Classification of AD and bvFTD using neuropsychological and neuropsychiatric variables: a machine learning study

Machine learning (ML) is increasingly used for clinical classification of Alzheimer's disease (AD) and related dementias. Prior studies identified useful diagnostic features for AD and behavioral variant frontotemporal dementia (bvFTD), though they often lack pathological verification. We applied ML to classify AD and bvFTD autopsy status using initial visit neuropsychological and neuropsychiatric data.

Data from the National Alzheimer's Coordinating Center Uniform Data Set and Neuropathology Data Set were analyzed using logistic regression, support vector machines, random forest, and artificial neural networks to classify autopsy-confirmed diagnosis based on symptom and cognitive data.

Among 1616 participants (AD = 1498, bvFTD = 118), all algorithms achieved high accuracy (80% to 90%) and discriminatory ability (AUC = 0.89 to 0.95). Apathy, disinhibition, and digit-symbol substitution were the most important classification features.

Findings emphasize the value of specific clinical disease markers to support differential diagnosis of AD and bvFTD.

Four ML algorithms were used for the classification of AD and bvFTD. Neuropsychological subtests and neuropsychiatric symptoms were input features. Models had high classification accuracy and discrimination. We identified important and accessible clinical features for classification.