acceptodds
Withdrawn from ICLR 2027

CauTwin: Counterfactual Contrast Supervision for Interventional Prediction in Clinical Time Series

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Estimating how a patient would respond to a treatment that was not given is central to clinical decision support, and electronic health records make it possible to learn such responses from observational data. Existing methods for treatment effects over time adjust for time-varying confounding through balanced representations, inverse propensity weighting, G-computation or meta-learners, and have made effect estimates far more reliable. In intensive care, however, two problems remain. First, treatments are given to patients who are already deteriorating, so models confuse this association with the effect of the treatment and may conclude, for example, that a vasopressor lowers blood pressure. Second, each patient is observed under one treatment only, so the data never show how the effect differs between patients, and models tend to give nearly the same answer to everyone. We propose CauTwin, which learns the effect of a treatment on each patient directly. For every patient, it finds similar patients who received the other treatment and uses this comparison to guide the model's own interventional prediction. Because the comparison remains biased on average by severity that the records do not capture, CauTwin uses only how it differs between patients, which recovers how the effect varies across patients, and sets the average effect with adversarial balancing. We show that the two components compete over how the treatment may depend on severity, and that their weights control this trade-off. On three ICU cohorts, CauTwin is the only method whose vasopressor effects both point in the pharmacologically correct direction and vary with the patient, and it attains the lowest counterfactual error on a semi-synthetic tumour growth benchmark. We further show that predictive accuracy is a misleading criterion for selecting such models.

Primary area:
learning on time series and dynamical systems
Keywords:
clinical time series, counterfactual prediction, intervention modeling, personalized medicine

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_h7jgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CauTwin to estimate patient-specific treatment effects from observational data. It uses propensity score matching to construct treatment effect targets and introduces a batch-centered contrast loss to learn differences in treatment effects across patients. The paper evaluates the method on three ICU datasets and the results show improved treatment effect direction and patient dependence on ICU data.

Critical strengths

The combination of contrast supervision and adversarial balancing is novel.

The experiments include three ICU datasets from different hospital systems and provides detailed implementation settings.

Critical weaknesses

Some key assumptions and design choices need further clarification, including the shared matching bias assumption and the different treatment definitions used in matching and counterfactual rollout.

Although modeling ICU treatments is challenging, the authors should better consider real clinical practice, including differences in vasopressor types and treatment duration. The current evaluation provides limited evidence of treatment effect accuracy and clinical utility. Moreover, the semi-synthetic experiments do not use the proposed contrast supervision.

Questions for authors
  1. Counterfactual rollout and treatment definition. Appendix C.1 defines the counterfactual rollout as sustained vasopressor treatment over 12 hours, and Appendix C.2 defines a patient as treated if vasopressors were administered at any point during the window (median duration of 7 hours). How does this difference affect the estimated treatment effects? Batch centering removes a shared offset, but may not account for differences in treatment duration across patients.

  2. Treatment conditioning. In Section 4.1, the authors state that an additive treatment input would shift every patient's prediction by the same amount and therefore cannot represent patient-dependent treatment effects. A nonlinear Transformer should already be able to model interactions between treatment inputs and patient representations.

3.Assumption on matching bias. The contrast loss assumes that matching errors are approximately shared across patients. What evidence supports this assumption in the ICU cohorts?

4.Clinical treatment setting. Have the authors considered the differences among vasopressor types?

5.Evaluation metrics. The ICU evaluation mainly relies on Correct Direction and Patient Dependence. Could the authors provide additional analyses, such as examining whether patients who received treatments consistent with the model's recommendations had better clinical outcomes than those who did not?

6.Other concerns. Appendix E.2 shows that the tumour benchmark did not use the proposed contrast supervision. Could the authors explain why the main method was not used in this experiment?

Reviewer_GBVQOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CauTwin for interventional prediction in clinical time series, addressing the incorrect direction and limited patient dependence of predicted treatment effects. The method introduces centered counterfactual contrast supervision based on propensity-score matching to learn the shape of treatment effects, while using adversarial balancing to anchor their level. The paper further analyzes the interaction between these two objectives and evaluates the method on three ICU cohorts and a semi-synthetic tumour growth benchmark.

Critical strengths

S1. The paper identifies a meaningful problem in individualized treatment effect estimation and provides a clear level–shape perspective.

S2. The proposed counterfactual contrast supervision offers an interesting approach to directly supervising patient-level treatment differences.

S3. The ablation studies demonstrate the distinct roles and trade-off between contrast supervision and adversarial balancing.

