Benchmarking Hybrid Learning Models for Personalized Risk Classification and Prediction in Oncology
Abstract
In the field of oncology, and specifically regarding prostate cancer, recent breakthroughs in Artificial Intelligence (AI) have opened fascinating possibilities for risk classification and prediction. However, moving from theoretical success to actual clinical practice remains a significant challenge. In this study, we chose to evaluate several architectures (generative, discriminant, and hybrid models) to test their real-world relevance when faced with longitudinal medical data, which is often noisy or incomplete. Our analysis does not merely measure raw performance; it seeks to understand the critical balance between statistical precision and the interpretability that clinicians require for daily decision-making. By examining the inherent strengths and limitations of each method, we developed an advanced hybrid approach designed as a genuine support tool for the patient care pathway. Our findings are particularly encouraging: these deep hybrid models significantly outperform traditional methods, showing a notable improvement of 6% to 9% in both AUC-ROC and C-index scores. Beyond these metrics, we observed a superior ability in the model to adapt to the temporal variations of patient data over time. Ultimately, our goal is to demonstrate that more refined predictions, when properly integrated into the clinical context, can simplify the decision-making process. By making these models more transparent and robust, we hope to facilitate their adoption by the medical community, leading to more personalized and effective management for patients battling prostate cancer.