vEpiSet

An EEG dataset for interictal epileptiform discharge with spatial distribution information

Abstract Interictal epileptiform discharge (IED) and its spatial distribution are crucial for the diagnosis, classification and treatment of epilepsy. Manual annotation by electroencephalography (EEG) experts led to a lack of publicly available datasets from multiple epilepsy centers, impeding the development of automatic IED detection. We present an EEG database containing annotated interictal epileptic EEG data from 84 patients. We extracted 20-minute continuous raw EEG data from each patient, amassing 28 hours in total. IEDs and the state of consciousness (wake/sleep) were meticulously annotated by at least 3 EEG experts. Based on the occurrence regions, the discharges are categorized into five types: generalized IED, frontal IED, temporal IED, occipital IED, and centro-parietal IED. All EEG data were segmented into 4-second epochs. This resulted 2,516 IED epochs and 22,933 non-IED epochs totally. We develop a VGG model for IED detection trained and validated on the present dataset. The integration of consciousness information improves model performance, particularly at high levels of sensitivity. Furthermore, our dataset is demonstrated to serve as a robust tool for validating existing IED detection models and for automatically classifying IEDs types based on spatial distribution.

Age distribution of patients in the dataset.

Schematic Diagram of EEG electrode placement, including the five spatial region for classifying IEDs.

Percentage and count of each IED type classified by spatial location.

Generalized IED Frontal IED Temporal IED Centro-parietal IED Occipital IED Precision Sensitivity
Generalized IED 504 48 8 5 8 0.913 0.880
Frontal IED 40 329 68 8 13 0.760 0.718
Temporal IED 7 41 631 5 18 0.812 0.899
Centro-parietal IED 0 5 13 348 0 0.935 0.951
Occipital IED 1 10 57 6 343 0.898 0.823

The table above exhibits a confusion matrix derived from five-fold cross-validation, tailored for the classification of five types of IEDs. The VGG model has demonstrated remarkable performance, achieving over 70% in both precision and sensitivity. This result further validates the high applicability and effectiveness of our dataset in training and classifying IED types based on spatial distribution.



[1] Nan Lin, Weifang Gao, Heyang Sun, Junhui Chen, Yisu Dong, Zi Liang, Haibo He, Peng Hu, Liying Cui, Qiang Lu,
An EEG dataset for interictal epileptiform discharge with spatial distribution information,
Scientific Data, Volume 11, Article 211, 2024,
https://doi.org/10.1038/s41597-024-03047-x

vEpiNet

Multimodal Interictal Epileptiform Discharge Detection Method Based on Video and Electroencephalogram Data

Abstract enhance deep learning-based automated interictal epileptiform discharge (IED) detection, this study proposes a multimodal method, vEpiNet, that leverages video and electroencephalogram (EEG) data. Datasets comprise 24 931 IED (from 484 patients) and 166 094 non-IED 4-second video-EEG segments. The video data is processed by the proposed patient detection method, with frame difference and Simple Keypoints (SKPS) capturing patients' movements. EEG data is processed with EfficientNetV2. The video and EEG features are fused via a multilayer perceptron. We developed a comparative model, termed nEpiNet, to test the effectiveness of the video feature in vEpiNet. The 10-fold cross-validation was used for testing. The 10-fold cross-validation showed high areas under the receiver operating characteristic curve (AUROC) in both models, with a slightly superior AUROC (0.9902) in vEpiNet compared to nEpiNet (0.9878). Moreover, to test the model performance in real-world scenarios, we set a prospective test dataset, containing 215 hours of raw video-EEG data from 50 patients. The result shows that the vEpiNet achieves an area under the precision-recall curve (AUPRC) of 0.8623, surpassing nEpiNet's 0.8316. Incorporating video data raises precision from 70% (95% CI, 69.8%-70.2%) to 76.6% (95% CI, 74.9%-78.2%) at 80% sensitivity and reduces false positives by nearly a third, with vEpiNet processing one-hour video-EEG data in 5.7 minutes on average. Our findings indicate that video data can significantly improve the performance and precision of IED detection, especially in prospective real clinic testing. It suggests that vEpiNet is a clinically viable and effective tool for IED analysis in real-world applications.

