PaperlayerView exact revision on arXiv

Source-linked AI summary

Physiological World Models for Human State Transitions

Chongyang Zhang, Rendong Wang, Hao Zheng, Hanwen Zhang, Yang Liu, Xiaolong Wei, Bin Chong

arXiv:2608.15309v1cs.AI

TL;DR

Health AI commonly recognizes current states or predicts risks without directly modelling how whole-person physiology changes in response to real-world events, behaviours, contexts, and interventions. This paper proposes an event-conditioned Physiological World Model using structured transition tokens, capability levels, validation protocols, and benchmark tasks. Its supported scope is a framework for forecasting and simulating physiological transitions while keeping causal claims, uncertainty, safety, and deployment boundaries explicit.

  • Problem

    Most health AI recognizes states, detects anomalies, or predicts risks, while event-, behaviour-, context-, and intervention-conditioned changes in individual physiology remain insufficiently modelled.

  • Method

    The paper defines an event-conditioned PWM and introduces quality-scored HumanState Transition Tokens, an L1–L4 capability ladder, P1–P4 data protocols, and benchmarks for representation, forecasting, individualization, simulation, planning, and reliability.

  • Results

    The paper presents a framework in which aligned transition data support learning, forecasting, multi-trajectory simulation, and bounded planning for individual physiological state transitions.

  • Takeaways & Limitations

    PWMs provide a proposed path toward personalized health management, behavioural intervention design, and clinician-supervised decision support within explicit causal, uncertainty, safety, and governance boundaries.

  • Takeaways & Limitations

    PWM development is constrained by heterogeneous, irregular observations, sparse and inconsistent event or intervention labels, and substantial variation in individual baselines.

Abstract

from arXiv · show

Continuous multimodal sensing now allows human physiology to be observed throughout daily life rather than only during occasional clinical visits. However, most health artificial intelligence systems are designed to recognize current states, estimate risks or analyse individual biomarkers. They do not directly model how physiological states change in response to real-world events, behaviours, contexts and interventions. Here we propose the Physiological World Model (PWM), an event-conditioned framework for learning these changes at the level of the whole person. We introduce the HumanState Transition Token, a structured, quality-scored unit that connects the physiological state before an event with the event or action, relevant context and intervention information, the physiological trajectory after the event, observed outcomes and data quality. We describe four capability levels, from state representation to bounded intervention planning, together with four data acquisition and validation protocols. We also propose six benchmark tasks covering HumanState representation, forecasting across multiple timescales, individualized response prediction, simulation of alternative interventions, bounded planning and reliability under distribution shift. Together, this framework provides a practical path towards personalized health management, behavioural intervention design and clinician-supervised decision support, while clearly separating prediction from causal inference and making uncertainty, safety, governance and limits of use explicit.

1 Introduction

Continuous multimodal sensing has expanded physiological monitoring into daily life, but most health AI still recognizes states or predicts risks rather than modelling event-driven physiological transitions. The PWM proposal extends world-model principles to integrated individual physiology by representing and learning these transitions.

  • Daily-life sensing now combines wearable measurements of cardiovascular, activity, temperature, electrodermal, and sleep-related signals.Miniaturized sensors, wireless connectivity, and cloud infrastructure enable multimodal monitoring beyond episodic clinical visits.
  • Most current health AI systems perform state recognition, anomaly detection, or risk prediction rather than modelling dynamic changes caused by events, behaviours, and context.The unresolved questions include why states change, why responses differ between individuals, and how trajectories might differ under alternatives.
  • World-model approaches provide task-relevant state representations and transition dynamics for forecasting, simulation, and planning, but event- and intervention-conditioned modelling of integrated human physiology remains underexplored.Existing biomedical applications include molecular, cellular, physiological-signal, and clinical-trajectory modelling.
  • The proposed PWM models transitions in an individual’s macro-level physiological state and is positioned at the integrated individual-physiology layer of a multiscale taxonomy.The taxonomy places broader population and One Health models above PWM and tissue, organ, cellular, and molecular models below it.

2 Defining Physiological World Models

A Physiological World Model models how an individual’s integrated physiological state changes under events, actions, contexts, and interventions. It represents this state across observation, interpretable-physiology, and latent layers, and uses event-anchored transition tokens to support forecasting, simulation, and intervention comparison.

