Integration of Machine Learning for the Automated Analysis of MEG Recordings in Non-invasive Seizure Mapping
Abstract number :
1.323
Submission category :
4. Clinical Epilepsy / 4B. Clinical Diagnosis
Year :
2024
Submission ID :
943
Source :
www.aesnet.org
Presentation date :
12/7/2024 12:00:00 AM
Published date :
Authors :
Presenting Author: Arya Shetty, BS – The University of Texas Health Science Center
Yash Vakilna, MS – The University of Texas Health Science Center
Nolan Rizzo, BS – The University of Texas Health Science Center
Ahmed Massoud, MD – The University of Texas Health Science Center
Caroline Hanan, BS – The University of Texas Health Science Center
Michael Funke, MD, PhD – McGovern Medical School, The University of Texas Health Science Center at Houston
Indira Kommuru, MD – The University of Texas Health Science Center
Jeremy Lankford, MD – McGovern Medical School, The University of Texas Health Science Center at Houston
Shelley Varnado, MD – The University of Texas Health Science Center
Gretchen Von Allmen, MD – McGovern Medical School, The University of Texas Health Science Center at Houston
Michael Watkins, MD – McGovern Medical School, The University of Texas Health Science Center at Houston
John Mosher, PhD – The University of Texas Health Science Center
Manish Shah, MD – McGovern Medical School, The University of Texas Health Science Center at Houston
Rationale: Pediatric epilepsy affects 1 in 300 children. Despite advances in medical management, approximately 30% of cases require surgical intervention. Surgical success is dependent on accurate localization of seizure foci, with success rates dropping by 20-40% in the absence of an MRI-distinguishable (lesion-negative) seizure foci. At our institution, standard of care calls for manual analysis of MEG recordings for identification of potential seizure foci. This study aims to assess whether automated analysis of MEG recordings can provide a feasible avenue for the localization of seizure foci in lesion-negative epilepsy.
Methods: The Freesurfer segmentation pipeline was used to generate anatomical references from patient MRIs. Once segmentations were generated, MATLAB was used to implement the Brainstorm toolbox in the following pipeline. First, a patient’s MEG data was integrated with their anatomic reference provided by Freesurfer segmentation. Next, dipole scanning was used to model epileptic activity as dipoles. Modeled dipoles that met the following criteria were selected for further analysis. Dipoles were then filtered by amplitude, goodness of fit, and reduced chi-squared. Filtered dipoles were clustered by k-means clustering and resulting clusters were overlaid over patient MRI. Patients chosen for this analysis had already received surgical intervention. The results predicted in this analysis were compared to the surgical resection site for assessment of accuracy.
Results: Clustering results were graded as “concordant” if they exhibited overlap between the surgical resection site and the clusters, “near concordant” if they lacked overlap, but closely approximated the resection site, and discordant if there was no association with the surgical resection site. In total, 12 patients were analyzed for this investigation. Concordant predictions were generated for 9 out of the 12 patients (75%), 1 patient had a near concordant prediction (8%) and 2 patients had discordant predictions (17%).
Conclusions: This study suggests that automated analysis of MEG data, alongside the incorporation of machine learning algorithms, holds significant potential as an avenue for unsupervised, non-invasive seizure mapping.
Funding: None
Clinical Epilepsy