Chai et al., 20 Nov 2025). This reduces the parameter count from $k$2 to $k$3 while supporting new-patient predictions via cluster assignment (Chai et al., 20 Nov 2025).

Finally, some HSMs are organized as modular digital-twin systems rather than a single predictor. OmniBioTwin proposes a System-of-Twinned-Systems (SoTS) architecture in which each twin $k$4 is an autonomous computational subsystem, and heterogeneous twins are coupled through explicit interaction operators in a seven-layer architecture spanning data integration, twin modeling, cross-scale coupling, temporal synchronization, and human-in-the-loop decision support (Wang et al., 9 Jun 2026). This framework addresses what the paper identifies as structural fragmentation in current health digital twins.

4. Intervention, control, and health-system response

A distinguishing property of HSMs is that they can represent not only disease evolution but also intervention. In CDTM, preventive interventions are implemented by modifying transition probabilities to reduce acquisition of a diagnosis $k$5: $k$6 where $k$7 corresponds to a 5% reduction in new cases (Ledebur et al., 2024). The model then propagates the intervention through the transition network to estimate long-term effects on multimorbidity and mortality (Ledebur et al., 2024).

DiffDT implements counterfactual simulation at the patient level. At inference, a subject’s event history up to age $k$8 is encoded, a conditional diffusion model generates age-$k$9 multi-organ digital twins given history and SDoH, and a predictive model estimates the next disease event from the generated twin (Wei et al., 10 May 2026). The paper describes “do(healthy)” experiments in which an exposure event in history is replaced with healthy, and the generated twin becomes significantly closer to ground-truth healthy subjects than to diseased subjects by FID and WD, with $$h_k(t \mid \mathbf{x}_{i,k}) := \lim_{\Delta t \to 0} \frac{1}{\Delta t} P\left(t \leq T < t + \Delta t, o \mid T \geq t, \ell, \mathbf{x}_{i,k}\right),$$0 (Wei et al., 10 May 2026). This is a specific implementation of intervention reasoning through generated mediators.

Clinical treatment policy is addressed by hierarchical multi-agent reinforcement learning. HMARL decomposes multi-organ sepsis management into a Root Agent $$h_k(t \mid \mathbf{x}_{i,k}) := \lim_{\Delta t \to 0} \frac{1}{\Delta t} P\left(t \leq T < t + \Delta t, o \mid T \geq t, \ell, \mathbf{x}_{i,k}\right),$$1, organ-specific agents $$h_k(t \mid \mathbf{x}_{i,k}) := \lim_{\Delta t \to 0} \frac{1}{\Delta t} P\left(t \leq T < t + \Delta t, o \mid T \geq t, \ell, \mathbf{x}_{i,k}\right),$$2, a Mixture Agent $$h_k(t \mid \mathbf{x}_{i,k}) := \lim_{\Delta t \to 0} \frac{1}{\Delta t} P\left(t \leq T < t + \Delta t, o \mid T \geq t, \ell, \mathbf{x}_{i,k}\right),$$3, and leaf-level dosage agents (Tan et al., 2024). Inter-agent communication is explicit: organ mixture agents consult treatment-specific sub-agents within a system, and $$h_k(t \mid \mathbf{x}_{i,k}) := \lim_{\Delta t \to 0} \frac{1}{\Delta t} P\left(t \leq T < t + \Delta t, o \mid T \geq t, \ell, \mathbf{x}_{i,k}\right),$$4 recursively passes recommendations across organ systems so that, for example, neuro decisions can depend on communicated cardio and renal dosages (Tan et al., 2024). The state representation is dual-layer, with a unified root representation built from 48 physiological variables, dense embeddings with $$h_k(t \mid \mathbf{x}_{i,k}) := \lim_{\Delta t \to 0} \frac{1}{\Delta t} P\left(t \leq T < t + \Delta t, o \mid T \geq t, \ell, \mathbf{x}_{i,k}\right),$$5, higher-order interaction terms, exponentially decayed temporal context, and targeted organ-level refinements (Tan et al., 2024). The learning formulation combines hierarchical Q-learning, options, and QMix-style monotonic mixing (Tan et al., 2024).

