Refining epileptogenic high-frequency oscillations using deep learning: a reverse engineering approach.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Interictal high-frequency oscillations (HFOs) are potential biomarkers for the epileptogenic zone in epilepsy.
• Visual verification of HFOs is time-consuming and lacks inter-rater reliability.
• Distinguishing epileptogenic HFOs from non-epileptogenic HFOs remains a challenge.

Purpose of the Study:

• To develop a deep learning algorithm for accurate HFO detection and classification.
• To differentiate epileptogenic HFOs from non-epileptogenic HFOs using a novel weakly supervised model.
• To identify key features of epileptogenic HFOs for improved surgical targeting.

Main Methods:

• Constructed a deep learning algorithm using chronic intracranial EEG data from 19 children with drug-resistant neocortical epilepsy.
• Employed a weakly supervised model to discover and label epileptogenic HFOs.
• Utilized 'purification power' of deep learning for automatic HFO relabeling and feature extraction.

Main Results:

• The model achieved 96.3% accuracy in artefact detection and 86.5% accuracy in classifying HFOs with or without spikes.
• Resection of identified epileptogenic HFOs positively correlated with post-operative seizure freedom (AUC=0.87, P=0.01).
• Epileptogenic HFOs exhibit distinct signal intensity patterns (inverted T-shaped) and spike-like features.

Conclusions:

• Deep learning provides a reliable method for HFO classification, replicating expert performance.
• The developed algorithm successfully distinguishes epileptogenic HFOs, aiding in precise localization of the seizure onset zone.
• Identifying specific features of epileptogenic HFOs enhances understanding and potential treatment of focal epilepsy.