Ear-Seizure Detection (EarSD) Study

Recordings of Bioelectrical signal of subjects with the wearable device and simultaneous continuous EEG data is collected for the duration of hospitalization of participants. Outcome measures reported include number of seizure events per participant.

Recordings of Bioelectrical signal of subjects with the wearable device and simultaneous continuous EEG data is collected for the duration of hospitalization of participants. Outcome measures reported include average duration of each seizure in minutes and seconds and total recording time in hours aggregated to arrive at one reported value seizure classification.

Recordings of Bioelectrical signal of subjects with the wearable device and simultaneous continuous EEG data is collected for the duration of hospitalization of participants. Outcome measures reported include reported value seizure classification. Seizure classification includes Unclassified (UC), Focal Onset Aware (FOA), Focal Onset Impaired (FOIA), Focal to Bilateral Tonic-Clonic (FBTC).

EarSD extracted EEG signals from the log file plotted alongside EDF files from cEEG are measured and compared to detect seizure onset and offset times for data interpretation. Two-minute segments of cEEG European Data Format (EDF) consisting of non-seizure signals from periods before and after the seizures, and non-seizure signals from periods of daily activities like talking, eating, and walking are involved in the comparison to detect seizure onset and offset times. Prediction measurement of Seizure Sensitivity (SS) and False Positivity Rate per hour (FPR/h) are measured from the recorded data signals. Seizure Sensitivity (SS) is the ratio between the (number of predicted seizures)/(total number of seizures) = (number of true alarms)/(total number of seizures). FPR/h is the number of alarms that do not correspond to seizures raised in one hour. FPR/h = ((Number of false alarms/Interictal Duration) - (Number of False Alarms × Refractory period)).

EarSD recordings from each electrode are separated and filtered to eliminate noise and artifact and results in 12 output signals (6 signals/ear) for comparison against cEEG EDF files for accuracy and precision. Mean, standard and average deviation, skewness, kurtosis, lowest and highest value, and the root mean square amplitude are measured from the dataset and are normalized between 0 and 1 then passed into the seizure detection and prediction Machine Learning (ML) model. ML model consisting of algorithms using deep neural networks (DNN), recurrent neural networks (RNNs) and Long Short-Term Memory networks (LSTM), classifies whether the signals are a seizure signal vs non-seizure signal, the focal type (left side/right side) and predicts the accuracy of seizures a minute ahead with the goal of achieving 96 percent or better accuracy and reducing the number of false positives.

At the end of the study, patients' experience and perception of the EarSD device are collected using a paper-based 7-question survey measured on a 5-point Likert scale ranging from Strongly Disagree to Strongly Agree. A maximum total point score of 35 represents a better reported satisfactory score from participants and having a good experience with the device and its comfortability for daily activities. The survey is a self-administered report, and participants will be asked about the comfortability and perceived utility of the device.