Authors :
Presenting Author: Mitchell frankel, PhD – Epitel
Avi Kazen, MS – Epitel; Zoe Tosi, PhD – Epitel; Tyler Newton, PhD – Epitel; Lillian Voke, BS – Boston Children's Hospital; Jay Jeschke, BS – New York University Langone Medical Center; Michele Jackson, BS – Boston Children's Hospital; Michelle Sandoval, BS – University of Colorado Anschutz Medical Center; Trey Jouard, BS – University of Colorado Anschutz Medical Center; Lauren McCall, BS – University of Colorado Anschutz Medical Center; Chris Mizenko, BS – University of Colorado Anschutz Medical Center; Mackenzi Moore, BS – University of Colorado Anschutz Medical Center; Claire Ufongene, BS – Boston Children's Hospital; Kristal biesecker, BS – University of Colorado Anschutz Medical Center; Meagan Watson, BS – University of Colorado Anschutz Medical Center; Mark Spitz, MD – University of Colorado Anschutz Medical Center; Laura Strom, MD – University of Colorado Anschutz Medical Center; Daniel Friedman, MD, MS – New York University Langone Medical Center; Tobias Loddenkemper, MD – Boston Children's Hospital; Mark Lehmkuhle, PhD – Epitel
Rationale:
Epitel has developed a wireless, wearable EEG monitoring system (REMI) that is currently US FDA-cleared for use in healthcare settings for up to 48h. Epitel is pursuing use of REMI in ambulatory settings over prolonged periods. To support review of this extended-duration data, a novel algorithm, diverse across patients and seizure types, was designed as a clinical decision support system to detect and highlight regions of REMI EEG that are indicative of self-limiting electrographic seizures. Methods:
Data was collected from patients in U.S. Epilepsy Monitoring Units. Patients wore REMI wireless sensors at F7/8 and Tp9/10 electrode locations alongside standard-of-care 19+ wired video-EEG. The EEG was preprocessed to account for noise, artifacts, and cross-patient differences. The data was then windowed into seconds-long segments, and features were extracted in time, frequency, and complexity domains. Additional features using pooling and correlations were created. A subset of patients was held out, and the segmented-data features from the remaining patients were used to train a machine learning classifier. The output of the classifier, the probability that each data segment contains seizure activity, was then fed to a post-processor trained to stitch the segment probabilities into complete seizure events (e.g., discrete event start/stop times). The algorithm was optimized for high-sensitivity because it is more important to ensure potential seizures are highlighted for clinician review than reduce false positives. Algorithm performance was evaluated on the held-out data by determining the True Positive Rate (Sensitivity) and False Alarm Rate (FAR = False Positives per hour - FP/hr) of the discrete seizure events. Confidence intervals were determined using a Bias-Corrected and Accelerated method.
Results:
There were 234 patients in the overall data set, with 30 (13%) held-out for testing. Ages ranged from 2 to 73 (median: 23) in the overall data set and 6 to 61 (median: 21) in the held-out set. Nine of the 30 held-out patients (30%) were noted as having no abnormal EEG characteristics and the 21 other held-out patients experienced a total of 60 electrographic seizures (range: 1-9, median: 2). The recording durations ranged from 4 to 201 hours with a median of 55 hours. The algorithm was 80% Sensitive across all seizures with a mean per-patient Sensitivity of 88% and a lower 95% confidence interval bound (CI) of 77%. The False Alarm Rate (FAR) was 0.16 FP/hr across all data with a mean per-patient FAR of 0.15 FP/hr and an upper 95% CI of 0.20 FP/hr. For patients who had noted seizures, the mean per-patient FAR was 0.16 FP/hr with an upper 95% CI of 0.24 FP/hr. For patients who had no noted EEG abnormalities, the mean per-patient FAR was 0.11 FP/hr with an upper 95% CI of 0.19 FP/hr. Five patients had no false detections, including two non-seizure patients.
Conclusions:
With low False Alarm Rates and high sensitivity, the discrete seizure detection algorithm demonstrates the potential to drastically reduce the time required for epileptologists to review extended-duration REMI EEG without sacrificing clinical accuracy.
Funding:
NIH 1U44 NS121562