Recently, the EEG and neurophysiology team under the Department of Neurology, PUMCH, in collaboration with the Technology Department, NetEase Media Group, published their original findings on an automatic detection model of interictal epileptiform discharge (IED) in the journal Neural Networks (a tier 1 journal, or among the top 5%, as ranked by the Chinese Academy of Sciences). The study proposed an AI-powered multimodal detection method, vEpiNet, that innovatively incorporates video data into IED detection. Compared with traditional models, vEpiNet demonstrated higher accuracy and efficiency. The research was supported by the National High-Level Hospital Clinical Research Funding and the Key Research and Development Program under the Ministry of Science and Technology.

Epilepsy is a common neurological disorder, and electroencephalogram (EEG) is a crucial tool for its diagnosis. IED, as a characteristic indicator, is essential for the diagnosis, classification, and medication management of epilepsy. Currently, the detection of IED relies heavily on the interpretation and manual annotation by epilepsy specialists, which is time-consuming and subjective. AI can improve EEG interpretation efficiency and reduce human error. In recent years, AI-based capabilities for IED detection and analysis have rapidly evolved. Most existing models at home and abroad, though, rely solely on EEG data and have low specificity.

The employed datasets comprise 24,931 IED video-EEG segments from patients at the PUMCH Epilepsy Center. EEG data were processed through short-time Fourier transform and the EfficientNetV2-S for electrophysiological feature extraction. Video data were first processed by a patient detection model YOLOv5-patient for target localization to minimize environmental and bystander interference. Then the frame difference algorithm was used to discern the body movements and address issues of body covered, and key point detection for subtle facial movements, transforming full-body and facial movements into video features. This approach significantly reduces video data volume, and ensures accurate feature extraction and fast processing speed. Finally, the video and EEG features were fused via a multilayer perceptron, and the model was named vEpiNet.

▲Illustration of the multimodal IED detection model

Results show that this model achieves high levels of sensitivity, specificity, and accuracy, with a lower per-minute false positive rate than previous studies. Video features significantly enhance accuracy and reduce judgement errors, particularly in real clinical settings.

▲Testing results of vEpiNet and nEpiNet which relies on EEG only

The model can be embedded in a compatible EEG machine operating system, where analyzed video EEG data is directly displayed with an annotated detection line on the EEG machine’s playback software. Currently, processing one hour’s video EEG data takes only 5.7 minutes.

▲Detection example: The vertical line is the software detection result, indicating the occurrence of IED within a two-second window before and after (as shown by the red dashed lines)

The integration of AI plays a crucial role in standardizing and normalizing EEG reports to make them more consistent. This study validates the effectiveness and feasibility of using a multimodal approach in the automatic detection of IEDs, providing a new direction for the application of AI in EEG analysis.

Translated by Liu Haiyan

Reviewed by Yang Xunzhe and Wang Yao