Longitudinal methods for Alzheimer's cognitive status prediction with deep learning

Area of Science:

• Neurology
• Artificial Intelligence
• Biostatistics

Background:

• Machine learning for cognitive decline prediction typically focuses on short-term (1-3 years) forecasting.
• Long-term prediction (3-10 years) of cognitive status remains a significant challenge in dementia research.

Purpose of the Study:

• To develop and validate a novel deep learning technique for extending the prediction horizon of cognitive status.
• To improve the accuracy of predicting amnestic mild cognitive impairment (aMCI) and Alzheimer's disease (AD) over a 3- to 10-year period.

Main Methods:

• Leveraged deep learning on longitudinal neuropsychological data and patient history.
• Introduced techniques to separate baseline features and deviations, and a linear attention-based imputation method.

Main Results:

• Achieved high prediction accuracy: 81.65% for Control, 72.87% for aMCI, and 86.52% for AD over a 3- to 10-year horizon.
• The new method outperformed previous approaches for long-term cognitive status prediction.

Conclusions:

• Presents a novel set of deep learning techniques for analyzing longitudinal dementia data.
• Highlights the feasibility and improved accuracy of long-term prediction for aMCI and AD, even for cases progressing to AD.