Alzheimer's Disease Stage Classification via Multimodal CNN on EEG Spectrograms and Cube-Drawing Images

BackgroundClinicians currently lack reliable tools to determine, at the point of mild cognitive impairment (MCI), which individuals will progress to Alzheimers disease (progressive MCI, PMCI) versus remain stable (SMCI). Early, patient-specific prognosis is therefore difficult using routine clinical evaluation alone. MethodsWe propose a dual-branch Convolutional Neural Network (CNN) that fuses two low-cost bedside measurements: short-time Fourier transform (STFT) spectrograms derived from 19-channel resting-state EEG and static images of the ACE-R cube-drawing task. Separate CNN sub-networks process each modality, and their latent representations are concatenated for four-class classification (healthy control, stable MCI, progressive MCI, AD). DataThe model was trained and five-fold cross-validated on a multimodal cohort of 114 participants (36 AD, 37 PMCI, 18 SMCI, 23 HC) recruited in Tehran between 2017 and 2022. ResultsThe system achieved 99.6% accuracy, 99.6% precision, 99.5% recall and an area under the receiver-operating-characteristic curve of 0.999. ConclusionsTo our knowledge, this is the first study to pair EEG spectrograms with cube-drawing images for Alzheimers-stage classification, and the proposed architecture outperforms recent multimodal baselines for early MCI detection. These findings indicate that inexpensive electrophysiological and visuospatial measures, when analysed jointly, can approach the diagnostic performance of resource-intensive neuroimaging and are therefore suitable for routine screening and timely intervention. However, the dataset used in this study is currently not publicly available, which limits immediate reproducibility.