Artificial intelligence and machine learning for predicting immune checkpoint inhibitor–associated cardiotoxicity: A systematic review and narrative synthesis.

Abstract

e24028 Background: Immune checkpoint inhibitor (ICI) associated cardiotoxicity, including myocarditis and major adverse cardiovascular events (MACE), is uncommon but associated with high mortality. Scalable risk prediction models may enable targeted surveillance and early intervention. Methods: We searched PubMed/MEDLINE, Embase, Web of Science, and Google Scholar from inception through January 2026 for human studies validating artificial intelligence and machine learning (AI/ML) models to predict ICI related cardiac events. We extracted data on demographics, type of cancer, ICI agents, follow-up duration, outcome definitions, model family, and discrimination metrics. Meta-analysis could not be performed due to heterogeneity in endpoints, feature sets, and validation strategies. Results: Six studies (n = 56,721) comprised 5 retrospective cohorts (2 single-center: n = 419 and 4,960; 2 multicenter: n = 615 and 2,258; 1 nationwide claim: n = 48,446) and 1 case series (n = 23). Age (reported in 4/6) ranged from 64.4 years to 72 years; sex (3/6) ranged 24.7–46.4% female and 53.6–75.3% male. Lung cancer was most common (4 studies; 30.6–82.7%, including 59.4% in claims), followed by melanoma/skin (3 studies; 2.0–14.0%) and renal/genitourinary cancers (3 studies; 0.16–22.0%); GI tumors were also frequent (3 studies; 3.6–47.4%). Pembrolizumab was administered in 26.3–48.0%, nivolumab in 30.0–65.9%, atezolizumab in 5.7–17.0% and durvalumab in 2.1–6.9% as reported in 3 studies. Follow-up spanned from 30 days to 299 days. Across cohorts, cardiotoxicity incidence varied markedly by endpoint definition and sampling frame: composite cardiac events occurred in 8.4% of single center cohort (418/4,960), immune-related cardiac adverse events in 23.9% of a multicenter cohort (147/615), and composite of myocarditis/MACE in 11.7% (264/2,258) with myocarditis alone 3.0% (59/2,258). In contrast, nationwide claims-defined myocarditis was rare (0.24%; 117/48,446), while a case series reported myocarditis in 65.2% (15/23), including 4 fatalities. Models included gradient boosting (3 studies), random forests (2), soft-voting two-stage ensemble (1), multimodal fusion deep learning (1), and neural networks (1). AUROC ranged from 0.635–0.851 with lowest in the LightGBM model and highest in the two-stage ensemble. Conclusions: Current AI/ML models demonstrate moderate-to-high discriminatory performance for predicting ICI-associated cardiotoxicity. However, evidence remains limited by retrospective study designs, heterogeneous endpoints, insufficient external validation and calibration reporting. Larger prospective studies incorporating standardized outcome definitions and clinically actionable thresholds are required for real world adoption.