Identifying bedrest using waist-worn triaxial accelerometers in preschool children

Abstract

Purpose: To adapt and validate a previously developed decision tree for youth to identify bedrest for use in preschool children.

Methods: Parents of healthy preschool (3-6-year-old) children (n = 610; 294 males) were asked to help them to wear an accelerometer for 7 to 10 days and 24 hours/day on their waist. Children with ≥3 nights of valid recordings were randomly allocated to the development (n = 200) and validation (n = 200) groups. Wear periods from accelerometer recordings were identified minute-by-minute as bedrest or wake using visual identification by two independent raters. To automate visual identification, chosen decision tree (DT) parameters (block length, threshold, bedrest-start trigger, and bedrest-end trigger) were optimized in the development group using a Nelder-Mead simplex optimization method, which maximized the accuracy of DT-identified bedrest in 1-min epochs against synchronized visually identified bedrest (n = 4,730,734). DT's performance with optimized parameters was compared with the visual identification, commonly used Sadeh's sleep detection algorithm, DT for youth (10-18-years-old), and parental survey of sleep duration in the validation group.

Results: On average, children wore an accelerometer for 8.3 days and 20.8 hours/day. Comparing the DT-identified bedrest with visual identification in the validation group yielded sensitivity = 0.941, specificity = 0.974, and accuracy = 0.956. The optimal block length was 36 min, the threshold 230 counts/min, the bedrest-start trigger 305 counts/min, and the bedrest-end trigger 1,129 counts/min. In the validation group, DT identified bedrest with greater accuracy than Sadeh's algorithm (0.956 and 0.902) and DT for youth (0.956 and 0.861) (both P<0.001). Both DT (564±77 min/day) and Sadeh's algorithm (604±80 min/day) identified significantly less bedrest/sleep than parental survey (650±81 min/day) (both P<0.001).

Conclusions: The DT-based algorithm initially developed for youth was adapted for preschool children to identify time spent in bedrest with high accuracy. The DT is available as a package for the R open-source software environment ("PhysActBedRest").

Conflict of interest statement

The authors have declared that no competing interests exist.

Figures

Fig 1. Graphical presentation of heuristic guidelines for visual identification of accelerometer recordings to estimate each bedrest period's start and end.

The smallest period to be considered a separate episode was 30 min for bedrest and 10 min for awake. Identification guidelines for marking bedrest-start required: (i) 10 consecutive minutes with ≤100 counts/min; (ii) followed by a 20-min period with consistent readings with ≤100 counts/min, but allowing for two minutes >100 counts/min. The first minute with ≤100 counts/min, identified in (i), was marked bedrest-start. Bedrest-end identification required: (i) identifying a minute with ≥500 counts/min; (ii) followed by 5 out of the next 9 minutes with ≥500 counts/min. The minute preceding the first minute with ≥500 counts/min, identified in (i), was marked bedrest-end. The period between bedrest-start and the next bedrest-end was marked bedrest, and the period between bedrest-end and the next bedrest-start was marked wake. The agreement between raters was assessed using kappa-statistics. To handle a disagreement between the raters in identifying 1-min epochs as bedrest or wake, we generated two separate sets of data named True1 and True2, respectively, for both the development and the validation group. In True1, the epochs in which there was disagreement were assumed to be truly wake, and in True2, they were assumed to be truly bedrest. Separate analyses were conducted for both True1 and True2. We report the True2 analysis as the algorithm was less accurate relative to True1. Thus, we report the more conservative estimate of identifying bedrest.

Fig 2. Simplified decision tree (DT) to identify triaxial accelerometer recordings (counts/epoch) as bedrest or wake.

The plots contain ~48 h data from a representative child (4-year-old, male) before (Step 1—left panel), during (Step 1 –right panel, Step 2, and Step 3), and after (Step 4) bedrest marking by the DT. The DT uses four algorithm parameter values (block length, threshold, bedrest-start trigger, and bedrest-end trigger) and has a four-step process to run through the data.