DeepADNet: A CNN‐LSTM model for the multi‐class classification of Alzheimer’s disease using multichannel EEG

Abstract Background Early and accurate diagnosis of Alzheimer’s disease (AD) – an incurable and progressive brain disorder, along with a sharp upward trend in the incidence rate, is significantly important for patients to take prevention and appropriate treatments. However, the pathophysiological mechanism at the levels of AD severity is still poorly understood in spite of the worldwide financial and research efforts. Using an inexpensive and noninvasive modality, such as Electroencephalography (EEG) coupled with Deep Learning (DL)‐derived diagnostic tool to retain the accuracy and versatility, therefore, has gained much attention in AD multi‐class classification. Methods The present study proposes DL model, named DeepADNet, for the AD multi‐class classification using multichannel EEG. First, we extracted the event‐related spectral perturbation (ERSP) features measuring the average dynamic changes in amplitude of three main EEG frequency bands relative to a specific experimental event. Second, the hybrid deep Convolutional Neural Network (CNN) ‐ Long Short‐Term Memory Network (LSTM) was built to exploit the wide range of ERSP patterns on a time‐frequency domain and to generate discriminant features to classify healthy controls (HC) with two AD subject groups. To tackle the imbalance problem and improve the classification accuracy, we also applied different over‐sampling techniques Results The EEG data were collected from the Oddball – a cognitive ability test with 63 subjects (23 HC, 17 Pre‐symptomatic AD (aAD) and 23 Prodromal AD (pAD) subjects) and the N‐back – a memory ability test with 36 subjects (13 HC, 11 aAD and 12 pAD subjects). We demonstrated that the ERSP patterns displayed significant differences among three subject groups during the two experimental stages. Classification results also reveal that a CNN‐LSTM model could overcome the existing methods utilizing the hand‐engineered features which demand on prior knowledge of AD analysis and achieved the highest accuracy during the Oddball (71.95% ± 0.019 and 75.95% ± 0.017) and during the N‐back (69.40% ± 0.003 and 73.70% ± 0.010) corresponding to original and SVMSMOTE‐based ERSP features. Conclusions These findings demonstrate the capability of the EEG systems to better underlie the AD progression through spatiotemporal‐dynamic regions in the brain and the potential of DL‐based models for further AD classification studies.