Abstract
BACKGROUND: Voice is a potential biomarker of cognitive impairment because mild cognitive impairment (MCI) can cause changes in speech patterns and tempo. Artificial intelligence (AI) can deliver voice biomarkers as prediction features, leading to a timely, noninvasive, and cost-effective detection of cognitive impairment. This study aimed to develop and test prediction models utilizing voice biomarkers to detect cognitive impairment, which AI derived from voice data of unstructured conversations in community-dwelling adults in Japan. METHODS: This observational study with a cross-sectional design, included 1461 community-dwelling adults. The outcome was cognitive impairment assessed by the Memory Performance Index score from the MCI screen. Voice data was collected from 3-min open-question interviews and extracted voice biomarkers based on acoustic and prosodic features as a 512-dimensional vector of individual voice information using the voice generator, Wav2Vec2. Other considerable predictors were age, sex, and education. We developed cognitive impairment prediction models by applying the extreme gradient boosting decision tree algorithm and a deep neural network model using 979 participants. Prediction performances were tested by area under the curves (AUCs) in 482 participants who were not used for model development. FINDINGS: We had 967 women (66·2%), 526 cognitive impairment (36·0%) participants with mean (standard deviation) age and education years of 79·5 (6·3) years old and 11·6 (2·2) years, respectively. The inclusion of voice biomarkers significantly improved AUCs (95% confidence intervals), from 0·80 (0·76, 0·84) to 0·88 (0·84, 0·91) for the age sex model and from 0·78 (0·73, 0·82) to 0·89 (0·86, 0·92) for the age sex and education model (p < 0·0001 for both comparisons by DeLong test). INTERPRETATION: Our prediction models for cognitive impairment using voice biomarkers can provide significantly timesaving MCI screening with high prediction performances (AUC = 0·89). Voice biomarkers significantly contributed to improving prediction performance. FUNDING: Small Business Innovation Research (SBIR Phase 3 Fund), the Intramural Research Fund of Cardiovascular Diseases of the National Cerebral and Cardiovascular Center, and JSPS KAKENHI.