Application of a novel machine learning framework for predicting cancer-specific mortality: analysis of 171,942 men with non-metastatic prostate cancer from the Surveillance, Epidemiology, and End Results dataset

Abstract

Abstract Background Accurate prognostication is vital for treatment decisions for men diagnosed with non-metastatic prostate cancer. Current models rely on pre-specified variables and hence have limits to their performance. We aimed to investigate a novel machine learning (ML) approach to develop an improved prognostic model for 10-year prostate cancer-specific mortality (PCSM) and compare performance to existing validated models. Methods We derived and tested a ML-based model using Survival Quilts, an algorithm that automatically selects and tunes ensembles of survival models, utilizing clinico-pathological variables. Our study involved a population-based cohort of 171,942 men diagnosed with non-metastatic prostate cancer between 2000 and 2016 from the prospectively-maintained United States Surveillance, Epidemiology, and End Results (SEER) program. Our primary outcome was prediction of 10-year PCSM. Model discrimination was assessed through concordance index (C-index), and calibration through Brier scores and the use of Decision Curve Analysis. Comparison was made to 9 other currently used models. Findings Discrimination improved with greater granularity, and multivariable models outperformed group-based systems. The Survival Quilts model showed good discrimination (C-index 0.829, 95% confidence interval (CI) 0.820-0.838) for PCSM. Discrimination was similar to the top-ranked multivariable models: PREDICT Prostate (C-index 0.820, 95% CI: 0.811-0.829), and Memorial Sloan Kettering Cancer Center (MSKCC) nomogram (C-index 0.787, 95% CI: 0.776-0.798). All 3 models were well-calibrated (Brier scores: Survival Quilts 0.036, 95% CI: 0.035-0.037; PREDICT Prostate 0.036, 95% CI: 0.035-0.037; MSKCC 0.037, 95% CI: 0.035-0.039). C-indices for other models ranged from 0.711 (95% CI: 0.701-0.721) to 0.761 (95% CI: 0.750-0.772). Discrimination for the Survival Quilts model was maintained when stratified by age and ethnicity. Decision Curve Analysis further showed an incremental net benefit from applying the Survival Quilts model in comparison to currently used models. Interpretation A novel ML-based approach produced a prognostic model with discrimination comparable to the best existing prognostic models for 10-year PCSM with using standard clinico-pathological variables only. Further integration of additional data will likely improve ML-generated model performance and accuracy for personalized prognostics. Limitations of our study include relatively few death events and bias from missing data.