Artificial intelligence approaches for early prediction and risk stratification of immune checkpoint inhibitor-related myocarditis: a systematic review

Background Immune checkpoint inhibitor (ICI)-related myocarditis is a rare but potentially fatal complication of cancer immunotherapy, with a mortality rate of up to 50%. Early identification of high risk individuals is critical to improving clinical outcomes, yet it remains challenging using conventional clinical approaches. Recent advances have explored artificial intelligence and machine learning models to enhance prediction through integration of clinical, laboratory, and imaging data. Purpose This systematic review aims to enable early risk stratification and identification of immune checkpoint inhibitor-related myocarditis in order to facilitate earlier intervention and improve clinical outcomes. Furthermore, the purpose is to evaluate the role of artificial intelligence-based models within modern cardiovascular and oncological practice. Methods MEDLINE and EMBASE were systematically searched without date restrictions in accordance with PRISMA guidelines. Primary studies evaluating artificial intelligence or machine learning models for predicting immune checkpoint inhibitor-related myocarditis were included. Two independent reviewers screened 119 records, of which 22 underwent full-text review. Six studies met inclusion criteria and were included in quantitative synthesis. Extracted data included model architecture, input features, validation strategies, and performance metrics, including area under the curve, accuracy, sensitivity, and specificity. Results Across six studies (n=22,700 ICI-treated patients), pooled model performance demonstrated strong discriminatory ability for myocarditis prediction, with a summary area under the curve of 0.86 (95% confidence interval 0.79-0.91). Multimodal machine learning frameworks integrating clinical, laboratory, and imaging data achieved the highest predictive performance (area under the curve 0.88, 95% confidence interval 0.83-0.92), followed by text-based natural language processing models (area under the curve 0.84) and clinical-only models (area under the curve 0.81, 95% confidence interval 0.77-0.86). Model interpretability techniques were applied in 75% of studies, consistently identifying cardiac biomarkers and echocardiographic strain parameters as key predictors. External validation was performed in only two studies. Conclusions Artificial intelligence-based models demonstrate strong performance for predicting immune checkpoint inhibitor-related myocarditis, particularly when employing multimodal data integration. These approaches have potential to support earlier detection and risk stratification, allowing for timely treatment and improved outcomes. However, heterogeneity in study design and limited external validation constrain generalisability. Prospective, multi-centre validation is required to support safe implementation in routine clinical care.