403 Machine learning-based prediction of adverse events in cancer patients treated with PD-1/PD-L1 inhibitors using FAERS data

Background Immune checkpoint inhibitors targeting PD-1/PD-L1 have been approved over the last decade for a range of malignancies, including melanoma and non-small cell lung cancer. 1 While effective, these agents are frequently associated with adverse events (AEs), such as colitis and pneumonitis.2 Accurately identifying patients at risk for severe toxicity remains a major challenge in oncology. The FDA Adverse Event Reporting System (FAERS), a pharmacovigilance database with over 30 million reports, provides a valuable resource for large-scale drug safety analysis. Leveraging this data, we developed machine learning models to identify cancer patients at elevated risk for severe AEs related to PD-1/PD-L1 inhibitors.Methods Demographic, clinical, comorbidity, and AE data for cancer patients treated with PD-1/PD-L1 inhibitors between January 2011 and March 2025 were extracted from FAERS. After data cleaning and deduplication, the dataset was randomly divided into training (70%), validation (15%), and testing (15%) sets. Machine learning pipelines using the LightGBM and CatBoost algorithms were developed to predict severity. AEs were considered low severity for non-serious outcomes, and high severity for events involving hospitalization, death, or other serious outcomes without hospitalization. Model performance was evaluated on the held-out test set using area under the receiver operating characteristic curve (AUC), accuracy, precision, recall, and F1-score. Performance was compared to baseline models, including random forest and logistic regression.Results A total of 208,971 reports involving PD-1/PD-L1 inhibitors were included. The median age was 67.0 (IQR: 59.0-74.0) years; 72,267 (34.6%) patients were female and 128,203 (61.3%) experienced severe reactions. The five most reported drugs were nivolumab (92,753; 44.4%), pembrolizumab (62,096; 29.7%), atezolizumab (31,427; 15.0%), durvalumab (13,485; 6.5%), and tislelizumab (2,666; 1.3%). The most prevalent cancer types were lung (62,590; 30.0%), renal (16,123; 7.7%), liver (9,513; 4.6%), breast (8,416; 4.0%), and gastric (7,026; 3.4%). In predicting severity, CatBoost achieved an accuracy of 0.75, precision of 0.83, F1-score of 0.79, and AUC of 0.836 (95% CI = 0.831-0.840). LightGBM achieved an accuracy of 0.76, precision of 0.76, F1-score of 0.82, and AUC of 0.825 (95% CI = 0.821-0.830). Shapley Additive Explanations (SHAP) analysis identified polypharmacy, treatment duration, and patient weight as the most influential predictors of AE severity.Conclusions Machine learning models can effectively predict AE severity in cancer patients receiving PD-1/PD-L1 inhibitors. Future studies may focus on external validation, utilizing a non-binary severity scale, and implementing a more clinically relevant probability threshold. These models offer promising tools to improve drug safety monitoring and advance precision medicine within oncology.References Ai L, Chen J, Yan H, He Q, Luo P, Xu Z, Yang X. Research Status and Outlook of PD-1/PD-L1 Inhibitors for Cancer Therapy. Drug Des Devel Ther. 2020 Sep 8;14:3625-3649. doi: 10.2147/DDDT.S267433. PMID: 32982171; PMCID: PMC7490077.Yang F, Shay C, Abousaud M, Tang C, Li Y, Qin Z, Saba NF, Teng Y. Patterns of toxicity burden for FDA-approved immune checkpoint inhibitors in the United States. J Exp Clin Cancer Res. 2023 Jan 5;42(1):4. doi: 10.1186/s13046-022-02568-y. PMID: 36600271; PMCID: PMC9814433.Ethics Approval IRB approval is not required as the data are de-identified and publicly available.