SGM: a novel time-frequency algorithm based on unsupervised learning improves high-frequency oscillation detection in epilepsy

Abstract

Objective. We propose a novel automated method called the S-Transform Gaussian Mixture detection algorithm (SGM) to detect high-frequency oscillations (HFO) combining the strengths of different families of previously published detectors. Approach. This algorithm does not depend on parameter tuning on a subject (or database) basis, uses time-frequency characteristics, and relies on non-supervised classification to determine if the events standing out from the baseline activity are HFO or not. SGM consists of three steps: the first stage computes the signal baseline using the entropy of the autocorrelation; the second uses the S-Transform to obtain several time-frequency features (area, entropy, and time and frequency widths); and in the third stage Gaussian mixture models cluster time-frequency features to decide if events correspond to HFO-like activity. To validate the SGM algorithm we tested its performance in simulated and real environments. Main results. We assessed the algorithm on a publicly available simulated stereoelectroencephalographic (SEEG) database with varying signal-to-noise ratios (SNR), obtaining very good results for medium and high SNR signals. We further tested the SGM algorithm on real signals from patients with focal epilepsy, in which HFO detection was performed visually by experts, yielding a high agreement between experts and SGM. Significance. The SGM algorithm displayed proper performance in simulated and real environments and therefore can be used for non-supervised detection of HFO. This non-supervised algorithm does not require previous labelling by experts or parameter adjustment depending on the subject or database considered. SGM is not a computationally intensive algorithm, making it suitable to detect and characterize HFO in long-term SEEG recordings.

Detection of high-frequency oscillations in electroencephalography: A scoping review and an adaptable open-source framework

“…16 studies incorporated time-frequency analysis into their HFO detection algorithms [10,14,[35][36][37][38][39][40][41][42][43][44][45][46][47][48]. 12 studies leveraged iEEG, 2 studies leveraged both iEEG and MEG, and 2 studies used scalp EEG.…”

Improving the ripple classification in focal pediatric epilepsy: identifying pathological high-frequency oscillations by Gaussian mixture model clustering

“…The SGM algorithm was used to detect HFOs in the ripple band (80-200 Hz). SGM is already published and was validated using simulated and real data assessed by two experts [28]. Briefly, the algorithm consists of: (a) baseline estimation using the entropy of the autocorrelation; (b) S-Transform [29] to calculate time-frequency features; and (c) GMM-based clustering of the features to decide if events are HFO-like activity.…”

Choosing Strategies to Deal with Artifactual EEG Data in Children with Cognitive Impairment

“…Two classification methods were used to analyze the group distribution of RTT patients: The Kernel Density Estimation (KDE), which represents the data using a continuous two-dimensional probability density curve that is analogous to a histogram and the Gaussian Mixture Model (GMM), which depicts the density representation as the weighted sum of Gaussian distributions. The GMM algorithm was applied to the dataset for fitting three mixture-of-Gaussian models and to assign each record to the Gaussian model it mostly belongs to [20]. As it is a probabilistic model, we filtered by probabilities to keep only those records with a probability (p) greater than 95% of belonging to its group, obtaining clearly differentiated data.…”

“…The most common method to remove artifacts from the EEG has been to rule out the activity that exceeds a certain threshold (usually ± 150 µV). An alternative to this method is based on the same idea of using a threshold but calculated from the data distribution, with the mean and standard deviation (SD) of each EEG channel [20]. An adaptive threshold is obtained using a k-factor: mean + k-factor × SD.…”