Explainable and visualizable machine learning model development and validation for 5-year postoperative survival prediction in prostate cancer patients aged ≥ 65 years

Abstract Background Machine Learning (ML) models have achieved outstanding performance in predicting post-surgical survival. However, the “black-box” nature of ML models restricts their clinical application. This study aims to develop and validate a clinically feasible, machine learning-based online prediction system for predicting the 5-year survival status of patients with prostate cancer (PCa) after surgery. Methods This study conducted a retrospective analysis of clinical data from 300 older adults with PCa aged ≥ 65 years. LASSO regression, random forest (RF), and recursive feature elimination (RFE) were employed to screen for clinical parameters. Subsequently, the cohort was split into a training set and a test set at a 7:3 ratio, and 25 machine learning approaches were utilized for comparative assessment to determine the optimal model. Calibration curves were applied to evaluate the performance of each model, while decision curve analysis (DCA) was adopted to assess their clinical usefulness. In addition, SHAP (Shapley Additive exPlanations) values were used to interpret the model features. Finally, the Shiny framework was employed to develop an online prediction system. Results The intersection of the three feature selection algorithms identified 13 clinical parameters: Age, ALB, BUN, CRE, HB, PLT, PT, PSA, GS, PNI, PSM, pT and pN. Comparing 25 machine learning algorithms, LightGBM gave the best results. Its performance metrics were: Accuracy 0.9328, Sensitivity 0.9074, Specificity 0.9538, Positive Predictive Value (PPV) 0.9418, Negative Predictive Value (NPV) 0.9118, AUC 0.9778, Recall 0.9074, F1 Score 0.9143. SHAP values revealed the contribution of each feature in the LightGBM model. Conclusion This study successfully developed a 5-year postoperative survival prediction model for prostate cancer patients. The model demonstrated favorable predictive performance in the test set, which may provide a reference for clinical decision-making. Further multi-center external validation is required to clarify its clinical application value in the future.