Machine learning for prediction of survival outcomes with immune‐checkpoint inhibitors in urothelial cancer

Machine learning (ML) may enhance the efficiency of developing accurate prediction models for survival, which is critical in informing disease prognosis and care planning. This study aimed to develop an ML prediction model for survival outcomes in patients with urothelial cancer-initiating atezolizumab and to compare model performances when built using an expert‐selected (curated) versus an all‐in list (uncurated) of variables. Gradient‐boosted machine (GBM), random forest, Cox‐boosted, and penalised, generalised linear models (GLM) were evaluated for predicting overall survival (OS) and progression‐free survival (PFS) outcomes. C‐statistic (c) was utilised to evaluate model performance. The atezolizumab cohort in IMvigor210 was used for model training, and IMvigor211 was used for external model validation. The curated list consisted of 23 pretreat-ment factors, while the all‐in list consisted of 75. Using the best‐performing model, patients were stratified into risk tertiles. Kaplan–Meier analysis was used to estimate survival probabilities. On external validation, the curated list GBM model provided slightly higher OS discrimination (c = 0.71) than that of the random forest (c = 0.70), CoxBoost (c = 0.70), and GLM (c = 0.69) models. All models were equivalent in predicting PFS (c = 0.62). Expansion to the uncurated list was associated with worse OS discrimination (GBM c = 0.70; random forest c= 0.69; CoxBoost c= 0.69, and GLM c = 0.69). In the atezolizumab IMvigor211 cohort, the curated list GBM model discriminated 1‐year OS probabilities for the low‐, intermediate‐, and high‐risk groups at 66%, 40%, and 12%, respectively. The ML model discriminated urothelial‐cancer patients with distinctly different survival risks, with the GBM applied to a curated list attaining the highest performance. Expansion to an all‐in approach may harm model performance.