LSTM과 Xception의 다차원 특징 융합 기반 심전도를 이용한 스트레스 분류에 관한 연구

Abstract

With the advancement of scientific technology and the rapid growth of the IT industry, human exposure to stress in the dynamic information society is on the rise. Prolonged exposure to stress poses a significant risk factor for various health issues such as cardiovascular diseases, compromised immune systems, and mood disorders. Therefore, research on stress classification is crucial in mitigating the adverse health outcomes associated with stress. Current research on stress classification employs various approaches, ranging from self-report questionnaires to the utilization of biosignals. While self-report methods rely on subjective personal experiences and may suffer from bias and inaccuracy, utilizing biosignals allows for accurate and objective analysis through quantitative data assessment. Commonly employed biosignals for stress classification include electrocardiograms(ECGs), electrodermal activity(EDA), blood volume pressure(BVP), pupil dilation(PD), and cortisol hormone levels. Among these, studies utilizing electrocardiograms, which offer insights into the body's autonomic nervous system response through heart rate variability analysis, have gained significant attention. In existing research focusing on stress classification using electrocardiograms, digital filters are employed to mitigate signal noise. Additionally, stress is classified by analyzing feature information in a single-dimensional format. However, using digital filters to eliminate morphologically damaged electrocardiogram signals presents challenges. In the case of using only single-dimensional feature information, detailed and comprehensive analysis is not possible because there is a feature information area that cannot be analyzed in the corresponding dimension. With these issues being addressed, this paper presents a novel stress classification method that addresses the limitations of existing techniques by leveraging the fusion of multi-dimensional features, employing Long Short-Term Memory(LSTM) and Xception networks, and effectively removing outlier signals from ECGs. We apply a two-fold approach to overcome the challenges associated with conventional stress classification methods utilizing ECGs. Firstly, outlier signals caused by motion artifacts are eliminated by detecting the R wave peak in the ECG and comparing the similarity between RR intervals. This innovative outlier removal technique surpasses the limitations of conventional noise removal methods relying on digital filtering. Secondly, we introduce multi-dimensional feature fusion to overcome the drawbacks of single-dimensional feature information. While one-dimensional features suffer from capturing frequency variations over time, two-dimensional features need help in analyzing overall signal patterns, averages, and trends. To address these issues, we extract nine heart rate variability(HRV) features from the ECG signal after outlier signal removal and input them as a one-dimensional feature vector into an LSTM network. Additionally, we utilize Short-Time Fourier Transform(STFT) to analyze the time-frequency information of the ECG signals and input the resulting spectrograms as two-dimensional images into an Xception network. Feature-level fusion is then applied according to feature information learned and analyzed by both networks. The proposed multi-dimensional feature fusion method effectively leverages the advantages of each dimension, enabling precise analysis and exhibiting high reliability in stress classification performance. To evaluate the performance of the proposed stress classification method, we conducted experiments to assess the impact of window length variations for signal segmentation(5 seconds, 10 seconds, and 60 seconds), the effectiveness of outlier signal removal, and the influence of feature-level fusion. Ultimately, by removing outlier signals and utilizing a window length of 5 seconds for signal segmentation, along with multi-dimensional feature fusion employing a weighted average method, we achieved a stress classification performance of 99.51%, representing a significant improvement of over 1.25% compared to previous studies that relied solely on single-dimensional feature information from electrocardiograms. The results present the efficacy of the proposed stress classification method, demonstrating its potential to address existing limitations and achieve high-performance stress classification.