Cardiovascular disease is the leading cause of death in humans, and heart failure is the primary cause of death among cardiovascular diseases, significantly impacting patients’ quality of life and life expectancy. This paper proposes corresponding heart sound signal analysis methods based on the physiological structure of the cardiovascular circulatory system, combining three aspects: cardiac reserve indicators, energy characteristics, and complexity characteristics. Subsequently, experimental studies were conducted to investigate the differences in short-term heart sound features between chronic heart failure patients and healthy individuals, as well as the relationship between short-term heart sound features and the staging of chronic heart failure. Finally, the MSCNNMGU heart failure prediction model was established by combining MSCNN with MGU. The results of this study indicate that as the severity of heart failure increases, the D/S and S1/S2 ratios in the time-domain features of heart sounds show a decreasing trend. In performance comparison experiments among different neural network models, the performance of EfficientNet-B2 was as follows: Acc=94.13, Pre=92.38, Rec=81.93, F1=84.5, AUC=0.945, and the inference speed with BatchSize set to 128 was 917ms. Thus, the model achieves high balanced performance while ensuring fast inference speed.