Extracting RDoC Constructs from EHR through Natural Language Processing to Predict Suicide in Youth

项目摘要/摘要 自杀是青春期第二大最常见的死亡原因，而 青少年自杀人数继续增加。尽管有50年的研究和努力， 自杀和自杀想法和行为(STB)的预测仍然很困难。近期 研究表明，电子健康记录(EHR)数据分析可以帮助预测 机顶盒。研究领域标准(RDoC)提供了调查跨诊断的框架 反映正价、负价和睡眠觉醒的区域 觉醒和调节领域内的元素失眠。这三个RDoC结构具有 显示出与抑郁症和焦虑症有很强的联系。然而，存在一种 关于从eHR中提取的RDoC措施的关系和影响的知识差距 关于青少年自杀未遂(SA)。考虑到积极情绪和认知能力的显著变化 在儿童和青少年时期，我们的总体目标是评估RDoC的积极作用 青年STBS患者失眠的价态、负价态和睡眠觉醒因子 使用基于机器学习和深度学习的自然语言处理电子病历数据。 这项研究将利用在以前的项目中投入的努力和资源 基于自然语言处理和机器的SA预测研究 从匹兹堡大学医学中心(UPMC)的EHR数据中学习(n=7670名青年) 医院；费城儿童医院(CHOP)SA研究的数据收集 有567,091名青年(n=3,125人)。具体目标是：1)开发和验证 使用深度神经网络语言模型从电子病历中提取汇总变量 正价、负价与睡眠-觉醒元素失眠 唤醒和调节领域；2)比较目标1中开发的ML模型的性能 用传统NLP从EHR中提取RDoC正负价和失眠 方法；3)RDoC正负效价和失眠的测试效用 对自杀行为的预测。这项拟议的研究如果成功，将是实现以下目标的第一步 从青年SA上的EHR中提取的其他RDoC域和结构，并将 将获得的模型应用于临床环境。徐博士(Chop)和瑞安博士(UPMC)有着悠久的历史 在使用ML和NLP进行心理健康研究方面的合作经验，并在 在各种研究中担任PI。总体而言，我们的研究具有推动SA领域发展的潜力 有助于及时干预并最终减少青少年自杀的预测。

PROJECT SUMMARY/ABSTRACT Suicide ranks as the second most frequent cause of death in adolescence, and the rate of suicide among adolescents has continued to increase. Despite 50 years of research and efforts, the prediction of suicide and suicidal thoughts and behaviors (STBs) remains difficult. Recent studies indicate that electronic health record (EHR) data analytics can help predict the risk of STB. The Research Domain Criteria (RDoC) provides a framework to probe transdiagnostic domains reflecting the positive valence, the negative valence, and the sleep-wakefulness element insomnia within the arousal and regulatory domain. These three RDoC constructs have shown strong association with depression and anxiety disorders. However, there exists a knowledge gap regarding the relationship and impact of RDoC measures extracted from EHR on youth suicide attempts (SAs). Given significant changes in positive affect and cognitive systems during childhood and adolescence, our overall goal is to assess the RDoC positive valence, the negative valence, and the sleep-wakefulness element insomnia in youth with STBs using machine learning and deep learning based natural language processing of EHR data. This study will leverage the effort and resources that have been invested in previous projects from two sites: a study on SA prediction using natural language processing and machine learning from EHR data (n=7,670 youths) in the University of Pittsburgh Medical Center (UPMC) hospitals; and the data collection for SA study in the Children’s Hospital of Philadelphia (CHOP) with 567,091youths (n=3,125 attempters). The specific aims are to 1) develop and validate extraction of summary variables from EHR using deep neural network language models for the positive valence, the negative valence, and the sleep-wakefulness element insomnia within the arousal and regulatory domain; 2) compare performance of the ML models developed in Aim 1 to extract the RDoC positive and negative valence and insomnia from EHR with traditional NLP approaches; and 3) test utility of the RDoC positive and negative valence and insomnia in prediction of suicidal behaviors. This proposed study, if successful, is the first steps towards other RDoC domains and constructs extracted from EHR on youth SAs and translating the obtained models to clinical settings. Dr. Tsui (CHOP) and Dr. Ryan (UPMC) have a long history of collaboration in mental health studies using ML and NLP and have strong experience in serving as PIs in various studies. Overall, our study has a potential to advance the field of SA prediction that facilitates timely intervention and ultimately reduces youth suicides.