A Multicenter Study of a Precise Automated Epileptiform Discharges Detection based Video and EEG Data
Abstract number :
1.197
Submission category :
2. Translational Research / 2B. Devices, Technologies, Stem Cells
Year :
2024
Submission ID :
1016
Source :
www.aesnet.org
Presentation date :
12/7/2024 12:00:00 AM
Published date :
Authors :
Presenting Author: Nan Lin, MD – Peking Union Medical College Hospital
Qiang Lu, MD – Peking Union Medical College Hospital
Liying Cui, MD – Peking Union Medical College Hospital
Weifang Gao, MD – Peking Union Medical College Hospital
heyang Sun, MD – Peking Union Medical College Hospital
yisu dong, BS – Netease Inc.
Lian Li, BS – Netease Inc.
Zi liang, BS – Netease Inc.
Rationale: There's a demand of automatically electroencephalogram (EEG) data analysis. The present study proposes a multi-modal method on video and EEG data and validates the detection model with prospective multicenter datasets.
Methods: The training dataset is comprised of 20,235 IED 4s video-EEG segments extracted from 240 patients from Peking Union Medical College Hospital (PUMCH), and 243,736 non-IED segments. We employed You Only Look Once version 5 (YOLOv5) to detect and annotate patients and patients’ face from the video data. Frame difference and Simple Keypoints (SKPS) were subsequently used to present movements of body, head and face. After being translated by the short-time Fourier transform, EEG data were put into the EfficientNetV2. EEG and video features were then fused through a multilayer perceptron. The test dataset independent from training dataset contains video-EEG data from three hospitals, including PUMCH. It comprises of 420 hours of raw continued video-EEG data from 140 randomly selected patients, without preprocessing and manual selection.
Results: The test data contains 5,484 IEDs (PUMCH 2,101, BTTH 1,521 and CHASU 1862) and 27,892 non-IEDs 4s video-EEG segments. The average processing time for one hour of video-EEG data was 5.5 minutes (range 5.0-5.8). Due to the imbalanced dataset, the precision (true positive / (true positive + false positive)) was calculated. With a sensitivity of 80%, the corresponding precisions were 71.4% for PUMCH, 47.4% for BTTH and 41.0% for CHASU. Comparison between models with or without video feature integration was performed, showing a superior precision-recall (sensitivity) characteristics (AUPRC) in all three datasets, surpassing the EEG model without video. The video feature eliminated nearly one third false positives, especially in wakefulness.
Conclusions: The integration of video data enhances diagnostic performance and reduces the false positive rate. The multi-modal deep learning approach for IED detection demonstrates generalizability and clinical applicability.
Funding: National Key Research and Development Project (grant number：2022YFC2503800)
National High Level Hospital Clinical Research Funding (grant number: 2022-PUMCH-A-168)
CAMS Innovation Fund for Medical Sciences (CIFMS) (grant number: 2023-I2M-C&T-B-023)
Translational Research