Participants
Sleep was monitored over a minimum of 7 nights in 203 athletes providing a total of 1975 nights of sleep. Professional elite athletes from four team sports (Netball, Australian Rules Football, Rugby League, Soccer) (age range = 19–36 years; female, n = 79; male, n = 124) gave informed consent to participate in the study. Participants were excluded if they were training or sleeping at altitude, if they were injured, if they reported a clinical diagnosis of a sleep disorder, or if they had undertaken transmeridian travel in the 2 weeks prior to data collection. All data collection occurred outside of competition periods and in the athlete’s typical training and sleep environment. Information on bed partners was not collected. The study was conducted in accordance with the standards of ethics outlined in the Declaration of Helsinki and was approved by the Australian Institute of Sport Human Research Ethics Committee.
Procedures
Athletes’ sleep/wake behaviour was monitored using wrist activity monitors in conjunction with self-report paper sleep diaries. Each athlete wore an activity monitor on the same wrist throughout the data collection period, except when showering, swimming, or training. The sleep diaries were used to record two pieces of information for each night-time sleep: start date/time and end date/time. Daytime naps were not recorded. Athletes were instructed to complete their sleep diary each morning within 30 min after waking. There was no experimental manipulation of the athletes’ training schedules or sleep/wake behaviours and the athletes were free to consume nutritional supplements, caffeine or alcohol during the data collection period. Information regarding medication use (including sleeping pills) was not collected.
Sleep Measurement
Two different models of activity monitor—produced by a sole manufacturer—were used in this study (Actiwatch-64 and Actical Z-series; Philips Respironics; Oregon, USA). Devices were configured to sum and store data in 1-min epochs based on activity counts from a piezoelectric accelerometer with a sensitivity of 0.05 g and a sampling rate of 32 Hz. Data from the sleep diary and activity monitor were used to determine when participants were awake and when they were asleep. Essentially, all time was scored as wake unless: (i) the sleep diary indicated that the athlete was lying down attempting to sleep and (ii) the activity counts from the monitor were sufficiently low to indicate that the athlete was immobile [26]. When these two conditions were satisfied simultaneously, time was scored as sleep. In this study, sensitivity was set at medium, which corresponds to a threshold activity count of 40. This scoring process was conducted using the Philips Respironics’ Actiwatch algorithm. Validation studies comparing wrist activity monitors with polysomnography report high levels of agreement in healthy adults (88%) [21] and well-trained athletes (81–90%). [22]
For each athlete, the following variables were derived for each sleep period:
-
Bedtime (h:min): time at which the athlete attempted to initiate sleep;
-
Sleep onset (h:min): the time at which an athlete first fell asleep after going to bed;
-
Sleep offset (h:min): the time at which an athlete last woke before getting up;
-
Sleep period (h): the time between sleep onset and sleep offset;
-
Total sleep time: (h): the amount of sleep obtained during a sleep period;
-
Sleep efficiency (%): total sleep time expressed as a percentage of the sleep period;
-
Midpoint of sleep (h:min): time of the day at which the middle of the sleep period occurred;
-
Sleep onset latency (min): the period of time between bedtime and sleep start.
Data Analysis
All data and statistical analyses were conducted in RStudio (Version 1.1.463) using the R programming language (Version 4.0.5, Shake and Throw). Due to data being non-normally distributed, evidenced by significant Shapiro–Wilk tests (p < 0.05) and visual inspection of Q–Q plots, the median and interquartile range were used as measures of central tendency and dispersion, respectively.
The sleep regularity index (SRI) was calculated to reflect the night-to-night shifts in sleep cycles by accounting for changes in sleep onset and sleep offset over the longest interval of whole number of weeks (i.e. 7 days, 14 days, 21 days). The SRI is a metric that calculates the likelihood of sleep–wake cycles matching from 1 day to the next, which is then aggregated over a given period [13], using the following formula:
$${\text{SRI}} = - 100 + \frac{200}{{M\left( {N - 1} \right)}}\mathop \sum \limits_{j = 1}^{M} \mathop \sum \limits_{i = 1}^{N} \delta \left( {s_{i.j} , s_{i + 1.j} } \right)$$
where N is the number of days and M is the number of epochs per day (1-min epochs in this study). The function \(\delta \left({s}_{i.j}, {s}_{i+1.j}\right)\) is equal to one, when the sleep–wake state is the same 24 h apart, otherwise zero. For example, if sleep occurred at 22:00 h on Monday and occurred at 22:00 h on Tuesday, then 1 was coded for the minute epoch. Only diurnal shifts in sleep onset and offset times were used to calculate SRI (i.e. daytime naps and wake after sleep onset were not assessed). Scores of 100 for SRI indicate that sleep–wake cycles are identical between days over the period, whereas a score of 0 would indicate no overlap between consecutive sleep–wake cycles. Participants were then classified as regular (n = 42) or irregular (n = 46) sleepers if they were in the top or bottom quintile based on their SRI score, respectively. Additionally, variability in sleep onset, offset, sleep efficiency and total sleep time were captured in two ways for each athlete; firstly, the median absolute deviation (MAD) was calculated to reflect individualised variation, relative to each individual’s median values. Secondly, absolute variation was calculated as the difference to the previous night for each variable. For example, if an individual had a sleep onset time of 22:00 h on night one and a sleep onset time of 21:30 h on night two, they would have a sleep onset variation of 30 min, these scores were then aggregated over the study period.
Statistical Analyses
Aggregated data across all days were analysed for all participants. Differences in sleep characteristics between regular and irregular sleepers were assessed using the Wilcoxon rank sum test. The magnitude of differences were interpreted with effect sizes (r) and 95% confidence intervals as trivial ≤ 0.10 small, ≤ 0.3; medium ≤ 0.5; and large, > 0.5 [18] using the rcompanion package.
The influence of sleep behaviour over the data collection period on total sleep time and sleep efficiency was assessed using a number of machine learning algorithms to determine the most effective model. The algorithms included random forest regressions, elastic net regressions, boosted generalised additive models and multiple linear regressions using the caret package in R. Data were split into a training set (80%) and a testing set (20%). Subsequently, tenfold cross-validation with 5 repeats was used to train each model, with the final, tuned model being tested on the hold-out set. Separate models were built for total sleep time and sleep efficiency. Predictor variables included SRI, sleep onset time, sleep offset time, sleep midpoint time, individualised variation of sleep onset and offset (captured by the MAD), absolute sleep onset and offset variation: total sleep time and efficiency were also included in the efficiency and total sleep time models, respectively. The linear regression offered the best fit and accuracy on testing data, highlighted by the highest coefficient of determination (R2) and lowest normalised root mean square error (NRMSE; %), respectively. Variance inflation factor (VIF) was used to assess collinearity issues between variables, with a score of ≥ 10 used to remove variables; normality of the residuals was checked via a density plot.