Overview of the proposed multimodal method. The data is collected from two primary sources: electrical and video data. The proposed vEpiNet also contains two branches: the above branch processes and extracts features from the electrical signal, and the bottom branch processes and extracts the motion features from the video data. Finally, the features of two sources are fused together, and the classifier makes decisions based on them.

Comparisons of different models. (a) Precision-Recall (PR) curves of the vEpiNet and nEpiNet models on the retrospective and prospective test sets. (b) Receiver operating characteristic (ROC) curves of the vEpiNet and nEpiNet models on the retrospective and prospective test sets. (c) Precision-Recall (PR) curves of nEpiNet, ResnEpiNet, and VggnEpiNet. (d) Precision-Recall (PR) curves of vEpiNet, ResvEpiNet, and VggvEpiNet.



[1] Nan Lin, Weifang Gao, Lian Li, Junhui Chen, Zi Liang, Gonglin Yuan, Heyang Sun, Qing Liu, Jianhua Chen, Liri Jin, Yan Huang, Xiangqin Zhou, Shaobo Zhang, Peng Hu, Chaoyue Dai, Haibo He, Yisu Dong, Liying Cui, Qiang Lu,
vEpiNet: A multimodal interictal epileptiform discharge detection method based on video and electroencephalogram data,
Neural Networks, Volume 175, 2024, 106319,
https://doi.org/10.1016/j.neunet.2024.106319

vEpiNetV2

Development and Validation of a Multimodal Automatic Interictal Epileptiform Discharge Detection Model: A Prospective Multi-Center Study

Abstract Visual identification of interictal epileptiform discharge (IED) is expert-biased and time-consuming. Accurate automated IED detection models can facilitate epilepsy diagnosis. This study aims to develop a multimodal IED detection model (vEpiNetV2) and conduct a multi-center validation. We constructed a large training dataset comprising 26,706 IEDs and 194,797 non-IED 4-second video-EEG epochs from 530 patients at Peking Union Medical College Hospital (PUMCH). The automated IED detection model was constructed using deep learning based on video and electroencephalogram (EEG) features. We proposed a bad channel removal model and patient detection method to improve the robustness of vEpiNetV2 for multi-center validation. Performance was verified in a prospective multi-center test dataset from three epilepsy centers: PUMCH, Children's Hospital Affiliated to Shandong University (SDQLCH), and Beijing Tiantan Hospital (BJTTH). The multimodal IED detection model, which integrates video and EEG features, demonstrated high precision and robustness. The large multi-center validation confirmed its potential for real-world clinical application and the value of video features in IED analysis.

Overview of vEpiNetV2. The model contains two key components: the EEG model (left) and the video processing model (right). EEG and video data are extracted and processed to generate electrical and video feature vectors. These features are then fused together, and the classifier makes the decision based on the combined features.

Center Patients Age (mean) IEDs AUPRC AUC Precision@80%Sens
PUMCH 54 27 (9-70) 2,680 0.76 0.98 61.5%
SDQLCH 51 9 (0.1-19) 3,669 0.80 0.96 66.9%
BJTTH 44 30 (9-65) 2,883 0.76 0.98 53.2%

Multi-center validation results. vEpiNetV2 demonstrated favorable accuracy for IED detection across three epilepsy centers (PUMCH, SDQLCH, BJTTH), with a total of 377 hours video-EEG data from 149 patients containing 9,232 IEDs. Video features improved precision by 5-9% across centers.

Sample detections of vEpiNetV2 across three epilepsy centers. The vertical line denotes the presence of an IED within a 2-s window. AI prediction confidences are displayed at the base of the detection lines (maximum at 1.000).



