Assessment of machine learning approaches for detecting brief arrhythmias: Short-duration AF and NSVT

Background and objectives: Non-sustained ventricular tachycardia (NSVT) is a well-recognized risk marker in hypertrophic cardiomyopathy, and the coexistence of atrial fibrillation (AF) further increases the risk of sudden cardiac death. However, current implantable cardiac monitors (ICMs) only store AF episodes longer than two minutes and are unable to detect NSVT, thereby missing relevant prognostic information. This study aimed to design a computationally efficient pipeline for detecting short AF and NSVT episodes using lightweight features, with long-term goal of enabling on-device integration.Methods: Six classifiers were trained and evaluated for (i) AF detection, (ii) NSVT detection, and (iii) AF and NSVT multi-class detection. Hand-crafted features from multiple domains were extracted from three PhysioNet databases. Long-Term AF and Ventricular Tachycardia alarms from ICUs (VTaC) databases were used for training, while MIT-BIH Arrhythmia database was reserved for independent testing. Segments of 2-minutes, 1-minute, 30-second, and 10-second were analyzed to assess the effect of recording length on model performance.Results: Models were trained with inter-patient cross-validation. Random forest achieved the best performance for AF detection, AdaBoost for NSVT detection, and both performed well in multiclass settings. For AF, the optimal window was 30-second (performance metrics above 0.90). For NSVT, the best results were obtained with 10-second windows (performance metrics above 0.83). In multiclass tasks, 30-second window provided the best trade-off between accuracy and computational cost.Conclusions: Shorter windows improve NSVT detection, whereas AF detection benefits from longer segments. The proposed lightweight models are computationally efficient and represent a step toward integration into battery-constrained ICMs.