Optimizing the Cost-Benefit of Diagnostic Tests in Clinical Decision-Making
In clinical practice, physicians rely on predictive models to assess disease risk and determine optimal treatments based on patient-specific features, such as demographic data and lab results. While more comprehensive models incorporating a wide range of features tend to offer higher accuracy, the associated costs—whether financial, temporal, or health-related—pose significant constraints. This paper addresses the challenge of optimizing the cost-benefit of diagnostic tests by introducing a novel approach that adapts to the unique characteristics of each patient. We propose a reinforcement learning-based methodology, framed within the Q-learning framework, to optimize the selection and sequencing of diagnostic tests, balancing the need for accurate predictions with the cost of feature collection. Additionally, our algorithm effectively handles informative missing data through a novel importance weighting procedure, ensuring robust performance even when critical predictors are not fully observed. This approach enhances the efficiency and effectiveness of clinical decision-making by minimizing costs while maintaining high predictive accuracy. Through theoretical development, practical applications, and empirical validation, we demonstrate the advantages of this cost-benefit optimization strategy in clinical settings.