Deviance-based logistic tree model and its applications

This study introduces the Deviance-based Logistic Tree (DbLT), a novel hybrid model integrating the interpretability of decision trees with logistic regression's predictive capability at each node. We explore the model's ability to unveil variable interactions, typically obscured in logistic regression. Our evaluation, using simulated and real datasets, focuses on predictive performance, computational efficiency, and variable selection, employing pruning methods such as Event Per Variable (EPV), Akaike Information Criterion (AIC), and Bayesian Information Criterion (BIC). In comparative analyses with standard classifiers using OASIS-3 data, DbLT demonstrated superior performance in identifying clinically validated SARA variables.