Critical weaknesses

W1. Reliance on the Shared Matching Bias Assumption. The core idea of the proposed method is to remove systematic bias in the matched pseudo-labels through centering. As stated in Assumption 2, this theoretical guarantee relies on the assumption that the matching bias is constant across patients. In practice, the matching bias may vary with patient characteristics, and such patient-dependent bias cannot be eliminated by centering and may be incorrectly learned as treatment effect heterogeneity. Therefore, the applicability of this assumption and the conditions under which the proposed method can reliably recover the shape require further validation.

W2. Insufficient Direct Experimental Validation of the Core Method. Related to W1, I think the current experiments do not sufficiently validate the effectiveness of centered counterfactual contrast supervision when the shared matching bias assumption is violated. The real-world ICU experiments mainly rely on indirect metrics, which cannot directly measure the error in recovering the true shape. Therefore, the experimental evaluation lacks direct validation of the core method, particularly a systematic analysis of its ability to recover the shape when matching bias varies across patients.

Questions for authors

My main questions concern the weaknesses discussed above. If the authors can address these concerns, I would be willing to reconsider my score.

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

This paper tackles the classical problem of counterfactual inference in time series data and propose a method for this setting that directly supervises a treatment effects training of the same model used to predict the potential outcomes (in this case, these are trajectories). The authors motivate their setup using clinical examples where confounding by indication, which can happen in this time series setup because patients who are deteriorating are more likely to be treated (e.g. vasopressors in the ICU).

The proposed methods constructs the training labels for treatment effects by comparing propensity matched patients receiving opposite treatments, and then centers these targets and predictions to learn variation between patients while removing the "level bias" shared across patients. The method also adds an adversarial balancing loss term. The authors provide a theoretical justification for this method, and report some clinically plausible vasopressor responses across multiple ICU dataset. They also show that selecting models for accurate factual predictions can favor less plausible treatment effect estimates.

Critical strengths

The paper focused on a clinically meaningful setting with comprehensive experiments and detailed reporting of those experiments.

Critical weaknesses

Overall, this paper tackles a really classical problem since the issue of regularization bias

The paper introduces a method to address this problem with a very modest and incremental that mostly recycles existing ideas. In their rebuttal, it would be great if the authors address the following points.

(1) The idea of creating pseudo-outcomes to target the training process towards the treatment effects instead of the potential outcomes is in fact a very standard approach in causal effect estimation, e.g., through meta-learner approaches as the authors acknowledge in lines 119-131. The authors make a specific choice of creating these pseudo-outcomes through propensity matched averages, but this idea itself is also not new. There are many methods that also use similar matching strategies such as Perfect Match (2018), SyncTwin (NeurIPS 2021), and PairNet from ICML 2024, which is not cited in this submission. It is not really clear where the novelty in the proposed method is and the authors argument about their method as "This trading efficiency for robustness to a bias shared across patients." in line 131 is also not very clear.

(2) There are also some technical issues with the rationale and assumptions behind the proposed method. The proposed method assumes that its matched patient targets equal the true treatment effect plus the same bias for every patient so that subtracting the average then removes that bias. However, there is not reason to believe this would be true under the proposed propensity matching approach and it is easy to construct counterexamples where the proposed method induces wrong heterogeneity when subgroups of patients have different baseline untreated outcomes but the same propensity scores and treatment effects.

Additional, I think that the the supervision target/label and evaluated intervention are not the same causal quantity (Appendix C.1-C.2, pp. 19–20). The authors use a matching target that compares patients receiving any vasopressor during a time window against patients receiving none. The treated patients receive vasopressor for a median of seven hours, but the model instead predicts 12 continuous hours of vasopressor treatment with the other treatment inputs set to zero. Wouldn't the differences in duration/timing of treatments change the vasopressor effects differently across patients? In this case, why would subtracting one average value correct that bias?

(3) I am unsure if the authors ever demonstrate the validity of their methods in a controlled experiment with semi-synthetic data and known counterfactual as is standard in this literature. It seems that the main experiment with known counterfactual outcomes does not test the proposed method as described in the main text (appendix E.2-E.3, pp. 26–27.) On the tumour benchmark, the CauTwin method uses a substantially different configuration from the one described in the main paper with separate effect heads, an additional effect-boost objective, and an R-learner training stage, and more importantly it seems to omit the contrast supervision loss which is the key contribution of the paper.

Questions for authors

Please address the points highlighted in the critical weaknesses section.

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

Related papers

Discussion (0)

Sign in to comment.