Comparative Evaluation of Global and Personalised Machine Learning Models for Precision Medicine

This paper elaborates a comparative analysis of two of the most important paradigms in modeling, namely global models, which are fitted to aggregate population-level data, and personalized models, which are fitted to individual or neighborhood-level similarities. A Random Forest classifier (global) and K-Nearest Neighbors model (personalized) were modeled and evaluated using a dataset of 500 simulated patients that included clinical and genomic variables. The evaluations of the performance included discrimination (ROC–AUC), calibration (Brier score), and interpretability analysis. Findings showed that the global model was more accurate (0.712) and has better calibration (Brier score = 0.237) than the personalized model (accuracy = 0.693; Brier score = 0.254). Nonetheless, the personalized model showed enhanced flexibility to subgroups with rare patterns, which features the character of its use in personalized prediction of treatment responses. The paper concludes that although global models are still applicable to the large-scale implementation, hybrid approaches to combine both paradigms can achieve the best performance, interpretability, and personalization, which are consistent with next-generation AI-based healthcare systems.