Abstract
High-frequency oscillations (HFOs) in intracranial EEG (iEEG) recordings are critical biomarkers for localizing the seizure onset zone (SOZ) in patients with focal refractory epilepsy. Despite their clinical significance, HFO analysis is often compromised by high-frequency artifacts that bypass conventional detectors, resulting in false-positive events that dilute the reliability of the HFO pool. To address this challenge, this study aimed to develop an automated method to accurately identify and eliminate false-positive events, ensuring more robust and artifact-free HFO analysis for clinical applications. Using iEEG data from 15 patients with focal epilepsy, we implemented an attention-based cascaded residual dictionary learning framework coupled with a random forest classifier. Events passing an initial amplitude detector underwent a second-stage refinement to remove artifacts and non-neural noise that mimicked HFOs. This was achieved by evaluating event reconstruction quality using a dictionary learned from genuine HFOs. Compared to visual assessments by three human experts, the proposed method demonstrated 92.14% classification accuracy in distinguishing real HFOs from pseudo-HFOs. Additionally, the method improved SOZ localization accuracy in noisy iEEG data by 20% (p=6e-5) and in clean iEEG data by 4% (p=3.3e-3). The learned dictionary effectively captured raw HFO morphology in shallow layers, while deeper layers identified ripple and fast ripple components, all without human supervision. These findings highlight the algorithm's effectiveness in detecting pseudo-HFOs in corrupted iEEG data, thereby enhancing the clinical utility of HFOs as biomarkers for SOZ in epilepsy.