KEST: Knowledge-Enhanced Spatio-Temporal Network for Cardiopulmonary Exercise Testing

Abstract:
The cardiopulmonary exercise test (CPET) is a clinical examination method used to assess the cardiopulmonary function of the human body during exercise with increasing load. However, research specifically targeting machine learning for CPET data remains notably constrained, resulting in a limited exploration of the extensive spatio-temporal and medical characteristics embedded within CPET. To address this issue, we propose a two-view fusion network named KEST-Net that incorporates spatio-temporal information and medical prior knowledge to enhance the disease classification for CPET data. We design an adjacency matrix integrating the connectivity between different modalities and apply graph structure to capture spatial features. On this basis, time-based characteristics are extracted by temporal transformer. Then, systemlevel tokens are introduced using channel-wise attention, followed by a global feature extraction with a cardiopulmonary system transformer, where the correspondence between the physiological signals and the human system is considered. Finally, features from two views are fused. Experimental results on public datasets show that the performance of our model is better than other models, and adding prior medical knowledge is an effective way to improve the model performance in many aspects.