Comparison of Machine Learning Algorithms for Somatosensory Functional Mapping
Iktimal Alwan1, Dennis Spencer2, Rafeed Alkawadri1
1Human Brain Mapping Program and Department of Neurology, University of Pittsburgh medical center, 2Yale University School of Medicine
Objective:

Using free-running intracranial EEG (icEEG) data, we compare the performance of machine learning algorithms with a binary electrocortical stimulation (ECS) outcome to map functional somatosensory (SM) and nonfunctional areas over the frontoparietal regions.

Background:

Artificial intelligence (AI) is the “science and engineering of making intelligent machines.” Despite advances in AI and its transformative role in medicine, it is lagging in the field of epilepsy surgery. IcEEG data are an ideal medium for AI applications.

Design/Methods:
The study included consecutive cases with intractable epilepsy that underwent icEEG monitoring in a single center between 2013-2018, in which standard functional evaluation of the SM cortex was performed.  We evaluated the performance of support vector machines (SVMs), simple multilayer perceptron (MLP), bootstrap forests (BF), boosted trees (BT), decision trees (DT), and logistic regression (LR) to identify the SM cortex using previously validated characteristics from 6-10 minutes of resting icEEG data, applying standard common parameters, 10-fold cross-validation, and no expert supervision.  
Results:

Seven subjects and 376 contacts were identified. The median threshold for the SM mapping was 2.5 mA and a range of 1.5 to 7 mA. Function-to-other ratio: 1:2.35.  SM cortex showed higher EEG band power than non-SM cortex (medians normalized power 0.22 [IQR 0.11–0.93] vs 0.12 [IQR 0.08–0.24] P < 0.0002). The mean (maximum) AUC for MLP, SVM, BF, LR, BT and DT in classifying function versus the other was 87.4% (95.2%), 84.6% (89%), 84.1% (90.6%), 82% (93.8%), 81.7% (94.2%) and 80.13% (90.7%) respectively.

Conclusions:

When analyzing binary SM functional classification without expert input, the performance of AI algorithms is comparable, with MLP then SVM consistently outperforming the others. Our findings suggest that future research and applications should emphasize clinical features and indications rather than algorithms. To translate our findings into clinical practice, experts should investigate new features, refine parameters within an algorithm, and conduct prospective validations.

10.1212/WNL.0000000000203115