Epilepsy is a chronic neurological disorder characterized by seizures. Although most patients respond favorably to medications, some patients continue having seizures and require surgery or alternative treatments. Recently, high frequency oscillations (HFOs) have been proposed as a biomarker of epileptic tissue providing a seizure onset zone (SOZ) localization and relate to surgery outcomes.
Visual HFO identification, a gold standard HFO marking, has limitations, such as, subjective, and time consuming; therefore, automatic detection algorithms have been developed. However, the automatic detections suffer from complex optimization and specific to recording. We present an algorithm with amplitude threshold as single parameter that requires optimization tuned by an iterative procedure. Algorithm is used to study HFOs in intracranial (iEEG) and scalp EEG. In the iEEG, our detector achieved 99.6% sensitivity with 1.1% false positive rate (FPR), and 37.3% false detection rate. Furthermore, the algorithm was used to detect HFOs in scalp EEG. Of the marked candidate events, 40% and 60% were visually confirmed to be ripples and fast ripples by three reviewers.
As all HFO study rely on an empirical, derived from visual observation, rather than physiological definition, we introduce the anomaly HFO detection algorithm (ADA). The algorithm integrates machine learning techniques, including anomaly detection, pattern matching, and clustering and classification to identify anomalous patterns in high frequency signals without prior assumption of the shape, amplitude, or duration. The events detected by ADA are the different population to the conventional HFOs. The amplitude of detected events is a superior candidate as a SOZ biomarker with area under the receiver operation characteristic curve (AUC), sensitivity and FPR at 0.959, 93.6% and 5.6% when comparing to the rate of conventional HFO, which was exclusively used as a biomarker, (AUC:0.912, sensitivity:86.0% and FPR:13.3%). Moreover, the amplitude is more robust to the additional events, and stable across recording segments.
We believe ADA and simple detection algorithm will be powerful tools for the assessment and localization of epileptic activity providing unbiased estimation of HFO properties. Furthermore, the amplitude of HFO can become a superior candidate as the SOZ biomarker for epilepsy patients comparing to the rate of HFO.
