Improved multifactorial prediction of suicidal behavior through integration of multiple datasets

自杀是美国第十大死因，超过4万人死于自杀 每年一次。尽管不断努力减轻自杀和自杀行为的负担，但自杀比率仍然保持不变 在过去的半个世纪里相对稳定。预测自杀行为的尝试几乎依赖于 独家报道自杀念头和意图。这是有问题的，因为众所周知 报道的偏见和事实，即许多高危人群都有动力否认自杀想法以避免 住院治疗。即使大多数自杀死者都与医疗保健专业人员有过接触 在他们死亡前的一个月里，这种情况下很少检测到自杀风险。识别风险因素的努力已经 也受到自杀是低基率事件的事实的阻碍，因此需要非常大的样本来 测试可能导致风险的因素的复杂组合。被广泛采用的 纵向电子健康记录(EHR)为以下方面创建了强大但仍未得到充分利用的资源 检测和预测重要的健康结果。在以往的工作中，使用机器学习的方法进行分析 结构化的EHR数据，我们开发的预测模型可以检测到高达45%的首发自杀 行为，平均提前3年。在这里，我们的目标是系统地扩展和改进我们的EHR 在大型医疗系统(N=460万患者)中通过纳入1)外部公众的预测方法 记录数据集(LexisNexis社会经济健康属性数据)， 和生活事件信息；2)自然语言处理(NLP)，以利用非结构化的EHR文本，包括 基于文本的分数，捕捉RDoC域；3)导出时间风险包络的新方法 捕捉单个风险因素的时间依赖效应；以及4)临床风险轨迹 风险因素的有序时间序列。我们将系统地比较这三种方法的性能 确定最佳策略以加强医疗保健环境中的风险监控和预测的方法。 这些目标的完成将代表着朝着新颖、临床可部署和潜在的方向迈出的关键一步 为那些有自杀和自杀行为风险的人改善结局的变革性工具。

Suicide is the tenth leading cause of death in the United States, accounting for more than 40,000 deaths annually. Despite ongoing efforts to reduce the burden of suicide and suicidal behavior, rates have remained relatively constant over the past half century. Attempts to predict suicidal behavior have relied almost exclusively on self-reporting of suicidal thoughts and intentions. This is problematic because of well-known reporting biases and the fact that many people at high risk are motivated to deny suicidal thoughts to avoid hospitalization. Even though the majority of all suicide decedents have contact with a healthcare professional in the month before their death, suicide risk is rarely detected in such cases. Efforts to identify risk factors have also been stymied by the fact that suicide is a low-base rate event so that very large samples are needed to test the complex combinations of factors that are likely to contribute to risk. The widespread adoption of longitudinal electronic health records (EHRs) has created a powerful but still under-utilized resource for detecting and predicting important health outcomes. In prior work using machine learning methods to analyze structured EHR data, we have developed predictive models that detect up to 45% of first-episode suicidal behavior, on average 3 years in advance. Here we aim to systematically extend and improve our EHR prediction methods in a large healthcare system (N = 4.6 million patients) by incorporating 1) external public record datasets (LexisNexis SocioEconomic Health Attribute data) that include environmental, socioeconomic, and life event information; 2) natural language processing (NLP) to leverage unstructured EHR text, including text-based scores that capture RDoC domains; 3) a novel method of deriving temporal risk envelopes to capture the time-dependent effects of individual risk factors; and 4) clinical risk trajectories that incorporate ordered temporal sequences of risk factors. We will systematically compare the performance of each of these approaches to identify optimal strategies for enhancing risk surveillance and prediction in healthcare settings. Completion of these aims would represent a crucial step towards novel, clinically deployable, and potentially transformative tools for improving outcomes for those at risk for suicide and suicidal behavior.