Abstracts

Deep Learning-Based Classification of High-Frequency Oscillations Identified by Ripplelab

Abstract number : 3.028
Submission category : 1. Basic Mechanisms / 1C. Electrophysiology/High frequency oscillations
Year : 2023
Submission ID : 879
Source : www.aesnet.org
Presentation date : 12/4/2023 12:00:00 AM
Published date :

Authors :
Presenting Author: Krish Pai, High School Student – USC Stevens Neuroimaging and Informatics Institute

Alexis Bennett, MS – USC Stevens Neuroimaging and Informatics Institute; Serena Ramde, BS – USC Stevens Neuroimaging and Informatics Institute; Kseniia Kriukova, MD – USC Stevens Neuroimaging and Informatics Institute; Tuba Asifriyaz, MS – USC Stevens Neuroimaging and Informatics Institute; Celina Alba, MS – USC Stevens Neuroimaging and Informatics Institute; Paree Merchant, High School Student – USC Stevens Neuroimaging and Informatics Institute; Samayan Bhattacharya, BS – USC Stevens Neuroimaging and Informatics Institute; Patricia Saletti, Ph.D. – Albert Einstein College of Medicine; Solomon Moshe, MD – Albert Einstein College of Medicine; Matt Hudson, Ph.D. – Monash University; Rhys Brady, Ph.D. – Monash University; Glenn Yamakawa, Ph.D. – Monash University; Emma Braine, Ph.D. – Monash University; Idrish Ali, Ph.D. – Monash University; Juliana Castro, Ph.D. – Monash University; Nigel Jones, Ph.D. – Monash University; Sandy Shultz, Ph.D. – Monash University; Pablo Casillas-Espinosa, MD – Monash University; Aristea Galanopoulou, MD – Albert Einstein College of Medicine; Terence O'Brien, MD – Monash University; Dominique Duncan, Ph.D. – USC Stevens Neuroimaging and Informatics Institute

Rationale: Post traumatic epilepsy (PTE) is a symptomatic epilepsy that might occur after traumatic brain injury (TBI). One aim of the Epilepsy Bioinformatics Study for Antiepileptogenic Therapy (EpiBioS4Rx) is to identify and validate PTE biomarkers from electroencephalography (EEG) data. A promising biomarker is high frequency oscillations (HFOs), which are brain patterns observed in EEG in the 80-500 Hz frequency range, appearing in PTE patients. However, there is a lack of automated tools that can accurately differentiate between artifact or noise activity and HFO activity in an EEG. This study proposes a deep-learning solution increasing specificity and sensitivity of RIPPLELAB HFOs detecting algorithms.

Methods: EEG data from 38 TBI induced rodents were collected from two EpiBioS4Rx (Project 1 and 2) data collection sites (Albert Einstein College of Medicine, PI: Aristea S. Galanopoulou and Monash University, PI: Terence J. O’Brien). After performing the EEG data wrangling process the resulting EEG were run through a short-time energy (STE) filter in RIPPLELAB, a MATLAB application (Mathworks®, Natick, MA, USA), developed by Navarrete M, et al., and converted into spectrograms. The HFOs detected by RIPPLELAB were then categorized by expert reviewers from the EpiBioS4Rx Research Group into artifact and true positive HFO categories. A custom HFO dataset was created using 761 manually detected EEG events recorded in the time-frequency plot between 200-500 Hz as EEG spectrograms (537 artifact/noise events and 224 true positive HFO events). The window size was 0.5 second and each spectrogram had the Jet Colormap filter applied. Prior to automated HFO classification, the data were preprocessed by applying data augmentation, which is random transformations applied to each spectrogram. The data were then divided into training and testing frameworks with a 80% to 20% ratio respectively. A custom-built hybrid transfer learning based convolutional neural network (CNN) using the ResNet-50 architecture and recurrent neural network (RNN) was used to recognize patterns and accurately classify HFOs based on the input spectrogram. To streamline the classification processes, a video classification algorithm was developed based on a sequence of spectrograms recorded in a video. Each video frame was automatically segmented, and subsequently applied to the HFO Classifier giving a real-time classification.

Results: The deep learning framework achieved a 96.05% validation accuracy, 98.3% precision, 98.36% recall, and 99.89% AUC score on the test set.

Conclusions: The HFO Classifier’s high accuracy, AUC, and precision scores are similar to an expert reviewer from the EpiBioS4Rx group, indicating that this novel technique is a reliable alternative to manual detection. By performing automated detection over intracranial EEG recordings, we can successfully differentiate HFOs from artifact or noise activity in the brain, providing valuable insights into PTE.

Funding: Funded by NINDS U54 NS100064 (EpiBioS4Rx) and R01 NS127524.

Basic Mechanisms