EEG biomarkers for the diagnosis and treatment of infantile spasms

Abstract

Early diagnosis and treatment are critical for young children with infantile spasms (IS), as this maximizes the possibility of the best possible child-specific outcome. However, there are major barriers to achieving this, including high rates of misdiagnosis or failure to recognize the seizures, medication failure, and relapse. There are currently no validated tools to aid clinicians in assessing objective diagnostic criteria, predicting or measuring medication response, or predicting the likelihood of relapse. However, the pivotal role of EEG in the clinical management of IS has prompted many recent studies of potential EEG biomarkers of the disease. These include both visual EEG biomarkers based on human visual interpretation of the EEG and computational EEG biomarkers in which computers calculate quantitative features of the EEG. Here, we review the literature on both types of biomarkers, organized based on the application (diagnosis, treatment response, prediction, etc.). Visual biomarkers include the assessment of hypsarrhythmia, epileptiform discharges, fast oscillations, and the Burden of AmplitudeS and Epileptiform Discharges (BASED) score. Computational markers include EEG amplitude and power spectrum, entropy, functional connectivity, high frequency oscillations (HFOs), long-range temporal correlations, and phase-amplitude coupling. We also introduce each of the computational measures and provide representative examples. Finally, we highlight remaining gaps in the literature, describe practical guidelines for future biomarker discovery and validation studies, and discuss remaining roadblocks to clinical implementation, with the goal of facilitating future work in this critical area.

Keywords: Detrended Fluctuation Analysis (DFA); entropy; epilepsy; functional connectivity; high frequency oscillation (HFO); hypsarrhythmia; long-range temporal correlation (LRTC); phase amplitude coupling.

Figure 1

Clinical application of the 2021 Burden of AmplitudeS and Epileptiform Discharges (BASED) score. A 4-month-old infant born at 39 weeks gestational age, with known developmental delay due to hypoxic ischemic encephalopathy, presented with new onset infantile spasms (IS). Onset of IS to treatment initiation was 1 day. (A) Day 0 (diagnosis and treatment), just <50% of 1 s (s) bins within a 5-min (m) epoch of sleep included at least one spike and there were grouped multifocal spikes (GMFS, arrows) as well as paroxysmal voltage attenuations (PVA, arrowheads). The presence of either GMFS or PVA indicated a BASED score of 4 (EEG findings to suggest a probable epileptic encephalopathy [EE]). (B) Day 14, IS and PVA resolved with high-dose prednisolone, with a subjective reduction in the burden of spikes. However, because of persistent GMFS (now less well-formed), the BASED score remained 4. At this point, there was no electrographic remission (pretreatment scores of 4 or 5 must improve to a 3 or less for remission). With the resolution of IS and subjective improvement on the EEG, the treating clinician did not pursue additional treatment at that time (despite the lack of electrographic remission by BASED score criteria). (C) Day 36, still no IS but the EEG showed a higher burden of spikes, now with >50% of 1-s bins having at least one spike in a 5-m epoch of sleep, indicating a BASED score of 5 (EEG findings to suggest a definite EE). Better formed GMFS (arrows) were present. Notice that the background wave amplitude assessment is not reliable when the burden of spikes is ≥50%. High-dose adrenocorticotropic hormone (ACTH) was started at that time. (D) Day 51, all spikes resolved with ACTH. The background was slow and disorganized indicating a BASED score of 1 (Any definite non-epileptiform abnormality). This score, with the persistent remission of IS, indicated electro clinical remission. Longitudinal bipolar montage, Sensitivity 10 mv/mm, LFF 1 Hz, HFF 70 Hz, 15 s per page. D, days; NA, not applicable; Pred, high-dose prednisolone.

Figure 2

Examples of Shannon entropy (SE). (A) The SE of white noise is high, as all values occur with equal probability. (B) The SE of a sine wave is almost as high as white noise, despite the drastically different visual appearance. The SE is calculated based only on the signal's histogram, and both a sine wave and white noise have fairly flat histograms, with almost equal probability for all values. (C) Normally-distributed noise has lower SE, as it is more likely to have values near zero. (D) Normally-distributed noise with outliers has even lower SE, as the probability of having values near zero is increased, relative to the range of possible values. (E) Noise with a 1/f power spectrum, similar to EEG, generally has relatively high values of SE.

Figure 3

Linear vs. non-linear connectivity measures. (A) EEG signals recorded concurrently from two electrodes, F3 and Fz, are shown on the left. In the panel on the right, each data point represents one time point from the EEG signals, with the value of F3 on the horizontal axis, and the value of Fz on the vertical axis. In this case, the signals exhibit a linear relationship and thus have high values of both correlation and mutual information (MI). Recall that correlation is sensitive only to linear relationships, while MI is sensitive to both linear and non-linear interactions. (B) When the two EEG signals are unrelated, no trend can be seen in the right panel, and the values of correlation and MI are low. (C) When the relationship between the signals is non-linear, the MI is high, while the correlation is low. (D) A second example of a non-linear relationship, with medium correlation and MI.

Figure 4

Scalp HFO and artifact examples. (A) Scalp HFO in bipolar channel T6–O2 in a child with IS. (B) Sharp artifact in channel T3–T5. (C) Muscle artifact in Channel Fp1–F7. (D) DC shift artifact in Channel Fp1–F7. Sharp artifacts, muscle activity, and DC shifts can all cause false positive HFO detections. In each subfigure, 5 s of broadband EEG is shown (left), with the red portion of the signal indicating the detected event (also marked by a red asterisk). On the right is the broadband EEG, with the detected event in red (top), the bandpass filtered EEG in the ripple frequency band (middle), and the time-frequency decomposition for the same segment of EEG (bottom).

Figure 5

Detrended fluctuation analysis. (A) To start, the broadband EEG signal (top panel) is bandpass filtered (black line, middle panel), and the amplitude envelope is determined (red line). Then the cumulative sum of the amplitude envelope is calculated (bottom panel). The remaining calculations use this cumulative sum. (B) The cumulative sum is divided into windows of a fixed length; typically, these windows overlap by 50%, but no overlap is shown in this figure for clarity. Within each window, a linear trend is fit to the data (red line). Then, for each window, the linear trend is subtracted, and the standard deviation of the residual is calculated (gray dashed line). This process is repeated for windows of varying lengths. Here, we show examples of 5-s and 1-s windows. (C) For each window length, the mean standard deviation across all windows is calculated. This is plotted on a logarithmic scale vs. the window length. For EEG data, this will typically result in data points with a linear relationship. A linear trend line is fit to the data points on this graph; the slope of the line is the DFA exponent, and its intersection with the vertical axis is the DFA intercept.

Figure 6

Examples of phase-amplitude coupling. (A) A signal with high phase-amplitude coupling. (B) A signal with low phase-amplitude coupling. Each subfigure shows the original signal (top), the bandpass filtered signal in a low frequency band (2nd from top), the phase of the low frequency bandpass filtered signal (3rd from top), and the bandpass filtered signal in a high frequency band (bottom). The amplitude envelope of the high frequency signal is also shown (bottom, red). In subfigure (A) the bursts of high amplitude, high frequency activity always occur at the peak of the low frequency signal, represented by a phase of zero. This consistent relationship will lead to a high value of phase-amplitude coupling. In contrast, in subfigure (B), the bursts of high frequency activity occur randomly with respect to the low-frequency phase, which will result in a low value of phase-amplitude coupling.