Leveraging deep learning to classify sitting posture and measure sedentary patterns from accelerometer data in diverse cohorts - PROJECT ABSTRACT High sedentary behavior (SB) increases risk for all-cause mortality, cardiovascular disease (CVD), cancer, and type 2 diabetes. However, mixed evidence on how much to limit SB or how to break up SB to reduce its negative health impacts has inhibited specific quantified SB guidelines The ubiquity of wearable sensors, able to collect data at fine granularity (e.g., 30Hz), enables rich, nuanced SB assessment. Computational methods to accurately quantify SB accumulation patterns (e.g., long uninterrupted bouts of SB versus fragmented SB with numerous breaks) are needed. Building on our previous work, in this project we will implement deep learning methods to derive posture-based SB measures for the widely used ActiGraph sensor from raw accelerometer (at 30 Hz granularity) or processed count outputs from hip- or wrist-worn devices across a broad range of population groups (Aims 1, 2). We will develop a deep-learned convolution neural network (CNN) bidirectional long short-term memory (Bi-LSTM) model. We will use extensive training and held-out testing to reduce overfitting and improve accuracy and reproducibility on future samples. To develop our models, we will leverage data from seven existing separately funded studies comprising 6390 unique free-living individuals with hip- or wrist-accelerometer data (>200,000 hours of device wear), and concurrent criterion posture assessment. We will apply deep transfer learning, novel in SB research, which exploits the basic neural architecture of an existing model, and finetunes it for a new cohort or application. Deep transfer learning can markedly reduce computational complexity and avoids the need for large criterion datasets. We will apply our new classifiers to quantify SB from hip- and wrist-worn ActiGraphs for participants from the NHANES and WHISH Studies (N= 42,496), and evaluate cross-sectional and longitudinal associations between SB patterns and cardiometabolic health in youth and adults (Aim 3). Enabling use of a single device, i.e., the ActiGraph, to obtain posture-based SB metrics, as well as energy-expenditure based physical activity measures, will obviate the need for multiple devices, and improve participant compliance. Our publicly available algorithms and accurate SB metrics will advance SB research, allowing researchers to transfer our models to other cohorts, enabling specific quantified public health SB guidelines (similar to physical activity guidelines) and laying the foundation for future development of self- monitoring tools for clinicians or interventionists to personalize SB goals for their patients, and facilitate SB change.
