Machine Learning-Based Interpretation and Visualization of Nonlinear Interactions in Prostate Cancer Survival.

Abstract

Shapley additive explanation (SHAP) values represent a unified approach to interpreting predictions made by complex machine learning (ML) models, with superior consistency and accuracy compared with prior methods. We describe a novel application of SHAP values to the prediction of mortality risk in prostate cancer. Patients with nonmetastatic, node-negative prostate cancer, diagnosed between 2004 and 2015, were identified using the National Cancer Database. Model features were specified a priori: age, prostate-specific antigen (PSA), Gleason score, percent positive cores (PPC), comorbidity score, and clinical T stage. We trained a gradient-boosted tree model and applied SHAP values to model predictions. Open-source libraries in Python 3.7 were used for all analyses. We identified 372,808 patients meeting the inclusion criteria. When analyzing the interaction between PSA and Gleason score, we demonstrated consistency with the literature using the example of low-PSA, high-Gleason prostate cancer, recently identified as a unique entity with a poor prognosis. When analyzing the PPC-Gleason score interaction, we identified a novel finding of stronger interaction effects in patients with Gleason ≥ 8 disease compared with Gleason 6-7 disease, particularly with PPC ≥ 50%. Subsequent confirmatory linear analyses supported this finding: 5-year overall survival in Gleason ≥ 8 patients was 87.7% with PPC < 50% versus 77.2% with PPC ≥ 50% (P < .001), compared with 89.1% versus 86.0% in Gleason 7 patients (P < .001), with a significant interaction term between PPC ≥ 50% and Gleason ≥ 8 (P < .001). We describe a novel application of SHAP values for modeling and visualizing nonlinear interaction effects in prostate cancer. This ML-based approach is a promising technique with the potential to meaningfully improve risk stratification and staging systems.