Development and clinical interpretation of machine learning emergency department suicide prediction algorithms using electronic health records and claims

项目摘要 预防自杀是美国医疗保健面临的重大公共卫生挑战之一 系统寻求精神健康投诉紧急护理的人短期内 非致命性自杀事件和自杀的风险。然而，识别高风险患者具有挑战性， 风险波动的方式知之甚少。评估风险尤其困难， 紧急情况下，获取患者心理健康史的机会往往有限。的 拟议的项目旨在通过配对数据挖掘和 机器学习方法与丰富的数据源，以发展短期预测 精神健康的ED患者的非致命性自杀事件和自杀模型 问题 本研究的具体目标是：1）应用先进的机器学习数据分析 电子健康记录（EHR）数据的技术，以开发临床上丰富的艾德描述 预测自杀和非致命性自杀事件的精神健康患者特征， 天的随访期; 2）使用纵向和时间特征的EHR和索赔数据， 艾德心理健康访视前180天，以产生临床可解释的自杀， 自杀事件风险评分;和3）召集艾德医师以增强模型开发， 临床可解释性和自杀风险评估临床决策支持工具的实用性。 我们将通过利用几种不同的复杂机器学习来实现这些目标 现有纵向临床和服务使用信息的分析方法。我们寻求 制定自杀症状和自杀死亡的时间点，短期风险评分， 驱动该风险的临床特征，可用于告知临床风险评估， 管理向急诊科提出精神健康投诉的患者。风险算法将 使用来自大型组合EHR和索赔的健康信息进行开发和验证 超过2400万商业保险患者的数据集，与全国死亡率相关 指数.研究结果将产生关于患者特定风险因素和潜在目标的新见解 进行干预。通过利用大多数医疗保健系统通用的数据源， 有效的计算机算法，这种方法有潜力开发临床解释 在艾德评估时和处置后的自杀风险评分。这将有助于前- 一线临床医生在高风险期将精力集中在高风险患者身上， 关于自杀风险的干预决定。

Project Summary Preventing suicide is one of the great public health challenges facing the US health care system. People who seek emergency care for mental health complaints are at high short-term risk of non-fatal suicide events and suicide. Yet identifying high-risk patients is challenging as risk fluctuates in a poorly understood manner. It is especially difficult to evaluate risk in emergency settings, where access to the patient's mental health history is often limited. The proposed project seeks to address this critical knowledge gap by pairing data mining and machine learning methods with rich data sources in order to develop short-term prediction models of non-fatal suicidal events and suicide for patients presenting to EDs with mental health problems. The specific aims of this study are to 1) apply advanced machine learning data analytic techniques to electronic health record (EHR) data to develop a clinically rich description of ED mental health patient characteristics that predict suicide and non-fatal suicidal events over a 90- day follow-up period; 2) use longitudinal and temporal features of EHR and claims data from the 180 days preceding the ED mental health visit to generate clinically interpretable suicide and suicidal event risk scores; and 3) convene ED physicians to enhance model development, clinical interpretability, and utility of a suicide risk assessment clinical decision support tool. We will achieve these aims by leveraging several different sophisticated machine learning analytic methods of existing longitudinal clinical and service use information. We seek to develop point-in-time, short-term risk scores for suicidal symptoms and suicide death and the clinical features that drive that risk that may be used to inform clinical risk assessment and management of patients who present to EDs with mental health complaints. Risk algorithms will be developed and validated using health information from a large combined EHR and claims dataset with over 24 million commercially insured patients, which is linked to the National Death Index. Findings will yield new insights regarding patient-specific risk factors and potential targets for intervention. By drawing on data sources common to most health care systems and using efficient computer algorithms this approach has the potential to develop clinically interpretable suicide risk scores at the point of ED evaluation and following disposition. This will help front- line clinicians focus their efforts on high risk patients during high risk periods to inform intervention decisions about suicide risk.