Novel Framework for Alzheimer Early Diagnosis using Inductive Transfer Learning Techniques

The early diagnosis of Alzheimer's disease (AD) is gaining popularity due to massive applications in the medical field. The estimated 50 million people affected by AD worldwide are increasing rapidly, and the condition has an approximately 1 dollar trillion economic cost. This necessitates developing scalable, economical, and reliable methods for detecting Alzheimer's disease. We describe a unique architecture that uses auditory, cognitive, and linguistic aspects to create a multimodal ensemble system. Scores from the Mini-Mental State Examination are used to assess AD severity, which is detected using specialized artificial neural networks with temporal properties. We first test it using the ADReSS challenge dataset, a balanced dataset available through Dementia Bank that contains no biases against any patient. Our system achieves state-of-the-art test accuracy, precision, recall, and F1-score of 93.30% each, as well as a state-of-the-art test Root Mean Squared Error (RMSE) of 4.60, when it comes to the classification of AD. To our knowledge, the algorithm achieves cutting-edge AD classification accuracy of 88.0 % when tested against the benchmark Dementia Bank Pitt database. In order to build a robust inductive transfer learning model, our study highlights the applicability and transferability of spontaneous speech and demonstrates generalizability by utilising a task-agnostic feature space.