Comparison of Machine Learning Algorithms for the Passive Identification of Sensorimotor Cortex During Intracranial EEG Evaluation: Cross-Validation and Extended Evaluation
Abstract number :
1.17
Submission category :
3. Neurophysiology / 3G. Computational Analysis & Modeling of EEG
Year :
2023
Submission ID :
130
Source :
www.aesnet.org
Presentation date :
12/2/2023 12:00:00 AM
Published date :
Authors :
Presenting Author: Rafeed Alkawadri, MD – Unviersity of Pittsburgh Medical Center
Iktimal Alwan, MBBS – UPMC; Cigdem Isitan Alkawadri, MD – University of Pittsburgh Medical Center; Dennis Spencer, MD – Yale University
Rationale: Despite its advanced spatial and temporal resolution, only 3% of intracranial (ic-) EEG data are used in clinical settings. This highlights the potential of machine learning (ML) to improve data interpretation. This study compared ML algorithms in the passive mapping of the sensorimotor cortex
Methods: Consecutive cases of refractory epilepsy with comprehensive sensorimotor mapping were included during icEEG evaluations from 2013 to 2018. Mapping was performed using standard electrical stimulation (ECS) and median somatosensory evoked potentials (SSEP). We assessed the performance of Support Vector Machines (SVMs), Random Forest (RF), Decision Trees (DT), Single Layer Perceptron (SLP), and Multilayer Perceptron (MLP), against standard Logistic Regression (LR) to identify the sensory-motor cortex and central sulcus. This was achieved using validated characteristics derived from six-minute icEEG NREM sleep data. We applied standard 10-fold cross-validation and extended evaluation to detect the best-performing models with one-electrode-contact sensitivity. Our analyzes were performed on four vetted features, according to the classical univariate analysis at p< 0.05, and extended to 17 features, incorporating a combination of power/coherence in different frequency bands, entropy, and distance-based metrics. These analyzes were performed before and after weight adjustment for unbalanced data with the most prevalent contacts and negative functional values. We evaluated performance based on accuracy, precision, recall, F1 scores, specificity, and area under the curve (AUC).
Neurophysiology