Prediction of Appropriate Implantable Cardioverter-Defibrillator Therapy Using Machine Learning and Routinely Available Clinical Data

Abstract Background and Aims Risk stratification for appropriate implantable cardioverter-defibrillator (ICD) therapy remains imprecise when based on conventional clinical variables alone. We aimed to develop and geographically validate a machine-learning model that integrates routinely available clinical, electrocardiogram, and device interrogation/programming parameters, and quantify the incremental value of non-sustained ventricular tachycardia (NSVT). Methods We retrospectively analysed 514 ICD recipients implanted between 2020 and 2025 at two hospital sites (development cohort) and an independent cohort of 220 patients from a third site (external validation). The endpoint was appropriate ICD therapy (anti-tachycardia pacing or shock). Models were trained using nested stratified cross-validation in the development cohort, and the final refitted model was applied to the external cohort without recalibration. Results In the development cohort, 77/514 patients (15%) experienced appropriate ICD therapy (follow-up 404 days). A base model (Histogram-based gradient boosting) excluding NSVT achieved modest discrimination (ROC-AUC 0.624; average precision 0.198). Adding NSVT improved performance (ROC-AUC 0.805; average precision 0.432). Logistic regression with NSVT achieved comparable discrimination (ROC-AUC 0.78; average precision 0.43), indicating a largely NSVT-driven gain. External validation (event rate 51/220, 23%) confirmed good discrimination (ROC-AUC 0.815; average precision 0.583) with acceptable calibration (Brier score 0.137). At prespecified threshold 0.16 in external validation, sensitivity was 0.78 and specificity 0.74. Conclusion A machine-learning model integrating routinely available clinical, electrocardiogram, and ICD programming/interrogation data may enable prediction of appropriate ICD therapy with preserved performance on geographic external validation. NSVT was the dominant contributor to predictive performance across modelling approaches.