Energy-Efficient Spiking-CNN-Based Cross-Patient Seizure Detection
Abdul Muneeb, Hossein Kassiri
Abstract
A neuromorphic spiking convolutional neural network (SCNN) is presented for cross-patient seizure detection using multi-modal features from multi-channel electroencephalogram (EEG) data. A mixture of spectral, temporal, and spatial features is employed for building robustness against domain-specific noise/artifacts, hence boosting detection sensitivity and specificity. The feature set is converted to temporally-coded spikes before being fed to the SCNN classifier. Thanks to the asynchronous spike-based multiplier-less operation, the SCNN significantly reduces the classification computational cost without sacrificing accuracy. The developed algorithm was validated on a publicly available dataset and an average sensitivity of 83.02%, a specificity of 86.31%, and a false positive rate of 0.69/hr were achieved for cross-patient seizure detection. Our results show that a 1-bit Integer-Net leads to less than 2% drop in sensitivity compared with a 32-bit real-value resolution CNN model while offering more than 27× improvement in memory efficiency. The SCNN achieves an estimated energy efficiency of <tex>$1.28\mu\mathrm{J}$</tex> /classification, which translates into a 98.6% improvement compared to a conventional CNN implementation with the same accuracy.