Work

Calibration Robustness and Event-Level Performance of Patient-Specific ECG Anomaly Detection: An Evaluation Using the MIT-BIH Arrhythmia Database

Year: 2026
Type: preprint
Abstract: Patient-specific electrocardiogram (ECG) anomaly detection offers an individualized approach to cardiac monitoring, but its reliability depends on baseline calibration, feature representation, and temporal decision-making. This study evaluates a multi-feature patient-specific ECG anomaly detection framework implemented in MATLAB/Simulink using the MIT-BIH Arrhythmia Database. The framework integrates a four-band filter bank, six-dimensional feature extraction, baseline comparison, and a temporal safety-state mechanism. An offline evaluation pipeline was verified against the original Simulink implementation, demonstrating numerical equivalence on three test records. Experiments were conducted across 21 test records containing 1,416 target ectopic events to investigate calibration strategies, calibration duration, feature contributions, and event-level detection performance. Median–median absolute deviation (MAD) calibration achieved 1.44% macro event sensitivity with 5.26 false alarms per hour, while mean–standard deviation calibration achieved 2.50% sensitivity with 13.37 false alarms per hour. Feature ablation showed that excluding sample entropy improved macro event sensitivity in the evaluated configuration. Calibration duration also influenced performance, although longer calibration did not consistently improve detection. Overall, the findings demonstrate the importance of calibration quality and event-level evaluation while highlighting the limited detection capability of the current framework and the need for further improvement before clinical application.
Institution: Varendra University
Language: English
Field: Medicine
Open Access status: green
OpenAlex ID: W7237355361
Cites:
Cited by:
Related to: