A Lightweight EEG-Based Alzheimer's Disease Detection Framework on an Edge-Computing Platform Toward Future Intelligent Digital Twin Systems

Alzheimer's disease (AD) progresses slowly through several stages until the patient reaches full cognitive decline. The electroencephalography (EEG) alterations associated with AD often emerge before symptom onset, suggesting that accessible and repeated EEG-based assessment could facilitate earlier detection and longitudinal monitoring. However, this type of assessment requires an AD detection model operating on edge-computing platforms within the Internet of Medical Things (IoMT). At present, many existing AD detection models rely on complex deep learning architectures with numerous parameters, which significantly limit their practicality for on-edge deployment.

We propose an on-edge AD detection framework comprising a lightweight dilated hybrid deep learning (DHDL) model with a channel selection approach to reduce the computational and power demands on the evaluated edge-computing platform. A key architectural component is the inclusion of dual-granularity dilated temporal convolutions. This method expands the receptive field at two scales to capture both short- and long-term temporal features from EEG signals without increasing the number of parameters. Subsequently, several hybrid lightweight layers extract distinct temporal and spatial features and are integrated with attention mechanisms. Additionally, our channel selection approach first forms an initial subset of EEG channels guided by individual channel classification accuracy, and then iteratively refines the selected channels using hill-climbing optimization. Here, we evaluated our framework on two publicly available EEG datasets recorded for the same group of subjects (n = 48 subjects): a conventional closed-eye (CE) dataset and a more challenging open-eye (OE) dataset. Validation followed a subject-independent Leave-2-Subjects-Out (L2SO) cross-validation protocol. Real-time deployment was tested on an edge-computing platform.

The proposed framework achieved 92.56% accuracy for the CE dataset and 75.15% accuracy for the OE dataset. Inference required 10.45 ms and 8.23 ms per segment for the CE and OE datasets, respectively, and consumed 9.69 mJ and 6.95 mJ of energy per segment, respectively.

These results demonstrate the feasibility of the proposed framework as a real-time, on-edge inference module for AD detection. Although a complete intelligent digital twin system is not implemented in this work, the proposed framework can serve as an inference component for future digital twin systems, supporting early AD detection and progression monitoring.