<p>Wearable devices are essential for the long-term monitoring of cardiovascular diseases (CVDs). However, the diagnosis of CVDs typically relies on large-scale neural networks, which, due to their significant hardware overhead and high energy consumption, pose challenges for deployment on wearable devices. This work proposes a low-power solution for wearable heart sound diagnosis using a lightweight convolutional neural network (CNN) design method, based on algorithm-hardware co-design. Quantization-aware training (QAT), along with weight/activation quantization and operator fusion, simplifies computations with minimal accuracy loss. A hybrid dataflow and dynamic multiplier-accumulator (HDF-DMAC) architecture is presented, combining weight stationary (WS) and output stationary (OS) dataflows with dynamic MAC reuse. This architecture reduces cache size, minimizes memory accesses, and enhances logic utilization. Field-programmable gate array (FPGA) verification shows that the proposed method decreases the model’s computational workload, reducing floating-point operations (FLOPs) from 0.594 M to 0.147 M and compressing weight storage from 80.01 KB to 20.18 KB, while maintaining a diagnostic accuracy of 93.74%±0.56%. The system operates at a power consumption of 4.362 mW, with a clock frequency of 8.192 MHz, demonstrating its potential for application-specific integrated circuit (ASIC)-based wearable heart sound diagnosis.</p>