Machine learning enables accurate prediction of patient outcomes for immune checkpoint blockade using real-world clinical data

Background: While immune checkpoint blockade (ICB) therapy can produce durable clinical responses and substantially improve patient outcomes, accurate predictors of outcomes are critical as up to half of patients with advanced disease derive limited or no benefit. Accurate ICB outcome prediction will improve treatment stratification, reduce unnecessary toxicity, and enhance outcomes for cancer patients. Methods: We used state-of-the-art machine learning survival models to accurately predict patient survival after ICB therapy in our large multi-cancer institutional cohort. This retrospective study included 2,090 patients with advanced melanoma (n=908), advanced non-small cell lung cancer (NSCLC, n=878), or metastatic renal cell carcinoma (RCC, n=304) who underwent anti-PD-1/PD-L1 and/or anti-CTLA-4 ICB therapy at Moffitt Cancer Center from 2011-2025. Over 50 pre-treatment clinical and laboratory features were abstracted from electronic health records and analyzed against overall and progression-free survival. We trained and tested survival support vector machine models to predict patient outcomes with a 75/25 random split. Results: Cox PH analysis identified 11-44 statistically significant features per data set for inclusion into each model, including serum albumin, neutrophil-to-lymphocyte ratio, blood pressure, heart rate, and ECOG scores. Our models achieved AUC values of up to 0.83 in melanoma, 0.80 in NSCLC, and 0.85 in RCC, with enhanced performance in progression models trained on multi-cancer data. Our models outperformed PD-L1 and TMB at each time point where data is available. Conclusions: Our work demonstrates the promise of machine learning with accessible clinical and laboratory features to predict ICB patient outcomes with improved performance versus current biomarkers. Time-dependent receiver operating characteristic area under the curve (AUC) for our model vs PD-L1 Real-world data set Sample size Model AUC at 6mo, OS PD-L1 AUC at 6mo, OS Model AUC at 24mo, OS PD-L1 AUC at 24mo, OS Model AUC at 6mo, PFS PD-L1 AUC at 6mo, PFS Model AUC at 24mo, PFS PD-L1 AUC at 24mo, PFS Melanoma n = 908 Melanoma: 0.80 , Multi: 0.79 0.60 Melanoma: 0.72 , Multi: 0.70 0.51 Melanoma: 0.72, Multi: 0.79 0.64 Melanoma: 0.73, Multi: 0.83 N/A Non-small cell lung cancer n = 878 NSCLC: 0.77 , Multi: 0.70 0.56 NSCLC: 0.71 , Multi: 0.62 0.59 NSCLC: 0.66, Multi: 0.69 0.57 NSCLC: 0.67, Multi: 0.80 0.61 Renal cell carcinoma n = 304 RCC: 0.79, Multi: 0.85 0.62 RCC: 0.80 , Multi: 0.74 0.56 RCC: 0.78 , Multi: 0.77 N/A RCC: 0.82 , Multi: 0.68 N/A