Enhancing Alzheimer’s disease prediction through multimodal fusion-enabled ensemble learning using deep learning algorithms

Abstract

Alzheimer’s Disease is not just a challenge within the health sector in the world today, but a global burden that has only increased with the aging of the world population. The main objective of this research was to explore state-of-the-art deep learning architectures for the purpose of Alzheimer’s Disease prediction. The primary aim of this research is to provide healthcare professionals with dependable predictive models that assist in making clinical decisions within a few seconds consisting of diagnosis, prognosis, and treatment selection. This study focuses on a new strategy for Alzheimer’s Disease recognition. Alzheimer's Disease Neuroimaging Initiative (ADNI) database and the Open Access Series of Imaging Studies (OASIS) database were used to obtain the Magnetic Resonance Imaging (MRI) scans and patients’ clinical data. Several models that have performed well when performing individually were selected from the literature review to develop the ensemble models. The selected models are Convolution Neural Network (CNN), Recurrent Neural Network (RNN), FeedForward Neural Network (FFNN), Multi-Layer Perceptron (MLP), InceptionV3, ResNet50, and VGG16. CNN, RNN, FFNN and MLP were considered when developing ensemble models for clinical data modality and InceptionV3, ResNet50, VGG16, CNN, and a simple neural network were considered when creating ensemble models for the MRI scan dataset. Three ensemble models were developed for clinical data using CNN, RNN, and FFNN for base models and MLP for the meta model. For image data which consists of preprocessed MRI scans, two ensemble models were developed using CNN, InceptionV3, and ResNet50 for base models and a simple neural network for the meta model. After training all five ensemble models, the highest accuracy models from both data modalities were selected. Late fusion was then performed on these selected ensemble models, as it was found to provide better performance. Late fusion was selected due to its ability to deliver enhanced performance. All the models were trained and error-handled individually, resulting in more effective fusion performance. The ensemble model of CNN, RNN, and MLP for clinical modality and the ensemble model of InceptionV3, CNN, and Neural Network for the imaging modality led to outstanding outcomes. The accuracies of the ensemble model of CNN and InceptionsV3 as base models with a simple neural network as the meta model for the imaging dataset are found to be 99.80% and the ensemble model of CNN and RNN as base models with MLP as the meta model for clinical data found to be 99.51%, respectively. Finally, the involvement of the late fusion technique increased the accuracy of combined results of clinical and image modalities up to 99.87 %. A web application was developed for the use of people who have a risk of getting Alzheimer’s Disease to check their health status. These findings are a glimmer of hope for improving and enhancing the capacity to deal with Alzheimer’s Disease.