Fast and fine: NLP methods for near real-time and fine-grained overdose surveillance

这项研究是美国国立卫生研究院帮助结束长期成瘾（HEAL）倡议的一部分，该倡议旨在加速科学解决全国阿片类药物公共卫生危机。NIH HEAL倡议支持NIH的研究，以改善阿片类药物滥用和成瘾的治疗。及时和准确地估计过量（OD）事件发生率是减轻持续的OD流行病造成的损失不可或缺的监测组成部分。获得非致命性用药过量的快速更新对于减少用药过量死亡的进一步升级至关重要。传统的用药过量监测方法目前依赖于疾病预防控制中心的综合征监测系统和汇总急诊科（ED）计费数据。然而，国家层面的估计受到严重延误的困扰。因此，越来越多的人推动监测（次）州一级的数据集，包括急诊科和紧急医疗服务（EMS）记录。同时，由于现有的基于诊断代码的OD定义具有较低的召回率（灵敏度），这些记录中叙述性数据的作用正在被认识到为OD和导致它们的药物提供补充信号。即使是在叙述中搜索术语的基于规则的定义也缺少叙述数据中的顺序语义上下文。为了解决这些缺陷，我们建议设计和实现最先进的自然语言处理（NLP）模型，使用深度神经网络（dnn）进行OD分类和细粒度识别导致OD的药物术语。为此，我们将首先使用ED和EMS叙述为这些任务创建并传播第一个公共黄金标准手工标记数据集。我们去识别的笔记将被用来建立DNN模型，也将

This study is part of the NIH’s Helping to End Addiction Long-term (HEAL) initiative to speed scientific solutions to the national opioid public health crisis. The NIH HEAL Initiative bolsters research across NIH to improve treatment for opioid misuse and addiction. Timely and accurate estimation of overdose (OD) event rates is an indispensable surveillance component to mit-igate the toll of the ongoing OD epidemic. Getting fast updates for nonfatal ODs is crucial in decreasing further escalations in OD deaths. Traditional approaches to OD surveillance currently rely on CDC's syndromic surveil-lance system and aggregated emergency department (ED) billing data. However national level estimates are plagued by substantial delays. Hence, there is an increasing push to monitor (sub)state level datasets including ED and emergency medical service (EMS) records. Meanwhile, the role of narrative data in these records is being recognized to offer complementary signal for ODs and drugs leading to them because existing diagnosis code based OD definitions are shown to have lower recall (sensitivity). Even rule-based definitions that search for terms in narratives are missing the sequential semantic context in narrative data. To address these shortcom-ings, we propose to design and implement state-of-the-art natural language processing (NLP) models using deep neural networks (DNNs) for OD classification and fine-grained recognition of drug terms leading to ODs. To this end, we will first create and disseminate the first of their kind public gold standard hand-labeled datasets for these tasks using ED and EMS narratives. Our de-identified notes will be used to build DNN models that will also be shared publicly to the wider OD surveillance community. Our models are expected to improve recall substantially and lead to better nonfatal OD surveillance in a timely manner. We will also develop domain adaption methods to enhance the application of models developed with data from a site to datasets from a different site. Overall, our project will create novel public resources (data, code, models) for the OD surveillance community to leverage latest advances in NLP methods.