Aim: Ventricular arrhythmias (VA) remain a major public health concern and are frequently managed in high-risk patients using implantable cardioverter-defibrillators (ICDs). Beyond their therapeutic function, ICDs continuously capture physiological data that provide early warning of impending arrhythmic events.This study aimed to assess the feasibility of predicting VA occurrence from routine collectedly ICD-derived data.
Method: We retrospectively analyzed two patient groups: patients who experienced at least one VA episode and a control group without documented VAs. Eleven daily physiological parameters, including mean heart rate and shock impedance, were extracted, and deviations from a reference follow-up were quantified using linear mixed-effects models, yielding 44 candidate predictive features. The 10 most informative variables were identified through clinical review and Morris sensitivity analysis.
Result: A gradient-boosting machine-learning model was then trained, with performance evaluated via 10-fold cross-validation. The model achieved a correct classification rate of 77% and an area under the receiver operating characteristic curve of 0.76.
Conclusion: This study provides proof of concept that routinely collected ICD data can enable short-term VA prediction, opening a potential avenue for timely, preventive clinical interventions.