Related Experiment Video
Updated: Aug 6, 2026
09:32
Cortical Source Analysis of High-Density EEG Recordings in Children
Published on: June 30, 2014
Unmasking data leakage in EEG-ADHD literature: a rigorous, interpretable SOTA framework (DSAEN)
Tushar Das1, Himanshu Kumar Pathak1, Koushlendra Kumar Singh1
1Machine Vision & Intelligence Lab, National Institute of Technology Jamshedpur, Jamshedpur 831014, Jharkhand, India.
Biomedical Physics & Engineering Express
|July 24, 2026
Summary
This study addresses data leakage in electroencephalogram-artificial intelligence (EEG-AI) by introducing a novel Dual-Stream Attention-Enhanced Network (DSAEN). DSAEN achieves high accuracy in subject-wise validation, outperforming existing models and offering interpretable, efficient clinical deployment.
Area of Science:
- Neuroscience
- Artificial Intelligence
- Biomedical Engineering
Background:
- Translational electroencephalogram-artificial intelligence (EEG-AI) research faces significant methodological challenges, particularly epoch-wise data leakage, which inflates performance metrics.
- Existing EEG-AI models often fail rigorous subject-wise validation, limiting their clinical applicability.
- A critical need exists for robust, interpretable, and clinically deployable EEG-AI diagnostic tools.
Purpose of the Study:
- To quantify the impact of epoch-wise data leakage on EEG-AI model performance.
- To propose and validate a novel Dual-Stream Attention-Enhanced Network (DSAEN) architecture for reliable subject-wise EEG analysis.
- To establish a trustworthy benchmark for diagnostic AI in clinical settings.
Main Methods:
- A methodological critique of existing EEG-AI models was performed, focusing on data leakage during validation.
- A novel Dual-Stream Attention-Enhanced Network (DSAEN) was developed, processing raw EEG tensors and neurophysiological features in parallel.
- The DSAEN architecture integrates 3D CNNs, LSTMs, Transformers, and Bahdanau Attention for spatiotemporal feature extraction.
- A replication study on a public ADHD dataset was conducted using rigorous subject-wise 5-fold cross-validation.
Main Results:
- Performance of leading models degraded significantly (e.g., 98.03% to 78.96%) when transitioning from epoch-wise to subject-wise validation.
- DSAEN achieved state-of-the-art subject-wise accuracy of 88.4% (AUC: 92.5%), outperforming baseline models.
- Dual-stream processing extracted orthogonal features, demonstrating complementarity and computational efficiency (2.04 MB footprint).
- Explainable AI confirmed neurophysiological plausibility, highlighting clinically relevant frontal brain regions.
Conclusions:
- Epoch-wise validation in EEG-AI is unreliable and leads to overestimated performance.
- DSAEN provides a robust, interpretable, and computationally efficient solution for subject-wise EEG analysis.
- The proposed methodology and DSAEN architecture set a new benchmark for trustworthy diagnostic AI in clinical neurology.
