Using Machine Learning to Predict Clinical Deterioration in Hospitalized Children

使用机器学习预测住院儿童的临床恶化

• 批准号: 9805481
• 负责人: Anoop Mayampurath
• 金额: $ 14.76万
• 依托单位: UNIVERSITY OF CHICAGO
• 依托单位国家: 美国
• 财政年份: 2019
• 资助国家: 美国
• 起止时间: 2019-07-15 至 2024-05-31
• 项目状态: 已结题
• 来源: https://reporter.nih.gov/project-details/9805481
• 关键词: Adult; Advisory Committees; Applications Grants; Area; Cardiopulmonary; Caring; Chicago; Child; Childhood; Clinical; Clinical Data; Clinical Trials; Complex; Data; Data Analyses; Data Element; Data Set; Detection; Deterioration; Doctor of Philosophy; Early Intervention; Ecosystem; Electronic Health Record; Environmental Risk Factor; Event; Frequencies; Future; Goals; Grant; Health; Heart Arrest; Hospital Costs; Hospital Mortality; Hospitalized Child; Hospitals; Hour; Hylobates Genus; Intervention; Knowledge; Laboratories; Lead; Location; Machine Learning; Maps; Mentorship; Methodology; Modality; Modeling; Monitor; Natural Language Processing; Neurological outcome; Outcome; Patients; Pediatric Hospitals; Performance; Pharmaceutical Preparations; Physiological; Process; Publishing; Randomized; Research Personnel; Risk; Risk Factors; Savings; Statistical Models; Structure; Techniques; Testing; Text; Time; Training; Universities; Validation; Work; base; career development; clinical decision support; clinical predictors; clinical risk; clinically relevant; cohort; control trial; cost; electronic structure; experience; high risk; improved; improved outcome; mortality; mortality risk; pediatric patients; prediction algorithm; predictive modeling; prevent; preventable death; research and development; risk prediction model; tool; trend; ward

PROJECT SUMMARY Children who are admitted to the hospital and experience deterioration have a high risk of mortality and poor long-term health. Current warning early scores indicating risk of deterioration are subjectively derived and have not reduced in-hospital mortality. In recent work, I developed a vital sign based statistical model that demonstrated improved accuracy over current risk scores at predicting clinical deterioration in hospitalized children 24 hours in advance. Within adults, the combination of longitudinal data analysis techniques, machine learning, and electronic health record (EHR) data have led to highly accurate early warning scores. Therefore, my aim in this grant proposal is to utilize longitudinal integration of EHR data in a machine learning framework to develop a model for predicting clinical deterioration in hospitalized children as early as possible. I will do this by first deriving and validating a prediction model using structured EHR data collected from three pediatric hospitals (Aim 1). Using the same cohort, I will then build and validate a prediction model using features derived from unstructured clinical notes (Aim 2). I will also compare if the addition of unstructured features improves the prediction accuracy of the model derived in Aim 1. Finally, I will determine the association between non-patient level environmental variables within the hospital ecosystem and risk of clinical deterioration in hospitalized children (Aim 3). I will also determine if the addition of these environmental risk factors improves performance of the prediction model derived through Aims 1 and 2. Completion of this proposal will result in a validated pediatric risk prediction model that will enable clinicians to recognize early signs of deterioration in hospitalized children. This will facilitate timely intervention, thereby saving lives and improving long-term health. In addition, this grant will also provide me with crucial data for a future R01 trial aimed at assessing the impact of the prediction model in reducing mortality, decreasing costs, and improving long-term outcomes in hospitalized children. To establish myself as an independent investigator in pediatric prediction modeling, I propose a training plan that includes comprehensive didactics and mentorship in the areas of longitudinal data analysis, advanced machine learning, natural language processing, and concepts in pediatric care. I have assembled a first-class mentorship team comprised of national experts in longitudinal data analysis techniques and EHR-based machine learning (Robert Gibbons PhD and Matthew Churpek MD, PhD). My advisory team is comprised of experts in natural language processing (Dmitriy Dligach PhD), clinical decision support around deterioration events (Dana Edelson MD, MS and Priti Jani MD), and pediatric early warning scores (Christopher Parshuram MB., ChB., D. Phil., FRACP). By completing my research and career development goals, I will develop into an independent expert investigator in developing pediatric prediction models for ultimately improving outcomes in hospitalized children.

项目总结 住进医院并经历病情恶化的儿童死亡和贫穷的风险很高 长期的健康。指示恶化风险的当前预警早期分数是主观得出的，并具有 没有降低住院死亡率。在最近的工作中，我开发了一个基于生命体征的统计模型 在预测住院患者临床恶化方面，显示出比当前风险评分更高的准确性 儿童提前24小时入场。在成人内部，结合纵向数据分析技术，机器 学习和电子健康记录(EHR)数据导致了高度准确的早期预警得分。因此， 我在这项拨款提案中的目标是在机器学习框架中利用电子病历数据的纵向集成 目的：建立早期预测住院儿童临床恶化的模型。我会这么做的 通过首先使用从三个儿科医院收集的结构化EHR数据来推导和验证预测模型 医院(目标1)。使用相同的队列，然后我将使用要素构建和验证预测模型 源自非结构化临床笔记(目标2)。我还会比较一下是否增加了非结构化功能 改进了目标1中导出的模型的预测精度。最后，我将确定关联 医院生态系统内非患者层面的环境变量与临床风险之间的关系 住院儿童情况恶化(目标3)。我还将确定这些环境风险的增加是否 因素提高了通过目标1和目标2得出的预测模型的性能。 提案将产生一个经过验证的儿科风险预测模型，使临床医生能够及早认识到 住院儿童病情恶化的迹象。这将有助于及时干预，从而挽救生命和 改善长期健康状况。此外，这笔赠款还将为我未来的R01试验提供关键数据 旨在评估预测模型在降低死亡率、降低成本和改善 住院儿童的长期结局。在儿科领域确立自己独立研究员的地位 预测建模，我提出了一个培训计划，其中包括全面的教学和导师在 纵向数据分析、高级机器学习、自然语言处理和 儿科护理。我组建了一支由国家纵向专家组成的一流导师团队 数据分析技术和基于EHR的机器学习(Robert Gibbons博士和Matthew Churpek医学博士， 博士)。我的顾问团队由自然语言处理(德米特里·德利加赫博士)、临床 关于恶化事件(Dana Edelson MD、MS和Priti Jani MD)和儿科早期的决策支持 警告分数(Christopher Parshuram MB.，ChB.，D.Phil.，FRACP)。通过完成我的研究和事业 发展目标，我将发展成为发展儿科预测的独立专家调查员 最终改善住院儿童预后的模式。