Recently two independent research groups published influential studies suggesting an extraordinary role of VHFOs (500–2000 Hz) in human SEEG recordings for more precise delineation of the epileptogenic zone. These oscillations seem to be preferentially observed within mesiotemporal regions and were promisingly described as more specific biomarkers than ripples and FRs. [13, 14] Similarly to red and green spikes, where the red spike zones are recognized as zones generating seizures and the green spike zones are recognized as zones generating only sporadic, isolated, spatially confined discharges [15], the examined channels could be divided into two groups of VHFO and HFO channels based on the VHFO information. The division could possibly resolve which HFOs are physiological, seemingly the HFO channels, and which HFOs are not physiological - the VHFO channels. However, high-quality, high-resolution, and low-noise recordings are critical for the identification of VHFOs. These requirements usually cannot be fulfilled in traditional clinical settings and, therefore, practical use of VHFO detection in clinical practice can be hardly expected in the near future. Thus, a question arose if the presence of VHFO could be estimated from standard SEEG recordings (sampled at 1 kHz) utilizing any other specific feature/s.
Our study investigated 1 kHz SEEG recordings, previously manually assessed as with/without VHFOs, for shadows of VHFOs in lower frequencies, focusing on the possibility of determining their presence or absence.
This study included 141 SEEG recordings, from a cohort of 13 patients, undersampled from 5 kHz to 1 kHz with a focus on VHFOs. 126 out of the total 141 (89%) of the contacts with present VFRs and/or UFRs were contacts that were also resected in the surgery. Contacts that were not resected were found in four patients: patient 15 contact C’2, patient 26 contacts B1-3, C1-2, patient 71 contacts B1-5 and C’1–2, and patient 82 contacts C1-2 (Table 1). Even if the contacts were not resected they were still used in the analysis and are included in the results.
Our results suggest that based on the calculated features it is feasible to differentiate between HFO and VHFO channels even in lower sampling frequencies. The calculated features, Maximum amplitude, Shannon entropy, and TKEO were calculated in the time domain of a signal while the 75th percentile was calculated from power spectral density estimates proposing that VHFOs to some extent directly affect the power spectrum below 450 Hz of the recorded signal. When the recordings were filtered into seven distinctive frequency bands, bands above theta also show significant differences in Amplitude Maximum, 75th percentile, and TKEO features, suggesting that VHFO activity affects the signal recording in certain frequency bands. On the contrary, Shannon Entropy is non-significant in most of the seven respective ranges, providing significant values in theta, alpha, and ripple bands. Shannon Entropy is expressing a value unrelated to the remaining features, describing the information and variability of the signal. Filtered EEG signal becomes more deterministic and causes the entropy to lower even in HFO channels, which were in the previous analysis proven to have significantly higher Shannon Entropy than VHFO channels, causing the medians of these two groups to overlap after the filtration into narrow frequency ranges.
Alternatively, the significant differences could be justified by the presence of HFO - ripples and fast ripples. The analysis of HFO rates and the non-significant p values validates that the HFO rates detected on the original 5 kHz are sufficient to differentiate between HFO and VHFO channels, nonetheless, ripple rates detected on 1 kHz data are not sufficient. Accuracies of the three trained models confirm that the proposed features bring new information compared to rates of HFOs only. The model trained on the calculated features and HFO rates has the best accuracy and F1 score, while the model trained in HFO rates has lower accuracy and F1 score than the model trained on feature values. The p-value of the Wilcoxon rank-sum test after Bonferroni correction between accuracies of proposed models describes significant differences between accuracies of Model 1 and 2, and Model 1 and 3, and no significant differences were confirmed between accuracies of Model 2 and 3. This is also confirmed by analyzing a model trained on ripple rates detected on the downsampled 1kHz data, similarly to the calculated features being calculated on 1 kHz data.
To conclude, shadows of VHFOs are present in lower frequencies of the SEEG recordings. The results suggest that it is possible to distinguish SEEG signals with or without the presence of VHFOs, using recordings with the sampling frequency of 1 kHz. The differences between the accuracies and F1 scores also suggest that our features provide additional information to traditional ripple and fast ripple rates. Introducing a computation of Amplitude Maximum, 25th Percentile, Shannon Entropy, and Teager-Kaiser energy operator into the analyses of SEEG recordings with standard sampling frequency improves our ability to differentiate HFO and VHFO channels and hence grossly improves epilepsy surgery outcomes.