Abstracts

Automated Determination of Seizure Onset Focality and Spread Using Machine Learning Approaches

Abstract number : 3.169
Submission category : 3. Neurophysiology / 3G. Computational Analysis & Modeling of EEG
Year : 2023
Submission ID : 3
Source : www.aesnet.org
Presentation date : 12/4/2023 12:00:00 AM
Published date :

Authors :
Presenting Author: Nikola Bolt, MS – Massachusetts General Hospital

Sydney Cash, MD, PhD – Massachusetts General Hospital and Harvard Medical School; Pariya Salami, PhD – Massachusetts General Hospital and Harvard Medical School

Rationale:
Current surgical treatments in patients with epilepsy include resection and neuromodulation. For many patients, these are tremendously useful approaches. For others, there may be little change in their seizures. Unfortunately, we are still struggling to understand the reasons for this discrepancy. The main hypothesis driving this study is that different patterns of seizure onset and spread have different outcomes after surgery. Exploration of this possibility requires annotated data in which seizure onset and spread regions are identified. To facilitate this labor-intensive process, we have developed a machine learning classifier that can detect epileptic seizures and identify the start of ictal activity on a single-channel basis with high temporal resolution. This allows us to identify new propagation patterns that will be informative of surgical outcomes.



Methods:
Seizures (n=523) from 81 patients who received intracranial electrodes as part of their presurgical evaluation were manually annotated. First, we used a subset (n=58) of these seizures and surrounding non-seizure data. The recordings were standardized and split into non-overlapping one-second segments. Time domain (line length, AUC, standard deviation) and spectral domain features (δ, θ, α, β, γ powers and δ-to-θ, θ-to-α, δ-to-α ratios) were extracted and normalized to a non-seizure baseline. To conserve some temporal dependency, ratios to the previous time-segment were included, resulting in a total of 22 features. Finally, various kinds of binary classifiers were trained to distinguish between seizure and non-seizure segments and evaluated using leave-one-patient-out cross-validation. The best classifier was then applied to the additional data to classify each channel into onset, propagation and no ictal activity.



Results:

With an average balanced accuracy of 89.3%, recall of 82.4% and precision of 95.1% on the held-out patient segments, a random forest classifier consisting of 280 decision trees showed the best performance. Analysis of feature importance showed that line length was by far the most important, followed by θ-power and δ-to-θ ratio. Most features of the ratio to the previous segment, however, were found to be unimportant in classification. By applying the classifier to the additional recordings, we could identify onset and propagation channels and facilitate the analysis of the corresponding seizures, of which 37% were found to have a focal onset without spread, 50% were focal onset with spread and 13% had a broad onset. Interestingly, some patients (36%) had multiple types of patterns.



Conclusions:
The proposed classifier demonstrated high performance in differentiating between seizure and non-seizure activities and was generalizable across patients. This approach allowed us to successfully group seizures into subtypes based on their spread into different brain regions. This novel classification scheme will be of great importance in future research to identify whether seizure onset and spread patterns can be determining factors in the patient’s treatment outcome.



Funding: R01-2NS062092, DoD CDMRP W81XWH-22-1-0315

Neurophysiology