  • 2 Defining Physiological World Models: PWMs model event-conditioned transition dynamics for an individual’s integrated, macro-level physiological state rather than only recognizing current states or predicting risks.They connect internal physiology with external events, actions, contexts, and interventions.
  • 2 Defining Physiological World Models: A PWM represents physiology through raw observations, interpretable intermediate variables, and a latent HumanState layer while keeping context as a distinct conditioning variable.This layered representation supports heterogeneous measurements without assuming devices measure identical variables or have equal fidelity.
  • 2 Defining Physiological World Models: Future HumanState is modeled conditionally on current state, actions, context, time, and prediction horizon, allowing multiple plausible trajectories through multi-step rollouts.Intervention comparisons are causal only when supported by appropriate study designs and assumptions.
  • 2 Defining Physiological World Models: HumanState is a latent macro-level summary inferred from integrated physiological observations, disease-related status, prior changes, behaviours, treatments, and sometimes subjective variables.It represents overall physiological condition rather than any single observed measure.
  • 2 Defining Physiological World Models: A HumanState Transition Token is an event-anchored, quality-assessed record containing baseline and response windows, context, interventions when applicable, outcomes, and data-quality information.Non-applicable fields are explicitly encoded as absent, and observed outcomes can evaluate predicted transitions.
  • 2 Defining Physiological World Models: High-confidence transition tokens, rather than sheer recording volume, determine useful supervision for transition learning.Carefully documented events, contexts, windows, interventions, and outcomes can make smaller datasets richer than millions of unlabeled wearable-data hours.

3 Model Architecture and Learning Paradigm

The proposed PWM architecture combines latent HumanState representations with event-conditioned transition dynamics, uncertainty estimation, and structured evidence requirements. Its capability ladder progresses from state representation through transition prediction and simulation to bounded intervention planning, with increasingly demanding data and validation needs.

  • Architecture: A functional PWM architecture includes a state encoder, conditional transition model, uncertainty estimation, and encoders for events, actions, interventions, and context.The architecture is presented as an illustrative system informed by latent-dynamics and JEPA-style predictive learning.
  • Architecture: The transition model represents future latent physiology as conditioned on current state, world events, human actions, interventions, and context over a process-dependent horizon.Different physiological processes may require different temporal resolutions or hierarchical transition components.
  • Learning challenges: Real-world PWM learning must address heterogeneous, irregular observations, sparse and inconsistent event labels, and substantial variation in individual baselines.These challenges motivate complementary learning strategies across PWM development stages.
  • Learning paradigms: Latent-dynamics and JEPA-style methods are proposed as promising early-stage starting points because they predict future representations while reducing emphasis on device noise and other low-level fluctuations.Observation-level generative models can capture complex temporal patterns but may prioritize local morphology over intervention-relevant physiological dynamics.
  • Learning paradigms: Model-based reinforcement learning is positioned mainly as a later-stage route toward L4 planning because health data are observational and offline, interaction is costly, rewards are difficult to define, and actions are confounded.The paper treats reinforcement learning as a conceptual reference rather than the sole starting point for early training.
  • Capability ladder: The development path proceeds from multimodal self-supervised HumanState representation learning to token-based transition learning, then to intervention comparison and planning.The four-level ladder characterizes capabilities from L1 state representation through L4 bounded intervention planning, with progressively stronger evidence requirements.

4 Data Protocols for Physiological World Models

PWM development requires protocol-based temporal alignment and structured transition evidence, not merely larger volumes of continuous recordings. Four acquisition and validation levels progressively strengthen event annotation, intervention control, prospective feedback, and state-transition evidence.

  • Data alignment: Fragmented records distribute physiological signals and contextual annotations across devices and systems, obscuring cross-modal relationships needed for event-anchored transitions.Protocol-based alignment maps observations, events, actions, and interventions onto a shared timeline with pre-event baselines and post-event response windows.
  • Data limitations: Sparse, incomplete, or imprecise event and intervention records limit the construction of event-anchored transition samples from extensive physiological recordings.Examples include unlogged meals, inaccurate self-reported timing, and physiological changes lacking reliable records of concurrent stressors or activities.
  • Minimum Data Protocol: Nine minimum protocol components define the precision, recording requirements, and failure modes needed to pool HumanState Transition Tokens across studies, devices, and populations.The Minimum Data Protocol also determines whether a token is suitable for training or evaluation at a given capability level.
  • P1–P2: P1 passively collected data characterize baseline physiology but may not produce high-confidence transition tokens because events and context are often incompletely labelled.P2 adds structured event and action logging in free-living settings to construct comparable event-anchored samples supporting L2 transition learning.
  • P3: P3 assigns behavioural, environmental, or clinical interventions under controlled or semi-controlled conditions to measure their physiological effects.Assigned interventions provide stronger evidence than passive or observational recordings for evaluating physiological responses.
  • P4 and progression: Synthetic trajectories may support rare-scenario exploration but cannot replace empirical observations, causal evidence, prospective validation, or intervention-safety evidence.Progress toward L4 additionally requires target populations, explicit safety constraints, prospective intervention evidence, failure-case analysis, long-term follow-up, and oversight.

5 Benchmarking and Applications

The paper proposes six complementary benchmark tasks that progress from HumanState representation and transition forecasting to individualized prediction, intervention simulation, bounded planning, and reliability assessment. Applications require task profiles matched to risk, with higher-risk uses needing prospective validation, safety constraints, uncertainty calibration, and clinical oversight.

