LAUSR

Research has found activity and light data continuously captured via wearable devices predict psychological outcomes. Understanding which sources are most predictive could suggest which biological or environmental factors most impact outcomes. Existing work has largely relied on summary measures of these sources. This may overgeneralize important nuances in the sources that are correlated to the outcomes. We utilized deep learning to predict daily affected based on raw actigraphy and light data to determine which sources are most predictive of daytime positive affect. We modeled the daily mean (DMPA) and standard deviation (DSDPA) of positive affect using actigraphy and light data from a cohort of 172 adolescents who completed an ecological momentary protocol for 7-8 days. Positive affect was measured via PANAS-SF and administered via smartphone 5-6 times daily. The outcomes are predicted using the actigraphy and light data during the environmental night (time between astronomical twilight and dawn) prior to each outcome. Convolutional neural networks are used to predict the outcomes using all permutations of actigraphy, white light, and red-green-blue (RGB) light data. The mean (standard deviation) of the DMPA and DSDPA outcomes were 6.53(1.60) and 1.28(1.15) respectively. Using white and RGB light we could predict DMPA and DSDPA with a root mean squared error (RMSE) of 1.59 and 1.08 respectively (i.e., within 0.99 and 0.93 deviations respectively). Our work indicates that white and RGB light best predict positive affect. This may indicate that light-based environmental night factors have a larger impact on daily affect than activity-based factors. Most results were within a few hundredths of a deviation from each other, and the channels that best predicted affect differed depending on whether DMPA or DSDPA was modeled. Our methods provide a way to predict affect from raw data and suggest which environmental and biological data have clinical implications towards affect. Important next steps include using time periods related to individual’s dim light melatonin onset and comparing findings from deep learning to those based on summary measures. CARRS Pilot Grant (Wheeler), R01-AA025626 (Hasler), R01-DA044143 (Hasler), RF1-AG056311-04 (Wallace)