What happened
A study accepted at Machine Learning for Healthcare 2026 evaluated five families of offline reinforcement-learning algorithms and 14 reward designs for antithrombotic treatment after acute ischemic stroke. Using 44,894 post-2018 patients from a nationwide registry of 129,033 people, the researchers found that standard evaluation initially suggested a small policy improvement. The signal became larger when the reward included a penalty for early neurological deterioration. After reward deconfounding, however, the estimated benefit fell and was no longer statistically distinguishable from zero.
The paper, submitted to arXiv on August 31, 2026, evaluates offline reinforcement learning for antithrombotic treatment in acute ischemic stroke. Its data come from a nationwide registry containing 129,033 patients; the main analysis covers 44,894 patients treated after 2018. The evaluation compares five offline reinforcement-learning algorithm families across 14 reward designs. The paper is accepted at Machine Learning for Healthcare 2026 and is scheduled to appear in the Proceedings of Machine Learning Research, volume 340.
The initial results produced a positive policy-improvement estimate of +0.0069 under standard Fitted Q-Evaluation, or FQE. When the reward design added a penalty for Early Neurological Deterioration, the apparent improvement increased to +0.0101. The authors argue that this signal was not a clean measure of treatment efficacy because the terminal reward also captured baseline disease severity and prognosis. In other words, the reward could partly reflect which patients were more likely to have poor outcomes, independently of the treatment decision being evaluated.
A 2-by-2 factorial analysis attributed 218.6% of the observed signal change to terminal-reward confounding; the authors note that simply removing that component overshot the null. After a DML-inspired gradient-boosting-machine reward residualization, the FQE estimate declined to +0.0033, with p = 0.132. Under full reward deconfounding, it declined further to +0.0025, with p = 0.291. The abstract says FQE-based diagnostics, T-learner analyses and direct recurrence analyses all moved away from a clinically meaningful aggregate improvement, and that a one-year modified Rankin Scale factorial analysis reproduced the attenuation.
Why it matters
The paper identifies a specific way clinical AI evaluations can make a treatment policy look better than it is: a reward can encode patients’ baseline severity and prognosis as well as the effect of treatment. That matters because systems trained and evaluated on observational medical records may be judged on outcomes they did not cause. The study’s results suggest that apparent gains in offline reinforcement learning should not be treated as evidence of clinical benefit without careful checks for confounding.
The study’s central implication is about evaluation validity, not a new treatment recommendation. An offline reinforcement-learning system can be assessed against historical clinical records, but the outcome signal in those records may combine treatment effects with patients’ starting conditions. If a reward function carries forward baseline severity or prognosis, an algorithm can appear to have improved outcomes because it is being scored partly on information about who was already more or less likely to recover. This is a methodological warning about evaluation validity, not a treatment recommendation.
That distinction is consequential for medical AI because a positive retrospective estimate can be mistaken for evidence that an automated policy should guide care. The paper shows that the estimated advantage changed materially as the researchers addressed reward-embedded confounding: from +0.0069 under standard FQE to +0.0025 after full deconfounding. The source does not claim that offline reinforcement learning is useless; it reports that the aggregate improvement in this evaluation was not clinically meaningful after the confounding analysis.
The research also illustrates why a single evaluation metric is insufficient for high-stakes clinical systems. The authors used several analyses, including FQE diagnostics, T-learner analyses, direct recurrence analyses and a one-year modified Rankin Scale factorial analysis. Their abstract says these methods converged away from a meaningful aggregate improvement. That convergence strengthens the paper’s methodological warning, although it remains the authors’ analysis of one registry-based study rather than independent confirmation across datasets or clinical settings.
What to watch next
The authors propose a six-step evaluation checklist and report that several diagnostics converged away from a clinically meaningful aggregate improvement after deconfounding. The source does not establish whether any policy would improve patient outcomes in prospective clinical use, and it does not report a deployment or randomized trial. NIHSS-stratified differences are described as hypothesis-generating for future prospective research, while hospital-level disagreement did not persist after full reward deconfounding.
The paper provides an empirically motivated six-step checklist for evaluating offline reinforcement-learning policies in clinical settings. The source does not list the six steps in the arXiv record’s abstract, so their exact contents and implementation details require review of the full paper. A practical next question is whether researchers evaluating other medical decisions can reproduce the same confounding pattern when rewards incorporate prognosis, severity or deterioration measures.
The authors report NIHSS-stratified heterogeneity, but explicitly characterize it as hypothesis-generating for prospective trial design. That means the subgroup pattern should not be treated as evidence that a particular stroke severity group will benefit from an AI-guided policy. The source also says hospital-level disagreement did not persist after full reward deconfounding, reducing support for an interpretation based on persistent differences between hospitals.
The main unknown is whether any evaluated policy would improve patient outcomes when used prospectively. The source reports no randomized trial, prospective deployment, clinical adoption, independent replication or patient-level safety assessment. It also does not establish how the findings generalize beyond this nationwide registry, the post-2018 subset, the antithrombotic-treatment decision or the reward designs studied. Those questions should be resolved before the results are used to justify clinical implementation.