Machine Learning for Predicting Human Drug-Induced Cardiotoxicity: A Scoping Review

Abstract

Background: Drug-induced cardiotoxicity poses a major challenge in drug development and clinical safety. Although machine learning (ML) methods have shown potential in predicting cardiotoxic risks, prior research has largely focused on specific mechanisms such as human Ether-& agrave;-go-go-Related Gene (hERG) inhibition. This scoping review systematically examined studies applying ML models to predict a broad range of drug-induced cardiotoxicity outcomes. Methods: A systematic search of PubMed, EMBASE, SCOPUS, and Web of Science identified studies developing ML models for cardiotoxicity prediction. Extracted data included sources, feature types, algorithms, and performance metrics, categorized by evaluation method (training, testing, cross-validation, or external validation). Results: Twenty-five studies met inclusion criteria, addressing outcomes such as arrhythmia, cardiac failure, heart block, hypertension, and myocardial infarction. Structured resources such as SIDER (Side Effect Resource) were the most common data sources, with features including molecular descriptors, fingerprints, and occasionally, target-based or transcriptomic data. Support vector machines (SVM) and random forest (RF) were the most common algorithms, showing robust predictive performance, with externally validated area under the receiver operating characteristic curve (AUC-ROC) values above 0.70 and accuracy exceeding 0.75 in several studies. Despite variability and limited external validation, ML approaches demonstrate substantial promise for predicting diverse cardiotoxic outcomes. Conclusions: This review underscores the importance of integrating heterogeneous data and rigorous validation for improving cardiotoxicity prediction.