Hybrid compressed lightweight epilepsy detection network based on greedy strategy

Epilepsy is a disorder of abnormal discharges in the brain, which is sudden and may lead to death in severe cases. Tens of millions of patients suffer from it every year. EEG is an effective tool for recording brain electrical activity. However, manual observation of the EEG to diagnose whether a person has epilepsy is time-consuming and labour-intensive. Thus automated epilepsy testing by EEG has become popular in recent years. In this paper, we propose a brief and efficient lightweight inverted triangular seizure detection network (LITEDN), which takes 1D channel-selected electroencephalogram (EEG) signals as input. The baseline achieves an inference accuracy of 96.08%, with a sensitivity of 95.69%, a precision of 96.44%, a specificity of 96.46%, and an F1 score of 96.06%. Based on the baseline, filter pruning is applied. Result in a mere 1.3% loss in network accuracy while reducing the number of parameters by 96.78%. Finally, a hybrid compression algorithm based on greedy strategy is proposed to further compress the bit width on the basis of filter pruning. The number of network parameters after hybrid compression is only 7.504 K, which is not only 124 times more compressed than LITEDN, but also the number of parameters in the epilepsy detection network after hybrid compression has reached a more advanced level in recent years.