Authors :
Presenting Author: Patrick Myers, MSE – Johns Hopkins University
Kristin Gunnarsdóttir, PhD – Biomedical Engineering – Johns Hopkins University; Adam Li, PhD – Columbia University; Vladislav Razskazovskiy, BS – University of Pittsburgh Medical Center; Dale Wyeth, REEGT – Thomas Jefferson University Hospital; Edmund Wyeth, REEGT – Thomas Jefferson University Hospital; Alana Tillery, BS – Johns Hopkins University; Jennifer Hopp, MD – University of Maryland Medical Center; Babitha Haridas, M.B.B.S. – Johns Hopkins Hospital; Jorge Gonzalez-Martinez, MD, PhD – University of Pittsburgh Medical Center; Anto Bagić, MD, PhD, FAES, FACNS – University of Pittsburgh Medical Center; Joon-Yi Kang, MD – Johns Hopkins Hospital; Michael Sperling, MD – Thomas Jefferson University Hospital; Nirav Barot, MD – Beth Israel Deaconess Medical Center; Sridevi Sarma, PhD – Johns Hopkins University; Khalil Husari, MD – Johns Hopkins Hospital
Rationale:
Scalp EEG is standard of care for evaluating potential seizures for the possibility of epilepsy. Clinicians search for multiple signature epileptic events, such as interictal epileptiform discharges (IEDs), in the EEG data. However, IEDs are an unreliable marker as they are only detectable in 29-55% of scalp EEG records.
The misdiagnosis rate of epilepsy remains high at nearly 30%. A reliable marker for epilepsy derived from scalp EEG would be ideal since it is a relatively simple and cheap neurophysiological test.
Methods:
Routine scalp EEGs with no detectible epileptiform abnormalities from a multi-center (Johns Hopkins Hospital, University of Pittsburgh Medical Center, Thomas Jefferson University Hospital, and University of Maryland Medical Center), retrospective study of 203 patients were collected. All time segments where the patient was asleep or drowsy were marked. Information about the patient’s age, sex, and medications were collected.
For each scalp EEG record, a dynamic network model (DNM) that estimates the interactions of different brain regions over time was generated. From the DNM, two metrics were calculated. Fragility analyzes the stability of the network by applying simulated perturbations to try to find the most “fragile” regions. Source-sink searches for unique patterns in the connection strength between functional neighbor regions. Frequency based features - power in delta, theta, alpha, and beta bands – were also calculated. All features were calculated using a sliding window per channel to capture the dynamics over time. A greedy search was implemented to find the best set of features to detect epilepsy.
Results:
A logistic regression model was generated, and a threshold was applied to classify each EEG record as epileptic or not. The model’s receiver operating characteristic curve had an area under the curve of 0.85±0.08. When a threshold that optimizes accuracy was applied, the model produced an accuracy of 0.78, sensitivity of 0.86, and specificity of 0.72 as shown in Figure 1. These patients were then stratified by the metadata described above and the results were replotted in Figure 2.
Conclusions:
Since only records containing no detectable epileptiform activity was included in the study, we are confident that our model provides new information that could be informative for diagnostic purposes. The model was mostly agnostic to the patient’s sex, age, the number of anti-seizure medications the patient was taking during the time of the recording, and the percentage of the recording the patient was asleep.
Funding:
Alvin and Fanny B. Thalheimer Foundation; National Science Foundation SBIR Phase I
Translational Research