Automated Ecological Video Identification of Physical Activity (E-VIP) Software

 描述(由申请人提供)：本提案通过推进对普通环境中的人口PA的评估，解决了PAR-12-197“改善饮食和体力活动(PA)评估”的多个优先事项。缺乏体力活动每年导致美国≈200,000人死亡，全球500万人死亡。大约10%的乳腺癌和10%的结肠癌可归因于PA不足，而缺乏运动导致总过早死亡的9%。公园、学校和青年体育是与大多数人口相关的PA的重要贡献者。在这些环境中支持PA将有助于提高PA的人口水平以及疾病预防和控制。我们将开发一个新的视频分析软件系统，名为E-VIP(生态视频识别PA)，它将从PA相关环境的视频记录中估计PA的人数和聚集量。E-VIP将提供自动化和持续(而不是瞬时)评估，这将允许持续的反馈，为适应性干预和决策提供信息。E-VIP基于计算机科学方法，用于统计人群数量和识别行为。我们将对来自2个公园、2所学校和2个青少年体育场所的32,400秒视频训练E-VIP机器学习算法。参与者将在佩戴加速度计以新陈代谢当量(METS)测量PA强度的同时进行各种活动。测试阶段包括在4个公园、4所学校和4个青少年体育场所捕获900个5分钟的自由生活行为视频，其中一半的场所/区域与培训阶段相同。将在培训和测试阶段捕获一系列活动和人员密度 以最大限度地覆盖更困难的高密度情况。将对每个时间样本进行系统的直接观察，为检验效度提供标准测量。将测试的指标包括平均人数、每个活动类别(久坐、轻度、适度、剧烈、非常剧烈)的平均比例，以及在5分钟(或任何给定)时间段内设置的PA Meet-Minents总和。将探索性别和年龄组分类器。将体育专家授课的体育课与课堂教师的体育课进行比较，以检验E-VIP的结构效度。将使用Bland Altman方法和组内相关系数来评估一致性。将使用主持人分析来评估潜在的误差源，如遮挡(例如，树木、阴影、其他人)。通过将生态PA评估自动化，E-VIP将具有广泛使用的可行性。E-VIP的持续评估能力将通过收集比现有直接观测工具收集的更高分辨率的数据来提高精度。当通过常用的安全摄像头或网络摄像头嵌入到特定环境中，或通过故意放置录像机，E-VIP将能够进行持续的评估，为公共卫生监测、干预和基于证据的决策提供信息(例如，优化干预策略、监测学校体育任务、在环境修改之前和之后提供需求评估)。

 DESCRIPTION (provided by applicant): This proposal addresses multiple priorities of PAR-12-197 "Improving Diet and Physical Activity (PA) Assessment" by advancing assessment of population PA in common settings. Physical inactivity is responsible for ≈200,000 deaths in the US and 5 million deaths worldwide annually. About 10% of breast and 10% of colon cancers are attributable to insufficient PA, and inactivity causes 9% of total premature mortality. Parks, schools and youth sports are important contributors of PA that are relevant to a majority of the population. Supporting PA in these settings would contribute to improvements in population levels of PA and disease prevention and control. We will develop a novel video analysis software system, named E-VIP (Ecological Video Identification of PA), which will estimate the number of people and aggregated volume of PA from video recordings of PA- relevant settings. E-VIP will provide automated and continuous (rather than momentary) assessment, which will allow ongoing feedback to inform adaptive interventions and decision making. E-VIP is based on computer science methods used to count crowds and identify behaviors. We will train the E-VIP machine learning algorithm on 32,400 seconds of video from 2 parks, 2 schools, and 2 youth sports settings. Participants will engage in a variety of activities while wearing accelerometers to capture PA intensity in Metabolic Equivalents (METs). The testing phase involves capturing 900, 5-minute videos of free-living behavior across 4 parks, 4 schools, and 4 youth sports settings, with half of the settings/zones being the same as from the training phase. A range of activities and density of people will be captured in both the training and testing phase to maximize coverage of more-difficult high-density situations. Systematic direct observation will be conducted on each time sample to provide the criterion measure for testing validity. The metrics that will be tested include the average number of people, average proportion of people in each activity category (sedentary, light, moderate, vigorous, very vigorous), and sum PA MET-minutes in the setting over the 5-minute (or any given) time period. A gender and age group classifier will be explored. PE classes taught by PE specialists vs classroom teachers will be compared to test construct validity of E-VIP. Bland Altman methods and intraclass correlation coefficients will be used to assess agreement. Potential sources of error such as occlusions (e.g., trees, shadows, other people) will be assessed using moderator analyses. By automating ecological PA assessment, E-VIP will be feasible for widespread use. E-VIP's capability of continuous assessment will improve precision by collecting higher resolution data than collected by existing direct observation tools. When embedded in specific settings through commonly-used security cameras or webcams, or by purposefully placing video recorders, E-VIP will be capable of ongoing assessment which will inform public health surveillance, intervention, and evidence-based decision making (e.g., optimizing intervention strategies, monitoring school Physical Education mandates, providing needs assessment prior to and evaluation after environmental modifications).