BioMIFormer: Biologically inspired spatiotemporal transformer for motor imagery EEG interpretation

Abstract

Electroencephalography (EEG) provides a high temporal resolution into brain activity, yet many existing decoding models overlook its neurophysiological basis, limiting the ability to capture both temporal and spatial dynamics effectively. To address this, we propose BioMIFormer, a Biologically Inspired Spatiotemporal Transformer that integrates neuroscientific insights into the functional architecture of motor imagery (MI)-related brain regions. BioMIFormer simulates functional divisions of the motor cortex through three parallel branches for temporal, spatial, and spatiotemporal feature encoding. It incorporates a biologically inspired temporal modeling mechanism, a multi-scale spatial extraction strategy, and a Transformer-based attention fusion module to capture long-range dependencies and cross-modal relationships in EEG signals. Different from the traditional feature fusion methods, BioMIFormer aligns its modular design with the brain’s functional architecture, enhancing both interpretability and decoding accuracy. Experimental results on two public MI-EEG datasets demonstrate that BioMIFormer achieves state-of-the-art performance, validating the effectiveness of biologically inspired modeling in EEG analysis.