Data-Driven Counterfactual Optimization For Personalized Clinical Decision-Making

Chronic diseases have a significant impact on global mortality rates and healthcare costs. Notably, machine learning-based clinical assessment tools are becoming increasingly popular for informing treatment targets for high-risk patients with chronic diseases. However, using these tools alone, it is challenging to identify personalized treatment targets that lower the risks of adverse outcomes to a clinically acceptable range and are realistic, actionable, and robust to changes in these clinical assessment tools. Accordingly, we propose a data-driven approach called Distributionally Robust Selection of Clinical Role Models via Counterfactual Optimization (DISC²O). DISC²O overcomes computational intractability issues arising in a more general distributionally robust counterfactual optimization problem by its data-driven nature via two crucial components. First, it leverages a personalized database to tailor treatment targets and ensure that they are realistic and actionable. Second, it incorporates historical information from clinical assessment tools to mitigate the risk of failing to achieve treatment objectives. We prove that DISC²O is polynomial-time solvable using a greedy algorithm, although this algorithm may be computationally prohibitive for large datasets. Consequently, we develop an Active Learning Algorithm for Robust Role Model Selection (ALAR²MS), and establish its exponential convergence rate. Utilizing a large and granular clinical trial dataset of patients with co-morbid Type-2 Diabetes Mellitus and hypertension, we illustrate that DISC²O outperforms medical practice and can reduce patients’ risks of adverse outcomes in over 91% of patients, improving 32% over other benchmarks. Our findings highlight the effectiveness and practical relevance of DISC²O  in personalized treatment and clinical decision-making.