Improving Suicide Risk Prediction with Social Determinants Data

摘要 2019年美国自杀死亡人数为47511人，自杀率自1999年以来增长了39%。 自杀预防是NIMH研究的优先事项。估计机器学习算法预测的最新研究进展 自杀风险已经非常成功了。这些模型已经作为常规预防的一部分进行了实施 医疗系统中的项目，如Kaiser Permanente Washington、HealthPartners和退伍军人健康 行政部门。尽管取得了这些成功，但现有的模型仍存在重大缺陷。相当大比例的 自杀发生在预计风险较低的医疗保健就诊(以及可能发生干预的地方)之后 否则)。这些模型目前不包括任何关于自杀的社会决定因素的信息(例如，独居， 经济压力)或负面生活事件(NLE)，如离婚、破产和刑事逮捕。添加社交网络 将决定因素数据和NLE数据添加到模型中可以提高预测精度。本研究的具体目的是：(1) 使用1500多个描述社会决定因素的带有日期戳的变量来扩展和增强风险预测数据集 利用社会决定因素和NLE数据构建和评估自杀风险预测模型 (3)使用社会决定因素、NLE和医疗保健数据构建和评估自杀风险预测模型 共同评估社会决定因素、NLE和医疗保健预测因素之间的交互作用术语。一个例子 “抑郁症诊断”与“最近30天内提出离婚”之间存在互动关系。这将是第一个大规模的研究 将个人层面的、带有日期戳的社会决定因素和NLE测量纳入机器学习自杀风险 预测模型。在成功完成这项研究后，我们希望知道这些新数据在多大程度上被纳入 有助于自杀风险预测模型的准确性。这将是下一步实现 更好的自杀预防计划和降低总体自杀率。

ABSTRACT Suicide accounted for 47,511 deaths in the United States in 2019 and the suicide rate has increased by 39% since 1999. Suicide prevention is an NIMH research priority. Recent research in estimating machine learning algorithms to predict suicide risk has been tremendously successful. The models have been implemented as part of routine prevention programs in health systems such as Kaiser Permanente Washington, HealthPartners, and the Veterans Health Administration. Despite these successes, existing models have important shortcomings. A significant proportion of suicides followed healthcare visits where the predicted risk was low (and where an intervention might have taken place otherwise). The models do not currently include any information about social determinants of suicide (e.g., living alone, financial stress) or negative life events (NLE), such as divorce, bankruptcy, and criminal arrest. Adding social determinants data and NLE data to models may improve predictive accuracy. The specific aims of this study are: (1) expand and enhance the risk prediction dataset with over 1500 date-stamped variables describing social determinants of suicide risk and NLE; (2) construct and evaluate suicide risk prediction models using social determinants and NLE data alone; (3) construct and evaluate suicide risk prediction models using social determinants, NLE and healthcare data together and estimate interaction terms between social determinants, NLE, and healthcare predictors. An example would be “depression diagnosis” interacted with “divorce filing in last 30 days”. This will be the first large scale study to incorporate individual-level, date-stamped measures of social determinants and NLE into machine learning suicide risk prediction models. Upon successful completion of this study we expect to know how much incorporating these new data contributes to the accuracy of suicide risk prediction models. This will be an important next step towards implementing better suicide prevention programs and reducing overall suicide rates.