Pressure wrist pulse signal analysis by sparse decomposition using improved Gabor function
Zhixing Jiang,Chaoxun Guo,David Zhang
DOI: https://doi.org/10.1016/j.cmpb.2022.106766
Abstract:Background and objective: In traditional Chinese medicine and Ayurvedic medicine, wrist pulse wave fluctuations are an important indicator for distinguishing different health states. Owing to the development of modern sensing technology, computational methods have been used in the analysis of pulse wave signals. The description and quantification of the peaks in the pulse wave is significant for the identification of health status. Methods: In this study, we decomposed the pressure pulse waveform of the radial artery into several components by sparse decomposition with an improved Gabor function. To better represent the position, shape, and relationship of the peaks, we designed an improved Gabor function structure based on the characteristics of the pulse waveform to generate a time-frequency dictionary. Compared with conventional representation methods, the shape of the Gabor function is more variable. In addition, owing to the limitation of windowing, the Gabor function can reduce the influence on other positions when it represents a specific position. Feature vectors consisting of decomposed components can be used for computerized pulse signal analysis and disease diagnosis. Results: In the binary classification of healthy and diseased pulse signals, the proposed method achieved the best results for health/diabetes, health/cardiac disease, health/hypertension, and health/nephropathy with accuracies of 93.54%, 73.42%, 88.42%, and 82.28%, respectively. The multi-classification performance of the different types of features was evaluated by six classifiers, and the proposed method obtained the highest classification performance with support vector machine-radial basis function for both balanced and imbalanced data. Conclusions: The results indicated that the proposed method enabled to obtain a smaller representation error and exhibited superior performance in distinguishing between the signals collected from patients and healthy individuals. Moreover, for the multi-classification of the pulse signals, the proposed method performed better than the state-of-the-art methods.