Using Artificial Intelligence to Optimize Delivery of Weight Loss Treatment

摘要 70%的美国成年人超重或肥胖，这对美国人来说是一个前所未有的挑战。 国家的卫生系统。有效的行为方案是存在的，但这些方案是密集的，长期的， 需要训练有素的临床医生，使其过于昂贵，从而限制了传播。 降低成本的方法包括用辅助专业人员取代训练有素的临床医生， 接触频率和/或自动干预。然而，尽管这些替代干预措施导致 平均重量损失相当低，重量损失的可变性高。具体来说，与 支持性问责模式，在高强度干预措施的参与者中， 接受低强度干预的一部分人实现了临床上显著的体重减轻。 一个理想的减肥治疗系统将提高结果，并通过匹配每个降低成本 参与者需要的干预措施，从而适应参与者的需求并节约资源 在不需要它们的地方。阶梯式护理就是这样一种系统，但其成功和失败都有。 从一些缺点。强化学习的创新人工智能（AI）策略 (RL)提供快速和重复变化的干预特征，不断“学习”哪些特征 为参与者提供最佳响应。我们的团队最近完成了人工智能减肥的试点 超重的成年人接受了简短的面对面减肥干预，然后随机 被分配接受3个月的非优化干预（即，12-分钟电话）或优化的 电话、短信和自动消息的组合，根据每个参与者的 根据体重和行为数据确定的每种干预措施的反应。正如假设的那样，我们实现了 以一小部分的时间成本获得相同的重量损失。这项研究计划招募320名超重的成年人， 提供1个月的基于小组的行为减肥治疗，然后将参与者随机分配到 继续以远程形式接受基于小组的行为减肥11个月（BWL-S），或 基于强化学习的治疗（BWL-AI）。根据我们的支持性问责制模型，BWL-AI 将基于持续监测的参与者数字数据改变模式、强度和顾问技能。的 拟议中的研究--这是同类研究中的第一个--将在几个方面扩大我们的试点，包括样本大小，持续时间， 以及人工智能系统选择的干预特征。这个项目的目的是测试假设， BWL-AI的减肥结果将等同于或优于BWL-S，并且每位参与者的成本 BWL-AI的每千克失重率低于BWL-S。其他包括表征AI系统（在 选择的干预措施），评估改进的人工智能系统的可行性和可接受性，评估 人工智能干预选择的心理和人口统计学预测因素，并调查 响应者和非响应者如何分配资源。

Abstract Seventy percent of American adults are overweight or obese, presenting an unprecedented challenge to the nation’s health systems. Effective behavioral programs exist, but these programs are intensive, long-term and require highly-trained clinicians, making them prohibitively expensive and thus limiting disseminability. Approaches to decreasing costs include replacing highly-trained clinicians with paraprofessionals, reducing contact frequency, and/or automating intervention. However, although these alternative interventions result in considerably lower average weight losses, variability of weight loss is high. Specifically, and consistent with a Supportive Accountability Model, a substantial minority of participants in high-intensity interventions receive no benefit, while a subset of those receiving low-intensity interventions achieve clinically significant weight loss. An ideal weight loss treatment system would enhance outcomes and reduce costs by matching each participant to the intervention he/she needs, thus adapting to participants’ needs and conserving resources where they are not needed. Stepped care represents one such system, but has had mixed success and suffers from a number of shortcomings. The innovative artificial intelligence (AI) strategy of reinforcement learning (RL) provides rapidly and repeatedly-varying features of intervention, continuously "learning" which features provide optimal responses for which participants. Our team recently completed a pilot of an AI weight loss system in which overweight adults received a brief in-person weight loss intervention and then were randomly assigned to receive 3 months of non-optimized interventions (i.e., 12-minute phone calls) or an optimized combination of phone calls, text exchanges, and automated messages, selected based on each participants’ response to each intervention as determined by weight and behavioral data. As hypothesized, we achieved equivalent weight losses at a fraction of the time cost. The proposed study would recruit 320 overweight adults, provide 1 month of group-based behavioral weight loss treatment and then randomize participants to either continue to receive group-based behavioral weight loss in a remote format for 11 months (BWL-S) or to reinforcement learning-based treatment (BWL-AI). In line with our Supportive Accountability model, BWL-AI would vary modality, intensity and counselor skill based on continuously-monitored participant digital data. The proposed study--the first of its kind--would expand on our pilot in several ways including sample size, duration, and features of intervention selected by the AI system. Aims of this project are to test the hypotheses that weight loss outcomes in BWL-AI will be equivalent to or better than BWL-S, and that the cost per participant and per kg of lost weight will be less in BWL-AI than in BWL-S. Other include characterizing the AI system (in terms of interventions selected), assessing feasibility and acceptability of the refined AI system, evaluating psychological and demographic predictors of AI intervention selection and investigating differences between responders and non-responders in how the AI system allocates resources.