Diagnosis of Alzheimer's disease using non-linear features of ERP signals through a hybrid attention-based CNN-LSTM model

Biological signals have a dynamic and non-linear nature, and hence nonlinear analysis is important for understanding the signals. In this study, a hybrid Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) model is proposed for the diagnosis of Alzheimer’s disease (AD) from the Event-Related Potential (ERP) signals obtained from the Electroencephalogram (EEG) data. The P300 component of the ERP signal, derived from acoustic stimulation, is a key indicator of AD, and its amplitude and latency are characterized. By using nonlinear features such as phase diagrams, correlation dimension, entropy, and Lyapunov exponents, the proposed model classifies AD stages. The hybrid CNN-LSTM architecture, enhanced by an attention mechanism, captures both spatial and temporal dependencies in the ERP signals, achieving high accuracy: For healthy people, 95%, for mild AD patients, 92.5%, and for severe AD patients, 97.5%. The model achieves 75% accuracy in recall mode for healthy individuals, 72.5% for mild AD, and 87.5% for severe AD. Results show that the proposed model outperforms traditional methods and provides a robust and accurate diagnostic framework for AD. The result of this approach is to show that the combination of non-linear EEG analysis with advanced deep learning methods could provide early and precise AD detection.