A Computationally Efficient Model for Predicting Successful Memory Encoding Using Machine-Learning-based EEG Channel Selection

Abstract

Computational cost is an important consideration for memory encoding prediction models that use data from dozens of implanted electrodes. We propose a method to reduce computational expense by selecting a subset of all the electrodes to build the prediction model. The electrodes were selected based on their likelihood of measuring brain activity useful for predicting memory encoding better than chance (in terms of AUC). A logistic regression prediction model was built using spectral features of intracranial electroencephalography (iEEG) from the selected electrodes. We demonstrate our method on iEEG data from 37 human subjects performing free recall verbal short-term memory tasks. The method achieves a 36.3% reduction in the number of electrodes used for prediction, resulting in a 64.9% reduction in inference computation time with just a 0.3% loss in prediction performance compared to the case when all electrodes were used. The electrodes selected using our method provided improved prediction performance compared to those electrodes that were not selected in 31 out of 37 patients. Building upon this observation, we also developed a method to identify the subjects for whom the proposed electrode selection method would be beneficial.