Learning response-aware patient dynamics for respiratory support

Xiaolei Lu; Shamim Nemati

cs.AISep 26, 2026

Learning response-aware patient dynamics for respiratory support

Authors: Xiaolei Lu, Shamim Nemati

Abstract

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.

Figures & tables

Figure 1: Clinical motivation for response-aware respiratory dynamics. (A) Respiratory support is delivered sequentially as patient physiology evolves. The respiratory support action space includes room air as a no-active-support category, since patients may experience periods without active respiratory support between support modalities in clinical practice. Following a respiratory support action, the observed physiological response provides new information that can update the patient state and inform the next support decision. (B) Example clinical pathway showing how short-term response after HFNC initiation may help determine whether to continue the current support or consider escalation.

Val
Model	MSE $\downarrow$	1-h RMSE $\downarrow$	3-h RMSE $\downarrow$	6-h RMSE $\downarrow$	MC-MAE $\downarrow$
Base dynamics	0.270	0.369	0.504	0.621	0.736
+ Treatment-associated dynamics	0.268	0.361	0.506	0.619	0.726
+ Response-aware state update (Full)	0.266	0.361	0.505	0.619	0.726
Temporal
Model	MSE $\downarrow$	1-h RMSE $\downarrow$	3-h RMSE $\downarrow$	6-h RMSE $\downarrow$	MC-MAE $\downarrow$

Table 1: Ablation study of the proposed patient dynamics model across three evaluation cohorts.

Model	Overall	1h	2h	3h	4h	5h	6h
Val
Persistence	0.549	0.370	0.469	0.535	0.591	0.619	0.659
Action-GRU	0.520	0.355	0.445	0.507	0.559	0.585	0.624
SepsisAgent-WM	0.519	0.355	0.445	0.508	0.559	0.585	0.624
MedDreamer-WM	0.516	0.344	0.435	0.502	0.556	0.584	0.624
Ours	0.516	0.361	0.440	0.505	0.550	0.578	0.619

Table 2: Comparison with clinical world models for multi-step physiological trajectory prediction (values are RMSE; lower is better).

	Val			Temporal			External
Model	MAE	RMSE	$\Delta$ (%)	MAE	RMSE	$\Delta$ (%)	MAE	RMSE	$\Delta$ (%)
Persistence	0.831	1.131	–	0.828	1.224	–	1.002	1.848	–
Action-GRU	0.738	1.036	+8.4	0.733	1.118	+8.7	0.911	1.771	+4.2
SepsisAgent-WM	0.742	1.041	+7.9	0.735	1.124	+8.2	0.911	1.777	+3.8
MedDreamer-WM	0.726	1.027	+9.2	0.718	1.106	+9.6	0.897	1.783	+3.5
Ours	0.726	1.016	+10.1	0.710	1.105	+9.7	0.906	1.771	+4.2

Table 3: Performance on physiologically meaningful changes (

|x_{t+h}-x_{t}|>0.25

SD).

\Delta

denotes relative RMSE improvement over persistence.

Figure 2: Relative RMSE improvement over persistence across respiratory-support modalities in the validation, temporal, and external cohorts. Positive values indicate lower RMSE than persistence. $N$ denotes the number of evaluation windows in each subgroup.

Figure 3: Relative RMSE improvement over persistence across respiratory-support modalities in the validation, temporal, and external cohorts. Positive values indicate lower RMSE than persistence. $N$ denotes the number of evaluation windows in each subgroup.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 4: Overview of the response-aware patient dynamics model. The model initializes a latent patient state from recent clinical history and predicts physiological change as the sum of state-dependent baseline dynamics and a respiratory-support-associated deviation. The predicted change produces the next physiological state through a residual update and is also explicitly encoded as the physiological response for latent-state updating. The resulting state is recursively propagated over the prediction horizon.

Characteristic	Dev	Val	Temporal	External
Patients, $N$	6,762	1,912	564	791
Age, years, mean $\pm$ SD	62.9 $\pm$ 16.4	62.1 $\pm$ 16.6	64.3 $\pm$ 16.5	64.3 $\pm$ 17.2
Female, $N$ (%)	3,171 (42.5%)	820 (41.7%)	257 (44.4%)	360 (44.4%)
Race, $N$ (%)
White	3,884 (52.1%)	963 (49.0%)	274 (47.3%)	393 (48.5%)
Black or African American	609 (8.2%)	146 (7.4%)	49 (8.5%)	19 (2.3%)

Appendix

Table 4: Characteristics of the study cohorts.

