Authors :
Presenting Author: Shingo Oana, MD, PhD – Devid Geffen School of Medicine at UCLA
Yipeng Zhang, MS – University of California; Lawrence Liu, MS – Department of Electrical and Computer Engineering – University of California; Yuanyi Ding, MS – Department of Electrical and Computer Engineering – University of California; Xin Chen, MS – Department of Electrical and Computer Engineering – University of California; Tonmoy Monsoor, MS – Department of Electrical and Computer Engineering – University of California; Atsuro Daida, MD, PhD, – Department of Pediatrics – Devid Geffen School of Medicine at UCLA; Shaun Hussain, MD, – Department of Pediatrics – Devid Geffen School of Medicine at UCLA; Raman Sankar, MD, PhD, – Department of Pediatrics – Devid Geffen School of Medicine at UCLA; Fallah Aria, MD, MS, – Pediatric Neurosurgery – University of California; William Speier, Ph.D. – Department of Radiological Sciences – University of California; Jerome Engel Jr., M.D., Ph.D. – Department of Neurology – University of California; Richard Staba, Ph.D. – Neurobiology – University of California; Vwani Roychowdhury, Ph.D. – Department of Electrical and Computer Engineering – University of California; Hiroki Nariai, MD, PhD, MS – Department of Pediatrics – Devid Geffen School of Medicine at UCLA
Rationale: Interictal high-frequency oscillations (HFOs) are considered one of the promising spatial neurophysiological biomarkers of the epileptogenic zone. Yet, manual analysis by human experts is not only time-consuming but also exhibits inconsistencies due to inter-rater variability. Automated detectors can provide a more objective approach, but their use is often limited by the scarcity of user-friendly applications available to clinicians or researchers who lack coding expertise. One such tool, RIPPLELAB, although user-friendly, falls short in terms of computational speed and classifying subsequent events, including detecting artifacts and HFOs with spikes through integrating deep learning-based algorithms. To tackle these limitations, we have developed PyHFO, a highly efficient, open-source, end-to-end application specifically designed for the streamlined analysis of HFOs.
Methods: Our application integrated two important processes of HFO analysis -- HFO detection, and Deep Learning (DL)-based HFO classification, which are executed swiftly via multiprocessing (
Figure 1). It accepts mainstream EEG data formats like EDF. In HFO detection, we implemented two automatic detectors, Short Time Energy (STE) and MNI, replicating RIPPLELAB's parameters and computation. Our design enhances detection speed by running the calculations for each channel on separate CPU cores, effectively capitalizing on computer hardware. The DL-based HFO classification utilized an annotated training dataset from 19 patients with medication-resistant focal epilepsy who underwent intracranial monitoring via strip/grid electrodes. The dataset consisted of a 10-minute EEG segment from each subject, 1,709 channels in total. We employed data augmentation to improve our DL model's generalizability. Lastly, we consolidated these capabilities into an accessible, multi-window GUI-based computer software designed in PyQt (
Figure 2).
Results: PyHFO, the HFO detection algorithm, reliably duplicated the detection algorithms of RIPPLELAB, with minor variations in data reading and filtering. Disparities between the two platforms were capped at 10% for the STE detector (out of 12,494 STE HFOs identified by RIPPLELAB's STE) and 14% for the MNI detector (out of 10,392 detected by RIPPLELAB's MNI). PyHFO outperformed RIPPLELAB in processing speed when analyzing HFOs, demonstrating this on both single-core and multi-core configurations. Particularly in multi-core utilization, our application demonstrated a nearly tenfold acceleration for both STE and MNI detection. The Deep Learning (DL)-based HFO classification achieved 98.6% and 89.1% accuracy in classifying artifacts and HFOs with spikes, respectively, in a 5-fold cross-validation compared to expert validation.
Conclusions:
Our efficient and user-friendly open-source application possesses significant potential for utilization in epilepsy research, particularly given its capabilities for deep learning-based classifications. This accessible tool could also pave the way for the integration of HFO analysis in clinical trials and everyday clinical practice.
Funding:
The National Institute of Health, K23NS128318