A Fast and Memory-Efficient CNN Accelerator for ECG Classification in Remote Healthcare Systems

Electrocardiogram (ECG) classification is crucial for addressing cardiovascular challenges in remote healthcare systems. Recent advances in artificial intelligence, particularly Convolutional Neural Networks (CNNs), have significantly improved the accuracy of ECG disease classification. However, existing CNN-based accelerators face challenges such as high area usage due to large parameter counts and slow processing speeds. This paper introduces the Tiny InceptionNet Accelerator (TINA), a fast and memory-efficient CNN accelerator for ECG classification, which reduces the parameter count by 43.42% compared to the smallest existing 1-D CNN. TINA incorporates three novel features: a dual Processing Element Array (D-PEA) for parallel processing of 80 multiply-accumulation (MAC) operations in the convolutional layer and other necessary operations in subsequent layers, a shared pixel distributor (SPD) for dynamic data coordination, and shared pixel memory between the PEAs to optimize area usage. TINA has been successfully implemented and verified on the ZCU102 FPGA. FPGA experiments show that TINA achieves improvements of at least 2.4 times in inference time and 3.06 times in area-delay product (ADP) compared to existing 1-D CNN accelerators.