Validating Machine-Learned Classifiers of Sedentary Behavior and Physical Activit

描述（由申请人提供）：久坐行为（SB）是一种健康风险，独立于中等到剧烈的身体活动（MVPA）。流行病学数据显示，SB与许多癌症、心血管疾病以及多种代谢功能障碍标志物之间存在一致的关联。减少SB被推荐为一种可行的新的癌症预防策略。髋部安装的加速度计是一种基于现场的PA标准测量，但不适用于测量SB。错误分类最常发生在躯干位移有限的运动中，在负载或倾斜下，在振动表面上（例如，在汽车中），或者当它们太小而无法与非磨损时间区分时。这是不可接受的，因为我们可能会错误地描述SB和PA的患病率。测量误差掩盖了行为与健康之间的真正关系，干预措施的影响可能无法被发现。[我们的研究将在实验室研究的基础上进行改进，这些实验室研究几乎没有在驾驶或看电视等自然环境中测试SB。]基于实验室的SB机器学习算法由人工起点和终点辅助。基于自由生活环境研究的算法将更加一般化，更适用于干预研究。在这项研究中，我们将使用加速度计、全球定位系统（GPS）和地理信息系统（GIS）数据，对从SB到MVPA的连续行为进行改进和验证新的机器学习分类算法。这项研究将重点关注最常被错误分类的行为：被编码为轻度的中等强度活动，被编码为久坐的轻度活动，以及被编码为轻度或非磨损时间的久坐活动。[我们将在目前自我报告和加速计估计的基础上改进30-50%。我们将着重于四种主要的行为类型：躺着、坐着、站着和移动运动。总共210名参与者（年龄6-10岁，n=70； 16-55岁，n=70； 65-85岁，n=70）将在2年的时间内被招募，并将佩戴3个ActiGraph加速计（两个髋关节，一个手腕）；一个GPS设备，一个SenseCam（一种自动图像捕捉设备），两个工作日和一个周末。[子样本将重复上述步骤3天]。每天有6小时的时间，参与者还将被训练有素的编码员直接观察，他们将使用为iPad开发的新型便携式行为评估系统记录自由生活行为。直接观察数据将为每隔一秒记录一次的注释数据文件提供行为的“基础真相”。对于没有直接观察到的自由生活行为，将使用SenseCam图像。机器学习将采用核方法。灵敏度、特异性、准确性和其他ROC图方法将用于比较来自：(a)单轴与多轴加速度计数据的分类器；(b)运动“计数”与原始加速度数据；(3)安装在臀部和手腕上的加速度计。当加入GPS和GIS数据时，分析将确定灵敏度和特异性的改进。我们将评估对特定人群分类器的需求，并量化目前使用的切割点的测量误差。

DESCRIPTION (provided by applicant): Sedentary behavior (SB) is a health risk, independent of moderate-to-vigorous physical activity (MVPA). Epidemiologic data reveal consistent associations between SB and numerous cancers, cardiovascular disease, as well as multiple markers of metabolic dysfunction. Reducing SB is recommended as a viable new cancer prevention strategy. Hip mounted accelerometry is a field-based criterion measure of PA but is inadequate for measuring SB. Misclassification most often occurs when movements are performed with limited trunk displacement, under load or on an incline, on a vibrating surface (e.g., in a car), or when they are too small to be distinguished from non-wear time. This is unacceptable because we may then mischaracterize the prevalence of both SB and PA. Measurement error obscures true relationships between behavior and health, and the effects of interventions may go undetected. [Our study will improve upon laboratory studies that do little to test SB in natural environments such as driving or watching TV. Lab-based machine learning algorithms for SB are aided by the artificial start and end points. Algorithms based upon research in free living settings will be more generalizable and applicable to intervention research]. In this study we will refine and validate new machine-learned classification algorithms for a continuum of behaviors from SB to MVPA using accelerometer, Global Positioning System (GPS), and Geographic Information System (GIS) data. This study will focus on behaviors that are most frequently misclassified: moderate intensity activities that are coded as light, light activities that are coded as sedentary, and sedentary activities that are coded as light or non wear time. [We will improve upon current self-report and accelerometer estimates by 30-50%.] We will focus on four primary behavioral classes: lying, sitting, standing, and ambulatory locomotion. A total of 210 participants (ages 6-10 yrs, n = 70; 16-55 yrs, n=70; 65-85 yrs, n= 70) will be recruited over a 2-yr period and will wear 3 ActiGraph accelerometers (two hip, one wrist); a GPS device, and a SenseCam (an automatic image capture device) for two weekdays and 1 weekend day. [A subsample will repeat the procedures for a further 3 days]. For a 6 hr period on each of these days, participants will also be directly observed by trained coders who will record free living behaviors using a novel portable behavioral assessment system developed for the iPad. Direct observation data will provide 'ground-truths' of behavior for an annotated data file recorded at one-second intervals. For free-living behaviors not directly observed, SenseCam images will be used. Machine learning Kernel methods will be employed. Sensitivity, specificity, accuracy and other ROC graph methods will be used to compare classifiers derived from: (a) single axis vs. multi axis accelerometer data; (b) movement 'counts' vs. raw acceleration data; and (c) hip vs. wrist mounted accelerometers. Analyses will determine the improvement in sensitivity and specificity when GPS and GIS data are added. We will evaluate the need for population specific classifiers and quantify measurement error in cut-points that are currently employed.