  • Benchmark framework: Six benchmark tasks evaluate progressively stronger modelling capabilities, while a sixth task assesses reliability across all capability levels.The tasks connect the capability hierarchy, data protocols, and potential applications.
  • Benchmark framework: T1 should combine validated measures, downstream transfer, missing-modality robustness, and cross-device or population generalization to assess HumanState representation.A single latent-space metric is insufficient because HumanState is not directly observable.
  • Benchmark framework: T2 forecasts event-, action-, or intervention-conditioned trajectories across multiple horizons, whereas T3 tests person-specific response heterogeneity through individualization.Forecasting from prior observations alone is only a baseline, not evidence of L2 PWM capability.
  • Benchmark framework: T4 distinguishes scenario simulation from counterfactual interpretation, with causal claims requiring randomized, crossover, or otherwise causally informative designs.Held-out intervention outcomes validate empirical targets but do not establish unobserved individual-level counterfactual ground truth.
  • Application readiness: No single task establishes application readiness; personalized recovery requires reliable state representation, multi-horizon prediction, individualized responses, and uncertainty calibration.Behavioural intervention design additionally requires credible comparison of alternative trajectories, while higher-risk applications require the complete benchmark profile.
  • Application horizons: Near-term applications target lower-risk recovery management, mid-term uses compare behavioural interventions, and longer-term uses support trajectory exploration and clinician-in-the-loop decisions.Examples include sleep and exercise recovery, meal timing and caffeine choices, chronic disease modelling, rehabilitation, and treatment-toxicity monitoring.
  • Application horizons: Simulated trajectories should not replace clinical trials or suffice for independent treatment decisions; higher-risk applications require prospective validation, safety gates, calibration, failure analysis, and accountability boundaries.Treatment selection, medication adjustment, and intervention planning additionally require regulatory oversight and continuing clinical supervision.

6 Challenges, Open Questions, and Scope Boundaries

The paper identifies evidence, data, causal inference, continual calibration, interpretability, privacy, equity, and governance as constraints on PWM development. Progress toward intervention planning must therefore be gated by stronger empirical validation, transparent governance, and explicit limits of use.

  • Evidence and data: The central challenge is establishing reliable evidence for state representation, transition prediction, and intervention comparison from heterogeneous, biased, and incomplete real-world records.The paper frames this as more than a data-volume problem.
  • Evidence and data: Higher PWM capability levels increasingly require aligned and verifiable HumanState Transition Tokens, despite differences among devices in sampling, algorithms, missingness, and annotation quality.Motion artefacts, non-wear periods, event ambiguity, and population differences complicate transition evidence.
  • Causal inference: Co-occurring events and delayed effects make observational trajectories insufficient for separating intervention effects from spontaneous recovery and confounding.Repeated within-person measurements, randomized N-of-1 or crossover designs, negative controls, sensitivity analyses, and free-living evaluation can strengthen evidence.
  • Continual adaptation: Changing baselines require continual calibration, but adaptation can introduce model drift and safety risks.The paper calls for separate validation of state updating, individual calibration, parameter updating, and high-risk recommendation generation, with version tracking and rollback mechanisms.
  • Interpretability and safety: PWM predictions and recommendations require calibrated uncertainty, applicability boundaries, and abstention or escalation criteria, with explanations linked to validated measures rather than treated as causal evidence.Interpretability should remain stable across small changes in inputs, devices, and model versions.
  • Privacy and governance: Multimodal combinations can reveal sensitive patterns, requiring privacy-by-design, auditable data-use policies, provenance tracking, and explicit handling of synthetic samples.Synthetic samples should remain distinguishable from empirical observations and should not serve as intervention or safety evidence.
  • Equity: Validation must cover devices, ages, socioeconomic groups, care settings, and relevant skin tones, with subgroup performance reported alongside aggregate results.This is intended to avoid privileging high-end devices or populations already overrepresented in digital-health datasets.
  • Scope boundaries: PWMs and digital twins overlap but differ: PWMs emphasize transition dynamics, conditional simulation, and intervention comparison, whereas digital twins add synchronization and broader system integration.Early PWMs can begin with lower-risk transition modelling, but fuller decision-support integration should follow prospective validation.

7 Outlook

The outlook positions PWMs as a staged framework for organizing multimodal physiology around event-conditioned state transitions rather than simulating the body in its entirety. The proposed progression moves from robust state representation to transition prediction, alternative-trajectory simulation, and bounded planning under increasingly strong validation.

  • Outlook: PWMs are intended to organize heterogeneous physiological and contextual data around state transitions, not to simulate the human body in its entirety.The paper argues that macro-level physiological applications do not require complete biological reconstruction.
  • Outlook: Development should proceed from multimodal state representation to event-conditioned prediction, alternative-trajectory simulation, and bounded planning, with calibrated uncertainty and progressively stronger validation.Longitudinal and event-anchored datasets must also span diverse populations and contexts.
  • Outlook: PWMs could provide shared infrastructure connecting physiological observation, state-transition modelling, and evidence-based evaluation.The supported scope includes estimating conditional future trajectories and comparing alternative actions when causally informative evidence exists.
Loading 2608.15309v1…