HSMs that explicitly model health system production move from treatment recommendation to delivery feasibility. TLO’s health system module models healthcare services as outputs of joint inputs—worker time, consumables, and facility assets—using a Leontief production function such as

$$h_k(t \mid \mathbf{x}_{i,k}) := \lim_{\Delta t \to 0} \frac{1}{\Delta t} P\left(t \leq T < t + \Delta t, o \mid T \geq t, \ell, \mathbf{x}_{i,k}\right),$$6

It also uses a Modes framework: Mode 0 imposes no requirement for specific workers, Mode 1 requires relevant worker presence, and Mode 2 enforces both presence and time-budget constraints (Chalkley et al., 15 Aug 2025). Workforce absenteeism, motivation, ownership category, management practice, and consumable stock-outs are treated as determinants of effective service output (Chalkley et al., 15 Aug 2025). In this formulation, an HSM does not merely forecast disease; it simulates whether interventions can actually be delivered under resource constraints.

5. Empirical performance and validation regimes

The empirical literature reports strong but heterogeneous performance, with metrics aligned to task structure rather than a single benchmark.

System	Task and metric	Reported result
DT-Transformer (Zhu et al., 14 May 2026)	Next-event prediction, median age- and sex-stratified AUC	0.871 across 896 disease categories, with all categories exceeding AUC 0.5
DiffDT-Qwen3 (Wei et al., 10 May 2026)	Next disease prediction, macro-average AUC / F1 macro-average	0.9171 / 20.92 vs Delphi-Qwen3 0.8931 / 18.17
HGDC-Fuse (Jiang et al., 19 Sep 2025)	Macro-PRAUC across 25 diseases	0.434 on full dataset; 0.470 on matched subset
HMARL (Tan et al., 2024)	Estimated mortality rate in sepsis management	8.81 ± 0.24 vs clinician 16.27
CDTM (Ledebur et al., 2024)	Model fit and intervention effect	Validation fit $$h_k(t \mid \mathbf{x}_{i,k}) := \lim_{\Delta t \to 0} \frac{1}{\Delta t} P\left(t \leq T < t + \Delta t, o \mid T \geq t, \ell, \mathbf{x}_{i,k}\right),$$7; 5% fewer new hypertensive cases yields 0.57% (0.06) lower all-cause mortality over 15 years
Gait foundation model (Gabet et al., 26 Mar 2026)	Phenotype prediction from gait embeddings	Age $$h_k(t \mid \mathbf{x}_{i,k}) := \lim_{\Delta t \to 0} \frac{1}{\Delta t} P\left(t \leq T < t + \Delta t, o \mid T \geq t, \ell, \mathbf{x}_{i,k}\right),$$8, BMI $$h_k(t \mid \mathbf{x}_{i,k}) := \lim_{\Delta t \to 0} \frac{1}{\Delta t} P\left(t \leq T < t + \Delta t, o \mid T \geq t, \ell, \mathbf{x}_{i,k}\right),$$9, VAT $$y_t = \mathrm{Embed}_\text{ICD}(s_t) + \mathrm{Embed}_\text{age}(\tau_t),$$0; 1,980 of 3,210 targets significant

Several studies emphasize not only discrimination but also mechanistic or structural validation. DiffDT reports that the best F1 for each disease category is always obtained using the matching organ’s digital twin as mediator, and that the generated mediators halve the mean absolute average treatment effect error relative to Delphi (0.004 vs 0.007) (Wei et al., 10 May 2026). HGDC-Fuse reports ablations showing that removing similar-patient connections, temporal CXR aggregation, or disease correlation-guided attention degrades performance, supporting the relevance of its graph construction and fusion design (Jiang et al., 19 Sep 2025). In HMARL, disabling either hierarchical state representations or cross-agent communication leads to substantial deterioration, with mortality increases reported as high as 33% higher than the full model in ablation experiments (Tan et al., 2024).

