Attention mechanism based IP design for dynamic channel selection in EEG epilepsy detect

Epilepsy, a prevalent neurological disorder requiring low-cost rapid detection in wearable devices, prompts this study to propose a channel attention mechanism(CAM) and convolutional neural network(CNN) hybrid framework that mitigates redundant electroencephalogram(EEG) channel interference and enhances cross-patient generalization through dynamic channel selection. The framework uses CAM to dynamically selects high-information EEG channels across patients/timepoints, eliminating redundancy through feature interaction analysis and feeds them into the CNN to carry out the feature extraction and classification.Then, lightweight processing means such as global average pooling(GAP) and dilated convolution are used to reduce the number of neurons and network complexity, and parallelisation means such as bitonic sorting and pulsed arrays are combined to achieve low-latency effects. Moreover, the hardware IP is designed under the 16-bit mixed-precision fixed-point coding and serial-parallel combination architecture. Implemented on TSMC 65nm process, the design achieves 97.1% accuracy with 77% storage reduction, 1.38mm² core area, and 37.28μs latency (5-10× faster than others) at 0.206μJ/class energy efficiency under 1V/20MHz conditions.