Learning Pulse Image with Deep Dynamic Frequency Network for Cardiovascular Diseases Diagnosis

Cardiovascular diseases (CVD) constitute a major component of the global public health burden. Traditional Chinese medicine practices prove a significant connection between CVD and pulse condition, and pulse diagnosis can be used as one of the effective indicators for diagnosing CVD. However, researchers in this field still face two major challenges: 1) Traditional pulse diagnosis in Chinese medicine is highly subjective. It lacks quantitative and qualitative judgments, which hinders its reliability and consistency. 2) There is a complex nonlinear relationship between the pulse signal and the diagnosis of CVD, making it difficult to model for prediction. To address the first challenge, this paper improves the original equipment and utilizes array sensors to establish two array pulse signal datasets and benchmarks for identifying CVD, including hypertension and coronary heart disease. To tackle the second challenge, first, we introduce the concept of “pulse image” to better model the relationship between pulse information and CVD. Second, we propose a deep dynamic frequency network (DDFNet) to model the frequency distribution of pulse signals for distinguishing CVD. Specifically, we introduce a dynamic frequency component module (DFC) to dynamically adjust different frequency components and extract distinctive features. To leverage the relationship between spatial and frequency features, we propose a cross-diagram-frequency fusion module (CDFF) to extract cross-domain features and improve the model’s recognition performance. Extensive experiments show that our method outperforms existing recognition methods on the proposed two datasets, where our approach achieves an accuracy rate of 86.8% for hypertension and 88.1% for coronary heart disease. The related code of our method is available at https://github.com/aisane-454/DDFNet