Classifying decision-based effector modalities from fragmented iEEG via dynamic recursive feature elimination

Presentation at conference

A key challenge in data science is dealing with structured missing data with little or no overlap of features between samples. This is especially pertinent in neuroscience, such as in high-density intracranial electroencephalography (iEEG), where different brain locations are recorded across participants, and making machine learning analysis less straightforward. To address this, we propose a dynamic recursive feature elimination with cross-validation (DyRFECV) approach that comprises dynamic step sizes. A Light Gradient Boosting Machine (LGBM) classifier acts as the base estimator, and stratified five-fold cross-validation is employed. The method, with its variants, was tested on an open iEEG dataset in which participants performed a perceptual decision-making task while reporting their decisions with different effector types - manual or vocal responses. We found DyRFECV could classify decision-based, effector-specific iEEG signals with high classification performance (F1-scores); 97.29% using the baseline approach without hyperparameter tuning, and 99.42% with hyperparameter tuning; identifying 320 and 416 iEEG contacts, respectively — a tradeoff between F1-score and fewer contacts. Furthermore, feature importance indicated that the main contributions came from frontal and parietal cortical regions, consistent with existing literature. Overall, this work demonstrates the promising application of DyRFECV on structured missing iEEG data.