PaperlayerView exact revision on arXiv

Source-linked AI summary

Learning response-aware patient dynamics for respiratory support

Xiaolei Lu, Shamim Nemati

arXiv:2609.32782v1cs.AI

TL;DR

Patient dynamics models often combine physiological progression with intervention-associated change and do not explicitly represent physiological change during state updating. This paper decomposes predicted change and feeds it into latent-state updates; across ICU cohorts, overall trajectory prediction is comparable to baselines, with more consistent improvements when physiological states are changing.

  • Problem

    Patient dynamics models typically combine progression with intervention-associated change and represent future physiological levels without explicitly modeling change during state updating.

    Source: §1 · p. 1

  • Method

    The model decomposes predicted physiological change into state-dependent baseline dynamics and respiratory-support-associated deviations, then uses the predicted change to update the latent patient state.

    Source: §1 · p. 2

  • Results

    Across internal, temporal, and external cohorts, overall performance was comparable to strong patient dynamics baselines, with more consistent improvements when physiological states were changing.

    Source: §5 · p. 8

  • Takeaways & Limitations

    The findings support explicitly representing physiological change in trajectory prediction, particularly when physiological states are changing.

    Source: §5 · p. 8

    The retrospective study addresses observational trajectory prediction, not causal treatment-effect estimation or treatment recommendation.

    Source: §5 · p. 8

Abstract

from arXiv · show

Respiratory support can shape the short-term physiological trajectory of critically ill patients, but patients receiving the same intervention may follow different physiological trajectories. Clinical patient dynamics models typically predict future states from recent physiology and recorded interventions, while physiological change is mainly represented through the predicted future state. We propose a response-aware patient dynamics model that explicitly represents physiological change during autoregressive state updating. The model decomposes predicted physiological change into state-dependent baseline dynamics and respiratory-support-associated deviations, with room air providing a reference for the decomposition. We provide a formal analysis of this reference-anchored formulation. A response pathway encodes the predicted physiological change and uses it to update the latent patient state across the forecast horizon. Across ICU cohorts from two independent institutions, the proposed model achieves comparable overall trajectory prediction to patient dynamics baselines, with more consistent improvements when physiological states are changing.

1 Introduction

Respiratory-support responses vary across patients, while conventional action-conditioned dynamics generally combine physiological progression with intervention-associated change and represent future levels rather than change explicitly. The paper proposes a response-aware formulation that decomposes predicted change and uses it during state updating.

1 Introduction

  • Acute respiratory failure may worsen despite initial support, and short-term physiological changes can describe patients’ evolving condition during respiratory support.Support may escalate from conventional oxygen therapy to HFNC, NIV, or IMV [Azevedo et al., 2015, Coudroy et al., 2022].Source: §1 · p. 1
  • Conventional action-conditioned dynamics generally combine progression and intervention-associated change and predict future physiological levels without explicitly representing change during state updating.Patients who later succeed or fail treatment may have similar baseline measurements but different early physiological changes.Source: §1 · p. 1
  • Respiratory support is delivered sequentially as physiology evolves, with room air included as a no-active-support category and observed response informing the next patient state and support decision.Source: Figure 1 · p. 2
  • The model decomposes predicted physiological change into state-dependent baseline dynamics and respiratory-support-associated deviations, using room air as a reference.The predicted change is also represented during autoregressive state updating.Source: §1 · p. 2
  • The study evaluates the model on real-world ICU data from two institutions, including temporal and external validation and analyses across forecast horizons, support states, and physiological variables.Source: §1 · p. 2

2 Related Work

Clinical world models condition predicted patient-state transitions on treatment actions, generally modeling physiological progression and intervention-associated change together. Research on respiratory-support response motivates explicitly representing physiological change in patient dynamics.

