Using Deep Learning to Classify Ictal Onset Patterns
Abstract number :
2.199
Submission category :
3. Neurophysiology / 3G. Computational Analysis & Modeling of EEG
Year :
2024
Submission ID :
1002
Source :
www.aesnet.org
Presentation date :
12/8/2024 12:00:00 AM
Published date :
Authors :
Presenting Author: Andrew Michalak, MD, MS – New York University - Langone
Edward Merricks, PhD – Columbia University Irving Medical Center
Adeen Flinker, PhD – New York University - Langone
Daniel Friedman, MD – New York University Grossman School of Medicine, NYU Langone Health
Werner Doyle, MD – NYU Langone Health
Sameer Sheth, MD, PhD – Baylor College of Medicine
Neil Feldstein, MD – Columbia University Irving Medical Center
Brett Youngerman, MD MS – Department of Neurosurgery, Columbia University Irving Medical Center
Guy McKhann, MD – Columbia University Irving Medical Center
Catherine Schevon, MD, PhD – Columbia University Irving Medical Center
Rationale: Ictal onset patterns (IOPs) on intracranial electroencephalography (iEEG) have utility in localizing the seizure onset zone and predicting outcomes after resection or ablation. However, prior IOP literature has relied on manual review which is time-intensive and subjective. Further, most studies have focused on the first pattern seen, and the pattern seen in the area of clinician-defined seizure onset. Subsequent ictal patterns and patterns outside of the seizure onset zone may have prognostic value. To address and make tractable these issues, we trained a convolutional neural network (CNN) to classify IOPs.
Methods: To train the CNN, iEEG seizure recordings were clipped and labeled using classic descriptions of low voltage fast activity, rhythmic slowing, repetitive spiking, and other/field. All channels were included, as well as primary and secondary patterns. Preprocessing consisted of applying 1 Hz highpass and 60 Hz notch filters then downsampling to 500Hz. Seizure clips ranged from 1-10 seconds. Randomly permutated one-second segments were selected from labeled clips, without replacement. Class imbalances were addressed with resampling. The CNN used sequences of 500 datapoints from 1 second clips as the input layer. Internal accuracy was assessed based on 80/20 training/testing of all IOP clips. External accuracy was assessed using iterative model training with “leave N patients out” by excluding all seizures from 1, 2, 3, and 4 patients.
For continuous classification, a one second sliding window gives a classification confidence for each IOP. Next, a 2 second sliding window provides an IOP classification based on the highest average IOP confidence in that window. The output is a sliding IOP classification as well as a confidence level for the classification (figure 1C).
Results: Twenty-seven seizures from 20 patients who underwent stereo EEG were used. After clipping, 32-174 unique one-second samples were available per IOP for model training. Hyperparameter tuning yielded models with an internal accuracy ranging between 80-90%. The selected model had a mean internal accuracy (± standard deviation) of 81.0% (± 1.6%) and external accuracy on “leave N out” testing of 74.7% (± 10%). Figure 1C shows the continuous readout of IOPs from a single seizure.
Conclusions: Convolutional neural networks can be used to classify iEEG ictal onset patterns across the entire recording array, enabling exploration of all (both spatial and temporal) patterns. Further directions include validating this with less commonly encountered IOPs and in a multicenter dataset with comparison to human interrater reliability.
Funding: R01NS084142, R01NS110669
Neurophysiology