Johns Hopkins Bloomberg School of Public Health, 415 N. Washington Street, Baltimore, MD, 21205, United States.
Johns Hopkins Bloomberg School of Public Health, 415 N. Washington Street, Baltimore, MD, 21205, United States.
Gait Posture. 2020 Sep;81:205-212. doi: 10.1016/j.gaitpost.2020.07.149. Epub 2020 Aug 7.
Step count monitors are frequently used in clinical research to measure walking activity. Systematically determining valid days and extracting informative measures of walking beyond total daily step count are among major analytical challenges.
We introduce a novel data-driven anomaly detection algorithm to determine days representing typical walking activity (valid days) and examine the value of measures of walking fragmentation beyond total daily step count.
StepWatch data were collected on 230 adults with severe foot or ankle fractures. Average steps per minute (SC), average steps per active minute (SCA), active to sedentary transition probability (ASTP) and sedentary to active transition probability (SATP) were computed for each participant. The joint distribution of these measures was used to identify and eliminate invalid days through a multi-step process based on the support vector machine. The value of SCA, ASTP and SATP beyond SC were assessed by regressing Short Musculoskeletal Functional Assessment (SMFA), a measure of self-reported function, on these measures and quantifying the increase in the adjusted R-squared. In an unsupervised comparison, the total joint variability of SCA, ASTP and SATP was decomposed into the variability explained by SC and the unique variability of these three measures.
Of the 4,448 days in the original data set, 39% were determined invalid. Individuals with higher average SC had higher SCA, lower ASTP and higher SATP. Measures of fragmentation (SCA, ASTP and SATP) explained 25% more of the variability in SMFA compared with SC alone. Approximately 41% of the variability in SCA, ASTP and SATP could not be explained by SC suggesting that these three measures provide unique information about walking patterns.
Applying SVM and quantifying fragmentation in walking bouts for step count data can help to more precisely assess activity in clinical studies employing this modality.
计步器常用于临床研究中以测量步行活动。系统地确定有效天数并提取总日步数之外的有意义的步行指标是主要的分析挑战之一。
我们引入了一种新颖的数据驱动异常检测算法，以确定代表典型步行活动的天数（有效天数），并研究总日步数之外的步行碎片化指标的价值。
在 230 名患有严重足部或踝关节骨折的成年人中收集了 StepWatch 数据。为每位参与者计算了平均每分钟步数 (SC)、活跃分钟内的平均步数 (SCA)、活跃到久坐的转换概率 (ASTP) 和久坐到活跃的转换概率 (SATP)。通过基于支持向量机的多步骤过程，使用这些指标的联合分布来识别和消除无效天数。通过将这些指标回归到短肌肉骨骼功能评估 (SMFA)，评估 SCA、ASTP 和 SATP 对 SC 的价值，SMFA 是一种自我报告功能的衡量标准，并量化调整后的 R 方的增加。在无监督比较中，SCA、ASTP 和 SATP 的总联合变异性被分解为 SC 解释的变异性和这三个指标的独特变异性。
在原始数据集的 4448 天中，39%被确定为无效。平均 SC 较高的个体具有较高的 SCA、较低的 ASTP 和较高的 SATP。与单独的 SC 相比，碎片化指标（SCA、ASTP 和 SATP）解释了 SMFA 变异性的 25%。SCA、ASTP 和 SATP 的变异性中约有 41%不能用 SC 来解释，这表明这三个指标提供了关于步行模式的独特信息。
在使用这种模式的临床研究中，应用 SVM 和量化步计数数据中的碎片化可以帮助更精确地评估活动。
