Summary: | 碩士 === 國立成功大學 === 醫學資訊研究所 === 107 === Wolff-Parkinson-White syndrome (WPWS) is a common heart disease in which an accessory pathway between the atrium and the ventricle is prone to tachycardia. WPWS diagnosis is conducted using an electrocardiogram (ECG). If WPWS is diagnosed, then the delta wave is detected using various leads to determine the position of the accessory pathway to facilitate further surgical treatment. The detection of delta waves is thus important. Existing algorithms for automatically diagnosing WPWS need to determine the delta wave pattern (positive, negative, or isoelectric), the duration of the PR interval, and the QRS complex in the ECG. However, such algorithms are difficult to implement and time-consuming, and the accuracy of their results is unsatisfactory.
This paper proposes an automatic delta wave detection algorithm based on the convolutional neural network (CNN) model. ECG data are first classified as WPWS or non-WPWS. Then, the onset position of the delta wave is detected. Finally, the delta wave is visualized in post-processing. We used the Taipei Veterans General Hospital ECG data set as an experimental data set to explore issues related to the automatic delta wave detection algorithm for one-dimensional (1D) and two-dimensional (2D) ECG data and balanced and imbalanced data. The experimental results show that the accuracy of the proposed 2D CNN can reach 99.52%, and that the average error between the delta wave onset position obtained using the proposed 1D CNN and the ground truth is 4.13 ms.
|