Evolutionary Synthesis of QRS Detection Algorithms for Wearable ECG-Based Heartbeat Classification

Existing evolutionary approaches to QRS detection are confined to the parametric tuning of rigid, human-designed pipelines or black-box neural architecture searches. This work applies evolutionary methods for the synthesis of QRS detection algorithms, evolving both the operational pipeline architecture and its underlying parameters simultaneously. A three-class heartbeat classifier (Normal, Ventricular, Other classes) based on the parallel operation of seven QRS detectors was designed. The detectors comprise algorithms from the literature and evolutionary synthesized QRS detectors. The differences in the R-peak detection times served as the feature signal for a decision tree classifier, which performed the final heartbeat classification. The overall accuracy of the proposed classification method for the test dataset was 95.82%, sensitivity was 93.05%, specificity was 97.50%, and the F1-score was 91.74%. The results achieved confirm that synthesized detectors provide valuable complementary timing information on relative R-peak detection times in the proposed heartbeat classifier with seven parallel QRS detectors. Careful detector selection can provide an effective compromise between classification quality and system complexity. The proposed method is applicable for real-time mobile HRV analysis and ischemia detection.