OBJECTIVES: Alzheimer's disease is a serious neurodegenerative disorder that is prevalent among the elderly. Progressive mild cognitive impairment (pMCI) has the potential to progress into Alzheimer's disease. In contrast, stable mild cognitive impairment (sMCI) refers to a milder form of cognitive decline that does not further deteriorate over time. We proposed a multimodal deep neural network approach to accurately classify pMCI and sMCI individuals based on MRI images and clinical information.

METHODS: We applied pretrained ImageNet like VGG19, ResNet18, MobileNet and Xception to train MRI images. Additionally, we utilized fully connected neural network and XGBoost to train clinical information. Considering the multimodal structure, we leveraged different weights, logistic regression, and random forest machine learning methods to combine the information from disparate data channels.

RESULTS: The unimodal training results demonstrated that the modified MobileNet architecture, which yielded the best classification performance of MRI images, with an accuracy, precision, sensitivity and F1 score of 0.9167,0.9167,0.7333 and 0.8148 respectively. For the clinical information, we achieved accuracy of 0.9500, precision of 0.8000, sensitivity of 1.0000 and F1 score of 0.8889 respectively trained by the fully connected deep neural network. Our multimodal neural network based on MRI images and clinical information achieved 0.9667,0.9500 and 1.0000 accuracy by applying three different combination methods on the testing set.

CONCLUSIONS: This multimodal neural network provides early diagnosis and accurate risk prediction of Alzheimer. This model can enable healthcare professionals to personalize patient management, intervene promptly, and develop tailored treatment plans. Leveraging this integrated approach has the potential to effectively slow down disease progression, improve clinical outcomes and enhance the quality of life for patients.