0264 Untangling Correlations Between Positive Affect and Light and Activity Data Using Deep Learning

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