2.1 Clinical Patient Dynamics and World Models

  • Clinical world models condition trajectory predictions on recorded treatment actions, but generally model physiological progression and intervention-associated change within the same transition process.Examples include MedDreamer’s latent-state action-conditioned rollout and SepsisAgent’s short-term state prediction under different actions [Xu et al., 2026, Wu et al., 2026].Source: §2.1 · p. 2

2.2 Physiological Response to Respiratory Support

  • Changes in oxygenation, respiratory rate, and related measurements after HFNC or NIV initiation have been used to assess subsequent treatment success or failure.The predictive value of the ROX index increases during the hours after HFNC initiation.Source: §2.2 · p. 2
  • Physiological change can carry information about evolving respiratory status beyond a single-time-point level, motivating its explicit encoding during autoregressive state updating.Source: §2.2 · p. 3

3 Methodology

The method forecasts future physiology autoregressively from recent measurements and recorded respiratory support. It decomposes predicted change, encodes that response for latent-state updating, and trains the rollout with room air as a soft reference condition.

3.1 Problem formulation

  • The model predicts the next H hours of physiology from a recent measurement history, recorded respiratory-support sequence, and static patient characteristics.The intervention representation includes modality and available settings such as oxygen flow, FiO2, or airway pressures.Source: §3.1 · p. 3

3.2 Response-aware physiological dynamics

  • Recent trajectory, measurement timing and availability, and respiratory-support history initialize the dynamic patient representation, which is combined with separately encoded static characteristics.Source: §3.2 · p. 3
  • The respiratory-support representation depends on both the current latent patient state and intervention, allowing the same intervention to have different representations across patient states.Source: §3.2 · p. 3
  • Predicted physiological change is decomposed into state-dependent baseline dynamics and a support-associated deviation; room air anchors this decomposition, which is a modeling construct rather than a causal treatment effect.The baseline term reflects the current latent state, while the deviation represents additional variation modeled from recorded support; concurrent treatments are not explicitly separated.Source: §3.2 · pp. 3–4
  • The response pathway encodes predicted change, and the latent-state update uses that encoded response together with predicted physiology and the support representation.During free-running prediction, the response equals the predicted change, providing a separate representation rather than additional information.Source: §3.2 · p. 4

3.3 Training

  • Training recursively generates future physiological states, carrying model-generated states forward without using future ground-truth states as rollout inputs.An initial response-guided period may supply observed responses to the state-update pathway, then decays to model-generated responses only.Source: §3.3 · pp. 4–5
  • Room air is the no-active-support reference, and training penalizes the effective support-associated deviation at room-air steps without requiring physiology to remain unchanged.The penalty is a soft reference; under the corresponding hard constraint, the additive decomposition is uniquely defined relative to room air.Source: §3.3 · p. 5
  • The trajectory-prediction loss is restricted to observed target measurements, while the complete objective includes a room-air regularization term.Source: §3.3 · p. 5

4 Experiments

The experiments evaluate short-term physiological trajectory prediction across internal, temporal, and independent external ICU cohorts. The model has similar overall RMSE to learning-based baselines, with more consistent advantages on changing targets and competitive performance across support states and physiological variables.

4.1 Experimental setting

  • After a 6-hour observed history, models recursively predict 6 hours without future physiological measurements; evaluation uses observed targets and includes RMSE, MAE, and a changing-target criterion.Targets qualify as changing when |x_t+h − x_t| > 0.25 SD; learning-based baselines share the splits, preprocessing, inputs, horizon, and selection objective.Source: §4.1 · pp. 5–6

4.2 Ablation and comparison with baselines

  • On validation, support-associated dynamics reduce MSE from 0.270 to 0.268 and MC-MAE from 0.736 to 0.726; the full model reaches MSE 0.266.Across temporal and external cohorts, support-associated dynamics alone do not consistently help, while the response-aware update recovers or improves several metrics, including MC-MAE.Source: §4.2 · p. 6, Table 1 · p. 6
  • On changing targets, our model improves over persistence by 10.1% on validation, 9.7% on temporal, and 4.2% on external cohorts; external RMSE ties Action-GRU at 1.771.MedDreamer-WM’s corresponding improvements are 9.2% and 9.6% on validation and temporal cohorts; it has the lowest external MAE (0.897), versus 0.906 for our model.Source: Table 3 · p. 7, §4.2 · p. 7
  • Learning-based models have similar overall RMSE; MedDreamer-WM is stronger early, while our model is lower at validation hours 4–6 and temporal hours 3–5, and SepsisAgent-WM has the lowest external overall RMSE.On the external cohort, our model has the lowest RMSE from 2–5 h.Source: §4.2 · p. 7, Table 2 · p. 7

