Research on a System Combining Improved Wavelet Thresholding with LightGBM for Intelligent Pulse Recognition

This is the authors' abstract. We don't add key points for this paper.

In response to the issues of subjective nature in traditional Chinese medicine’s pulse diagnosis, the lack of objective criteria, and the low level of intelligence and insufficient recognition accuracy of existing pulse diagnosis equipment, this paper proposes a method for intelligent pulse recognition that integrates improved wavelet threshold denoising with LightGBM. Firstly, an improved exponential wavelet threshold function is designed to address the shortcomings of traditional hard threshold functions, which are discontinuous, and soft threshold functions, which have constant biases. This function effectively removes baseline drift and high-frequency noise while preserving the key details of the pulse signal to the greatest extent. Secondly, multi-domain feature vectors are extracted from three dimensions: time domain, frequency domain, and time-frequency domain, to comprehensively characterize the morphological and dynamic characteristics of the pulse. Finally, a pulse classification model based on LightGBM is constructed, with bagging integration strategies enhancing the model’s generalizable discriminative ability. Experimental results show that the accuracy of pulse recognition using this method reaches 91.6%, with an average recall rate of 89.4%. Compared to traditional SVM and random forest methods, there is a significant improvement. The improved threshold function enhances the signal-to-noise ratio by approximately 4.7 dB, and the accuracy of distinguishing between "slippery pulse" and "sticky pulse" increases by 12.6% after incorporating multi-domain features. Based on this, a pulse intelligent recognition software system with a B/S architecture has been developed, validating the engineering feasibility of the proposed method.