Accurate Prediction of Alzheimer's Disease Progression Trajectory via a Novel Encoder-Decoder LSTM Architecture

With an increase in life expectancy, there has been an increase in the aged population globally, and around 10% of this population suffers from Alzheimer's disease. Alzheimer's hugely impacts the quality of life and well-being of older adults and their caregivers. Thus, it is an emerging challenge to improve the early diagnosis and prognosis of the disease. Detecting hidden patterns in complex multidimensional datasets using recent advancements in machine learning provides a tremendous opportunity to meet this crucial need. In this study, using multimodal features and an individual's clinical status on one or more time points, we aimed to predict the individual's cognitive test scores, changes in Magnetic Resonance Imaging features, and the individual's diagnostic status for the next three years. We presented a novel Encoder-Decoder Long Short-Term Memory deep-learning model architecture for implementing our prediction. We applied it to data from the Alzheimer's Disease Neuroimaging Initiative, comprising longitudinal data of 1737 participants and 12,741 instances. The proposed model was found to be competent, with a validation accuracy of 0.941, a multi-class area under the curve of 0.960, and a test accuracy of 0.88 in identifying the various stages of Alzheimer's disease progression in patients with an initially cognitively normal or mild cognitive impairment which is a significant improvement over previous methods.