Predicting suicide attempt in youth by integrating EHR, clinical, cognitive and imaging data

总结。青少年自杀是一个日益严重的健康问题，但目前的临床实践还不够及时。 识别有自杀企图(SA)风险的青少年。这项研究的总体目标是使用数据驱动 机器学习方法，以促进初级保健儿科环境中青少年SAS的初级预防。 临床指南建议从12岁开始筛查抑郁症，被认为是自杀风险的替代。 儿科环境。拟议的研究旨在确定可以在早期收集的变量(特征) 青春期，并有助于预测SA在青春期后期。这项研究将利用 投资于以前的项目：一项使用电子健康记录(EHR)预测SAS和死亡的研究 匹兹堡大学医学中心(UPMC)医院和费城神经发育队列 (PNC)，这包括对约9,500名青年的全面表型分析。这些先前的努力将被整合到 开发和优化来自费城儿童医院(CHOP)网络的青年SA预测，来自 我们有约40,000人的数据，这些人在2014-2018年间接受了SA病史筛查 (n~1500)。首先，在CHOP数据集中，我们将基于UPMC数据生成预测模型，测试其 青年人口预测效度，然后开发、优化和交叉验证这些预测 使用CHOP EHR数据作为训练集的模型(目标1)。其次，在PNC数据集中，我们将使用多个数据 对有自杀意念的青少年进行分类的类型(人口统计、行为、认知、成像)(SI，n~750)和 确定指示SI的特征(可能可修改)，并可能指出潜在的机制 基础青年SI(目标2)。最后，在936名年轻人(49名患有SAS)中，既有CHOP EHR数据，也有 研究PNC评估在平均年龄11岁(T1)，SA筛查前约5年(T2)进行，我们将 测试AIMS 1和AIMS 2中的模型的有效性，并确定在T1时收集的数据特征 改进/优化/超越仅依赖电子病历数据的SAS预测(目标3)。建议进行的研究 依靠由Barzilay博士(Pi)、儿童- 有自杀研究经验的青少年精神病学家和PNC自杀相关表型分析 数据；徐博士(Pi)，机器学习专家，开发了SA和死亡的预测算法 使用UPMC数据；以及对实现研究目标至关重要的合作者，拉克尔·古尔博士担任首席研究员 谁建立了PNC，鲁本·古尔博士开发了PNC神经认知评估工具，以及Dr。 他将提供自杀预测研究方面的专业知识。团队的访问权和熟悉度 Chop EHR和PNC数据资源，再加上其跨学科的专业知识，创造了一个独特的机会 确定可以优化青春期SA预测的童年特征。预期的结果最终可能 转化为现实世界的临床实践，整合到EHR中，并帮助标记儿科SA的风险青年 环境，使及时识别和干预，为减少青年自杀的使命作出贡献。

Summary. Suicide in youths is a growing health concern, yet current clinical practice falls short of timely identifying youths at risk for suicide attempt (SA). The overarching aim of this research is to use data driven machine learning methods to facilitate primary prevention of youth SAs in primary care pediatric settings. Clinical guidelines recommend screening for depression, considered a proxy for suicide risk, from age 12 in pediatric setting. The proposed study aims at identification of variables (features) that can be collected by early adolescence, and contribute to prediction of SA in later adolescence. This study will leverage the effort that has been invested in previous projects: a study using electronic health records (EHR) to predict SAs and deaths in University of Pittsburgh Medical Center (UPMC) hospitals; and the Philadelphia Neurodevelopmental Cohort (PNC), that included comprehensive phenotyping of ~9,500 youths. These previous efforts will be integrated to develop and optimize SA prediction in youth from the Children’s Hospital of Philadelphia (CHOP) network, from which we have data on ~40,000 who were screened for a history of SA between the years 2014-2018 (n~1500). First, in the CHOP dataset, we will generate predictive models based on UPMC data, test their predictive validity in CHOP youth population, and then develop, optimize, and cross validate these predictive models using CHOP EHR data as a training set (Aim 1). Second, in the PNC dataset, we will use multiple data types (demographic, behavioral, cognitive, imaging) to classify youths with suicide ideation (SI, n~750) and identify features (potentially modifiable) that are indicative of SI and may also point to potential mechanisms underlying youth SI (Aim 2). Lastly, in a subset of 936 youths (49 with SAs) with both CHOP EHR data and research PNC evaluation that was conducted at mean age 11 (T1), ~5 years before SA screening (T2), we will test the validity of models from Aims 1&2, and aim to identify data features that were collected at T1 and can improve/optimize/outperform the prediction of SAs that relies solely on EHR data (Aim 3). The proposed study relies on the expertise of a highly capable multidisciplinary team comprised of Dr. Barzilay (PI), child- adolescent psychiatrist experienced in suicide research and analysis of suicide related phenotypes in PNC data; Dr. Tsui (PI), an expert in machine learning who has developed predictive algorithms of SA and deaths using UPMC data; and collaborators critical for meeting study aims, Dr. Raquel Gur as the lead researcher who established the PNC, Dr. Ruben Gur who developed the PNC neurocognitive assessment tools, and Dr. Oquendo who will provide expertise in suicide prediction research. The team’s access and familiarity with CHOP EHR and PNC data resources, coupled with its interdisciplinary expertise, creates a unique opportunity to identify childhood features that can optimize later adolescent SA prediction. Expected findings can ultimately translate to real world clinical practice, be integrated in EHR, and help flag youths at risk for a SA in a pediatric setting, allowing timely identification and intervention, contributing to the mission of reducing suicide in youth.