Research indicates that mild cognitive impairment (MCI) progresses to Alzheimer’s disease at an estimated rate of 10–15% per year. A common symptom associated with this progression is memory loss. Traditionally, studies in this field have relied on a single type of data to predict the disease's occurrence. However, recent studies have shown pro4mising results when multiple data modalities are considered. Methodologically, standard convolutional neural networks (CNNs) are typically used in successful deep learning techniques. However, they often face challenges such as large parameter sizes, resulting in computationally expensive operations. In this study, we enhance the prediction performance of Alzheimer’s disease by augmenting patient electronic health records (EHR) with magnetic resonance images (MRI), focusing on metrics like accuracy, precision, and recall. To achieve this, we utilized a stacked denoising autoencoder to generate intermediate relevant features from the patient’s EHR. Concurrently, a depthwise separable convolutional network was employed to extract features from the patient’s MRI data. Experimentally, this multimodal approach demonstrated significant improvements, recording an accuracy of 95.74%, a recall of 95.21%, and a precision of 95.00%, compared to classical CNNs used in existing networks.