A - 53 A Digital Health Solution for Early Detection of Cognitive Impairment in Primary Care

Abstract Objective To determine which task or combination of tasks provided the most effective way to differentiate cognitively impaired (CI) from cognitively normal (CN) participants in under 5 minutes and to ensure that classification accuracy was equal to or better than a traditional brief cognitive screening task, the Quick Mild Cognitive Impairment (Qmci) screen. Method CN (n = 53) and CI (n = 51) participants completed a risk assessment task, a symbol matching (SM) task, and four speech-language tasks, followed by a second administration of SM to examine utility of practice effects administered on an iPad. Participants also completed the Qmci. Eleven models were tested using Bayesian adaptive regression trees. Results The top three models all included the two SM variables: the one with SM by itself (estimated c = 0.91), one with SM and features from a personal narrative task (c = 0.94), and one with SM and a counting backwards task (c = 0.90). Models with picture description and procedural discourse tasks performed the worst. For comparison, the QMCI-only model yielded c = 0.91. Conclusions A combination of working memory/processing speed and acoustic and linguistic variables from recalling a personal story achieved a high level of classification accuracy, slightly exceeding that of a traditional cognitive screening task. The inclusion of both verbal and nonverbal tasks may be an important feature, allowing for cognitive screening of individuals who are not able to do one type of task or the other. Future work is planned to examine this shortened tool in a pragmatic clinical trial in two primary care clinics.