acceptodds
Under review as a conference paper at ICLR 2027

Learning response-aware patient dynamics for respiratory support

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.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Patient dynamics, Respiratory support, Treatment response, Physiological trajectory prediction, Electronic health records

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_22XLOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper formulates short-term respiratory-support forecasting as autoregressive patient dynamics. It separates predicted physiological change into a state-dependent baseline component and a respiratory-support-associated deviation, uses room air as a reference condition, and feeds the predicted change back into the latent patient state through a response pathway. The model is evaluated on development, internal validation, temporal, and external cohorts against persistence, an action-conditioned GRU, and adapted clinical world models. However, the evaluation conditions on the recorded future respiratory-support sequence, whose availability at forecast time is unclear. The reported performance is also largely comparable to the baselines, without patient-level uncertainty estimates or a matched ablation isolating the response pathway.

Critical strengths

The paper aims to address a clinically significant task: treatment response prediction.

The paper proposes a method to decompose predicted physiological change into state-dependent baseline dynamics and respiratory-support-associated deviations, and to feed the predicted change back into the latent patient state during autoregressive forecasting.

Critical weaknesses
  1. Observed treatment assignment depends strongly on the current patient state, so treatment and physiological response are confounded, which can substantially bias response prediction. The proposed model does not address this confounding problem.

  2. The confounding also biases evaluation, yet the paper only evaluates factual prediction. Counterfactual validity should instead be assessed using simulated or semi-synthetic data with known potential outcomes, randomized trials, or a target-trial emulation with explicit identification assumptions.

  3. The first contribution claim is not established as novel. Similar problem formulations and definitions have already been proposed in prior work [1-3].

  4. The proposed model is just comparable to baselines.

[1] Matthieu Komorowski, Leo A Celi, Omar Badawi, Anthony C Gordon, and A Aldo Faisal. 2018. The artificial intelligence clinician learns optimal treatment strategies for sepsis in intensive care. Nature medicine 24, 11 (2018), 1716–1720. [2] Arne Peine, Ahmed Hallawa, Johannes Bickenbach, Guido Dartmann, Lejla Begic Fazlic, Anke Schmeink, Gerd Ascheid, Christoph Thiemermann, Andreas Schuppert, Ryan Kindle, et al. 2021. Development and validation of a reinforcement learning algorithm to dynamically optimize mechanical ventilation in critical care. NPJ digital medicine 4, 1 (2021), 1–12. [3] Lu Wang, Wei Zhang, Xiaofeng He, et al. 2018. Supervised reinforcement learning with recurrent neural network for dynamic treatment recommendation. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2447–2456.

Questions for authors

see above

Reviewer_Qo8kOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a response-aware patient dynamics model for forecasting physiological changes in ICU patients receiving respiratory support. The model separates baseline patient dynamics from respiratory-support-associated effects and uses predicted physiological changes to update the patient state. It is evaluated on retrospective EHR data from two institutions, predicting six-hour trajectories using six hours of history. The results show performance comparable to existing baselines, with modest improvements on changing-state targets.

Critical strengths

The paper addresses a clinically relevant problem. I appreciate the use of free-running trajectory evaluation rather than one-step predictions, as well as the inclusion of temporal and external test cohorts. The baseline comparisons also appear reasonably well controlled.

Critical weaknesses
  1. The benefit of the proposed response-aware design is not convincing. During inference, the response encoder takes the model's own predictions as input, so it is unclear what additional benefit this provides over a standard autoregressive model. The ablation results do not show consistent improvements, particularly on the external cohort. More importantly, the paper lacks analysis of whether the learned support-associated effects are physiologically meaningful. The theoretical uniqueness result also relies on a hard room-air constraint, while the actual implementation uses soft regularization. It is unclear whether the decomposition remains identifiable in practice, especially for patients rarely observed on room air.

  2. The overall performance improvement is limited. The proposed model does not outperform the strongest baselines on either the temporal or external test cohort. The gains on changing-state targets are small and do not generalize well externally, where a standard Action-GRU performs similarly. Without confidence intervals or statistical testing, it is difficult to judge whether these differences are meaningful. Given the added model complexity, stronger empirical evidence is needed.

