Detecting accelerometer non-wear periods using change in acceleration combined with rate-of-change in temperature

Abstract

Background: Accelerometery is commonly used to estimate physical activity, sleep, and sedentary behavior. In free-living conditions, periods of device removal (non-wear) can lead to misclassification of behavior with consequences for research outcomes and clinical decision making. Common methods for non-wear detection are limited by data transformations (e.g., activity counts) or algorithm parameters such as minimum durations or absolute temperature thresholds that risk over- or under-estimating non-wear time. This study aimed to advance non-wear detection methods by integrating a 'rate-of-change' criterion for temperature into a combined temperature-acceleration algorithm.

Methods: Data were from 39 participants with neurodegenerative disease (36% female; age: 45-83 years) who wore a tri-axial accelerometer (GENEActiv) on their wrist 24-h per day for 7-days as part of a multi-sensor protocol. The reference dataset was derived from visual inspection conducted by two expert analysts. Linear regression was used to establish temperature rate-of-change as a criterion for non-wear detection. A classification and regression tree (CART) decision tree classifier determined optimal parameters separately for non-wear start and end detection. Classifiers were trained using data from 15 participants (38.5%). Outputs from the CART analysis were supplemented based on edge cases and published parameters.

Results: The dataset included 186 non-wear periods (85.5% < 60 min). Temperature rate-of-change over the first five minutes of non-wear was - 0.40 ± 0.17 °C/minute and 0.36 ± 0.21 °C/minute for the first five minutes following device donning. Performance of the DETACH (DEvice Temperature and Accelerometer CHange) algorithm was improved compared to existing algorithms with recall of 0.942 (95% CI 0.883 to 1.0), precision of 0.942 (95% CI 0.844 to 1.0), F1-Score of 0.942 (95% CI 0.880 to 1.0) and accuracy of 0.996 (0.994-1.000).

Conclusion: The DETACH algorithm accurately detected non-wear intervals as short as five minutes; improving non-wear classification relative to current interval-based methods. Using temperature rate-of-change combined with acceleration results in a robust algorithm appropriate for use across different temperature ranges and settings. The ability to detect short non-wear periods is particularly relevant to free-living scenarios where brief but frequent removals occur, and for clinical application where misclassification of behavior may have important implications for healthcare decision-making.

Keywords: Accelerometry; Algorithm; Behavior classification; Non-wear detection; Physical activity; Remote monitoring; Wearables.

Fig. 1

Sample temperature profile for sleep versus non-wear

Fig. 2

Relationship between starting temperature (°C) and maximum negative rate-of-change (°C/minute) at 1, 3, 5, and 10 minutes (Fig. 2A-D, respectively) using data from the training dataset. Dashed line indicates zero slope. Shaded band represents the 95% confidence interval

Fig. 3

Distribution of non-wear periods by duration (minutes) with periods less than 60 minutes (top) and greater than 60 minutes (bottom)

Fig. 4

Average rate-of-change in temperature (°C/minute) by minute into non-wear period for data from the training dataset