Machine-learning survival models for predicting time to recurrence in epithelial ovarian cancer

To evaluate the effectiveness of machine learning survival models to predict time to recurrence using information from patient medical records known at the completion of frontline chemotherapy.

Five survival models - Penalized Cox Proportional Hazards (PenCoxPH), Random Survival Forest (RSF), Gradient Boosted Survival Analysis (GBSA), DeepSurv, and FastCPH - were trained on medical record data and used to predict time to recurrence. The models were trained on both the full set of patients and high-stage (III and IV) patients only. They were trained on the full-length (total time to recurrence) data as well as short-horizon (restricted to 15 months) recurrence data to increase prediction accuracy for the first year following completion of frontline chemotherapy. Feature hazard ratios for the PenCoxPH model were evaluated.

GBSA received the highest performance scores when predicting full-length time to recurrence, while the DeepSurv and RSF models did best on predictions for short-horizon recurrence. GBSA achieved CD-AUC (cumulative/dynamic AUC) measures above 0.8 at 2 and 3 years. Stage I, HRD Negative, NACT and a rise in CA125 over the course of frontline chemotherapy were significant predictors of recurrence. PenCoxPH and FastCPH achieved a 6-month CD-AUC of 0.74 the high-stage, high-grade serous cohort for full-horizon recurrence.

Machine learning survival models can predict time to recurrence with sufficient accuracy to be clinically useful. While confirmatory studies are needed to validate these findings, providers could potentially use this information to tailor treatment strategies in maintenance therapy and select patients for clinical trial enrollment.