4.3 Clinical Stratified Analysis

  • Across support states, our model is competitive and has the lowest RMSE in several subgroups; differences between learning-based models are generally small, especially for HFNC and smaller NIV subgroups.The reported subgroup wins span validation conventional oxygen and NIV, temporal room air and HFNC, and external conventional oxygen and NIV.Source: §4.3 · p. 8, Figure 2 · p. 8
  • On changing targets, our model has the lowest RMSE for 6/10 validation variables, 5/10 temporal variables with a tie for HR, and 8/10 external variables.Improvements are distributed across variables, with large gains for several respiratory variables, particularly FiO2, and gains also for MAP and DBP.Source: §4.3 · p. 8

5 Conclusion

The response-aware model achieved performance comparable to strong patient dynamics baselines across internal, temporal, and external cohorts, with more consistent improvements when physiological states were changing. The study is retrospective and observational, not an evaluation of causal treatment effects or treatment recommendations.

5 Conclusion

  • Across internal, temporal, and external cohorts, the model achieved overall performance comparable to strong patient dynamics baselines, with more consistent improvements when physiological states were changing.The model decomposes predicted physiological change into state-dependent baseline dynamics and respiratory-support-associated deviations, and represents that change during latent-state updating.Source: §5 · p. 8
  • The study is retrospective and focuses on observational trajectory prediction rather than causal treatment-effect estimation or treatment recommendation.Future work will investigate counterfactual trajectory analysis and respiratory-support decision modeling, with explicit treatment of confounding and prospective validation.Source: §5 · p. 8
  • Figure 3 reports relative RMSE improvement over persistence across respiratory-support modalities in validation, temporal, and external cohorts; positive values indicate lower RMSE than persistence.Source: Figure 3 · p. 9

AI Use Disclosure

Generative AI tools assisted with manuscript drafting and editing; the authors reviewed and verified AI-assisted text and take responsibility for the final content.

AI Use Disclosure

Ethics Statement

The study uses retrospective, de-identified clinical data under institutional review and data-governance requirements, and its decomposition is not a causal treatment-effect estimate. The model is a research framework, with generalization potentially limited by differences across populations and care settings.

Ethics Statement

  • The study uses retrospective, de-identified electronic health record data under applicable institutional review, data-use, and governance requirements; consent was waived where permitted.Source: Ethics Statement · p. 9
  • Underlying clinical data cannot be publicly released where institutional policies or data-use agreements restrict release.Source: Ethics Statement · p. 9
  • The model is intended as a research framework, not for autonomous clinical decision-making; predictions are not guaranteed responses or support-selection recommendations.Prospective clinical use would require further external and prospective validation, failure-mode assessment, clinical oversight, and regulatory review.Source: Ethics Statement · p. 9
  • Differences in practice patterns, patient populations, measurement processes, and support use may leave residual bias or under-representation and limit generalization.Performance may not generalize to all patient groups or care settings despite evaluation across multiple cohorts.Source: Ethics Statement · p. 9
  • The decomposition of physiological change is a modeling construct, not identification of a causal treatment effect.Interventions are observed rather than randomized, and counterfactual claims about alternative support strategies are outside the study’s scope.Source: Ethics Statement · p. 9

Reproducibility Statement

The study describes its trajectory construction, preprocessing, model, and evaluation procedures, including free-running autoregressive prediction without future ground-truth physiology. It plans to release implementation and experiment-generation materials subject to institutional and data-use restrictions, while the underlying clinical datasets cannot be publicly released where restricted.