Other validation regimes target generalizability, explainability, or transferability. The Austria–Denmark multimorbidity study reports Adjusted Rand Index 0.998 and Normalized Mutual Information 0.88, indicating strong alignment between independently learned clusterings despite different healthcare systems (Einsiedler et al., 8 Oct 2025). The explainable surveillance system for eight chronic diseases reports F1 scores >0.75 and AUROC >0.80 for most diseases and time points, using routine EHR variables rather than laboratory tests and integrating SHAP, surrogate models, and a rule-engineering framework into an EMR deployment pathway (Khan et al., 27 Jan 2025). ADH-MTL reports 15–17% F1 improvement for depression over single-task deep learning baselines and 10% F1 gain for diabetes plus 14% for depression over the best multi-task baseline, with smaller performance gaps across age, income, race, and gender (Chai et al., 20 Nov 2025). These results indicate that evaluation in multi-disease HSMs is multi-criterion: predictive accuracy, counterfactual validity, policy relevance, robustness under missingness, and system transferability are all treated as first-order concerns.

6. Limitations, open problems, and research directions

The literature repeatedly identifies limits in both data and model structure. Existing generative disease models “largely depend on event-level representations from hospital and registry data,” and the absence of explicit modeling of social determinants of health limits personalized disease modeling and clinical decision support; DiffDT addresses this only through ICD-coded proxies rather than direct SDoH measurements (Wei et al., 10 May 2026). The same study notes UKB’s healthy volunteer bias and that its imaging data are “mostly cross-sectional rather than longitudinal,” although the architecture is described as modular for future incorporation of direct SDoH, environmental data, and higher-resolution multimodal inputs (Wei et al., 10 May 2026).

Real-world multimodal prediction remains constrained by modality missingness, noise, temporal asynchrony, and evidentiary inconsistency across modalities for different diseases, which HGDC-Fuse treats as central obstacles to clinical deployment (Jiang et al., 19 Sep 2025). Statistical estimation for disease histories remains difficult even when the state space is well defined: PAM-based multi-state models can handle dependent left-truncation and multiple time scales, but multiple-time-scale models were found less robust to the data-generating process than stratified single-time-scale models, and baseline hazards were not well recovered under interval-censoring (Wiegrebe et al., 24 Sep 2025). These are methodological limits, not merely engineering details, because they affect the interpretation of estimated transition hazards and progression risks.

At the systems level, current health digital twins are characterized as structurally fragmented: monolithic models addressing a single organ or task lack cross-scale fidelity, while system-level twins lack a generalizable architectural framework (Wang et al., 9 Jun 2026). OmniBioTwin responds with a seven-layer SoTS architecture, but the paper presents a demonstration focused on GLP-1 signaling pathways in Alzheimer’s disease, so broader validation across multi-disease clinical settings remains to be established (Wang et al., 9 Jun 2026). Similarly, the TLO line of work is still extending its production-side modeling to account for ownership forms, management practices, and worker absence in greater detail (Chalkley et al., 15 Aug 2025).

A common misconception is that multi-disease HSMs are a single model class. The cited work instead shows a heterogeneous field: some models forecast next diagnoses in health-system EHRs, some estimate transition hazards under censoring and left-truncation, some simulate preventive counterfactuals through digital mediators, some optimize multi-organ treatment through hierarchical RL, and some model whether constrained health systems can deliver care at all (Zhu et al., 14 May 2026, Wiegrebe et al., 24 Sep 2025, Wei et al., 10 May 2026, Tan et al., 2024, Chalkley et al., 15 Aug 2025). A plausible implication is that future HSMs will increasingly combine these strands: multimodal state representation, explicit causal or mechanistic mediation, intervention and resource simulation, and modular cross-scale composition. The current literature already points toward that synthesis, even though each component remains under active development.