A PhD thesis proposes three methodological innovations and a design framework for bringing causal machine learning into clinical practice, testing continuous treatment models on a physiotherapy dosage case

The thesis identifies eight key decision points that directly influence treatment effect estimates and proposes a structured framework for designing causal models in clinical settings, to guide these choices and avoid unintended model behaviour. Causal models are often developed in theoretical settings that leave a gap with the requirements of real-world clinical practice; the framework makes the translation from theory to practice explicit and responds to the difficulty that key assumptions are often hard to verify and violations are not straightforward to address. Based on the thesis abstract's own account of the framework and the eight decision points; this is a methodological and process-level contribution, with no quantitative evaluation reported in the abstract.

The thesis proposes a novel method that grounds average model predictions in a group-level quantity that can be empirically verified using, for example, randomized or real-world clinical trial data, and shows it is possible to construct a model with a lower counterfactual error bound, with improved performance through information bottleneck regularization. Existing literature derives bounds on the counterfactual prediction error, but their practical relevance in a clinical context is not always clear, since even a small error bound may correspond to prediction errors large enough to make a model unsuitable for clinical use; this work grounds verification in an empirically checkable group-level quantity and uses information bottleneck regularization to learn more informative and robust representations. Based on the abstract's statements about the method and its effects; the abstract gives no specific error-bound values, dataset sizes, or control settings.

The thesis studies a concrete medical use case, the effect of a continuous treatment, physiotherapy dosage, on functional independence, to investigate whether continuous causal models perform well in practice and yield clinically meaningful outputs. Binary treatment models reduce treatment modelling to a yes-no decision, whereas dosage (continuous) models more realistically capture varying treatment intensities; continuous causal models are still relatively new and it remains unclear whether they perform well in practice or yield clinically meaningful outputs, and this study targets that gap. Based on the abstract's description of the use case; the abstract does not report the study's specific results, sample size, or effect sizes.

The thesis proposes a novel meta-reinforcement learning algorithm that leverages multiple datasets to learn more robust and generalizable policies, addressing generalization challenges arising from limited data and variability across hospitals or patient populations. Reinforcement learning aims to learn an optimal policy directly rather than first estimating treatment effects and likewise relies on counterfactual reasoning, but generalization remains a key challenge in clinical settings; meta-RL enables rapid adaptation to new contexts such as a new hospital setting, and this work advances that direction by learning from multiple datasets. Based on the abstract's account of the algorithm's approach and goal; the abstract gives no experimental benchmarks, comparison methods, or performance numbers.

The work targets researchers and practitioners who want to use causal machine learning models in real-world clinical practice, especially in settings involving binary treatment decisions, continuous dosage treatments, and generalization across hospitals or patient populations. Its design framework is meant to guide choices when building causal models in clinical settings; its group-level grounding method is meant to empirically verify average predictions using randomized or real-world clinical trial data; and its meta-reinforcement learning algorithm is meant to use multiple datasets to obtain more robust and generalizable policies. The continuous treatment part uses the effect of physiotherapy dosage on functional independence as a concrete use case, indicating that its conclusions apply first to that medical context and its data types.

A careful reader would still watch: how the eight key decision points are operationalized in concrete clinical projects and how usable the structured framework is under real team and data conditions; how individual-level counterfactual prediction error is constrained once average predictions are grounded in a group-level quantity, and whether that group-level quantity is equally verifiable in non-randomized real-world data; how the continuous causal model actually performs and what clinical meaning it yields in the physiotherapy dosage and functional independence use case; how large the reduction in the counterfactual error bound and the improvement in treatment effect estimation from information bottleneck regularization are; and how robust the meta-reinforcement learning algorithm is when transferred across hospitals and patient populations, and what effect differences in data sources have. In addition, because the loaded text is an abstract without figures or experimental details, these questions cannot be answered from the current text.