The University of Queensland, School of Public Health, Herston, Queensland, Australia.
Institute of Health and Biomedical Innovation at Queensland, Centre for Children's Health Research, Queensland University of Technology, South Brisbane, Queensland, Australia.
PLoS One. 2018 Mar 7;13(3):e0193971. doi: 10.1371/journal.pone.0193971. eCollection 2018.
Most wearable devices that measure movement in workplaces cannot determine the context in which people spend time. This study examined the accuracy of Bluetooth sensing (10-second intervals) via the ActiGraph GT9X Link monitor to determine location in an office setting, using two simple, bespoke algorithms.
For one work day (mean±SD 6.2±1.1 hours), 30 office workers (30% men, aged 38±11 years) simultaneously wore chest-mounted cameras (video recording) and Bluetooth-enabled monitors (initialised as receivers) on the wrist and thigh. Additional monitors (initialised as beacons) were placed in the entry, kitchen, photocopy room, corridors, and the wearer's office. Firstly, participant presence/absence at each location was predicted from the presence/absence of signals at that location (ignoring all other signals). Secondly, using the information gathered at multiple locations simultaneously, a simple heuristic model was used to predict at which location the participant was present. The Bluetooth-determined location for each algorithm was tested against the camera in terms of F-scores.
When considering locations individually, the accuracy obtained was excellent in the office (F-score = 0.98 and 0.97 for thigh and wrist positions) but poor in other locations (F-score = 0.04 to 0.36), stemming primarily from a high false positive rate. The multi-location algorithm exhibited high accuracy for the office location (F-score = 0.97 for both wear positions). It also improved the F-scores obtained in the remaining locations, but not always to levels indicating good accuracy (e.g., F-score for photocopy room ≈0.1 in both wear positions).
The Bluetooth signalling function shows promise for determining where workers spend most of their time (i.e., their office). Placing beacons in multiple locations and using a rule-based decision model improved classification accuracy; however, for workplace locations visited infrequently or with considerable movement, accuracy was below desirable levels. Further development of algorithms is warranted.
大多数可穿戴设备在工作场所测量运动时无法确定人们所处的环境。本研究通过 ActiGraph GT9X Link 监测器（蓝牙感应，10 秒间隔），使用两个简单的定制算法，检验了在办公室环境下确定位置的蓝牙感应的准确性。
在一个工作日（平均±SD 6.2±1.1 小时）中，30 名办公室工作人员（30%为男性，年龄 38±11 岁）同时佩戴胸部摄像头（视频记录）和手腕及大腿上的蓝牙监测器（初始化为接收器）。额外的监测器（初始化为信标）放置在入口处、厨房、影印室、走廊和佩戴者的办公室。首先，根据该位置的信号存在/缺失情况（忽略所有其他信号），预测参与者在每个位置的存在/缺失情况。其次，同时利用多个位置收集的信息，使用一个简单的启发式模型来预测参与者所处的位置。根据 F 分数，测试每个算法的蓝牙确定位置与摄像机的结果。
单独考虑各个位置时，办公室的准确性非常高（大腿和手腕位置的 F 分数分别为 0.98 和 0.97），但其他位置的准确性较差（F 分数为 0.04 至 0.36），主要原因是误报率较高。多位置算法对办公室位置具有很高的准确性（两个佩戴位置的 F 分数均为 0.97）。它还提高了其他位置的 F 分数，但并不总是达到表示良好准确性的水平（例如，两个佩戴位置的影印室的 F 分数均约为 0.1）。
蓝牙信号功能有望确定工人大部分时间（即办公室）所处的位置。在多个位置放置信标并使用基于规则的决策模型可以提高分类准确性；但是，对于很少访问或移动较大的工作场所位置，准确性低于理想水平。需要进一步开发算法。