Questions for authors

1 Since the response encoder uses predicted changes during inference, what is its advantage over directly feeding these changes into the state update? Could the authors include a comparison with a similarly sized GRU or residual dynamics model?

2 How well does the learned decomposition satisfy the room-air constraint? Could the authors examine its sensitivity to regularization and explain how baseline dynamics are identified for patients rarely observed on room air?

3 Could the authors analyze the learned support-associated effects across different respiratory-support modalities? For example, do changes in support settings produce clinically plausible changes in oxygen saturation and respiratory rate?

4 FiO2 appears to be both an input setting and a prediction target. Could the authors clarify their relationship and report results excluding FiO2 and other treatment-setting variables?

5 Could the authors provide confidence intervals or multi-seed results? Also, does truncating trajectories at the start of invasive mechanical ventilation exclude important deterioration episodes?

6 The meaningful-change target counts in Appendix Tables 4 and 6 appear inconsistent, with some individual feature counts in Table 6 exceeding the totals in Table 4. Could the authors clarify the denominators and which targets contribute to Table 3?

7 The caption of Figure 3 appears to repeat the description of Figure 2 and should be corrected.

Reviewer_o1yXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies 6-hour, action-conditioned forecasting of 50 ICU physiological variables across four respiratory support modes (room air, conventional oxygen, HFNC, and NIV) prior to invasive ventilation. At each step, the model predicts physiological change as the sum of a state-dependent baseline term and a gated treatment-associated deviation, regularized by a soft room-air penalty. A separate response encoder then feeds the predicted change back into the latent state update, trained with a four-epoch response-guided warm-up on observed changes. Evaluated under free-running 6-hour rollouts on internal validation, temporal, and external ICU cohorts against Persistence, Action-GRU, SepsisAgent-WM, and MedDreamer-WM, the method achieves overall RMSE comparable to the strongest baselines (0.516 on Val, 0.550 on Temporal, 0.880 on External), with modest gains on intermediate horizons and changing targets (∣x(t+h)−xt∣>0.25SD)(|x_{(t+h)}-x_t |>0.25 SD).

Critical strengths
  1. The clinical problem setup is well bounded and practical. Focusing on 6-hour trajectories across non-invasive respiratory support modes before intubation matches clinical decision timelines, and the authors appropriately avoid overclaiming causal treatment effects from observational EHR data.
  2. The evaluation spans multiple distribution shifts. Testing free-running 6-hour rollouts on internal, temporal, and external institution cohorts, e.g., broken down by horizon, support modality, and changing targets, provides a realistic test of model generalization.
Critical weaknesses
  1. There is a potentially serious action-target overlap or future-information leakage concern involving FiO2FiO_2, together with inconsistent reported statistics across tables. In Section 3.1 Eq. (1) and Eq. (3)–(6), predicting x ̂_{(t+1)} uses current action at(A(t−L+1:t+H−1))a_t (A_{(t-L+1:t+H-1)}), whereas Appendix B Eq. (24)–(26) pairs future states X(k+6:k+11)X_{(k+6:k+11)} with future actions A(k+6:k+11)A_{(k+6:k+11)}—an apparent one-step alignment discrepancy. Because FiO2FiO_2 is both an input action setting in a (Eq. 21) and a predicted target in x (Table 6), conditioning step k+6 on a_(k+6) could expose a same-hour recorded intervention setting overlapping with the prediction target FiO2FiO_2. Notably, FiO2FiO_2 shows a large RMSE advantage in Table 6 (1.878 versus 2.277–2.474 across comparators on Val), but on only 151 meaningful-change targets. Without timestamp-level implementation details, actual target leakage cannot be established from the manuscript alone. The reported statistics also require reconciliation: (i) in Table 4, race counts sum to 7,458, 1,965, 579, and 810 across the four cohorts instead of the reported patient totals (6,762, 1,912, 564, and 791); (ii) Table 4 reports 7,316 meaningful-change targets in total on Val, yet Table 6 lists 8,275 for O2SatO _{2}Sat and 8,524 for Resp alone; (iii) initial support subgroup counts N in Table 4 (e.g., 510/1,046/312/125 on Val) do not match Figure 2 and Table 5 (491/1,056/314/132); and (iv) Full model MC-MAE differs between Table 1 and Table 3 on both Temporal (0.723 vs. 0.710) and External (0.901 vs. 0.906).

