Novel computational techniques to detect the relationship between sitting patterns and metabolic syndrome in existing cohort studies.

摘要 代谢综合征是一组病症（血压升高、血糖升高、体脂过多 以及不正常的胆固醇或甘油三酯水平）一起发生，增加心脏病的风险 中风和糖尿病。流行病学研究表明，久坐对健康有害。 代谢指标，即使在调整身体活动（PA）后。急性实验室试验表明， 打破坐着的时间可以改善代谢因素。坐是所有人群中普遍存在的行为 按年龄、性别和种族分列。与代谢综合征因素的关联，如肥胖，也被认为是 在所有人群中。流行病学研究主要依赖于报告的坐着时间， 尤其是电视评论。最近的大型队列研究收集了来自臀部佩戴的加速度计的数据 并应用切割点（例如，每分钟100次计数）来估计久坐时间。等 许多研究中包括了这种装置，主要是因为它们测量PA的准确性 强度。主要用于干预试验以减少坐姿，大腿佩戴ActivPAL已被证明 更准确地评估姿势，并提供坐、站和坐-站转换的有效测量。到 迄今为止，很少有健康结局队列研究纳入ActivPAL。与ActivPAL和免费 对坐着时间的活体观察表明，100计数临界点低估了长时间的坐着 通过大大高估坐立转换。新研究表明我们如何积累坐姿 时间（即长或短的发作）与代谢健康结果相关，并且可能独立于总的 时间和PA。从加速度计数据中得出的久坐和代谢风险因素的研究结果如下： 不一致，并且可能是由于所采用的切割点的测量误差。在一小部分老年人中， 我们已经证明，新的机器学习方法可以大大改善女性、成年人和年轻人的生活。 估计长时间的坐姿和过渡。这些方法的进一步发展和测试将 支持对现有大型队列研究的有效应用，使用原始加速计数据，以改善 坐姿与代谢健康之间的关系。也有许多大型队列（如NHANES 2003/6），具有高质量的健康结果，但非原始加速度计计数数据，因此需要调整校准方法 还需要对坐着的时间进行非原始的估计，这对还不熟悉的研究人员来说是有吸引力的。 机器学习过程。我们建议使用7个现有的年龄和性别匹配的数据集（N= 20，000）， 跨越青年、成年人和老年人。我们将扩大训练规模， 算法来检测坐-站频率、长时间的坐、通常的回合持续时间和阿尔法（阿尔法的组合）。 持续时间和频率）。我们将测试算法对ActivPAL（地面实况）和新的性能 样本用客观健康结果评估预测有效性。我们将测试现有的 以及使用R2和预测均方误差（通过自举）和GEE技术的新技术。

Abstract Metabolic syndrome is a cluster of conditions (increased blood pressure, high blood sugar, excess body fat around the waist, and abnormal cholesterol or triglyceride levels) that occur together, increasing risk of heart disease, stroke and diabetes. Epidemiological studies have shown that prolonged sitting is deleterious to metabolic indicators, even after adjusting for physical activity (PA). Acute laboratory trials have shown that breaking up sitting time can improve metabolic factors. Sitting is a prevalent behavior in all population groups by age, gender and ethnicity. Associations with metabolic syndrome factors, such as obesity, have also been shown in all population groups. Epidemiological studies have mostly depended on reported sitting time, especially TV reviewing. More recently large cohort studies have collected data from hip worn accelerometers and applied a cut point (e.g., 100 counts per minute) on single axis data to estimate sedentary time. Such devices have been included in numerous studies, principally because of their accuracy to measure PA intensity. Primarily used in intervention trials to reduce sitting, the thigh worn ActivPAL has been shown to more accurately assess posture and provide valid measures of sitting, standing, and sit-stand transitions. To date, very few health outcome cohort studies have included the ActivPAL. Compared to the ActivPAL and free living observations of sitting time, the 100 count cut point has been shown to underestimate prolonged sitting by substantially overestimating sit-stand transitions. New studies are showing that how we accumulate sitting time (i.e. in long or short bouts) is associated with metabolic health outcomes, and may be independent of total sitting time and PA. Study results on prolonged sitting and metabolic risk factors from accelerometer data are inconsistent and may be due to measurement error in the cut points employed. In a small sample of older women, adults, and youth we have demonstrated that novel machine learned methods can greatly improve estimates of prolonged sitting and transitions. Further development and testing of these methods would support valid applications to existing large cohort studies with raw accelerometer data to improve estimates of associations between sitting patterns and metabolic health. There are also many large cohorts (e.g. NHANES 2003/6), with quality health outcomes, but non raw accelerometer count data, so calibration methods to adjust non raw estimates of sitting time are also needed and would be attractive to researchers not yet familiar with the machine learning process. We proposed to employ 7 existing data sets (N=20,000) matched for age and spanning youth, adults and older adults. We will scale up our training and test the performance of the refined algorithms to detect sit-stand frequencies, prolonged sitting, usual bout duration and Alpha (a combination of duration & frequency). We will test performance of the algorithms against ActivPAL (ground truth) and in new samples assess predictive validity with objective health outcomes. We will test differences between the existing and new techniques using R2 and mean-squared error of prediction (via bootstrapping) and GEE techniques.