An Automated Classifier of Harmful Brain Activities for Clinical Usage Based on a Vision-Inspired Pre-trained Framework

Yulin Sun, Xiaopeng Si, Runnan He, Xiao Hu, Peter Smielewski, Wenlong Wang, Xiaoguang Tong, Wei Yue, Meijun Pang, Kuo Zhang, Xizi Song, Dong Ming, Xiuyun LiuNpj Digit. Medicine 2025 · DOI

TL;DR

This paper introduces VIPEEGNet, a novel deep learning model that leverages transfer learning from computer vision to automatically classify harmful brain activities from EEG signals. By converting EEG data into an image-like format, VIPEEGNet achieves high accuracy and efficiency, offering a promising solution for real-time clinical monitoring and expanding access to expert-level EEG interpretation.

What Problem Does This Paper Address?

Timely identification of harmful brain activities, such as seizures and periodic discharges, is crucial for diagnosing and treating neurological conditions. However, manual EEG analysis is time-consuming, prone to inter-rater variability, and requires specialized expertise, limiting its widespread clinical application. While automated methods exist, they often struggle with generalization due to small, dataset-specific training and a lack of robust, pre-trained models, especially compared to the advancements seen in computer vision. This paper aims to bridge this gap by adapting powerful vision models for EEG analysis.

Methods

The core innovation of VIPEEGNet lies in its "EEG-to-image" conversion module. This module transforms multi-channel, one-dimensional EEG signals into a format that can be processed by pre-trained vision models. Specifically, it uses learnable convolutional kernels to create image representations where the horizontal axis encodes temporal information and the vertical axis encodes spatiotemporal information.

These image representations are then fed into a pre-trained EfficientNetV2-B3 model (trained on ImageNet) for feature extraction. The model leverages transfer learning, using only a small fraction of the parameters of the vision model, to capture complex patterns in the EEG data. After feature extraction, average pooling is applied to the central regions of the feature map, followed by a dropout layer and a fully connected layer with SoftMax activation to output probabilities for different brain activity classes.

Datasets

The study utilized two main datasets:

  • Development Cohort: EEG recordings from 1950 patients, with 106,800 EEG segments annotated by experts (ranging from 1 to 28 annotations per segment). This dataset was used for training and validation.
  • Online Testing Cohort: A subset of EEG segments from an additional 1,532 patients, each annotated by at least 10 experts. This dataset was used for external validation in a Kaggle competition.

The data was collected between 2006 and 2020 from Massachusetts General Hospital/Harvard Medical School.

Key Insights and Ideas

  • Vision Models for EEG: The paper demonstrates that the powerful feature extraction capabilities of ImageNet-pretrained vision models can be effectively transferred to the domain of EEG analysis, even with a novel EEG-to-image conversion.
  • Efficiency through Transfer Learning: By leveraging pre-trained vision models, VIPEEGNet achieves high performance with significantly fewer parameters than many state-of-the-art algorithms, making it more efficient and suitable for real-time deployment.
  • EEG-to-Image Conversion: The proposed EEG-to-image module is a key enabler, allowing the direct application of vision architectures to time-series EEG data without requiring extensive domain-specific feature engineering.
  • Importance of Pre-training and EEG-to-Image: Ablation studies highlight that both the pre-trained vision backbone and the EEG-to-image conversion component are critical for the model's high performance. Removing either significantly degrades accuracy.

Key Results

  • High Accuracy: VIPEEGNet achieved high AUROC values for binary classification of harmful EEG patterns, with scores ranging from 0.930 to 0.972.
  • Competitive Performance: On the online testing cohort, VIPEEGNet ranked second among 2,767 competing algorithms, achieving a KLD score of 0.223296 on the public leaderboard and 0.272543 on the private leaderboard.
  • Parameter Efficiency: VIPEEGNet uses only 2.8% of the parameters of the top-ranked algorithm, demonstrating remarkable efficiency.
  • Expert-Level Performance: The model's precision was found to be higher than human experts in some categories, and its overall performance was comparable to human experts.
  • Interpretability: T-SNE visualization showed that the model effectively differentiates between lateralized and generalized events, and its feature representations align with clinical understanding.

Limitations and Future Work

  • Data Complexity: The paper acknowledges that real-world EEG data can be more complex than the studied patterns, suggesting a need for more data and the inclusion of a wider range of EEG patterns.
  • Generalizability to Other Diseases: While VIPEEGNet is effective for epileptic and seizure disorders, its applicability to other neurological conditions (e.g., sleep disorders, metabolic diseases) remains to be explored.
  • Future Directions: The authors suggest that the EEG-to-image component can be adapted to other pre-trained visual models, opening avenues for future extensions and improvements. The modular design also facilitates future extensions and clinical deployment.

Summary updated

BibTeX
@article{sun2025automated,
  author = {Yulin Sun and Xiaopeng Si and Runnan He and Xiao Hu and Peter Smielewski and Wenlong Wang and Xiaoguang Tong and Wei Yue and Meijun Pang and Kuo Zhang and Xizi Song and Dong Ming and Xiuyun Liu},
  title = {{An Automated Classifier of Harmful Brain Activities for Clinical Usage Based on a Vision-Inspired Pre-trained Framework}},
  journal = {Npj Digit. Medicine},
  year = {2025},
  doi = {10.1038/s41746-025-02154-4},
}
Figure