Using Machine Learning to predict daily PTSD and cannabis use disorder symptoms among non-treatment seeking veterans

项目摘要 创伤后应激障碍（PTSD）是退伍军人中最常见的并发症， 大麻使用问题然而，许多退伍军人未能寻求或参与医疗保健服务， 因此，症状严重程度的增加和相应的风险可能未被发现， 无人管理。尽管越来越多的人有兴趣通过提供医疗服务来接触不寻求治疗的退伍军人， 通过移动的设备进行及时干预，这种干预需要清楚地了解退伍军人 患有创伤后应激障碍和有问题的大麻使用者的症状升级的风险更高。虽然各方不断 努力确定退伍军人谁需要支持的心理健康和物质使用困难的时候， 尽管这些努力在重返社会（从部署返回后）方面取得了很小的成功。机器学习 一种特殊形式的人工智能，有助于将个人分类为风险状况， 改善风险评估和症状升级。机器学习算法应用于被动- 从移动的和可穿戴设备收集的数据（例如，加速度计数据，看屏幕的时间， 睡眠数据、锻炼、GPS数据）可能是一种有前途的、负担最小的策略， 最终为及时干预提供信息。来自智能手机和可穿戴设备的被动数据已经被 用于机器学习算法以预测PTSD和其他状况的风险（例如，抑郁症），但 没有被应用于预测PTSD和大麻使用或理解之间的相互作用 了以下条件虽然过去的研究已经成功地让退伍军人参与被动的数据收集， 战略的负担将低于主动数据收集，目前还不清楚这是否是一个可行的方法， 临床应用。因此，本申请的目的是了解被动数据的效用， 结合自我报告数据或单独预测PTSD症状的临床显著升级， 最近从美国退伍军人协会出院的非寻求治疗的退伍军人中有问题的大麻使用 军方75名有创伤暴露史的男性和女性非寻求治疗的退伍军人， 将在网上招募在重返平民社会后六个月内使用过大麻的人。参与者 将获得一个FitBit，并在智能手机（HeadSmart）上安装被动和主动数据收集应用程序。 他们将完成一项基线调查和三项每月一次的后续调查。此外，在观察期间， 退伍军人将完成PTSD症状和大麻使用的简短每日调查，被动数据将被 记录。被动和每日日记数据将在机器学习算法中进行分析，以预测症状 升级和将来的情况（例如，存在临床显著增加）（目的1）并了解 每日/每周症状相互作用（目标2）。我们亦会评估这做法的可行性和可接受性 (Aim 3）。这项研究的结果将最终为预防或早期干预工作提供信息， 需要大量退伍军人。

PROJECT SUMMARY Posttraumatic stress disorder (PTSD) is the highest co-occurring disorder among veterans who report problematic cannabis use. However, many veterans fail to seek or engage with health care services for both conditions, and as a result, increases in symptom severity and corresponding risk may go undetected and unmanaged. Although there is increasing interest in reaching non-treatment-seeking veterans by delivering just-in-time interventions via mobile devices, such interventions require a clear understanding of when veterans with PTSD and problematic cannabis use are at heightened risk for escalating symptoms. Despite ongoing efforts to identify veterans who need support for mental health and substance use difficulties at the time of reintegration (upon return from deployment), these efforts have achieved minimal success. Machine learning-- a special form of artificial intelligence that aids in classifying individuals into risk profiles--may have promise in improving risk assessment and symptom escalation. Machine learning algorithms applied to passively- collected data from mobile and wearable devices (e.g., accelerometer data, time spent looking at screens, sleep data, exercise, GPS data) could be a promising, minimal-burden strategy to detect periods of risk and ultimately inform just-in-time interventions. Passive data from smartphones and wearable devices has been used in machine learning algorithms to predict risk for PTSD and other conditions (e.g., depression), but has not been applied to the prediction of PTSD and cannabis use or the understanding of the interplay between these conditions. Although past research has successfully engaged veterans in passive data collection and this strategy would be lower-burden than active data collection, it is unclear whether this is a feasible approach in clinical applications. Thus, the objective of this application is to understand the utility of passive data, in conjunction with self-report data or alone, in predicting clinically significant escalations in PTSD symptoms and problematic cannabis use among non-treatment seeking veterans who have recently discharged from the military. Seventy-five male and female non-treatment-seeking veterans with a history of trauma exposure and past-month cannabis use who are within six months of civilian reintegration will be recruited online. Participants will be given a FitBit and install the passive and active data collection app on their smartphone (HeadSmart). They will complete a baseline and three monthly follow-up surveys. Further, over the observation period, veterans will complete brief daily surveys of PTSD symptoms and cannabis use, and passive data will be recorded. Passive and daily diary data will be analyzed in machine learning algorithms to predict symptom escalation and future caseness (e.g., presence of clinically significant increase) (Aim 1) and understand daily/weekly symptom interplay (Aim 2). We will also assess the feasibility and acceptability of this approach (Aim 3). The results of this research will ultimately inform prevention or early intervention efforts among this high-need population of veterans.