PROJECT SUMMARY People vary in both the quantity and the quality of their sleep. Traditional sleep “stages” (for example, slow- wave sleep or REM sleep) can exhibit different neural correlates (as measured by electroencephalography, EEG) in different people and at different times during the night. Standard approaches for characterizing sleep (“staging”) do not take this variability into account, however. Here we suggest that attempts to use EEG and other data to infer sleep stages should in fact tackle the heterogeneity between and within individuals explicitly, to yield more accurate classification not just of an individual’s minute-by-minute sleep, but of the sleeper’s typical characteristics more generally. Quantitative, data-driven individual sleep profiles will be important prerequisites for emerging large-scale molecular genetics and other ‘omics studies of sleep, and for driving personalized sleep medicine more generally. Specifically, we propose to leverage more than ten thousand whole-night polysomnography (PSG) studies from the demographically diverse National Sleep Research Resource (NSRR), in order to optimally apply machine-learning methods to sleep signals. Most importantly, the NSRR dataset will enable approaches that take the fundamental heterogeneity of sleep into account: this is a rate-limiting step that has not been tackled by previous approaches. We hypothesize that it is not the large training set per se, but rather the increased prospects for better-matched training data, combined with the methods developed here, that will be beneficial. To this end, we will build and disseminate a novel, multi-level automated sleep classifier based on the world’s largest dataset of sleep signals.