Recently, the EEG and neurophysiology team from PUMCH's Department of Neurology, in collaboration with institutions including Beijing Tiantan Hospital, Capital Medical University and Children's Hospital Affiliated to Shandong University, published their multi-center validation results for vEpiNetV2, an automatic interictal epileptiform discharge (IED) detection model, in BMC Medicine (a Tier 1 journal, or ranked among the top 5% by the Chinese Academy of Sciences). The team conducted the first prospective large-scale multi-center validation, faithfully replicating the model's performance in real clinical applications. Results show that vEpiNetV2 can effectively assist physicians in EEG analysis, improving reading efficiency and reducing human error.

The diagnosis of epilepsy, a common brain disorder, depends on the detection of interictal epileptiform discharges (IEDs) as a key marker. While visual analysis by trained specialists remains the gold standard for IED detection, this approach is labor-intensive, time-consuming, and subject to inter-expert variability. While artificial intelligence (AI) has been gradually applied to IED detection, differences in patient populations, EEG equipment, and operational techniques across hospitals can compromise detection accuracy in real-world applications. Large-scale, multi-center prospective validation studies of intelligent IED detection models remain lacking both domestically and internationally.

Building on their previously developed vEpiNet model, PUMCH's neurology team expanded training samples and further optimized the technical architecture to launch the next-generation vEpiNetV2 model. The new model maintains the multimodal fusion design combining EEG and video data, while adding a bad channel removal model and a more broadly applicable patient detection method to better address multi-center data variability.

▲Architecture overview of vEpiNetV2

This study established the largest-scale test dataset for IED detection models to date, with all test data completely independent from training samples, to mimic real clinical settings to the largest extent possible. The research prospectively included raw video-EEG data from three epilepsy centers: PUMCH, Beijing Tiantan Hospital, Capital Medical University, and Children's Hospital affiliated to Shandong University. The test datasets span a wide range of age groups from infants to elderly, containing nearly 10,000 IED samples of different types.

Results demonstrate that vEpiNetV2 exhibits efficient and stable detection performance, achieving high levels of sensitivity, specificity, and precision across multi-center test datasets, with both sensitivity and specificity exceeding 90%. Video features represent a major innovation of vEpiNetV2. Through efficient extraction and fusion of video and EEG signals, detection precision improved significantly at all three epilepsy centers, with overall false positive rates reduced by one quarter.

Currently, vEpiNetV2 has been translated into software and hardware systems and deployed for clinical trials at over ten hospitals nationwide, including PUMCH, Peking University First Hospital, Affiliated Hospital of Chifeng University, First Affiliated Hospital of Air Force Medical University (Xijing Hospital), Hebei Children's Hospital, Second Affiliated Hospital of Hebei Medical University, General Hospital of Ningxia Medical University, Children's Hospital Affiliated to Shandong University, Beijing Anding Hospital, Capital Medical University, Beijing Tiantan Hospital, Capital Medical University, Xi'an Children's Hospital, First Affiliated Hospital of Xi'an Jiaotong University, Second Affiliated Hospital of Xi'an Jiaotong University, and Xi'an No. 3 Hospital. This technological application will help more medical institutions improve EEG diagnostic efficiency and accuracy.

This research was supported by the National Key R&D Program and the National High Level Hospital Clinical Research Funding. Chief Physician Cui Liying and Associate Chief Physician Lu Qiang from PUMCH's Department of Neurology are co-corresponding authors of this paper, and Attending Physician Lin Nan is the first author.

Written by and pictures courtesy of the Department of Neurology

Edited by Wang Jingxia

Chief editor Duan Wenli

Supervised by Wu Peixin