[1] Nan Lin, Lian Li, Weifang Gao, Peng Hu, Gonglin Yuan, Heyang Sun, Fang Qi, Lin Wang, Shengsong Wang, Zi Liang, Haibo He, Yisu Dong, Zaifen Gao, Xiaoqiu Shao, Liying Cui, Qiang Lu,
Development and validation of a multimodal automatic interictal epileptiform discharge detection model: a prospective multi-center study,
BMC Medicine, Volume 23, Article 479, 2025,
https://doi.org/10.1186/s12916-025-04316-3

EEG-VL

Integrating Visual Features with Large Language Models for Automated Seizure Detection

Abstract The increasing demand for accurate EEG-based epileptic seizure detection calls for more sophisticated and semantically informed methodologies. Traditional approaches often struggle to capture the complex spatiotemporal patterns inherent in EEG signals and typically lack high-level contextual understanding. In this study, we propose EEG-VL, a novel vision-language framework that treats EEG signals as visual patterns and integrates them with large language models to improve seizure detection. Specifically, a pretrained EfficientNet encoder is used to extract abstract visual features from EEG representations, which are embedded into structured prompts and processed by the Qwen language model. This design synergistically combines the spatial modeling capabilities of convolutional networks with the semantic reasoning strengths of large language models. Extensive experiments on the TUSZ and CHB-MIT datasets demonstrate that EEG-VL achieves state-of-the-art performance.

Overview of the proposed EEG-VL architecture for seizure detection. The framework integrates a pretrained 2D vision encoder that converts raw multichannel EEG signals into visual feature representations, with a Qwen2.5-0.5B language model decoder that processes these features alongside structured textual prompts to generate seizure/non-seizure classifications.

Model AUPRC AUROC Precision@60%Recall Specificity
EEGNet 0.6415 0.9181 0.6186 0.9767
EEG-TCNet 0.6111 0.9085 0.5285 0.9661
EEG2VIT 0.6662 0.9287 0.6865 0.9827
EEG-VL (Ours) 0.7599 0.9466 0.8214 0.9917

Results on TUSZ dataset. EEG-VL achieves state-of-the-art performance, attaining an AUPRC of 0.7599 and an AUROC of 0.9466, surpassing previous best results by 8.19% and 0.80% respectively.

Performance comparison on TUSZ. (a) Precision-Recall curves and (b) ROC curves for different models on the seizure detection task. EEG-VL consistently dominates across different operating thresholds.



[1] Zi Liang, Peng Hu, Lian Li, Zebang Cheng, Yisu Dong, Haibo He, Qiang Lu, Nan Lin,
EEG-VL: Integrating Visual Features with Large Language Models for Automated Seizure Detection,
2025 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), pp. 4620-4627,
DOI: 10.1109/BIBM66473.2025.11356689

vEpiSpy

A Multimodal AI-Based Interictal Epileptiform Discharge Detection System Based on vEpiNet

vEpiSpy is the first system to combine patient video and EEG data, fully simulating how physicians annotate EEG readings, thereby significantly improving the accuracy of AI-based EEG detection. In clinical trials, the system achieved specificity and sensitivity of 90% and 80% respectively, with a false positive rate below 40%, matching the reading level of human EEG experts. Additionally, vEpiSpy seamlessly integrates with hospital EEG software, allowing physicians to directly view AI-annotated EEG results without any additional operations. With vEpiSpy assistance, physicians' reading time has been reduced by one-third on average, improving efficiency by approximately 50% and greatly saving physician resources.
vEpiSpy configuration platform: automatic program execution, continuous EEG file reading and automatic annotation
Seamless integration with EEG software: physicians can directly view AI anomaly markers
Currently being deployed in major hospitals nationwide

Patents

14 Authorized Patents

ZL202311630571.9 ZL202311635333.7 ZL202311635642.4 ZL202311598803.7 ZL202311630555.X ZL202411162618.8 ZL202411281529.5 ZL202411281798.1 ZL202411293834.6 ZL202510230853.2 ZL202510232423.4 ZL202510233779.X ZL202510233871.6 ZL202411171359.5

News

September 7, 2023

vEpiSpy officially made its debut at the 26th National Neurological Academic Conference. Dr. Lin Nan, a neurologist from Peking Union Medical College Hospital, published a report on a multimodal epileptic like discharge detection system based on YOLOv5 and SKPS human motion capture models, which received unanimous praise from attending experts.