Authors :
Presenting Author: Glykeria Sdoukopoulou, Msc – Cook Children's Health Care System
Saeed Jahromi, MSc – Cook Children's Health Care System
Ludovica Corona, PhD – The University of Texas at Arlington
Christos Papadelis, PhD – Cook Children's Health Care System
Rationale:
Studies using simultaneous EEG and magnetoencephalography (MEG) recordings have shown that interictal spikes are often detected by only one modality. This may be due to the anatomical sensitivities of each modality: MEG is almost blind to gyral crest and sulcal bottom sources, while EEG detects both gyri and sulci sources. Here, we aim to assess the sensitivity profile and localization accuracy of electric, magnetic, and electromagnetic source imaging (ESI/MSI/EMSI) for interictal spikes using an unsupervised machine learning (ML) framework based on computational modeling.
Methods:
We modeled sources using MRIs from eight children with drug-resistant epilepsy (DRE). In total, 1,344 sources were placed in seven brain regions of both hemispheres. Orientation was defined as the angle between each source’s mean normal vector and the nearest inner skull surface. Sources at gyral crests and sulcal bottoms were radial, those on sulcal wall were tangential, and oblique sources had both components. Each source was individually activated using a spike waveform from the intracranial EEG of a patient with DRE. Simulated EEG and MEG signals were acquired by superimposing artifact-free background activity from eight age-matched healthy controls onto the modeled sensor-level data. Spikes were marked by an epileptologist and the signal-to-noise ratio (SNR) was calculated. We performed ESI, MSI, and EMSI and computed the localization error (LE) as the Euclidean distance between the estimated and ground-truth sources (Fig. 1A). We applied unsupervised ML to identify sources having similar biophysical properties and localization profiles. We defined a dataset for each source imaging technique, with the spatial extent, orientation, MEG/EEG spike SNR, and their LEs as features. Finally, we applied Principal Component Analysis on the normalized features of each dataset, performed k-means clustering, and identified the key features driving localization performance in ESI, MSI, and EMSI (Fig. 1B).
Results:
We found that MEG’s sensitivity increased with SNR for sulcal sources (p<