ECG 데이터 분석을 위한 DNN 모델에서 은닉층의 수가 신경망 판정의 정확도에 미치는 영향

초록

Electrocardiograms (ECGs) reflect various cardiac conditions, diseases, and individual characteristics. However, the development of automated ECG signal analysis programs remains challenging due to the presence of diverse types of noise. Consequently, there has been a growing interest in applying deep neural network (DNN) tech- nologies to automate ECG analysis. Existing automated ECG analysis techniques primarily focus on detecting the peak points of P, QRS, and T waves. However, in practical ECG analysis, distinguishing the start and end points of these waveforms (P, QRS, T) is more critical. Such tasks often require additional neural networks, underscoring the need to develop techniques that can reduce the cost and time associated with network development. This study evaluates the ability of neural networks to identify the boundaries of P, QRS, and T waves using the Rec- ognition Score as a performance metric. Six neural networks, each with two hidden layers (h2), were developed and trained over 2000 Epochs to distinguish waveform regions. However, the impact of increas- ing the number of hidden layers or adjusting the number of training Epochs on recognition performance has not been clearly identified. To address this, the number of hidden layers was increased to four (h4) and eight (h8), and training Epochs were set to 400 and 600 to compare recognition accuracy. For P and T waves, the training data demonstrated improved performance at 600 Epochs compared to 400 Epochs. However, the test data revealed a decrease in recognition accuracy by 13.3% and 1.1%, respectively. Addi- tionally, networks with two hidden layers outperformed those with four and eight hidden layers. Spe- cifically, the recognition performance for P, QRS, and T wave regions was 2.6, 1.2, and 1.4 times higher, respectively, in networks with two hidden layers compared to those with eight hidden layers. The results indicate that neural networks designed to distinguish the regions of P, QRS, and T waves can be effec- tively trained with a simple structure and fewer training Epochs. This finding suggests that even with the use of multiple neural networks, the development cost and time required for ECG analysis do not increase significantly.

키워드