Aim: With the integration of implantable cardiac monitors (ICMs) such as Medtronic's LINQ IITM in clinical practice, continuous monitoring has significantly improved arrhythmias detection. However, current devices only detect atrial fibrillation (AF) episodes lasting ≥ 2 minutes, potentially missing short but clinically relevant arrhythmias such as brief AF or non-sustained ventricular tachycardia (NSVT), particularly in hypertrophic cardiomyopathy patients. This study aimed to develop a machine learning algorithm capable of detecting short-duration arrhythmias while maintaining a balance between sensitivity, specificity, and computational efficiency.
Methods: A Random Forest classifier was trained using rhythm-labeled ECG data from two PhysioNet databases, Long Term AF and VTaC datasets, comprising 44 patients with paroxysmal AF and sinus rhythm (SR), and 777 with NSVT. The model was tested on MIT-BIH Ar-rhythmia Database (43 SR, 8 AF, and 13 NSVT). Different window lengths were evaluated (2-minute, 1-minute, 30-second, and 10-second) to assess the impact of segment duration on detection performance. Hand-crafted extracted features included RR-interval variability, QRS morphology, wavelet descriptors, high-order statistics, and Hermite coef-ficients. Feature selection combined correlation filtering and Least Abso-lute Shrinkage and Selection Operator regularization, resulting in a final set of 14 selected features.
Results: Using 2-minute segments, AF detection achieved an F1-score of 0.70 and a specificity of 0.95. In contrast, NSVT detection resulted in a lower F1-score of 0.44 and specificity of 0.70, primarily due to the pre-dominance of sinus rhythm within NSVT-labeled segments. Reducing the segment length significantly improved NSVT detection, with specificity rising to 0.82 using 10-second windows. Conversely, AF detection slightly declined in specificity (0.87).
Discussion: These results highlight contrasting trends: AF is better identified in longer segments, while NSVT benefits from shorter observation windows. The top-ranked features identified through MI are computationally efficient, supporting their integration into ICMs.