Causal inference and explainable machine learning for analyzing treatment side effect in metastatic castration-resistant prostate cancer patients

Optimal treatment recommendation for metastatic castration-resistant prostate cancer (mCRPC) are inherently diverse, being contingent upon individual patient response. Furthermore, treatment efficacy in specific patient cohorts can be influenced by confounding factors. Considering the substantial genetic heterogeneity among patients, generating population-level generalizations may compromise the precision and clinical applicability of predictive models. This study examines the prediction of treatment-induced adverse events in mCRPC patients using Explainable AI (XAI), focusing on both global and local levels of interpretability. Machine learning and other computational tools are often perceived as ”black-box” techniques, largely due to the challenge of linking their internal processes to the final model outputs. Consequently, XAI offers crucial insight into the specific features that the algorithms prioritize for prediction, thereby illuminating the opacity and decision-making intricacies of these ”black-box” models. Furthermore, causal inference was used to identify the attributes that specifically precipitate adverse events in patients with a smoking history. This analysis demonstrated that testosterone levels, prior analgesic use, and calcium levels act as confounders for adverse events within the smoking patients subgroup. The integration of causal inference and XAI establishes a robust and interpretable framework for making personalized treatment decisions in cancer care.