Abstract
Alzheimer's Disease (AD) is a life-threatening neurodegenerative disease with far-reaching global implications. Deep neural network-based techniques are being utilised to predict and diagnose AD at every stage. The Mini-Mental State Examination (MMSE) scores are essential for monitoring the onset and progression of the disease since they serve as numerical assessments of cognitive function. This paper introduces a novel multi-stage algorithm for predicting MMSE scores in patients with AD. Initially, a regression analysis was carried out to predict the patient's age on a combined MCI dataset created by including individuals with mild cognitive impairment (MCI), early cognitive impairment (EMCI), and late cognitive impairment (LMCI). Subsequently, transfer learning techniques were applied to incorporate knowledge from the regression model into an autoencoder. The autoencoder extracted significant and latent features from the combined MCI dataset, creating meaningful encoded representations. A regression analysis was employed to predict MMSE scores based on the encoded features, and to subsequently classify patients into two categories Mild and Moderate based on cognitive status. This strategy achieved an accuracy of approximately 73.26% with a 3.92% standard deviation. For comparison, a simple regression model without employing Transfer Learning and Auto Encoders was implemented. This simple regression model achieved an accuracy of 61.08% with 2.21% Standard Deviation, indicating an enhancement of approx. 12.18% accuracy due to transferred knowledge. This comparison illustrated the effectiveness of the proposed methodology. Cross-validation techniques were used to analyse the stability and applicability of the approach, confirming consistency and performance across subsets of the dataset. The results highlighted the potential of transfer learning and autoencoder-based feature extraction as validated by achieving improved predictions of MMSE scores for AD patients.