Decoding fatal toxic effects in checkpoint inhibitor therapy using real‐world pharmacovigilance data and machine learning

Immune checkpoint inhibitors (ICIs) improve cancer outcomes but are also associated with immune-related adverse events (irAEs), which pose significant challenges for clinical management. An observational pharmacovigilance analysis on FDA Adverse Event Reporting System was performed to identify ICI-related adverse event (AE) signals. Fatality kinetics simulation and multivariate logistic regression were used to investigate patterns of fatal AEs and multisignal involvement. A machine learning framework, SAFE-ICI, was developed to predict short-term risk and outcomes of fatal irAEs occurring within the first 90 days of ICI therapy. The analysis identified 358 significant AE signals associated with ICI therapies across 18 organ systems. PD-1/PD-L1 therapies were associated with 54 fatal irAEs, including 23 in non-small cell lung cancer (NSCLC), 5 in melanoma, 6 in renal cell carcinoma (RCC) and 20 in other cancers. Combination therapies were associated with 20 fatal irAEs, including 3 in NSCLC, 6 in melanoma, 7 in RCC and 4 in other cancers, with stable involvement of multiple AE signals. The SAFE-ICI model demonstrated robust performance in predicting fatal irAE risk, successfully stratifying patients into low- and high-risk phenotypes with significantly different survival benefits, in both the discovery and holdout validation cohorts. Our findings highlight the potential of machine learning to improve pharmacovigilance systems and aid clinicians in enhancing patient outcomes during ICI therapy.