Unreadable segment recognition of single-lead ECG signals based on hand-crafted features and lightGBM

Abstract:
Electrocardiogram (ECG) signals are often corrupted by severe noises and artifacts, especially for single-lead signals obtained from wearable or mobile devices. The interfered segments in signals are of low quality and unreadable for both doctors and automated ECG analysis methods. Therefore, it is necessary to identify and eliminate these unreadable segments in ECG signals before further clinical usage. In this paper, we propose an unreadable ECG segment recognition method based on various types of hand-crafted features and LightGBM. A total of 106 features were extracted as Set-A, within which 30 features, i.e., Set-B, were obtained after feature selection. The single-lead ECG segments for evaluation in this work was generated from PhysioNet/CinC Challenge 2017 database by dividing long signals into 5-second ECG segments. Then both the feature sets were trained and tested on the database with 54593 segments. An ac-curacy of 89.69±0.64% was obtained on Set-A and 89.46±0.60% on Set-B. The former showed a performance, and the latter was less time-consuming with an accuracy loss of 0.23%. Our results show that the proposed method could effectively identify unreadable ECG segments to enable efficient long-term single-lead ECG interpretation.