A Transformer Approach for Cognitive Impairment Classification and Prediction

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Informatics

Background:

• Early detection of Alzheimer disease (AD) and amnestic mild cognitive impairment (aMCI) is crucial but challenging due to data limitations.
• Noninvasive approaches and advanced modeling techniques are needed to improve diagnostic accuracy.

Purpose of the Study:

• To develop and evaluate a deep learning model for classifying current cognitive status and predicting future diagnosis of AD and aMCI.
• To assess the model's performance with sparse and incomplete datasets.

Main Methods:

• Utilized a masked Transformer-encoder model on the National Alzheimer Coordinating Center (NACC) dataset.
• Analyzed neuropsychological data, health history, or a combination, dynamically handling missing features.
• Fine-tuned the model for 1- to 3-year future diagnosis prediction and assessed feature importance.

Main Results:

• The Transformer-encoder successfully predicted cognitive status with sparse input data.
• Achieved high classification accuracy for current cognitive status (87% control, 79% aMCI, 89% AD).
• Demonstrated acceptable prediction accuracy for future cognitive status (83% control, 77% aMCI, 91% AD).

Conclusions:

• The proposed deep learning method effectively handles inconsistent and sparse data for cognitive status analysis.
• This approach offers a flexible and powerful tool for early AD and aMCI classification and prediction.