A Transformer Approach for Cognitive Impairment Classification

Early classification of Alzheimer’s disease (AD) and amnestic mild cognitive impairment (aMCI) using non-invasive approaches is a longstanding challenge. Furthermore, data needed to assist with addressing this problem is frequently sparse in features. To address this data sparsity challenge, we use a masked Transformer-encoder to learn the relationship among all input features, dynamically masking input features that are not provided for each given sample. We demonstrate that this masked self-attention scheme can achieve high multi-class cognitive status classification accuracy ( \(87\%\) control, \(78.9\%\) aMCI, \(89.1\%\) AD) without the use of instrumental activities of daily living (IADL) information or summative metrics designed for staging aMCI and dementia as input. In conclusion, we report here a new method for studying datasets with sparse input features via masked Transformers—providing a new venue for the analysis of cognitive status data.