Predicting immune checkpoint inhibitor cardiotoxicity using machine learning: a systematic review of model performance and methodological quality.

backgroundImmune checkpoint inhibitors (ICIs) significantly improve cancer outcomes but can cause rare, potentially fatal cardiotoxicity, including myocarditis and major adverse cardiovascular events (MACE). Machine learning (ML) and artificial intelligence (AI) models have emerged as promising tools for early detection and risk prediction; however, their performance, methodological quality, and clinical applicability remain unclear.

objectiveTo systematically evaluate ML-based prediction models developed to identify cardiotoxicity, myocarditis, or cardiac adverse events in patients receiving ICIs, and assess their methodological robustness using PROBAST.

methodsA systematic search of PubMed (18-20 November 2025) was performed using a multi-stage strategy (broad → refined → high-specificity). Seven studies met eligibility criteria. Data extraction covered population, predictors, outcomes, model types, validation, and performance metrics. Risk of bias was assessed using PROBAST. Supplementary files include full search strategies, extraction sheets, PRISMA flow diagram, and PROBAST tables.

resultsSeven ML models were identified, including XGBoost, multimodal deep learning, neural networks, ECG-based AI models, pharmacovigilance-driven ML, and two clinical nomograms. Sample sizes ranged from 23 to 4,282 patients. Reported AUC values varied: XGBoost, 0.92; Fusion AI, 0.88; FAERS ML, 0.83; ECG AI, 0.91; Time-series NN, predictive ranking only; Nomogram, 0.83-0.967. Troponin, ECG features, neutrophil-lymphocyte ratio, NT-proBNP, and clinical factors such as steroid dose and radiotherapy emerged as consistent predictors. PROBAST identified high risk of bias in three models (mainly due to small datasets, inadequate validation, and unclear handling of missing data). Two studies using independent validation cohorts achieved low risk of bias.

conclusionML-based models show promising discriminative ability for predicting ICI-related cardiotoxicity and myocarditis; however, most studies suffer from methodological limitations that restrict real-world clinical adoption. Larger multicenter cohorts, standardized definitions, multimodal integration, and adherence to TRIPOD-AI reporting guidelines are needed to advance clinically deployable models.