Abstract Details

Title The Sounds of Seizures: Audio-Triggered Detection by Convolutional Neural Network (CNN)

Topic Epilepsy/Clinical Neurophysiology (EEG)

Presentation(s) P4 - Poster Session 4 (5:30 PM-6:30 PM)

Poster/Presentation
Number 6-002

Objective To identify seizure-specific sound using end-to-end deep learning-based predictive modelling

Background Patients with epilepsy who have frequent generalized tonic-clonic (GTC) type seizures have an associated increased risk for Sudden Unexpected Death in Epilepsy (SUDEP), respiratory distress, and physical injuries. Concern over such events has prompted heightened interest in seizures detection aids.

Design/Methods Audio files were collected from the Epilepsy Monitoring Unit (EMU). Audio-segmentation was carried out and over 15 hours of sound data was subsequently labeled as either seizure (bilateral tonic-clonic seizures) or non-seizure by an epileptologist using both EEG, video and audio. The dataset contains about 30 minutes of seizure-related sounds (positive samples) and we created an equal sized set of randomly selected non-seizure sounds (negative samples) from the same recordings. All sounds were separated into slices of ½ second per segment with ¼ second overlap between segments and sampled at a rate of 48kHz which were further split into training and test sets. We converted the labeled time-domain seizure signals into spectrograms which were used to train CNN models for seizure detection.

Results Several CNN models were tested and network with > 90% accuracy was used for further training. 23,622 samples were trained and 5,904 samples tested. Of the 2911 test samples that our model predicted to be seizure sounds, 2665 of them were true positives, giving a precision of .915. And from the 2941 total positive test samples, our model correctly identified 2665 of them, giving a recall of .906

Conclusions Machine learning CNN models can accurately identify seizure sounds. Further optimization of prediction accuracy and off-line and online clinical testing in the EMU is needed.

Authors/Disclosures
Maysaa M. Basha, MD, FAAN (Wayne State University, Detroit Medical Center) PRESENTER	Dr. Basha has nothing to disclose.
	No disclosure on file
Hani Alhourani	No disclosure on file
Aashit K. Shah, MD, FAAN (Carilion Clinic)	Dr. Shah has stock in Abbot, Abbivie, Gilead, Johnson and Johnson, Pfizer. The institution of Dr. Shah has received research support from Xenon Pharma.
	No disclosure on file

��ɫ��

��ɫ��