High-Accuracy Alzheimer's Prediction Without Expensive Biomarkers: A Multi-Modal Deep Learning Approach

Abstract

The initial diagnosis of people with Mild Cognitive impairment (MCI) who are going to develop Alzheimer Disease (AD) is a significant issue in dementia treatment. Existing predictive technologies typically rely on invasive or costly biomarkers, which restricts their widespread clinical application. This paper presents a tested, low cost, computational model of MCI-to-AD conversion, three years to the future, based on data collected routinely. Multi-modal deep learning architecture was designed and it synergistically used structural T1-weighted Magnetic Resonance Imaging (MRI), longitudinal cognitive scores (MMSE, CDR) and basic demographics of the OASIS-3 data. Strict testing with a 10-fold stratified cross-validation procedure showed a mean predictive accuracy of 97.1%±1.2% and Area Under the Curve (AUC) of 0.98. The effectiveness of the model was validated by its ability to perform successfully on an external dataset which was not re-trained. Ablation experiments were used to measure the important predictive value of each of the data modalities, and fairness assessment was used to find equal performance between different demographic subgroups. These results demonstrate that a simple data solution can be used to reach the state of art in AD prediction and provides a possible avenue to creating accessible, scalable, and low-cost decision support solutions to support early risk stratification in a clinical environment.