Low-Power AI Chip for Atrial Fibrillation Detection

Atrial fibrillation (AFib) is a serious cardiac arrhythmia that can lead to embolism and heart failure. With widespread integration into wearable devices, photoplethysmography (PPG) provides a promising solution for AFib monitoring and diagnosis when combined with artificial intelligence models. Deep learning models, such as convolutional neural networks (CNNs), are often favored due to their high accuracy; however, they tend to consume large amounts of energy and computational resources when deployed. To address these challenges, this paper presents an ultra-lightweight, energy-efficient AI chip desigwwenabling real-time AFib detection. The proposed chip implements a Spiking Neural Network (SNN) model, providing an early hardware deployment of a PPG-based AFib detection system. The design exploits the Delta-PPI Spike Encoding and near-memory computing techniques to reduce the complexity and the power consumption. Two peak detection algorithms are proposed, including a dynamic thresholding method and an exponential moving average (EMA)-based approach, to enhance robustness across different signal characteristics. Experimental results on the MIMIC III and MIMIC PERform AF datasets, evaluated using the dynamic thresholding scheme, achieve an accuracy of 98.42% on AFib detection. The design was first validated on an FPGA and then synthesized and post-synthesis simulated using a low-power 40-nm CMOS technology. The ASIC implementation occupies only 0.027 mm2. Operating at 1 MHz, it consumes 0.018 mW, achieves an energy efficiency of 0.6 nJ per classification with a latency of 34 μs. Compared with previously reported ECG-based hardware accelerators and software-driven PPG solutions, typically operate at milliwatt-level power budgets and millisecond-scale latency, the proposed PPG hardware accelerator provides substantially lower power consumption while maintaining comparable or faster inference. These characteristics make the proposed AI chip particularly well suited for always-on, energy-constrained wearable AFib monitoring applications.