0074 Actigraphic Sleep-Based Machine Learning Prediction of Cognitive Impairment

Abstract Introduction Cognitive impairment is currently undetected in primary care in up to 80% of affected patients. Therefore, it is crucial to improve ways to easily screen for dementia. Identifying common risk factors and correlates, such as sleep health, could prompt targeted, easy to employ, dementia screening. The present study utilized machine learning to test if actigraphic sleep-based measures can distinguish older adults with cognitive impairment from those who are cognitively healthy, and to identify which actigraphy variables may be driving the predictive potential of the model. Methods 1,308 participants from the Rush Memory and Aging Project (MAP) contributed actigraphy, medical history, and cognitive data from N=4,839 annual visits (Mbaseline age=81.4±7.4 years, 999 women). At each visit, participants were categorized as cognitively healthy (N0=3,593) or cognitively impaired (Mild Cognitive Impairment, Alzheimer’s Disease, or other dementia; N1=1,246). In the model, 6 behavioral sleep, 7 rest-activity rhythm (RAR; estimated using extended cosine models), and 7 non-parametric RAR actigraphic-derived features were used. In addition, 14 demographics, 13 medical conditions, 106 medication usage, and 28 health behaviors features were used in the model. A machine learning classifier (XGBoost) was used for the cognitive impairment/no cognitive impairment prediction task. A 3-fold cross-validation framework was used to train and validate models. Feature importance was analyzed with Shapley values to identify the most influential features. Results The model achieved an aggregated area under the curve (AUC)-ROC of .77, an average precision of .58, and an accuracy of 70%. Shapley results revealed relative amplitude (RA) and interdaily stability (IS) are two of the top 25 most influential features in the model. Specifically, lower RA and IS are correlated with higher risk of cognitive impairment. Conclusion Overall, results show potential for actigraphic sleep-based measures to predict if an individual has cognitive impairment, when paired with other demographic and risk factors. Future studies should investigate if RA and IS are potential treatment targets for reducing cognitive impairment. Support (if any) T32HL082610; R01AG056331; R01AG17917