Methodological and data-driven approach to infer durable behavior change from mHealth data

摘要 穷 癌症， 冗长的， 实现 饮食和体力活动(PA)行为是心脏代谢性疾病最普遍的危险因素 可以通过治疗来预防疾病。然而，大多数节食、PA和减肥干预措施都很昂贵， 也很繁重。如果我们能知道人们什么时候进行干预，这些干预措施可能会更具成本效益 一种可持续的健康行为改变模式，使治疗可以逐渐减少，然后停止 没有行为上的复发。习惯养成理论可能被认为可以解决这个问题，但他们 已经证明不能用来指导治疗决定，因为它们没有具体说明可衡量的标准 可靠地检测持久行为模式的获取。因此，我们建议确定以下行为模式 在干预后持续维持目标水平的行为改善之前和预测 结束了。判断干预是否持久地影响了行为的衡量标准 改善是作为饮食、PA和减肥干预措施的一部分收集的。具体来说，参与者 持续以数字方式自我监控他们的行为，同时将评估反馈给他们，以告知他们 朝着目标前进。我们将分析在6项移动健康试验中收集的自我监测措施， 14年，超过1,600名参与者和超过147,000名日常观察，以评估何时 干预持久地牵涉到有针对性的行为，这一点得到了他们在治疗后的可靠持久性的验证。 干预。我们将使用位置比例模型来量化变化，不仅是在绝对水平(位置) 不仅是一种行为，而且还体现在其人内的可变性(尺度)上。我们假设耐久行为的诱导 改变需要改善地点(增加健康行为；减少不健康行为) 以及规模的降低(即行为一致性的提高)。目标1将应用现有的位置比例 方法检验有效干预将改善位置和缩小规模的假设。 所有试验中的有针对性的行为。因为现有的方法只测量组级别的规模，并且 我们无法衡量个性化治疗所需的个人行为一致性的变化 适应，目标2将扩展位置比例尺方法，以使个人能够估计 行为的一致性。从新方法得出的估计将被分析，以了解哪些参数 干预过程中的行为变化与治疗后的维持性最相关。最后，《目标3》将 将机器学习应用于扩展的位置-比例混合模型的估计，以建立范围和 预测治疗后行为维持的行为模式。这些结果将为行为干预科学提供信息，并通过指导所需剂量的实时决策来提高治疗效率 以及行为治疗的持续时间。

Abstract Poor cancers, lengthy, achieve diet and physical activity (PA) behaviors, the most prevalent risk factors for cardiometabolic diseases and can be treated to prevent disease. However, most diet, PA, and weight loss interventions are costly, and burdensome. Theseinterventions could be more cost-efficient if we could tell when people a sustainable pattern of health behavior change so that treatment could be tapered and then stopped without behavioral relapse. Theories of habit formation might be assumed to address this problem, but they have not proved actionable to guide treatment decisions because they do not specify measurable criteria to reliably detect acquisition of a durable behavior pattern. Hence, we propose to identify behavior patterns that precede and predict maintenance of target-level behavioral improvement that persist after an intervention ends. The measurements needed to tell whether an intervention has durably entrained behavioral improvement are collected as part of diet, PA, and weight loss interventions. Specifically, participants continuously self-monitor their behavior digitally while assessments are relayed back to inform them about progress toward goals. We will analyze self-monitoring measures collected in 6 mHealth trials, conducted over 14 years among over 1,600 participants and more than 147,000 daily observations, to assess when an intervention has durably entrained targeted behaviors, as validated by their reliable persistence post- intervention. We will use location scale modeling to quantify change not only in the absolute level (location) of a behavior but also in its within-person variability (scale). We posit that the induction of durable behavior change requires both improvement in location (increases for healthy behaviors; decreases for unhealthy ones) and decrease in scale (i.e., increased behavioral consistency). Aim 1 will apply existing location scale methods to test the hypothesis that effective interventions will improve the location and reduce the scale of targeted behaviors across all trials. Because existing methods only measure scale at the group level and cannot measure the change in an individual's behavioral consistency that we need to personalize treatment adaptation, Aim 2 will extend location scale methods to enable individual estimation of the rate of change in behavioral consistency. Estimates derived from the new method will be analyzed to learn which parameters of behavior change during intervention are most associated with maintenance post-treatment. Finally, Aim 3 will apply machine learning to estimates from the extended location-scale mixed models to establish ranges and behavioral patterns that predict behavioral maintenance post-treatment. These resultswill inform behaviorinterventionscience and improve treatment efficiency by guiding real-timedecisions about the needed dosage and duration of behavioral treatments.