LiMO: A Lightweight MambaOut system for end-to-end IoMT ECG Diagnosis with Fully Configurable Quantization Co-Design...

Deep learning frameworks such as Mamba have significantly improved inference speed and computational efficiency. However, deploying them on resource-constrained edge devices on the Internet of Medical Things (IoMT) poses challenges due to the demands for real-time processing, secure data handling, and ultra-low power consumption. To address these constraints, this paper introduces LiMO, a lightweight MambaOut architecture that runs entirely on an microcontroller unit (MCU) for real-time electrocardiogram (ECG) diagnosis. By integrating the proposed fully configurable quantization, scale-linear layer fusion, and a hardware-software co-design approach, LiMO substantially reduces computational loading while maintaining medical-grade accuracy. Notably, the entire model consists of just 1.8k parameters. When deployed on the nRF52840 MCU, LiMO occupies only 54.4 KB of Flash and 10.25 KB of RAM, achieves 98.75% accuracy on the MIT-BIH dataset, and consumes as little as 8.08 mJ per heartbeat classification at 3.3 V. These results demonstrate that LiMO successfully bridges the gap between state-of-the-art deep learning methodologies and the stringent requirements of real-time ECG diagnosis in MCU-based IoMT environments, offering a practical and energy-efficient solution for next-generation medical monitoring.