Quantifying sleep architecture dynamics and individual differences using big data and Bayesian networks

Abstract

The pattern of sleep stages across a night (sleep architecture) is influenced by biological, behavioral, and clinical variables. However, traditional measures of sleep architecture such as stage proportions, fail to capture sleep dynamics. Here we quantify the impact of individual differences on the dynamics of sleep architecture and determine which factors or set of factors best predict the next sleep stage from current stage information. We investigated the influence of age, sex, body mass index, time of day, and sleep time on static (e.g. minutes in stage, sleep efficiency) and dynamic measures of sleep architecture (e.g. transition probabilities and stage duration distributions) using a large dataset of 3202 nights from a non-clinical population. Multi-level regressions show that sex effects duration of all Non-Rapid Eye Movement (NREM) stages, and age has a curvilinear relationship for Wake After Sleep Onset (WASO) and slow wave sleep (SWS) minutes. Bayesian network modeling reveals sleep architecture depends on time of day, total sleep time, age and sex, but not BMI. Older adults, and particularly males, have shorter bouts (more fragmentation) of Stage 2, SWS, and they transition less frequently to these stages. Additionally, we showed that the next sleep stage and its duration can be optimally predicted by the prior 2 stages and age. Our results demonstrate the potential benefit of big data and Bayesian network approaches in quantifying static and dynamic architecture of normal sleep.

Fig 1. Hypnogram and corresponding stage proportions of fragmented sleep (top) and normal sleep (bottom).

Stage proportions are the minutes in each stage normalized by total sleep time. REM: rapid eye movement sleep; SWS: slow wave sleep; WASO: wake after sleep onset.

Fig 2. Static sleep statistics across age and sex.

Top row: Sleep Efficiency, Sleep Latency, Bottom row: minutes in each stage. To show the uncertainly in predictions, regression parameters are randomly sampled 100 times from each model’s joint parameter distribution and each is used to plot a regression line. REM: rapid eye movement sleep; SWS: slow wave sleep; WASO: wake after sleep onset.

Fig 3. Best fitting models to predict the current stage and duration from previous sleep architecture variables.

“In dataset” and “out of dataset” prediction accuracy and prediction error for current stage (top) and current stage duration (bottom) is shown. Model 1A) 1 back model (including t-1 variables), Model 1B) 2 back model (including t-2 variables), Model 1C) 3 back (including t-3 variables). When considering previous sleep archetecture only, Model 1B gave the best fit and states that the identity of the current stage is dependent on the identity of the previous 2 stages and the duration of the last stage (blue arrows). The duration of the current stage is dependent on the identity of the current stage (at t) and the previous one (at t-1) (red arrows).

Fig 4. Effects of Time of Day and Total Sleep Time.

Model 2A: When previous stages not included, Model 2B: 1 back included, Model 2C: 2 back included. Beside each model is the “in dataset” and “out of dataset” prediction accuracy and prediction error for current stage (top) and current stage duration (bottom). Time of Day influences 0^th order transition probabilities (2A) and 1^st order transition probabilities (2B). Total Sleep Time influences both 0^th order transition probabilities and duration distributions when no previous stage information is available (2A).

Fig 5. Effects on model parameters (Model 2A) across the night.

A) Stage proportions, B) Transition Probabilities (0^th Order—the probability of transitioning to a particular stage from any stage), C) Stage duration distributions as measured by expected duration. To calculate each statistic, we ran the model 14 times, each time removing (and then replacing) one of the datasets from the full set of 14 datasets used to train the model. Points are the mean and error bars are the standard deviation across these 14 runs (see Methods). REM: rapid eye movement sleep; SWS: slow wave sleep; WASO: wake after sleep onset.

Fig 6. Full model including individual factors.

Model 3A: No previous stage information, Model 3B: 1 back stage information, Model 3C: 2 back stage information. Beside each model is the “in dataset” and “out of dataset” prediction accuracy and prediction error for current stage (top) and current stage duration (bottom). BMI: Body Mass Index.

Fig 7. Model parameters, for each stage, age and sex group separately, as Time of Day and Total Sleep Time are increased across the night.

Expected durations do not depend on sex, and therefore are the same for both sex groups. To calculate each statistic, we ran the model 14 times, each time removing (and then replacing) one of the datasets from the full set of 14 datasets used to train the model. Points are the mean and error bars are the standard deviation across these 14 runs (see Methods). REM: rapid eye movement sleep; SWS: slow wave sleep; WASO: wake after sleep onset.

Fig 8. Distributions of continuous variables.

Red dashed lines indicated edges of discretization where relevant (see Discretization section). A) Age (per subject), B) Time of Day (per data point), C) Body mass index (BMI, per subject), D) Time Slept (per data point—e.g. a subject who slept for 400 mins will impact the histogram from 0 to 400).

Fig 9. Stage duration distributions.

Data from all subjects and across the whole night. REM duration distribution shows less fragmentation (longer bouts) than SWS. REM: rapid eye movement sleep; SWS: slow wave sleep; WASO: wake after sleep onset.