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

Area of Science:

• Machine learning applications in oncology
• Predictive modeling for cancer outcomes
• Explainable artificial intelligence in healthcare

Background:

• Prostate cancer risk stratification relies on clinical and pathological factors.
• Accurate prediction of mortality risk is crucial for treatment decisions.
• Interactions between prognostic factors can be complex and difficult to model.

Purpose of the Study:

• To apply Shapley additive explanation (SHAP) values for predicting mortality risk in prostate cancer.
• To utilize SHAP values to model and visualize nonlinear interaction effects between prognostic factors.
• To explore the utility of machine learning interpretability methods in refining prostate cancer staging.

Main Methods:

• Utilized the National Cancer Database to identify 372,808 patients with nonmetastatic, node-negative prostate cancer (2004-2015).
• Trained a gradient-boosted tree model incorporating features: age, prostate-specific antigen (PSA), Gleason score, percent positive cores (PPC), comorbidity score, and clinical T stage.
• Applied SHAP values to interpret model predictions and analyze feature interactions using Python 3.7 open-source libraries.

Main Results:

• SHAP analysis confirmed known interactions, such as low PSA and high Gleason score predicting poor prognosis.
• A novel finding demonstrated stronger interaction effects between PPC and Gleason score (≥8) compared to Gleason 6-7, especially with PPC ≥50%.
• Confirmatory analyses showed significantly lower 5-year overall survival in Gleason ≥8 patients with PPC ≥50% (77.2%) versus PPC <50% (87.7%), P < .001.

Conclusions:

• Shapley additive explanation (SHAP) values provide a powerful method for modeling and visualizing nonlinear interactions in prostate cancer.
• This machine learning-based approach shows promise for enhancing risk stratification and staging systems.
• The findings highlight the importance of considering the interplay between percent positive cores and Gleason score in predicting prostate cancer outcomes.