A Machine Learning Approach to Predicting Iatrogenic Withdrawal in Critically-ill Children

项目摘要 在美国接受镇静和止痛药物治疗的儿童中，医源性戒断影响高达57% 儿科重症监护病房(ICU)，导致康复延迟、患者和父母痛苦以及 据估计，每年有70000名儿童长期住院。由于样本量和 变量集，儿科ICU中医源性戒断的研究主要集中在 单一的危险因素、筛查工具和治疗方案，而不重视早期识别高危人群 孩子们。这项提议将利用一个由20多万人组成的全国性电子健康记录数据库 儿科ICU患者全面调查危险因素、患者概况和执业模式 与医源性停用镇静剂和止痛药有关，这可能会识别儿童 在戒断症状之前或疗程早期的风险。我将通过首先识别风险来实现这一点 与医源性戒断相关的因素、患者概况和实践模式 生物统计学技术。其次，除了时间依赖变量外，我还将使用已识别的风险因素， 例如生命体征和化验值，以开发动态模型来预测发生医源性疾病的风险 使用新的有监督的机器学习方法在儿科ICU患者中的戒断。第三，我会 从外部验证我的机构电子健康记录中本地数据集中的动态预测模型 以确定该模型是否能准确地预测那些临床确诊的医源性患者。 戒烟。成功完成这些目标将导致开发一种分析工具，以确定 ICU中使用电子资源的儿童医源性戒断，可操作为 临床实践。拟议的研究是可行的，因为1)我强大而富有成效的多学科 由临床医生和数据科学导师组成的团队，他们在我的导师团队的指导下每两周会面一次 其中包括儿科重症监护预测建模领域的领军人物Murray Pollack博士和Dr。 迈克尔·贝尔，神经危重护理领域的全国领导者，以及2)最近可靠的、大型的、多学科的 机构儿科数据库直接来自电子健康记录(EHR)。本次K23大奖 提案还将促进综合教学和导师主导的体验式培训计划，旨在 发展和完善我在大数据库研究、预测模型和发病率方面的知识和技能 与镇静和止痛药用药有关。职业发展与研究 提案将使我实现长期的职业目标，那就是成为一家独立资助的临床 调查人员专注于通过大数据研究和 预测性分析。

PROJECT ABSTRACT Iatrogenic withdrawal affects up to 57% of children who receive sedative and analgesic medications in the pediatric intensive care unit (ICU), contributing to delayed recovery, patient and parental distress and prolonged hospitalization in (an estimated) 70,000 children per year. Due to limitations in sample size and variable sets, studies on iatrogenic withdrawal in pediatric ICUs have primarily focused on the association of single risk factors, screening tools, and treatment regimens, without attention to early identification of at-risk children. This proposal will leverage a national, electronic health record derived database of over 200,000 pediatric ICU patients to investigate the full spectrum of risk factors, patient profiles, and practice patterns associated with iatrogenic withdrawal from sedatives and analgesic medications that could identify children at risk prior to withdrawal symptoms or early in their treatment course. I will achieve this by first identifying risk factors, patient profiles and practice patterns associated with iatrogenic withdrawal using traditional biostatistical techniques. Second, I will use the identified risk factors in addition to time dependent variables, such as vital signs and laboratory values, to develop a dynamic model to predict risk of developing iatrogenic withdrawal in pediatric ICU patients using novel supervised machine learning methodology. Third, I will externally validate the dynamic prediction model in a local dataset from my institution’s electronic health record to determine if the model can accurately predict those patients who develop clinically confirmed iatrogenic withdrawal. Successful completion of these aims will lead to the development of an analytical tool to identify iatrogenic withdrawal in children in ICUs using electronic-based resources which can be operationalized into clinical practice. The proposed studies are feasible because of 1) my strong and productive multi-disciplinary team of clinician and data science mentors who meet biweekly under the guidance of my mentorship team including Dr. Murray Pollack, a leader in the field of predictive modelling in pediatric critical care and Dr. Michael Bell, a national leader in neurocritical care, and 2) the recent availability of reliable, large, multi- institutional pediatric databases derived directly from the electronic health record (EHR). This K23 award proposal will also facilitate an integrated didactic and mentor-led experiential training program designed to develop and refine my knowledge and skills in big database research, predictive modelling, and morbidity associated with sedative and analgesic medication administration. The career development and research proposal will enable my long-term career goal, which is to become an independently funded clinical investigator focused on the prevention of healthcare-acquired morbidity through big data research and predictive analytics.