Machine learning (ML)-based prediction of fatal immune checkpoint inhibitor (ICI)–associated myocarditis (MC) with multi-cohort validation.

12019 Background: ICI-associated MC is a rare but highly fatal immune-related adverse event (irAE). Identifying clinical predictors of MC-specific fatality remains an unmet need. Herein, we develop and externally validate an ML model to predict ICI-MC fatality using clinical and laboratory features. Methods: The WHO Vigibase database was queried through December 22, 2024, to identify cases of ICI-MC with complete clinical data, including age, sex, cancer and ICI type, co-occurring irAEs, overlap status with myositis (MS) and/or myasthenia gravis (MG), MC timing from ICI start, and MC-specific fatality. Multivariable logistic regression was used to assess correlates of MC-specific fatality. An XGBoost ML model using all collected clinical data was trained to predict MC-specific fatality using an 80/20 split of data for training and testing. We augmented our clinical ML model with troponin (T or I) values at the time of ICI-MC diagnosis from two institutions (n = 37: OU n = 13, MGH n = 24), normalized to institutional upper limits of normal. Two additional external cohorts (n = 68: VUMC n = 18, MDACC n = 50) were used to validate our final clinical and augmented models. Results: We identified 822 cases from VigiBase with ICI-MC, with a median age of 69 years (IQR: 60-76), 59. 1% (n = 486) were males. Pts with lung cancer had the highest proportion of cases (23. 7%, n = 195). Most patients received ICI monotherapy (59%, n = 485). Median time to MC onset was 30. 4 days (IQR: 21-91. 3). MC occurred alone in 588 cases (71. 5%) and with MS and/or MG in the remainder. MC-specific fatality occurred in 147 cases (17. 9%). After multivariable adjustments, MC onset within the first month of ICI start was associated with higher MC-fatality odds vs. later onset (vs 1-3 months: ROR = 0. 38 0. 21-0. 67, vs 3-12 months: ROR = 0. 5 0. 25-0. 95). Co-occurring non-MC major cardiac adverse events (ROR: 0. 27 0. 17 - 0. 44) and pneumobronchitis (ROR: 0. 29 0. 11 - 0. 78) were also associated with increased MC-specific fatality. The optimized MC-fatality clinical ML model achieved AUCs of 0. 76 (training) and 0. 73 (testing), with early MC onset as the most contributing feature. External validation (n = 37) of our clinical model yielded an AUC of 0. 74. Augmenting our clinical ML model with baseline troponin (n = 37) significantly improved performance (AUC: 0. 79, likelihood-ratio p = 0. 046), with a further external validation (n = 68) AUC of our augmented ML model of 0. 67. Our ML model is available online for use through: https: //ohbiolab. shinyapps. io/ici-mcfatalityₚredictor/. Conclusions: ICI-MC timing and co-occurring cardio-pulmonary reactions are key determinants of MC-specific fatality. Our findings highlight the potential promise for early risk stratification to identify high-risk pts with ICI-MC who could benefit from closer monitoring or more tailored immunosuppressive interventions at ICI-MC diagnosis.