Abstract
This study aimed to advance the understanding of factors that predict mean sleep latency (MSL) on the multiple sleep latency test (MSLT) by applying machine learning methodology on a high-dimensional dataset from a large community sample. A cross-sectional analytic dataset of first visit clinical-protocol MSLTs (without shift workers) was developed from the Wisconsin Sleep Cohort Study, a community-based longitudinal study of middle-aged to older adults in Wisconsin, USA. Fifty predictors captured demographics, medical and psychiatric health, sleep (diary; polysomnography [PSG]) and circadian characteristics. The random forest (RF) algorithm identified the 10 most important predictors, which underwent subsequent regression analyses. Primary analyses focused on MSLT MSL, whereas secondary analyses centred on nap-specific sleep latencies. Post hoc analyses further explored the relationship between circadian preference and MSLT MSL. The primary sample (n = 301) of middle-aged adults (mean age = 57.5 ± 7.71 years) was predominantly non-Hispanic White (97%) and nearly equal across sexes (percentage female = 51.8%). RF model showed low explanatory value for MSLT MSL (R (2) = 12%) with PSG sleep onset latency, circadian preference, daily caffeine use and Epworth Sleepiness Scale emerging as the most important predictors of MSLT MSL in the dataset. Top predictors varied across nap-specific sleep latency. Morning preference displayed significantly longer MSLT MSL, relative to neither and evening preferences. The low explanatory value observed in our high-dimensional RF models seemingly reflects the complexity and variability of the MSLT. Additionally, our results underscore the importance and challenge of accounting for circadian characteristics when utilising the MSLT.