University of Aizu Researchers Develop Explainable AI Method for EEG-Based Alzheimer’s Disease Detection

A research team led by Professor Jungpil Shin at the University of Aizu has developed a deep learning framework designed to classify Alzheimer's disease from electroencephalography (EEG) recordings while using fewer EEG channels. The accepted paper is now available as a citable Article in Press in Scientific Reports (Nature Portfolio).

Research overview

The study introduces Fused-MSCA-2B, a two-branch multi-scale channel-attentive one-dimensional convolutional neural network that learns complementary temporal patterns directly from EEG signals. The researchers also used Integrated Gradients, an explainable AI technique, to identify informative EEG channels using validation data only, enabling the reduction of the electrode configuration from 19 channels to 12.
(Evaluation Using Publicly Available EEG Data)

Source：Shin et al., Scientific Reports (2026), DOI:10.1038/s41598-026-76382-3（CC BY 4.0）

pdf: https://u-aizu.ac.jp/files/EEG_AD_poster_V2.pdf

Key findings

After identifying and removing duplicate signals from two public EEG resources, the researchers evaluated 113 unique recordings (88 Alzheimer's disease and 25 healthy control recordings). The 12-channel configuration achieved 98.60% ± 1.18% mean recording-level classification accuracy, with 99.07% sensitivity, 96.60% specificity, and a ROC-AUC of 0.9991. The evaluation used recording-level grouped cross-validation, and held-out test recordings were excluded from channel ranking and selection.

The results suggest the potential for more compact EEG-based decision-support systems. However, the reported findings are research results, not evidence of clinical diagnostic readiness; larger independent and multicenter studies are needed before clinical deployment.

Publication details

Title: EEG-based Alzheimer's disease detection using two-branch multi-scale deep learning with integrated gradients-guided channel selection

Authors: Jungpil Shin, Najmul Hassan, Abu Saleh Musa Miah, and Satoshi Nishimura

Journal: Scientific Reports (Article in Press, 2026)  |  Accepted: October 13, 2026  |  Published online: October 20, 2026

DOI / Article link: https://doi.org/10.1038/s41598-026-76382-3

Link to paper: https://u-aizu.ac.jp/files/s41598-026-76382-3_reference.pdf
　　　　　　　Open Access（CC BY 4.0）

Editorial note: Article in Press is citable and will be replaced by the final Version of Record under the same DOI. The 98.60% accuracy is from research-dataset cross-validation, not a verified clinical screening performance.