Abstract
Background
Progression-free survival (PFS) is a critical end point in oncology, yet real-world applications of individualized, explainable machine learning (ML) predictions remain limited.
Objective
This study aimed to develop and validate explainable ML models to predict PFS using retrospective data from a national prostate cancer cohort in Brunei Darussalam.
Methods
We analyzed a retrospective cohort of 212 patients (478 longitudinal observations) treated at the Brunei Cancer Centre (January 2018 to December 2024). Clinical, laboratory, and treatment data were harmonized, with missing values imputed via extremely randomized trees. Longitudinal patterns were captured using a recurrent autoencoder to generate latent representations. We compared 4 modeling approaches: Cox proportional hazards, random survival forest (RSF), gradient boosting survival (GBS), and deep neural network survival models. Performance was evaluated using time-dependent area under the receiver operating characteristic curve (AUC), Harrell C-index, and integrated Brier score (IBS), with Shapley additive explanations (SHAP) used for interpretability.
Results
RSF demonstrated improved discriminative performance and balanced calibration, achieving a C-index of 0.906 and AUCs of 0.941 and 0.917 at 4 and 5 years (IBS=0.0698). In contrast, the traditional CPH model performed poorly (C-index 0.531 and AUC 0.706 at 4 years and 0.833 at 5 years). Deep survival (AUCs of 0.941 at 4 years and 0.917 at 5 years, C-index 0.719, and IBS=0.0887) and GBS (AUCs of 0.765 at 4 years and 0.833 at 5 years, C-index 0.844, and IBS=0.0590) models showed moderate performance. SHAP analysis identified sodium, alanine aminotransferase, mean corpuscular hemoglobin, platelet count, and specific treatment categories as key drivers of increased progression risk.
Conclusions
Tree-based ensemble approaches, particularly RSF integrated with SHAP, offer high accuracy for personalized risk stratification in prostate cancer. These findings highlight the potential of explainable ML to enhance clinical decision-making. However, external validation in a larger multi-institutional, multiomics dataset is required before routine clinical implementation.