Characterizing physiological high-frequency oscillations using deep learning - PubMed (original) (raw)

. 2022 Dec 7;19(6):10.1088/1741-2552/aca4fa.

doi: 10.1088/1741-2552/aca4fa.

Hoyoung Chung 1, Jacquline P Ngo 2, Tonmoy Monsoor 1, Shaun A Hussain 2, Joyce H Matsumoto 2, Patricia D Walshaw 3, Aria Fallah 4, Myung Shin Sim 5, Eishi Asano 6, Raman Sankar 2 7 8, Richard J Staba 7, Jerome Engel 7 9 10 11, William Speier 12 13, Vwani Roychowdhury 1, Hiroki Nariai 2 8

Affiliations

Characterizing physiological high-frequency oscillations using deep learning

Yipeng Zhang et al. J Neural Eng. 2022.

Abstract

_Objective._Intracranially-recorded interictal high-frequency oscillations (HFOs) have been proposed as a promising spatial biomarker of the epileptogenic zone. However, HFOs can also be recorded in the healthy brain regions, which complicates the interpretation of HFOs. The present study aimed to characterize salient features of physiological HFOs using deep learning (DL)._Approach._We studied children with neocortical epilepsy who underwent intracranial strip/grid evaluation. Time-series EEG data were transformed into DL training inputs. The eloquent cortex (EC) was defined by functional cortical mapping and used as a DL label. Morphological characteristics of HFOs obtained from EC (ecHFOs) were distilled and interpreted through a novel weakly supervised DL model._Main results._A total of 63 379 interictal intracranially-recorded HFOs from 18 children were analyzed. The ecHFOs had lower amplitude throughout the 80-500 Hz frequency band around the HFO onset and also had a lower signal amplitude in the low frequency band throughout a one-second time window than non-ecHFOs, resembling a bell-shaped template in the time-frequency map. A minority of ecHFOs were HFOs with spikes (22.9%). Such morphological characteristics were confirmed to influence DL model prediction via perturbation analyses. Using the resection ratio (removed HFOs/detected HFOs) of non-ecHFOs, the prediction of postoperative seizure outcomes improved compared to using uncorrected HFOs (area under the ROC curve of 0.82, increased from 0.76)._Significance._We characterized salient features of physiological HFOs using a DL algorithm. Our results suggested that this DL-based HFO classification, once trained, might help separate physiological from pathological HFOs, and efficiently guide surgical resection using HFOs.

Keywords: HFO; machine learning; physiological HFO.

© 2022 IOP Publishing Ltd.

PubMed Disclaimer

Conflict of interest statement

CONFLICTS OF INTEREST

The authors have no conflicts of interest to disclose.

Figures

Figure 1:

Figure 1:

Processing workflow (A): The detected HFO (by STE detector) is first filtered by the artifact detector that was developed in our previous study. The constructed image features along with the channel-wise clinical information, are fed into the convolutional neural network to train the model (ecHFO detector). In the real application of the model (inference), an unseen EEG signal is sent into the pipeline, and the STE detector is applied to detect all HFOs. Then we construct the image feature based on the detected HFO. The same artifacts detector is applied to reject the artifacts, and then the real HFO is filtered by the ecHFO detector. The detector will output a score from zero to one representing the confidence of each HFO being ecHFO. HFO feature representation and model architecture (B): We captured the time-frequency domain features as well as signal morphology information of the HFO window via three images. The time-frequency plot (scalogram) was generated by continuous Gabor Wavelets ranging from 10Hz to 500Hz. The EEG tracing plot was generated on a 2000 × 2000 image by scaling the time-series signal into the 0 to 2000 range to represent the EEG waveform’s morphology. We constructed another two images from the original EEG signal serving as the additional inputs to the network. The amplitude-coding plot was generated to represent the relative amplitude of the time-series signal: for every time point, the pixel intensity of a column of the image represented the signal’s raw value at that time. These three images were resized into the standard size (224 × 224), serving as the input to the neural network. We used ResNet 18 with a modified output layer for binary classification. The weights in the convolution layers are frozen and serve as feature extractors in the convolutional neural network.

Figure 2:

Figure 2:. Characteristics in the time-frequency plot of ecHFOs against non-ecHFOs.

(A) The time-frequency plot characteristics of the eloquent cortex and non-eloquent cortex HFOs for Pt 4, 5, 7, and 8. The yellow-colored regions in the figure stood for the pixels, where the power spectrum of ecHFOs is statistically lower than (p-value below 0.05 from the one-tailed t-test) non-ecHFOs. The figure showed one set of clearly interpretable distinguishing features between ecHFOs and non-ecHFOs: the ecHFOs generally have lower power at lower frequencies during the HFO event (center part along the time axis), Panel (B-Right) was generated by taking the average of the individual binary images from each of the 18 patients. It showed the distinguishing features are also significant at the population level. The feature can be assembled as a “Bell-Shape” if we take the region > 0.7 in the plot averaging all of the time-frequency plot characteristics (red color). (B-Left). (C) The model’s response to the Bell-shaped perturbation on the time-frequency plot. We provide two examples of perturbation for ecHFO events in Pt 3. Each row presents one example and the first column indicates the original time-frequency plot while the second indicates the perturbed time-frequency plot based on the Bell Shape perturbation. The prediction value of the model changed from 0.780 (therefore originally predicting it as non-ecHFO) to 0.299 (thus a change of 0.481), implying that the perturbed HFO would correspond to a non-ecHFO. (D) The change in model confidence in population level. Each column (along the y-axis) is a histogram of the percentage of change in confidence for one distinct patient. It shows the frequency distribution of confidence changes after adding the bell shape perturbation to the time-frequency plot to all classified ecHFOs for the given patient. The change in confidence level is significant, with an average of 0.532 with a 95% confidence interval [0.528, 0.535] noted as the solid red line in the histogram. HFO = high-frequency oscillation; ecHFOs = eloquent cortex HFOs; non-ecHFO = non-eloquent cortex HFOs.

