Sleep metrics from machine learning for Alzheimer's disease diagnostics

点击翻译按钮获取中文摘要

PROJECT SUMMARY This proposal is responsive to NIH solicitation PA-17-089 for projects involving secondary analysis of pre-existing geriatric datasets. While presently there is no cure for Alzheimer’s disease, existing literature indicates that early diagnosis in the preclinical stage, i.e., before the onset of clinical symptoms, will be key to treatments. There is a pressing need for noninvasive predictors of cognitive decline that can enable early identification of individuals at Alzheimer’s disease risk. A mounting body of scientific evidence suggests that sleep disturbances (including microarchitectural disruptions to non-rapid-eye-motion sleep and decline in sleep quality) might be the earliest observable symptoms of Alzheimer’s disease. On-the-go sleep and activity monitoring could address the need for noninvasive indicators of cognitive decline in subjects who are in the (asymptomatic or mildly symptomatic) preclinical stage of Alzheimer’s disease. Here, we will build on preliminary results that reveal a set of sleep features derived from polysomnography (PSG) that are predictive of cognitive performance. We are proposing to perform secondary analysis of sleep and cognition data from the Multi-Ethnic Study of Atherosclerosis (MESA) cohort using state-of-the-art deep learning tools to enable sleep-based prediction of cognitive impairment for early detection of Alzheimer’s disease. While PSG is the gold standard for sleep measurement, it is not well- suited for routine, day-to-day use. In comparison, wrist-based measurements (e.g. actigraphy, heart rate, ECG, and pulse oximetry) obtained from wearable devices allow “on-the-go” sleep monitoring. The combination of these on-the-go measures with the latest artificial intelligence tools is a feasible route to early Alzheimer’s diagnostics. We will use attention-guided long short-term memory autoencoders to identify overt and latent characteristics of the raw time-series datasets, which will allow us to more effectively mine the rich MESA data resource. Our deep learning framework will also take into account sociodemographic variables, indicators of health status, and medications. To ensure scientific rigor, secondary validation of the MESA-trained deep learning models will be performed on PSG and actigraphy data from the Harvard Aging Brain Study, which is a longitudinal study designed to further our understanding of what differentiates normal aging from preclinical Alzheimer’s disease. To address any concern about the “black-box” nature of deep learning models, we will compare the learned feature set with sleep microarchitectural features previously computed using classical statistical techniques. Previous data suggests that a subject’s apolipoprotein ε4 (ApoE4) allele carrier status influences the degree to which their sleep patterns impact their cognitive abilities. We will verify this by incorporating ApoE4 status as an additional input to the deep learning model. Literature shows that over 60% of patients with mild cognitive impairment and Alzheimer’s disease have at least one clinical sleep disorder. The on-the-go prediction paradigm using noninvasive sleep measurements to be validated in this project will have a significant impact on early Alzheimer’s diagnostics and facilitate ongoing clinical trials.