Explainable and visualizable machine learning models to predict biochemical recurrence of prostate cancer

Area of Science:

• Oncology
• Medical Informatics
• Biostatistics

Background:

• Machine learning (ML) models show promise in predicting prognosis but often lack clinical interpretability due to their 'black box' nature.
• Explainable AI (XAI) techniques are crucial for translating complex ML models into clinically actionable insights.

Purpose of the Study:

• To develop and validate explainable and visualizable ML models for predicting biochemical recurrence (BCR) in prostate cancer (PCa).
• To enhance the clinical utility of ML-based prognostic tools for PCa management.

Main Methods:

• Retrospective analysis of 647 PCa patients.
• Feature selection using LASSO regression to identify 11 BCR-related clinical parameters.
• Development of prediction models using Cox regression and five ML algorithms (RSF, SSVM, sTree, GBDT, XGBoost).
• Model performance evaluation via concordance index (C-index) and decision curve analysis (DCA).
• Interpretability analysis using Shapley Additive Explanation (SHAP) values.

Main Results:

• The Random Survival Forest (RSF) model achieved the highest C-index (0.846), outperforming other ML models and Cox regression.
• Decision curve analysis indicated superior clinical utility for the RSF model.
• SHAP values provided clear insights into the contribution of each feature within the RSF model.

Conclusions:

• The developed explainable ML models, particularly RSF, offer a reliable framework for predicting PCa BCR.
• These models can serve as a reference for identifying patients at risk of recurrence and inform personalized therapeutic strategies.