Discovering EEG biomarkers of Lennox-Gastaut syndrome through unsupervised time-frequency analysis.

OBJECTIVE

The discovery and validation of electroencephalography (EEG) biomarkers often rely on visual identification of waveforms. However, bias toward visually striking events restricts the search space for new biomarkers, and low interrater reliability can limit rigorous validation. We present a data-driven approach to biomarker discovery called scalp EEG Pattern Identification and Categorization (s-EPIC), which enables automated, unsupervised identification of EEG waveforms. S-EPIC is validated on Lennox-Gastaut syndrome (LGS), an epilepsy that is difficult to diagnose and assess due to its variable presentation and insidious evolution of symptoms.

METHODS

We retrospectively collected 10-min scalp EEG clips during non-rapid eye movement (NREM) sleep from 20 subjects with LGS and 20 approximately age-matched healthy controls. For s-EPIC, EEG events of interest (EOIs) were detected in all subjects using time-frequency analysis. The 11 705 EOIs were characterized based on 11 features and were collectively grouped using both k-means clustering and feature categorization. To provide clinical context, 1350 EOIs were visually reviewed and classified by three epileptologists.

RESULTS

s-EPIC identified four clusters as candidate biomarkers of LGS, each having significantly more LGS EOIs than control EOIs. Two clusters contained EOIs resembling known LGS biomarkers such as interictal epileptiform discharges and generalized paroxysmal fast activity. The other two LGS-associated EEG clusters contained short bursts of power in beta and gamma frequency bands that were primarily unrecognized by epileptologists. This approach also uncovered significant differences in sleep spindles between LGS and control cohorts.

SIGNIFICANCE

s-EPIC provides a quantitative approach to waveform identification that could be broadly applied to EEG from both healthy subjects and those with suspected pathology. s-EPIC can objectively identify and characterize relevant EEG waveforms without visual review or assumptions about the waveform's morphology and could therefore be a powerful tool for the discovery and refinement of EEG biomarkers.