Abstracts

Rationale: Despite considerable research in the field of automated EEG interpretation, the review process remains a largely manual endeavor. For EMU and ICU recordings, the burden on physician and technician time is significant. Fully automated seizure detection systems are not yet in widespread use due to a lack of adequate sensitivity and specificity. We propose that automated systems with human-in-loop customization and navigation could be a step towards bridging the gap. We demonstrate this with a tool that lets a reviewer search for specific patterns in an EEG. This method combines the strengths of human expertise with the strength of machine learning to boost efficiency. We implemented this using an EEG viewer developed in-house, evaluated on the Temple University Hospital (TUH) Seizure Dataset, and open-sourced the tool/viewer for the community.  Methods: Since a patient's seizures often appear similar, we sought to detect future seizures given past patterns in EEG recordings1. For quantifying similarity, we used a simple cross-correlation based methodology2, available in OpenCV as the matchTemplate API3. This process takes a template EEG and computes the normalized cross correlation over the whole EEG to compute a score for each timestep. We then present the results ranked by how closely they match the template. We used the public TUH Seizure dataset4 to test the performance of the seizure similarity algorithm. The algorithm was run on an EEG viewer developed in-house and open-sourced5. For this pilot study, we randomly selected 54 patient EEG segments with labeled seizures; only one EEG epoch was used per patient. All studies were reviewed to ensure accuracy of the original interpretation; if there were discrepancies with the original labeling, studies were removed. Additionally, if the EEG segment was status epilepticus, it was removed. Segments with 2 or fewer seizures, or with greater than 15 were also removed. After exclusions, 26 separate patient files were evaluated. We also did a separate evaluation in which we excluded files with a persistently abnormal background or multiple seizure types. Our epileptologist used the tool to pick the first labeled seizure and create a template. The template defined the set of electrode pairs most involved in the electrographic pattern and the time constraints of the seizure. The template was then used as input to the similarity algorithm, which returned the resulting matches in a ranked fashion. We evaluates up to 10 results or until all seizures in the file were identified correctly.  Results: The total number of seizures was 156, with a mean number of seizures per patient file of 6 (SD 3). The mean EEG file duration was 20.46 minutes (SD 9.88 minutes). The mean length of each seizure was 51.73 seconds (SD 38.98 seconds). Overall, the accuracy of the tool (percentage of matches that were true seizures) was 74.5%. If cases with a persistently abnormal background or multiple seizure types were removed, the accuracy improved to 81.9%. The tool performed perfectly (100% accuracy) in 46% of cases. Conclusions: This pilot study demonstrates the feasibility of human-in-loop seizure detection using an EEG similarity algorithm. The algorithm performed well on files with normal background patterns, low artifact burden, and a well-defined seizure onset. EEGs with a persistently abnormal background, heavy artifact, or multiple seizure types resulted in poorer performance. Existing seizure monitoring systems could add EEG similarity measures to incorporate expert feedback and adjust future predictions in real time.  Funding: No funding