Interpretable Clustering for Patient Phenotyping Using Advanced Machine Learning Models

Essential to personalized treatment in cardiology is identifying patient phenotypes and interpreting their association with post-interventional outcomes, such as in transcatheter aortic valve replacement.After obtaining the phenotypes via unsupervised clustering methods, one method is to use machine learning to predict cluster membership in order to interpret patient characteristics defining each cluster.In previous work, we used decision trees due to their ease of interpretability, which is limited in capturing complex relationships in the data.We mitigate this issue by integrating SHapley Additive exPlanations (SHAP) values with advanced machine learning algorithms.To demonstrate our approach, we identified six distinct patient clusters using single-center data from 581 TAVR patients using k-means and applied various classifiers to classify cluster memberships, with Light Gradient Boosting performing best (F1-score: 0.879).SHAP values were used to interpret cluster memberships, revealing key features influencing outcomes.For example, creatinine level and annulus area were significant predictors for high-risk phenotypes.This SHAP-based approach balances model complexity with interpretability, providing robust, detailed insights into patient phenotypes.