  2. The distinct benefit of the proposed response-aware pathway has not been isolated, the aggregate gains are small, and key implementation details are missing. At inference, r ̂_{t}=Δx ̂_{t} (Eq. 29), so feeding both x ̂_{(t+1)}=x ̃_{t}+Δx ̂_{t} and z_t^r=E_{response} (Δx ̂_{t} ) into UstateU_{state} (Eq. 9) introduces no independent information; its benefit as an architectural inductive bias is not distinguished from a capacity-matched residual formulation operating on Δx ̂_{t}. Proposition 1 (Appendix D.2) establishes uniqueness only under the hard reference constraint d(s,a_0 )=0; this guarantee does not directly extend to the trained model using the soft penalty LRAL_{RA} in Eq. (15). Table 1 does not isolate the response-guided warm-up (Eq. 13), the gate gtg_t, or λRAλ_{RA} ; adding treatment dynamics alone worsens Temporal MC-MAE (0.728 to 0.772) and External MSE (0.778 to 0.814). Overall RMSE is tied with MedDreamer-WM on Val (0.516) and Temporal (0.550) and worse than SepsisAgent-WM on External (0.880 vs. 0.878), with no error bars reported. Reproducibility is further limited by missing exact module dimensions, random-seed settings, and the horizon-weighting scheme mentioned in the Reproducibility Statemen.

Questions for authors
  1. How are action timestamps aligned with target timestamps during rollout, given the apparent one-step discrepancy between Eq. (1) (A(t−L+1:t+H−1))(A_{(t-L+1:t+H-1)}) and Appendix Eq. (25) (A(k+6:k+11))(A_{(k+6:k+11)})? Since FiO2FiO_2 is both an input setting in a and a predicted target in x, could you clarify whether the recorded same-hour FiO2FiO_2 at step t+h is supplied as an action input when predicting x ̂_{(t+h)}?
  2. Can you explain and, where appropriate, correct the apparent numerical inconsistencies across the tables—specifically: (i) why demographic counts in Table 4 exceed the patient totals across all four cohorts, (ii) why total meaningful-change targets in Table 4 (e.g., 7,316 on Val) are smaller than single-feature counts in Table 6 (e.g., 8,275 for O2SatO_2Sat), (iii) why subgroup counts N differ between Table 4 and Table 5 / Figure 2, and (iv) why the Full model's MC-MAE differs between Table 1 and Table 3 on Temporal (0.723 vs. 0.710) and External (0.901 vs. 0.906)?
  3. Since the response representation is deterministically derived from the predicted physiological change, what evidence demonstrates that the reported improvements stem from the proposed response-aware mechanism rather than increased model capacity or residual prediction?
Reviewer_fbt9OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Authors propose a response-aware patient dynamic model for a respiratory support system. They achieve this by modifying the underlying autoregressive model architecture and decomposing the state into baseline dynamics and respiratory-support-based evidence. They used the room air information to facilitate this process.

Critical strengths

The authors address an important problem of modelling vital signs and other clinically relevant features that are critical for patient survival, particularly in ICU settings. With a clearer articulation of the situations where the proposed conditioning can improve patient survival, the overall approach could have a greater impact.

Critical weaknesses
  • The paper needs substantial improvement in writing style; few parts of the method are very hard to follow and lack proper intuition.

  • The author should clearly articulate their methodological novelty.

  • Considering the complexity of the proposed method, the performance gain seems to be marginal.

Questions for authors
  • Why have the authors chosen room air as the reference? How reliably is this information recorded across different hospitals, and how likely is it to be available in real-world clinical settings? The authors should also clarify the extent to which their method depends on the availability and accuracy of this information.

  • line [047-052] I am still not clear on what challenges there are in modelling patient variation and respiratory support in existing methods. The authors should clarify (if possible, with some examples) the benefit of the proposed method. Similarly, what is the advantage of modelling changes from the previous state?

  • Empirically, the baselines still outperform the method in multiple scenarios; can the author clarify why the practitioner should use the given approach?

  • Has the author done any sensitivity analysis of the proposed method to different hyperparameters? How was the hyperparameter for the baseline chosen?

  • Why is the performance gain of the proposed method higher in long time horizon compared to sort (Table 2).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.