1011 A Patient-Specific Machine-Learning Model for Discriminating Between Behavioral States Using Intracranial EEG Recordings

INTRODUCTION: Patient-specific models using neural recordings to distinguish behavioral states can guide neuromodulation therapy design, specifically by identifying symptomatic versus asymptomatic states. Detecting symptomatic states could inform when to deliver neuromodulation therapy. We propose developing individual-subject specific models using cluster-based permutation testing (CBPT) on time-frequency maps for feature selection. METHODS: Intracranial recordings were collected from a patient (n=8 hippocampal electrodes) during a memory discrimination task with two states - successful and failed discrimination. Time-frequency analysis was performed, followed by CBPT analysis to identify spectral content differing between the two classes1. Power in the selected cluster was used as the classification feature. Class data was preprocessed with the synthetic minority oversampling technique to handle imbalances and hyperparameter tuning with RandomizedSearchCV. The XGBoost model trained with an 80/20 test/train split and 1000 permutations were tested for statistical significance. RESULTS: The model classified between the two behavioral states with 71% accuracy, 68% precision, 68% recall and an area under the curve of 73% indicating strong discriminatory power (p<0.01, permutation testing). CONCLUSIONS: We developed a predictive model that can distinguish between failed and successful memory discrimination states based on a CBPT obtained time-frequency cluster (hippocampal low frequency power) that is able to perform on an individual subject and trial basis. Next steps could expand on this by estimating patient-specific CBPT clusters and across sites, thereby using patient specific and larger feature vectors. Moreover, it is necessary to obtain such patient models for additional subjects. These findings highlight the potential of individualized computational approaches in understanding and predicting behavioral states.