acceptodds
Under review as a conference paper at ICLR 2027

Physiologically Aware Flow Matching: Jacobian Constraints for Dose-Consistent Forecasting

Abstract

Accurate forecasts under observed treatment plans do not guarantee physiologically consistent responses to altered ones: state-dependent dosing couples treatment to patient state, so models trained on such records learn associations that contradict physiology. We introduce Physiologically Aware Flow Matching (PAFM), a conditional flow-matching framework for probabilistic forecasting under planned treatments. PAFM regularizes the Jacobian of the velocity field with respect to pharmacokinetically smoothed treatment drivers, with physiological priors specifying the sign of each sensitivity and a band for its magnitude. The penalty is evaluated at sampled flow states and operates without a mechanistic outcome model or rollouts under alternative plans. We develop autoregressive (AR-PAFM) and joint-horizon (JH-PAFM) formulations, which constrain sensitivities at individual forecast steps and across forecast positions respectively, and evaluate both on two simulated closed-loop benchmarks. On glucose, every unregularized model integrates a glucose-raising response to added insulin; the penalty restores near-perfect directional and dose-ordering consistency in both formulations, and JH-PAFM achieves lower mean RMSE than all regularized baselines. On mean arterial pressure, both formulations reach full directional and dose-ordering consistency under propofol and norepinephrine infusions, with JH-PAFM lower than AR-PAFM in mean RMSE and calibration error.

Primary area:
learning on time series and dynamical systems
Keywords:
flow matching, probabilistic forecasting, physiological priors, Jacobian regularization, dose--response consistency

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_MiPaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Physiologically Aware Flow Matching (PAFM) addresses treatment forecast under OOD states that arise under altered treatments. To improve predictions under altered treatments, PAFM encodes domain knowledge of relations between physiological variables into a regularizing loss term over the derivative of the flow in the model. The paper measures the physiological consistency via the proxy metric of directional consistency. The paper demonstrates an improved tradeoff of directional consistency vs RMSE accuracy over 2 simulated benchmarks, compared to several baselines.

Critical strengths

The problem is well motivated. The idea of encoding domain knowledge via regularization of the flow direction is novel and interesting, and potentially useful.

Critical weaknesses

My main concern is the strength of the evidence. There are no theoretical results (which is understandable). If I understand correctly, the experiments present two simulated benchmarks that the authors devised, where the improvement is measured through a consistency metric that the authors defined. If I understand correctly, RMSE is measured only for the original treatment, As the motivation is about predicting well even when the treatment is modified - I would ideally expect to see a benchmark with "default" and "modified" treatments, where RMSE is improved explicitly for the modified treatments. Ideally on a public benchmark (to assure that the robustness to modified treatment isn't an overfitting of the method and simulated benchmark built with the same artifacts in mind).

Presentation: I'd use some visualization to make the paper more accessible - e.g. for the method and the results. nit: I'd also consider mentioning that the problem is essentially OOD prediction under control settings - this will be more familiar to some readers and more general that medical treatments alone.

Literature: The following two references proposed different approaches to address the same issue of prediction under modified treatment. I would either compare to them as baselines, or if for some reason they are not compatible with the benchmark - explain why. Forecasting Treatment Responses Over Time Using Recurrent Marginal Structural Networks (Lim et al, neurips 2018), Individualized Dosing Dynamics via Neural Eigen Decomposition (Belogolovsky, neurips 2023).

Questions for authors

Is my understanding of the experimental setup as described above correct?

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

This paper proposes Physiologically Aware Flow Matching (PAFM), to solve the problem of accurate model prediction but possible violation of physiological laws in treatment dose response in physiological time series forecasting. The authors improve the predicted response direction and dose ranking by constraining the Flow Matching velocity field to smooth the Jacobian treatment input, and propose autoregression (AR-PAFM) and joint prediction (JH-PAFM). The experiment is conducted on two simulated datasets, and the results show that this method can significantly improve physiological consistency while maintaining competitive prediction accuracy.

Critical strengths
  1. Significant practical significance of the research question. This paper focuses on the problem in physiological time series forecasting (TSF) that high prediction accuracy cannot guarantee treatment response conforms to physiological laws. It has clear practical application value and also points out the limitations of traditional predictive metrics.

  2. Reasonable method design. The authors introduce physiological priors through Jacobian regularization of velocity fields, without the need to generate multiple alternative treatment plans during training. Both models significantly improve dose response consistency on two simulated datasets, demonstrating the effectiveness of the methods in different physiological scenarios.

