A Machine Learning Approach for Stratifying Ictal EEG Patterns and Optimizing Epilepsy Surgery Outcomes
Abstract number :
1.438
Submission category :
9. Surgery / 9A. Adult
Year :
2024
Submission ID :
682
Source :
www.aesnet.org
Presentation date :
12/7/2024 12:00:00 AM
Published date :
Authors :
Fawzi Babtain, MBBS, MHSc, FRCPC, CSCN (EEG, EMG) – King Faisal Specialist Hospital and Research Centre- Jeddah
Presenting Author: Albatool Almelhem, MBBS – King Faisal Specialis Hospital and Research Center-Jeddah
Afnan Alkhotani, MBBS – King Faisal Specialis Hospital and Research Center-Jeddah
Saleh Baeesa, MD – King Faisal Specialist Hospital and Research Centre
Mohammed Bin Mahfoodh, MBBS, FRCS(C) – King Faisal Specialis Hospital and Research Center-Jeddah
Youssof Al Said, MBBS, FRCP(C) – King Faisal Specialis Hospital and Research Center-Jeddah
Rationale:
Machine learning (ML) applied to scalp EEG has advanced the prediction of surgical outcomes in refractory epilepsy. However, using ML to identify ictal EEG patterns to enhance surgical planning and outcomes needs further understanding.
Methods:
We conducted a retrospective evaluation of epilepsy surgery outcomes, applying the International League Against Epilepsy (ILAE) classification to refractory epilepsy cases from 2020 to 2023. We used unsupervised machine learning to identify ictal EEG patterns and predict surgical outcomes.
Results:
We studied 29 patients. Table 1 summarizes patient demographics. We reviewed 235 seizures, a mean of 10 recorded per patient, with a mean duration of 98 seconds. Seizure outcomes classification revealed that 52% were Class I, 10% were Class II, 3% were Class III,17% were Class IV, and none had Class V. The Elbow method trained 10 models to identify 4 clusters to use for the K-means model (figure 1). Cluster 0 exhibited seizures averaging 102 seconds and high onset frequencies (6.5 Hz). These seizures had well-defined onset zones in the left temporal/central regions. Cluster 1 demonstrated less defined onset zones, shorter seizures (89 seconds), and lower onset frequencies (3.7 Hz). Cluster 2 had the longest and most complex seizures (188 seconds), with primarily left temporal but not central origins and mixed surgical outcomes. Cluster 3 was the most homogenous, with moderate seizure durations (97 seconds), clear onset zones, and exclusive rhythmic theta patterns, potentially indicating a specific seizure subtype amenable to treatment. Finally, 52% of Cluster 0 and 60% of Cluster 3 had class I surgical outcomes (figure 2).
Conclusions:
The k-means clustering analysis identified distinct ictal EEG patterns associated with surgical outcomes. Such findings suggest the potential use of machine learning to utilize ictal EEG to predict surgical success and personalize treatment in epilepsy. Furthermore, this could pave the way for improved patient selection and tailored epilepsy surgery approach.
Funding: None
Surgery