Although the use of ambulatory EEG (aEEG) to detect seizures in the outpatient setting is associated with greater convenience and cost savings compared with inpatient video-EEG monitoring, it can be burdensome for clinicians to review prolonged recordings in their entirety. Automatic seizure detectors (ASDs), which are currently available in several commercial formats, can help reduce this burden. “ASDs can facilitate the review of long-term EEG recordings to detect potential epileptiform activity and assist in artifact suppression,” Stephan U Schuele, MD, MPH, chief of neurophysiology and epilepsy at Northwestern University Feinberg School of Medicine in Chicago, Illinois, told Neurology Advisor.

Thus far, these tools have demonstrated substantial variability in their yield of seizure detection, and aEEG services may provide “only software preselected clipped data to their reading physicians unless raw tracings are specifically requested,” according to a new paper published in Neurology by Dr Schuele and colleagues.1 They conducted a single-center retrospective study to investigate the effectiveness of such clipped data in detecting “at least one seizure per study, therefore bringing to the reader’s attention a potential ‘high-risk’ study or whether it is still needed to read the continuous data.”

Related Articles

The raw tracings from 1257 prolonged aEEG studies (>24 hours) were reviewed in their entirety and processed using 1 of 3 commercially available ASD programs. For the raw data readings, the reviewer was a fellowship-trained neurophysiologist, who confirmed detections of ictal events with a second neurophysiologist. When both readers marked the same time period as an ictal event, these were considered expert-detected seizures; in cases of disagreement between the neurophysiologists, a consensus was reached after further review of the events.

While electrographic seizures were identified by expert review in 5.6% of studies, the ASDs correctly identified at least 1 seizure in only 53% of these cases, with no significant differences in detection rates between the 3 programs. There were slightly more frequent detections with ASDs in patients with generalized vs focal seizures (70% vs 46%; P =.06). “The study shows that ASDs have not only a high false positive rate — which has been widely recognized before — but also limited sensitivity,” said Dr Schuele. Thus, they “are not a substitute for a trained technologist reader or qualified physician to review the raw EEG.”

He believes that ASD software will advance to the level of the trained human reader, but even then, we “need to continue to validate those advancements in the actual clinical setting compared to the gold standard of a qualified human reader, rather than blindly relying on the convenience and promises of computer assisted readings.”

In a related editorial published in the same issue of Neurology, Juhász and Berg wrote that these results “demonstrate the disappointing reality that automated detection accuracy is inferior on aEEG compared to expert human analysis, at least for the software algorithms included in this study.”2 They also noted that the 53% detection rate of the ASD on aEEGs is significantly inferior to that of ASD data obtained from long-term inpatient studies, which are associated with nearly 90% sensitivity.3

Other findings from the new study 1 offer some clues regarding ways in which the accuracy of ASD in aEEG could potentially be improved in the future. The addition of a pushbutton event marker or patient log sheet provided an incremental yield of 19% over seizures identified by ASDs, although 61% of patients who experienced a seizure did not report these events; this observation aligns with previous results indicating substantial seizure underreporting by patients.4

In addition, in a subgroup of 32 studies that were prescreened by a registered technologist before review by neurophysiologists, 22% more studies with a seizure were identified compared with studies identified by the ASD and patient (84.5% vs 62.5%, respectively; P =.047).1 However, Juhász and Berg note that the prescreening approach only shifts the human burden rather than reducing it, and the expertise of technologists may be variable.2

“Recently emerging advances in multimodal recording and advanced big data analytic approaches offer a glimmer of hope,” wrote Juhász and Berg.2 For example, the integration of scalp EEG data with other physiologic measures on wearable smart devices could lead to improved seizure detection with aEEG, and the further development of deep learning paradigms may help increase accuracy to 90% or greater, as observed with inpatient EEG recordings. “However, the true clinical value of these emerging technologies remains to be rigorously tested for massive aEEG recordings,” they stated. 2“Until then, the careful look of the trained human reader, however imperfect, will remain indispensable for accurate detection of seizures buried in the mass of aEEG data.”

References

  1. González Otárula KA, Mikhaeil-Demo Y, Bachman EM, Balaguera P, Schuele S. Automated seizure detection accuracy for ambulatory EEG recordings. Neurology. 2019;92(14):e1540-e1546.
  2. Juhász C, Berg M. Computerized seizure detection on ambulatory EEG: Finding the needles in the haystack. Neurology. 2019;92(14):641-642.
  3. Ulate-Campos A, Coughlin F, Gaínza-Lein M, Sánchez Fernández I, Pearl PL, Loddenkemper T. Automated seizure detection systems and their effectiveness for each type of seizure. Seizure. 2016;40:88-101.
  4. Elger CE, Hoppe C. Diagnostic challenges in epilepsy: seizure under-reporting and seizure detection. Lancet Neurol. 2018;17(3):279-288.

This article originally appeared on Neurology Advisor