Reproducibility Statement

  • Each modeling example uses a fixed observed history and prediction horizon, with free-running autoregressive evaluation that does not access future ground-truth physiology.Source: Reproducibility Statement · p. 9
  • The documented architecture and training details include latent-state initialization and updating, respiratory-support encoding, change decomposition, treatment gating, objectives, regularization, optimization settings, horizon weighting, model selection, and random seeds.Source: Reproducibility Statement · p. 10
  • The authors will release preprocessing, trajectory-construction, training, evaluation, baseline, and figure/table-generation code with configuration files, subject to institutional and data-use restrictions.Underlying clinical datasets cannot be publicly released where restricted by institutional policies or data-use agreements.Source: Reproducibility Statement · p. 10

A Model architecture overview

The study constructs six-hour observed histories and six-hour respiratory-support-conditioned prediction horizons, then compares representative recurrent and clinical world-model baselines. Preprocessing retains treatment-setting availability and uses training-derived parameters across validation and external cohorts.

B Data preprocessing and trajectory construction

  • Physiological imputation and normalization parameters are estimated from training data and reused without re-estimation in validation and external cohorts.Missingness indicators are retained, and time-since-last-measurement features are supplied separately to the initial state encoder.Source: B Data preprocessing and trajectory… · pp. 12–13
  • Treatment representations pair continuous settings with availability masks, distinguishing unavailable settings from observed values near zero.Setting normalization statistics are estimated from training observations where each setting is available.Source: B Data preprocessing and trajectory… · p. 12
  • The observed history and static information form the initial latent state, while respiratory interventions condition the autoregressive rollout.Trajectories use a six-hour observed history and six-hour prediction horizon.Source: B Data preprocessing and trajectory… · pp. 12–13

C Baselines

  • The comparison includes persistence, an action-conditioned GRU, and GRU-based adaptations of SepsisAgent and medDreamer world models.Methods share the same splits, preprocessing, inputs, treatment representations, horizons, and observation masks, with independently tuned hyperparameters.Source: C Baselines · pp. 13–14

D Additional Analysis of Response-Aware Dynamics

The analysis explains how predicted physiological change enters latent-state updating and how a room-air reference anchors the decomposition of conditional transitions. The formal uniqueness result applies to a hard reference constraint, while the implemented model uses a soft constraint and does not identify causal treatment effects.

D.1 Explicit response representation

  • The response pathway encodes predicted physiological change separately from predicted physiological level and supplies both representations to the latent-state update.The response pathway adds no observation at inference; observed change may be used only during an initial response-guided training period, after which model-generated responses are used.Source: §D.1 · p. 14

D.2 Reference-anchored decomposition of physiological dynamics

  • Under the hard room-air reference constraint, baseline dynamics equal the room-air conditional transition and the support-associated deviation is the difference from that reference.The constraint uniquely determines both components of the additive decomposition.Source: §D.2 · p. 15
  • The implemented model uses a soft room-air penalty, so it encourages but does not enforce zero support-associated deviation under room air.The proposition establishes uniqueness only under the corresponding hard constraint.Source: §D.2 · p. 15

D Additional Analysis of Response-Aware Dynamics

  • The decomposition describes observational conditional transitions and does not identify causal treatment effects because respiratory support is not randomized and assignment may depend on patient characteristics.The analysis concerns model parameterization and observational transition dynamics.Source: D Additional Analysis of… · p. 14, §D.2 · p. 15

E Additional Stratified Results

The section provides subgroup-level trajectory prediction and feature-level performance summaries for physiologically meaningful changes.

E Additional Stratified Results

  • Table 5 reports absolute RMSE for trajectory prediction by respiratory-support subgroup over the 1–6 h horizon; lower is better.The table also reports evaluation-window counts.Source: E Additional Stratified Results · p. 15, Table 5 · p. 16
  • Table 6 reports feature-level RMSE on changes exceeding 0.25 SD across internal and external cohorts; lower RMSE indicates better performance.The table reports the number of observed meaningful-change targets.Source: Table 6 · p. 16

Related papers

Loading 2609.32782v1…