Figure 3:

Figure 3:. The model’s responses to injecting a spike-like feature into the amplitude-coding plot.

Examples of introducing a downgoing (A) and upgoing (C) spike feature into classified non-ecHFO events. These demonstrate that on the introduction of a spike-like perturbation, the model predicts higher confidence towards non-ecHFOs (B, D). Subfigure (A) shows the amplitude encoding image before perturb, spike template, and the after-perturb (top row), and the corresponding time-series signal with downgoing spike perturbation (bottom row). Similarly, subfigure (C) shows the same information on a different classified ecHFO but with upgoing spike perturbation. For each patient, we compute a histogram for the distribution of the change in confidence (B) for downgoing spike perturbation. The same steps are repeated for upgoing spike perturbation, and the results are shown in (D). The percentage of change in confidence for both up-and-downgoing spike perturbation is significantly greater than zeros, with means downgoing: 0.268 (95% confidence interval [0.263, 0.272]) and upgoing: 0.320 (95% confidence interval [0.316, 0.324]), which are noted as solid red lines in each histogram. HFO = high-frequency oscillation; ecHFOs = eloquent cortex HFOs; non-ecHFO = non-eloquent cortex HFOs.

Figure 4:

Figure 4:. Traditional feature characterization of ecHFO and Non-ecHFO:

The normalized histogram of peak frequency (A), amplitude (B), and duration (C) of predicted ecHFO. (A) The ecHFO generally has a lower peak frequency than the non-ecHFO (p-value < 0.001). (B) The ecHFO generally has a smaller amplitude than the non-ecHFO (p-value < 0.001). (C) The ecHFO generally has the trend of having a longer HFO than the non-ecHFO (p-value = 0.07). However, there are no clear decision boundaries that can be drawn to clearly discriminate between ecHFO and non-ecHFO using each of these traditional features.

Figure 5.

Figure 5.. Inference of channel characteristics:

(A) The ratio of ecHFOs (eHFOs/total HFOs) in different types (physiological: behavior or gamma only; pathological: SZ/AD or spike only; both: both physiological and pathological) of channels from all patients (n=14) is plotted in box plots. The percentage of ecHFOs was higher in physiological channels than that in pathological channels. (B) The model confidence distribution of each individual eHFOs in channels with physiological, pathological, and both categories are shown. The distribution in both channels is closer to a uniform distribution.

Figure 6.

Figure 6.. The accuracy of prediction models incorporating HFO resection ratio.

We constructed postoperative seizure outcome prediction models using the HFO resection ratio derived from EEG data (n = 14). Each receiver-operating characteristics (ROC) curve delineates the accuracy of seizure outcome classification of a given model, using the area under the ROC curve statistics. (A) Unclassified HFO resection ratio was used as a single classifier. (B) non-ecHFO resection ratio was used as a single classifier. (C) non-ecHFO resection ratio was used as a single classifier. (D, E) A multiple regression model incorporating the resection ratio of HFOs (non-ecHFOs or spk-HFO) and complete removal of the SOZ (yes or no) was used, which demonstrated further improved predictive value of postoperative seizure outcomes. HFO = high-frequency oscillation; non-ecHFO = non-eloquent cortex HFOs; spk-HFO = HFO with spike; SOZ = Seizure onset zone.

References

    1. Bragin A, Engel J Jr., Wilson CL, Fried I, Buzsáki G. High-frequency oscillations in human brain. Hippocampus 1999;9:137–142. -PubMed
    1. Gotman J High frequency oscillations: the new EEG frontier? Epilepsia 2010;51 Suppl 1:63–65. -PMC -PubMed
    1. Bragin A, Engel J Jr., Wilson CL, Fried I, Mathern GW. Hippocampal and entorhinal cortex high-frequency oscillations (100−−500 Hz) in human epileptic brain and in kainic acid--treated rats with chronic seizures. Epilepsia 1999;40:127–137. -PubMed
    1. Staba RJ, Wilson CL, Bragin A, Jhung D, Fried I, Engel J Jr. High-frequency oscillations recorded in human medial temporal lobe during sleep. Annals of neurology 2004;56:108–115. -PubMed
    1. Jirsch JD, Urrestarazu E, LeVan P, Olivier A, Dubeau F, Gotman J. High-frequency oscillations during human focal seizures. Brain 2006;129:1593–1608. -PubMed

Publication types

MeSH terms

Grants and funding

LinkOut - more resources