A Hybrid Network Integrating 1DCNN, GRU, and ISENet for Pulse Signal Pattern Analysis.

objectivePulse diagnosis holds significant importance in traditional Chinese medicine (TCM). The accurate identification of pregnancy pulse patterns is crucial in TCM diagnosis, as it provides valuable insights into a patient's health status and enables the development of personalized treatment strategies. However, the complexity and variability of pulse characteristics across different stages of pregnancy pose a significant challenge for TCM practitioners.

methodsTo address the issue, a novel one-dimensional convolutional neural network (1DCNN) integrated with a gated recurrent unit (GRU) and an improved squeeze-and-excitation network (ISENet), named 1DCNN-GRU-ISENet, is proposed to identify the pregnancy pulse of three stages. Specifically, the 1DCNN is employed to construct parallel convolutional branches and series-stacked blocks. This architecture not only functions as a residual network but also enhances the network's depth, enabling effective extraction of pulse features. By incorporating this structure, the model is capable of capturing more intricate and subtle characteristics of the pulse within different frequency ranges, leading to improved feature representation and enhanced performance. Simultaneously, the integration of GRUs facilitates comprehensive learning of pulse temporal features. Moreover, the ISENet assigns varying weights to different feature components based on pulse trends, effectively focusing on and highlighting crucial spatial features.

resultsThrough rigorous evaluation with a locally collected research dataset, our network achieves accuracy rates of 90.8%, 91.6%, and 91.5% in identifying pregnancy pulse patterns for the three stages, respectively. Additionally, our model demonstrates good performance across various metrics, including average precision, sensitivity, F1 score, and AUC, with scores of 91.7%, 91.5%, 0.916, and 97.1%, respectively.

conclusionThese results underscore the effectiveness of our proposed network in accurately classifying pulse patterns and showcase its potential for practical application in assisting with auxiliary diagnosis in the context of TCM pulse evaluation.