Classification of mild cognitive impairment and normal cognition using an automated voice-based testing application

Abstract

An easy-to-use tool that can detect cognitive decline in mild cognitive impairment (MCI) is required. In this study, we aimed to construct a machine learning model that discriminates between MCI and cognitively normal (CN) individuals using spoken answers to questions and speech features. Participants of ≥50 years of age were recruited from the Silver Human Resource Center. The Japanese Version of the Mini-Mental State Examination (MMSE-J) and Clinical Dementia Rating (CDR) were used to obtain clinical information. We developed a research application that presented neuropsychological tasks via automated voice guidance and collected the participants' spoken answers. The neuropsychological tasks included time orientation, sentence memory tasks (immediate and delayed recall), and digit span memory-updating tasks. Scores and speech features were obtained from spoken answers. Subsequently, a machine learning model was constructed to classify MCI and CN using various classifiers, combining the participants' age, gender, scores, and speech features. We obtained a model using Gaussian Naive Bayes, which classified typical MCI (CDR 0.5, MMSE ≤26) and typical CN (CDR 0 and MMSE ≥29) with an area under the curve (AUC) of 0.866 (accuracy 0.75, sensitivity 0.857, specificity 0.712). We built a machine learning model that can classify MCI and CN using spoken answers to neuropsychological questions. Easy-to-use MCI detection tools could be developed by incorporating this model into smartphone applications and telephone services.