Abstracts

Rationale: Increasing power of oscillation in high frequency band on intracranial EEG has been accepted as a useful biomarker for detection of epileptic focus. Automated machine learning methods for EEG reading have been increasingly attracting attention recently. In this study, we investigated the accuracy of machine learning using some entropies and support vector machine (SVM) for detection of seizure onset zone (SOZ) in patients with epilepsy. Methods: Subjects of this study were 8 patients with FCD type 2 and Engel's class 1 surgical result. The ECoGs datasets of 30-minutes of interictal phase were used and splited into 20-s segments each to analyze. A band-pass filter was applied between 100 to 600Hz and divided them into 10 sub-bands with 50Hz interval. The following entropies were indicated to extract features as SOZ from non-SOZ; Approximate, Sample, Permutation, Shannon, Reny's, Phase 1 and 2, and Tsallis entropies. Sparse linear discrimination analysis (sLDA) was applied to select the prominent entropies, and subsequently 10-fold cross-validation techniques were adapted to evaluate the accuracy of this method using SVM. Ethical procedures: This study is approved by Juntendo university ethical committee (16-163, 18-143).  Results: Features of Sample, Permutaion and Approximate in fast ripple band (250-600Hz) resulted in higher sLDA weights in many cases. Using selection of entropies and adjusted sampling data number between SOZ and non-SOZ, the AUCs of ROC resulted from 0.63 to 0.99. Sensitivities of this methods were from 23.70% to 88.52%, and specificities were from 82.83% to 98.54%. Tendency of lower sensitivity and precision in pediatric patients existed in comparison to adults. Conclusions: Machine learning for epileptic focus detection using multiband entropy-based feature-extraction is valuable. Feature-extraction using some entropies in pediatric patients has a tendency to be less sensitive than in adults. Funding: This study is supported by JST CREST (JPMJMR 0784), Japan.