Pseudo-HFOs Elimination in iEEG Recordings Using a Robust Residual-based Dictionary Learning Framework

Behrang Fazli Besheli, Zhiyi Sha, Amir Hossein Ayyoubi, Chandra Prakash Swamy, Thomas Henry, Gregory A. Worrell, Kai J. Miller, Nuri F. Ince

Keywords: Artifact Rejection, Cascaded Dictionary Learning, Drug-resistant Epilepsy, intracranial EEG, Focal Epilepsy HFO, Neurobiomarkers

TL;DR: In this work, we proposed a machine learning method using dictionary learning and random forest classifier to distinguish between HFOs and those artifacts resembling HFOs (pseudo-HFOs)

Abstract: High-frequency oscillations (HFOs) in intracranial EEG (iEEG) recordings are crucial for localizing the seizure onset zone (SOZ) in patients with focal refractory epilepsy. While HFOs are essential for clinical assessment, high-frequency artifacts may pass conventional HFO detectors, resulting in false-positive events that contaminate the HFO pool. The main goal of this study is to automatically detect and eliminate those false positive events in the pool of initially detected candidate HFOs. We analyzed one hour of iEEG data from fifteen patients with focal epilepsy using an attention-based cascaded residual dictionary learning framework, coupled with a random forest classifier. This data-driven method employed sparse robust representation using the Huber loss to eliminate artifacts and noises with non-neural origins that mimicked HFOs by evaluating the quality of event representation using a dictionary learned from real HFOs. Compared with visual assessments by three human experts, the proposed method achieved a 92.14% classification accuracy in distinguishing real HFOs from pseudo-HFOs. Additionally, in noisy iEEG data, our method improved HFO-based SOZ localization by 20% (p=6e-5), while in clean iEEG data, the improvement was 4% (p=3.3e-3). The learned dictionary successfully captured the morphology of raw HFOs in shallow layers, while it captured ripple and fast ripple components in deeper layers without human supervision. Our work shows that the proposed algorithm effectively detects pseudo-HFOs and improves the clinical value of HFOs in SOZ localization.

Track: 4. AI-based clinical decision support systems

Registration Id: 94NXDKM7DZD

Submission Number: 88