TinyEEGConformer: An Attention-Based EEG Decoding Model for Embedded Systems

Brain-Machine Interfaces (BMIs) are crucial in prosthesis control and motor rehabilitation. Processing real-time electroencephalographic (EEG) data on edge devices, such as low-power microcontroller units (MCUs) positioned close to the sensors, is an emerging and important trend. In fact, this approach enhances user experience and preserves data privacy. However, due to the limited computational resources of the low-power processors, having adequate performance of AI-driven EEG decoding models on low-power devices is still challenging. This paper presents TinyEEGConformer, a novel attention-based EEG decoding model targeting resource-constrained embedded systems for real-time processing at the edge. The proposed model is designed to meet the computational constraints of a low-power processor while maintaining optimal classification performance. The model achieves an average accuracy of 79.63%, outperforming the baseline EEG Conformer model by 0.97% on the BCI Competition IV 2a dataset while using 21 times fewer parameters (i.e., 37,135 parameters corresponding to 145 kB). The proposed approach is evaluated on Arm Cortex-M processors demonstrating its feasibility for on-board deployment. TinyEEGConformer shows comparable classification accuracy as state-of-the-art (SoA) EEG decoding models but significantly reduces the computational burden, paving the way for online motor intention decoding on low-power wearable BMIs.