Critical weaknesses
  1. There is a lack of sufficient theoretical or experimental connection between local Jacobian constraints and the final predicted response. The constraint in this article is the local sensitivity of the generated velocity field to the treatment input, but this does not necessarily guarantee that the final predicted trajectory has a correct dose response. The paper lacks analysis of this relationship and does not fully validate the accuracy of the predicted effects using real intervention responses in simulators.

  2. Insufficient baseline comparison. The experiment mainly used models such as LSTM, TCN, S4D, and some mechanistic models, lacking comparison with TSF frontier models and recent conditional probability prediction models. More importantly, there is a lack of direct comparison with output level Ranking Loss or Jacobian Regularization under the same Flow Matching architecture, making it difficult to fully assess the unique advantages of the proposed regularization method.

Minor: Unclear writing presentation. The Method section of the main text takes up about four pages and contains many formulas and network components, but there is no overall framework diagram, which increases the difficulty of understanding.

Questions for authors

1.Can you explain the reason for choosing Mamba as the historical encoder and verify it through corresponding architecture ablation experiments? 2.Can you further explain how local Velocity Jacobian constraints affect the final predicted dose response? Can theoretical analysis be provided, or can the sensitivity of the final prediction to the therapeutic dose be directly evaluated? 3.Can you compare the regularization method with output level Ranking Loss or Jacobian Regularization under the same Flow Matching Backbone? 4.Can you supplement the validation on two suitable real clinical datasets, such as publicly available research datasets like OhioT1DM, DiaTrend, AZT1D, and VitalDB? All experiments are based on simulated physiological data, and it is unclear whether the proposed method can be generalized to real patients.

Reviewer_WMJBOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes Physiologically Aware Flow Matching (PAFM), a probabilistic forecasting framework that regularizes velocity-field derivatives with respect to pharmacokinetically smoothed treatment inputs using physiological sign and magnitude priors. It develops autoregressive and joint-horizon formulations without requiring alternative-treatment rollouts for the Jacobian penalty. Experiments on simulated glucose and blood-pressure benchmarks demonstrate substantial improvements in directional and dose-ordering consistency, with the joint-horizon model generally achieving better predictive accuracy than the autoregressive version and the tested regularized glucose baselines.

Critical strengths

NONE

Critical weaknesses

Overall, I read this paper as a combination of several machine learning methods motivated by treatment planning. However, I see little demonstrated practical value, which raises a fundamental question: what specific treatment-planning problem does the proposed method meaningfully solve?

  1. Practical utility is insufficiently demonstrated. Both applications are simulated, yet forecasts under altered doses are not compared with the simulators’ actual outcomes under those interventions. Agreement with prescribed signs and dose orderings does not establish accurate treatment responses or improved treatment decisions.

  2. The methodological complexity is insufficiently justified. The experiments do not clearly isolate the benefits of combining the proposed components. Moreover, the ranking baselines enforce weaker constraints than those assessed during evaluation. Comparisons with simpler methods using comparable constraints are needed to establish a meaningful advantage.

I encourage the authors to examine the practical motivation more carefully and critically assess whether the proposed framework offers a meaningful solution to a concrete real-world problem.

In addition, I have the following technical concerns:

  1. Joint flow-matching validity. The causal mask in Eq. (18) restricts access to noisy latent coordinates, potentially preventing recovery of the intended joint distribution. This construction requires justification.

  2. Local versus integrated sensitivity. Correctly signed velocity derivatives in Eqs. (8)–(9) do not guarantee correctly signed terminal treatment responses because effects propagate through unconstrained latent-state dynamics.

  3. Interpretation of sensitivity bounds. The pharmacokinetic smoother has learned gains, making the Jacobian bounds dependent on feature scaling. Their interpretation as physiological dose-response limits therefore needs clarification.

Questions for authors
  1. Can the authors identify a concrete clinical treatment-planning problem, supported by real-world evidence or a semi-synthetic study grounded in patient data, where this framework offers a meaningful advantage over existing approaches? Please specify who would use it, which treatment decision it would inform, and why simpler methods are inadequate. Are there documented applications of comparable methods supporting this use case?

  2. What evidence in the present study shows that improved sign and dose-ordering consistency translates into accurate responses to altered treatments or better dose selection? Please distinguish demonstrated benefits from anticipated future applications.

  3. Please address the technical concerns raised in points 3–5 above.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.