Causal Machine Learning in the Real World

Abstract

Causal inference has been studied for a long time. However, it is not always clear how current causal models can best be adapted for practical use. In particular, there remains a gap between the theoretical settings in which these models are developed and the requirements of real-world clinical practice. Furthermore, some well-known models do not fully meet the practical and theoretical requirements of clinical decision-making. Thus, on multiple fronts, further work is needed to make causal machine learning models suitable for real-world clinical practice. This dissertation therefore examines how existing models can be translated into practice and what improvements are needed for more experimental methods such as continuous causal models. We address this along three lines: (1) Binary Treatment Effect Estimation in Clinical Contexts; (2) Continuous Causal Machine Learning and (3) Meta-Reinforcement Learning. We examine what is required to bring causal models into clinical practice, identifying eight key decision points that directly influence treatment effect estimates. To guide these choices and avoid unintended model behaviour, we propose a structured framework for designing causal models in clinical settings. Such guidance is essential, as translating causal inference theory into practice is challenging: key assumptions are often difficult to verify, and violations are not straightforward to address. Furthermore, due to the fundamental problem of causal inference, verifying whether counterfactual predictions are correct is inherently difficult. Existing literature attempts to address this issue by deriving bounds on the counterfactual prediction error. While these bounds are theoretically useful, their practical relevance in a clinical context is not always clear: even a small error bound may still correspond to prediction errors large enough to make a model unsuitable for clinical use. For this reason, we propose 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. Binary treatment models reduce treatment modelling to a yes–no decision, whereas dosage (continuous) models more realistically capture varying treatment intensities. However, continuous causal models are still relatively new, and it remains unclear whether they perform well in practice or yield clinically meaningful outputs. To investigate this, we conduct a study based on a concrete medical use case, examining the effect of a continuous treatment, physiotherapy dosage, on functional independence. Next, we introduce a novel method and show that it is possible to construct a model with a lower counterfactual error bound. In addition, we demonstrate improved performance through the use of information bottleneck regularization, which encourages the model to learn more informative and robust representations for treatment effect estimation. Lastly, we examine an alternative paradigm: reinforcement learning (RL), which aims to learn an optimal policy directly rather than first estimating treatment effects. Like causal inference, RL relies on counterfactual reasoning to evaluate decisions or interventions. However, generalization remains a key challenge, especially in clinical settings with limited data and variability across hospitals or patient populations. Meta-RL addresses this by enabling rapid adaptation to new contexts (e.g., a new hospital setting). We contribute to this area by proposing a novel meta–reinforcement learning algorithm that leverages multiple datasets to learn more robust and generalizable policies. In conclusion, this thesis has made contributions to causal inference and reinforcement learning. Specifically, the goal was to develop better models and ensure that these models can be used in real-world clinical practice by taking clinical considerations into account. To this end, we presented three methodological innovations: two in causal inference and one in reinforcement learning. Furthermore, we proposed a framework to help translate models from the lab into clinical practice.