Training a Logistic Regression and Decision Tree machine learning model for identifying the epileptogenic zone by classification of intraoperative electrocorticography leads as epileptic or non-epileptic in patients with focal epilepsy undergoing epileptic surgery.

For people with focal epilepsy who do not respond to anti-epileptic medication, surgery to remove the epileptogenic zone (EZ) is considered the first-line treatment. However, the seizure-free rate is just 38% to 79% at ten years postsurgery. This can be due to the inaccurate delineation of the EZ during presurgical assessment, of which intraoperative electrocorticography (ECoG) is one. Nowadays, the physicians visually asses the ECoG for EZ biomarkers, but the inter-rater reliability of ECoG assessment is relatively low. For this reason Machine Learning (ML) was applied as it might help with delineation of the EZ. Two ML models, Logistic Regression (LR) and Decision Tree (DT), were developed and trained, using cross-validation, GridSearch and hyperparameter tuning. The dataset used for this process contained intraoperative ECoG observations from 96 patients. The DT model had an overall better performance in F1 score (0.609), accuracy (76.4%), sensitivity (55.7%) and AUC (0.80) compared to the LR model (0.348, 69.8%, 24.4%, 0.69). However, the specificity was higher in the LR model (92.1%) as to DT (86.5%). Because these measurements are relatively low for clinical practice, it can be concluded that both models cannot be used as support for surgeons with delineation of the EZ in patients with focal epilepsy.