Cohort	Support	$N$	Persistence	Action-GRU	SepsisAgent-WM	MedDreamer-WM	Ours
Val	Room air	491	0.546	0.517	0.517	0.510	0.515
	Conventional O 2	1,056	0.540	0.509	0.510	0.502	0.500
	HFNC	314	0.527	0.504	0.502	0.510	0.512
	NIV	132	0.665	0.637	0.633	0.634	0.630
Temporal	Room air	161	0.528	0.463	0.469	0.459	0.456
	Conventional O 2	346	0.510	0.479	0.477	0.474	0.474

Appendix

Table 5: Trajectory prediction performance stratified by respiratory support modality. RMSE is computed over all observed targets across the 1–6 h prediction horizon within each subgroup; lower is better.

N

denotes the number of evaluation windows.

Feature	$N$	Persistence	Action-GRU	SepsisAgent-WM	MedDreamer-WM	Ours
Val
O 2 Sat	8,275	1.057	0.914	0.905	0.872	0.850
Resp	8,524	1.029	0.898	0.902	0.819	0.893
FiO 2	151	2.474	2.426	2.462	2.277	1.878
PaO 2	272	1.440	1.422	1.416	1.460	1.396
PaCO 2	278	1.652	1.592	1.617	1.664	1.518

Appendix

Table 6: Feature-level RMSE on physiologically meaningful changes (

|x_{t+h}-x_{t}|>0.25

SD) across internal and external cohorts.

N

denotes the number of observed meaningful-change targets. Lower RMSE indicates better performance.

Explore similar work

Biological Amnesia in ICU Time-Series Prediction: A Drift-Adaptive Two-Stream Architecture with Temporal Retrieval
Jul 21, 2026Fatema Ferdous Tamanna, K. M. Merajul Arefin, Md. Abdul MasudClinical Decision Support Longitudinal EHR Modeling
Dept. of Computer Science and Information Technology, Patuakhali Science and Technology University, Bangladesh · Dept. of Computer Science and Engineering, University of Dhaka, Bangladesh
DRIFT: Direct-Recursive Intervention-Conditioned Forecasting of ICU Physiological Trajectories
Jul 28, 2026Weixin Liu, Juming Xiong, Congning Ni +4Multivariate Time Series Forecasting Time Series Forecasting
Vanderbilt University, Nashville, Tennessee, USA · Vanderbilt University Medical Center, Nashville, Tennessee, USA
Toward World Modeling of Physiological Signals with Chaos-Theoretic Balancing and Latent Dynamics
May 14, 2026Yunfei Luo, Xi Chen, Yuliang Chen +8Multivariate Time Series Forecasting World Model Learning
University of California San Diego · Dartmouth College · Amazon Web Services +3

Jul 21, 2026, cs.LGJ/K move · Enter open · S save

Biological Amnesia in ICU Time-Series Prediction: A Drift-Adaptive Two-Stream Architecture with Temporal Retrieval

Fatema Ferdous Tamanna, K. M. Merajul Arefin, Md. Abdul Masud

Dept. of Computer Science and Information Technology, Patuakhali Science and Technology University, Bangladesh · Dept. of Computer Science and Engineering, University of Dhaka, Bangladesh

Clinical Decision Support Longitudinal EHR Modeling

Open paper AskPDF arXiv

Background: Clinical decision support systems degrade silently as treatment protocols evolve, yet standard adaptation methods treat models as monolithic blocks, unable to distinguish stable patient physiology from shifting institutional practice. Methods: We propose an adaptive clinical intelligence architecture for ICU intervention prediction that structurally decouples physiological from treatment representations, confining parameter updates to the treatment stream upon a dual distributional and accuracy trigger. Automated audit logs record which treatment features drove each adaptation event and how their importance shifted. At inference, an attribution-driven Temporal RAG module grounds each prediction in patient-specific, era-matched PubMed evidence anchored to the patient's dominant physiological features. Experiments used 84,792 MIMIC-IV stays (2008-2022) under strict chronological split. Results: Drift localised entirely to the treatment stream, validating the structural prior. Selective adaptation improved vasopressor and septic shock discrimination and calibration over the static source model. A fully retrained baseline yielded marginally higher aggregate discrimination but missed 26 septic shock cases the framework correctly identified, with none in the reverse direction; retrieval consistency with the pre-adaptation source model was preserved by the framework but degraded substantially in the retrained baseline. Conclusions: Structurally constraining adaptation to drifting components while preserving stable physiological representations enables clinical AI to evolve with practice without distorting learned patient biology. This architecture offers a template for governable, interpretable deployment of adaptive models in high-stakes clinical environments.

Abstract

Explore similar work

Biological Amnesia in ICU Time-Series Prediction: A Drift-Adaptive Two-Stream Architecture with Temporal Retrieval

DRIFT: Direct-Recursive Intervention-Conditioned Forecasting of ICU Physiological Trajectories

Toward World Modeling of Physiological Signals with Chaos-Theoretic Balancing and Latent Dynamics