Abstracts

Rationale:
An accurate record of a patient’s seizures is important to inform clinical management and may identify circadian or multidian seizure patterns that can produce seizure risk forecasts. Seizure diaries are typically kept manually by patients or caregivers, are burdensome to maintain, and have been shown to be unreliable. Non-invasive biosensors capable of acquiring signals that represent physiology and behavior are available and may be able to provide an objective record of seizures. However, improvement is needed in the accuracy of currently available seizure detection systems. The Empatica E4 wristband is a medical-grade wearable device that offers real-time physiological data acquisition, including accelerometry (ACC), photoplethysmography (PPG), electrodermal activity (EDA) and skin temperature (TEMP). We aim to develop and optimize a deep neural network seizure detection algorithm for non-invasive wearable devices.
Method:
We developed and benchmarked 3 long short-term memory recurrent neural network (LSTM-RNN) architectures to classify 10-second ictal and interictal segments recorded with Empatica E4s during scalp or invasive EEG recordings in 11 patients with recorded focal motor or tonic-clonic seizures. Architecture 1 – The algorithm was trained on 8 channels of recorded Empatica E4 data, including three axes of ACC data, ACC magnitude, blood volume pulse (BVP), EDA, TEMP, heart rate (HR), with recursive node inputs along the time axis (Figure 1). Architecture 2 – The algorithm was trained on the 8 channels of recorded Empatica E4 data with input transposed, with the recursive inputs between channels (Figure 1). Architecture 3 – The algorithm was trained on 16 channels, including 8 recorded channels, as well as, the fast Fourier transform (FFT) of ACC magnitude, BVP, EDA, TEMP, HR, and 3 signal quality indices (SQI) for ACC magnitude, BVP and EDA. The algorithms were trained using 18 seizures and 2100 minutes of interictal data, and noise-added copies of the ictal data segments were generated to compensate for the unbalanced ictal/interictal ratio. The architectures were trained on 8 subjects and tested on 3 unseen subjects in a 3-fold cross-validation approach.
Results:
The area under the ROC curve (AUC-ROC) values were calculated. For architecture 1, the average AUC-ROC was 0.932. For architecture 2, the average AUC-ROC was 0.834. Training time for architecture 1 was 2-3 hours per epoch vs. 1-2 minutes for architecture 2. For architecture 3, training time was 2-3 hours per epoch and the average AUC-ROC was 0.949.
Conclusion:
Accurate seizure detection using signals from non-invasive wearable devices is possible with the LSTM architecture. These results show the importance of utilizing both time-series and spectral information in the signals tested.
Funding:
:This work was funded by the ‘My Seizure Gauge’ grant provided by the Epilepsy Innovation Institute, a research program of the Epilepsy Foundation of America.