Early Differentiation of Refractory Temporal Lobe Epilepsy Exposed to Anticonvulsant Monotherapy Based on EEG Signal Using Machine Learning Algorithm
Abstract number :
3.164
Submission category :
3. Neurophysiology / 3C. Other Clinical EEG
Year :
2022
Submission ID :
2204607
Source :
www.aesnet.org
Presentation date :
12/5/2022 12:00:00 PM
Published date :
Nov 22, 2022, 05:25 AM
Authors :
Kyung-Il Park, MD, PhD – Seoul National University Hospital Healthcare system Gangnam center; seon-Jae Ahn, MD – Seoul National University Hospital; Kon Chu, Prof – Seoul National University Hospital; yoonhyuk Jang, MD – Seoul National University Hospital; Keun-Hwa Jung, Prof – Seoul National University Hospital; ki-Young Jung, Prof – Seoul National University Hospital; Mi-Kyung Kang, MD – Seoul National University Hospital; Yong-Jeong Kim, PhD – Seoul National University Hospital; Young-Gon Kim, prof – , Seoul National University Hospital; Han Sang Lee, MD – Seoul National University Hospital; Seung-Bo Lee, Prof – Keimyung University; WanKiun Lee, MD – Seoul National University Hospital; yoonkyung Lee, MD – Seoul National University Hospital; Yong Woo Shin, MD – Seoul National University Hospital; Hyoshin Son, MD – Seoul National University Hospital; Sungeun Hwang, MD – Ewha Womans University Mokdong Hospital; Sang Kun Lee, Prof – Seoul National University Hospital
Rationale: Temporal lobe epilepsy (TLE) is one of the most common forms of focal epilepsy. Approximately 30% of patients are drug-refractory. Surgical resection has been proved as an effective treatment option for refractory TLE patients, but it takes years to judge if a patient is refractory. Therefore, differentiating refractory cases in the early stage of treatment could reduce unnecessary, unsuccessful antiepileptic drug (AED) treatment periods.
Methods: All EEGs following 10-20 system were collected from the same machine (NicoletOne, Natus). A total of 46 sleep-and-waking EEG sets from 46 TLE patients. Twenty-two patients were seizure-free, and 24 were refractory for the last one year of follow-up. All EEGs were the first EEG taken for each patient, and patients were taking one AED at the time of EEG acquisition. After preprocessing, four sets of time-domain features were extracted, and a random forest was utilized to investigate the best features. Classification performance of identifying refractoriness has been evaluated using XGBoost, CatBoost, and linear discriminant analysis for the best feature set.
Results: Feature set using kurtosis, maximum, mean, median, minimum, and skewness showed the best performance (AUC 0.748±0.163, accuracy 0.824±0.135, F1 score 0.767±0.213, true positive rate 0.750±0.274, true negative rate 0.880±0.160, positive predictive value 0.887±0.157 and negative predictive rate 0.842±0.145). Among various algorithms, XGBoost showed the best overall performance with AUC 0.765±0.179, accuracy 0.827±0.112, F1 score 0.798±0.126, true positive rate 0.760±0.159, true negative rate 0.880±0.160, positive predictive rate 0.880±0.160 and negative predictive value 0.820±0.099. Kurtosis and maximum using more than 120 seconds of EEG signal in low gamma band (30-50 Hz) showed optimal performance (P = 0.0038).
Conclusions: Using optimal time-frequency features, refractoriness in TLE patients was successfully predicted with EEG signals from the early stage of AED treatment. Early identification of refractory TLE patients can lead to early surgical intervention to minimize recurrent seizures.
Funding: Funded by Seoul National University Hospital
Neurophysiology