4. Principal methodologies to study and analyze ripples as a biomarker of memory consolidation
Studies of oscillations with frequencies above traditional electroencephalogram limits (greater than 40 Hz), have increased over the last decade due to increasing evidence that suggests that HFOs reflect a mechanism of epileptic phenomena and might serve as a biomarker of epileptogenesis and epileptogenicity. This has resulted in a growing interest in the detection and analysis of these events (Birot et al., 2013; Navarrete et al., 2016). The bandwidth of SWRs (120-250 Hz) (Buzsáki, 2015) falls between the one of HFOs (80-500 Hz) (Burnos et al., 2014), therefore, the methodologies for automatic identification of HFOs might be adapted to identify SWRs. In recent decades, there has been a substantial surge in the field of artificial intelligence, which has been extensively explored for enhancing the classification of HFOs. Although our subsequent discussion primarily focuses on conventional methodologies that do not utilize machine or deep learning, it is worth noting that readers may find the works of Wong et al., 2021 and Navas-Olive et al., 2020 intriguing in the context of HFO classification.
In general, HFO detection methods are classified in three main groups: manual review, supervised detection, and unsupervised detection. Manual review is a visual inspection performed by an expert. It is a time-consuming process and highly subjective to the perception of the reviewer; however, it is currently considered the gold standard when assessing the performance of automated algorithms for HFO detection. Supervised detection consists of methods with high sensitivity and low specificity detection, which is later reviewed manually to eliminate false positive events. On the other hand, unsupervised detection methods must have high sensitivity and high specificity, which is difficult to achieve (Birot et al., 2013; Navarrete et al., 2016).
In the last two decades, several methods for the automatic detection of HFOs have been developed. Broadly speaking, the supervised and unsupervised detection algorithms perform a series of common steps to identify the putative events. The first step is to emphasize the frequency of interest by filtering the raw signal. Then, a threshold handling detection method is implemented. This can be based on the energy, statistical or spectral characteristics of the filtered data. Finally, depending on the method, a supervised or unsupervised mechanism is implemented to distinguish the HFO from noise, artifacts, and spikes (Navarrete et al., 2016).
Staba et al. (2002) reported the first supervised method for the automatic detection of HFOs, which uses the root mean square (RMS) to calculate the threshold after applying a band-pass filter. The authors declared a sensibility of 84%. Later, other authors used the algorithm proposed by Staba et al. as the initial steps for theirs. For example, Burnos et al. (2014) decreased the number of standard deviations (SDs) established to classify the oscillation as an event of interest (EoI). Then, they utilized the instantaneous power spectra of the Fourier Transform representation to eliminate some false positive events. This methodology had a higher sensibility in 4 out of 5 accepted patients when compared to the methodology proposed by Staba et al (Burnos et al., 2014). Other authors that used the algorithm of Staba et al. as the initial steps of theirs include Crèpon et al. (2010), Ellenrieder et al. (2012), and Blanco et al. (2010).
Charupanit and Lopour (2017) reported a simple statistical detection method. The iterative algorithm uses an estimate of the amplitude probability distribution of the background activity to calculate the optimum threshold for identification of HFO. It described a sensitivity of 99.6%. Additionally, Shimamoto et al. (2018) computed an algorithm that uses independent component analysis to detect ripples. The events detected were further classified as true or false ripples by implementing a topographical analysis to the time-frequency plots. They declared a precision of over 91% and a sensibility of over 79%.
Another supervised algorithm to detect HFOs was implemented by Birot et al. (2013) . This algorithm uses the RMS amplitudes and a Fourier or Wavelet energy ratio to detect putative events. They implemented this algorithm in both human and animal datasets and had an area under the curve greater than 0.95. Another author that used wavelet entropy was Zelmann et al. (2010) with a reported sensitivity and specificity of 96%. Finally, Liu et al. (2021) computed in 2021 an algorithm that combines several reported features like Short-time Energy and Hilbert Transform with visual extracted features from the frequency spectrum. They also evaluated this algorithm with datasets from both patients and animal models and reported a sensibility greater than 91%.
However, despite the increasing methodologies for effective automatic identification of HFOs, there is currently no gold standard other than visual examination. Therefore, we suggest that a global consensus of the identification of true ripple activity and its analysis must be reached to evaluate more accurately and compare properly under different experimental conditions. It has been proposed as a solution to implement different algorithms across similar datasets to reach a consensus about the one method that performs best in all contexts (Navarrete et al., 2016).