Mitigating False Positives in ECG Segmentation Through Deep Learning-Based Postprocessing

The characteristics of P, QRS, and T ECG waves, as well as the intervals and segments between them, are crucial for the diagnosis of cardiac diseases. However, automatic ECG signal segmentation methods suffer from the False Positives (FP) issue caused by either those isolated artifacts that are mistakenly detected as events of interest or artifacts that hide target events of interest. In this paper, a shallow Deep Learning (DL)-based postprocessing stage is proposed to mitigate FP in ECG segmentation. The effectiveness of the proposed postprocessing stage is evaluated by applying it to a recent supervised approach, namely ConvLSTM, and two well-known unsupervised methods, namely ECGdeli and ECGKit, using the popular QT and LU databases of PhysioNet. The results obtained confirm the effectiveness of the proposed postprocessing in reducing the False Detection Rate (FDR). In addition, for a comprehensive analysis of segmentation quality, an additional evaluation criterion addressing the number of false detected waves is also suggested. The current study demonstrates the relevance of combining ECG segmentation approaches, especially supervised ones, with the proposed shallow DL-based postprocessing stage, in terms of both the accuracy and the number of detected waves.