Multimodal behavioral analysis for early detection of Alzheimer’s disease: A preliminary result

Abstract

AbstractBackgroundWith the rapid growth in the elderly population worldwide, early detection of Alzheimer’s disease (AD) has become increasingly important. In this respect, behavioral analysis of sensor data is expected to extend opportunities for assessments and help identify subtle changes resulting from AD at an early stage. Gait, speech, and drawing behaviors have been actively investigated and recognized as behavioral biomarkers that can be used for detecting individuals with AD and mild cognitive impairment (MCI). However, previous studies have mainly investigated single types of behavior. It remains unclear whether combining multimodal behavioral data could improve detection performance.MethodIn this study, we collected multimodal behavioral data related to gait, speech, and drawing from 73 Japanese seniors consisting of 36 healthy controls (HC), 26 MCI, and 11 AD. For gait data, we collected motion capture data during 9‐meter walks at the usual walking pace and extracted gait features related to pace, rhythm, variability, and postural control. Speech data were collected while performing tablet‐based cognitive tasks consisting of verbal fluency, calculation, and picture description, followed by extraction of acoustic, prosodic, and linguistic features. We also collected drawing behaviors with a drawing tablet during five tasks including clock drawing and trail making, and obtained features related to kinematic parameters of drawing trajectories and drawing pressure. We then used these features to build binary classification models for differentiating MCI or AD from HC using a support vector machine model with a feature selection method. The models were then evaluated by ten‐fold cross‐validation.ResultOur models using features of three types of behavior achieved 99.8% accuracy for HC vs. AD and 92.0% accuracy for HC vs. MCI. Compared to models using single modality behavioral data, we found that our model improved detection by up to 9.1% for HC vs. AD and 21.0% for HC vs. MCI.ConclusionWe demonstrated that multiple behavioral data could improve detection accuracy for patients with both MCI and AD by capturing different functional changes resulting from AD. Our approach based on multimodal behavioral analysis shows promise for timely diagnosis at an early stage of AD.