How people wake up is associated with previous night's sleep together with physical activity and food intake

Abstract

How people wake up and regain alertness in the hours after sleep is related to how they are sleeping, eating, and exercising. Here, in a prospective longitudinal study of 833 twins and genetically unrelated adults, we demonstrate that how effectively an individual awakens in the hours following sleep is not associated with their genetics, but instead, four independent factors: sleep quantity/quality the night before, physical activity the day prior, a breakfast rich in carbohydrate, and a lower blood glucose response following breakfast. Furthermore, an individual's set-point of daily alertness is related to the quality of their sleep, their positive emotional state, and their age. Together, these findings reveal a set of non-genetic (i.e., not fixed) factors associated with daily alertness that are modifiable.

Fig. 1. Experimental design.

The Personalized Responses to Dietary Composition Trial (or “PREDICT1”) is a two-country (UK, US) longitudinal study whose primary goal is to predict metabolic responses to foods based on the individual’s characteristics, including molecular biomarkers and lifestyle factors, as well as the nutritional composition of the food. PREDICT1 consists of one clinic baseline visit followed by a two-week home-based study. During the at-home phase, participants consumed multiple standardised test meals differing in macronutrient composition, while wearing an accelerometer wristwatch and a continuous glucose monitor. The former was used to determine sleep/wake activity during the night and physical activity during the day. The continuous glucose monitor was used to measure postprandial glucose response. Participants also recorded their dietary intake, satiety, mood, and exercise on the study app throughout the study. The app also prompted participants to report their alertness levels on a 0–100 visual analogue scale at t = 0 minutes (time of logging of a meal) and regular intervals following the logging of a meal (see “Methods”). Source data are provided as a Source Data file. The cupcake icon, the ruler icon and the smartwatch icon were purchased and downloaded from thenounproject.com. All other icons were purchased and/or downloaded from iconfinder.com. MZ monozygotic, DZ dizygotic.

Fig. 2. Alertness ratings throughout the day.

a Alertness as a function of time of day. The orange line shows a cubic regression of all the alertness ratings logged between 5am and midnight (n = 89,440). Alertness progressively increased in the first hours of the morning, reached a plateau during midday and progressively decreased in the evening. Sample size for each unique box is shown in panel B. Box plots show centre line as median, box limits as upper and lower quartiles. The notches represent confidence intervals around the median. The whiskers extend from the box limits by 1x the interquartile range. b Polar histogram of the number of alertness ratings as a function of time of day. c Alertness ratings within the first three hours after breakfast onset. Participants were instructed to rate their alertness at t = 0 min, t = 30, t = 90 and t = 150 min after breakfast start. During that ~3 hour period, they were also instructed to fast and avoid physical activity. Each black dot represents one alertness rating from one participant. The purple line shows a cubic regression of all the morning alertness ratings. Alertness immediately increased after breakfast, and then plateaued for the subsequent 2.5 hours. d Distribution of breakfast start time. By definition in the protocol, the first alertness rating of the day must coincide with breakfast onset. Source data are provided as a Source Data file.

Fig. 3. Predictors of day-to-day fluctuations in morning alertness.

Standardised regression coefficients and confidence intervals from a linear mixed effect model. Sex, BMI, zygosity and sunrise time were also included in the model but are not reported here for conciseness since none of them was a significant predictor of morning alertness. Unstandardised regression coefficients and raw p-values can be found in Supplementary Table 1. Sleep predictors were normalised using person-mean centering. The dependent variable of the model is morning alertness, which is calculated by averaging the alertness ratings that were made within the first three hours after breakfast start (n = 6,744 observations). Family ID and participant ID were defined as nested random effects of the linear mixed model. Predictors with a positive coefficient (i.e predicting higher morning alertness) are shown in blue, while those with a negative coefficient (lower morning alertness) are shown in red. Error bars represent 95% confidence intervals. Stars indicate significance. P-values are based on two-tailed Wald tests (degrees of freedom = 6717) and are not adjusted for multiple comparisons. *p < 0.05, **p < 0.01, ***p < 0.001. L5 = least active 5 hours of the day, M10 = most active 10 hours of the day, MCB metabolic challenge breakfast, OGTT Oral Glucose Tolerance Test, iAUC incremental area under the curve. Source data are provided as a Source Data file.

Fig. 4. Predictions of morning alertness from the hold-out validation.

a Hold-out validation strategy. For each participant, the multilevel model was trained on half the available days and then tested on the remaining half. Predictors of the multilevel model are shown in Fig. 3. The dependent variable of the model is morning alertness, which is calculated by averaging the alertness ratings that were made within the first three hours after breakfast start. Family ID and participant ID were defined as nested random effects of the linear mixed model. b Scatter plot showing the true and predicted values of morning alertness from a full model that included all the aforementioned predictors. Each dot in the scatter plot represents the morning alertness value from one day from one participant. c True and predicted values of morning alertness from a naive model that only included random effects. Predicted values are therefore, for each participant, the average of all the morning alertness values in the training set. Source data are provided as a Source Data file. The woman icon was downloaded from iconfinder.com.

Fig. 5. Predictors of trait alertness.

a Cross-validated performance of the gradient boosting algorithm. Predictions of the trait alertness are plotted against the ground-truth values. Each dot represents an individual’s predicted and actual trait alertness. Trait alertness was calculated by averaging, for each individual, all their alertness rating across the two weeks of the study. b Features of the gradient boosting algorithm, ranked by order of descending importance. Features importance was calculated using the SHAP method (see “Methods”). Each dot on the plot is a Shapley value for a given predictor and participant. Shapley values represent, for each participant, the exact contribution of a given feature on the output of the model. The colour of the dots represents the value of a given predictor from low to high (e.g. for age, younger = blue, older = pink). Global feature importance is shown in the rightmost horizontal grey bars and was calculated by averaging the absolute Shapley values of a given predictor across all participants. c Trait correlation between happiness and alertness. Happiness was the most important feature of the model. Higher trait happiness is associated with higher trait alertness. Source data are provided as a Source Data file.