文献速递儿科与产科急症消化、代谢与肾脏急症

开发电子健康记录模型预测儿科急诊就诊中执法人员在场情况

Development of an electronic health record model to predict law enforcement presence in pediatric emergency department encounters.
PubMed 文献 · Am J Emerg Med · 2025-12-15
一句话结论基于电子健康记录的多变量模型(含关键词算法、到达方式和主诉)可高精度识别儿科急诊中执法人员在场,整个队列敏感度90%、特异度92%、准确率91%、PPV 68%、AUC-PR 0.85。

本研究利用单家儿童医院2021年10月1日至2022年10月1日661次独特急诊就诊的电子健康记录数据,开发并验证预测执法人员(LEO)在场的模型。先在高 prevalence 队列(n=406,通过急救医疗服务或公共车辆到达)测试关键词确定性算法,敏感度95%、特异度93%、准确率93%、PPV 61%;应用于一般人群时敏感度100%、特异度54%、准确率66%、PPV 44%。最终多变量概率模型纳入关键词算法结果、到达方式和主诉,经Lasso收缩后从29个变量中保留21个,在整个队列(n=661)中敏感度90%、特异度92%、准确率91%、PPV 68%、AUC-PR 0.85。作者认为该模型可使医院就诊中LEO在场的回顾性分析、监测和报告以高准确度实施,从而分析LEO在治疗空间中的 prevalence 及其对临床护理和健康结局的影响。

要点

  • 最终多变量模型在整个队列(n=661)中检测LEO在场:敏感度90%、特异度92%、准确率91%、PPV 68%、AUC-PR 0.85。
  • 确定性关键词算法在高 prevalence 队列(n=406)中敏感度95%、特异度93%、准确率93%、PPV 61%。
  • 同一关键词算法应用于一般人群时敏感度100%、特异度54%、准确率66%、PPV 44%。
  • 模型纳入关键词算法结果、到达方式和主诉,经Lasso收缩后从29个变量中保留21个。
  • 数据来自单家儿童医院2021年10月1日至2022年10月1日661次独特急诊就诊,患者<22岁,排除直接由其他医院转运者。

适用人群与场景

  • 单家儿童医院急诊科就诊患者,年龄<22岁;排除直接从其他医院转运的患者。
  • 高 prevalence 队列(n=406)纳入通过急救医疗服务或公共车辆(如LEO车辆)到达者;排除步行、私家车和航空到达者。

研究设计与关键结果

  • 研究类型为回顾性分析,使用单家儿童医院2021年10月1日至2022年10月1日661次独特急诊就诊的电子健康记录数据开发并验证模型。
  • 高 prevalence 队列(n=406)中,确定性关键词算法检测LEO在场:敏感度95%、特异度93%、准确率93%、PPV 61%。
  • 一般人群(n=255)中,确定性关键词算法检测LEO在场:敏感度100%、特异度54%、准确率66%、PPV 44%。
  • 最终多变量概率模型纳入关键词算法结果、到达方式和主诉,经Lasso收缩后从29个变量中保留21个,在整个队列(n=661)中检测LEO在场:敏感度90%、特异度92%、准确率91%、PPV 68%、AUC-PR 0.85。
  • 关键词搜索包括由儿科急诊临床医生选择的14个独特词。

证据类型、等级与局限

  • 证据类型:回顾性分析,单家儿童医院数据。
  • 局限:原文未报告局限性。

对床旁实践的改变

  • 作者认为,基于电子健康记录的多变量模型可使医院就诊中LEO在场的回顾性分析、监测和报告以高准确度实施,从而分析LEO在治疗空间中的 prevalence 及其对临床护理和健康结局的影响。

提炼说明

  • 本条依据论文摘要整理;摘要未涉及「诊断标准与关键阈值」「处置流程与药物剂量」「红旗征与不能漏的情况」「指南或共识出处」,故未列出这些栏目;全文请点「阅读原文」。
英文原文(摘要)· 点击展开

BACKGROUND: Law enforcement intersects with emergency care often, but the prevalence of these interactions and impact on care delivery and health outcomes is unknown. Existing qualitative literature signals potential negative impact of law enforcement officer (LEO) presence on health outcomes, yet rigorous analysis has been limited by the inability to identify LEO presence reliably and accurately in the electronic health record (EHR). This study describes the development and validation of an EHR-based model that includes unstructured data to accurately predict LEO presence during emergency department (ED) encounters.

METHODS: EHR data from 661 unique ED encounters from a single children's hospital between 10/1/2021 and 10/1/2022 were used for model development and validation. Encounters were included if patients were < 22 years of age and excluded if they were transferred directly from another hospital. We initially identified a "high prevalence cohort" with higher likelihood of LEO presence for initial keyword deterministic algorithm testing (n = 406). The high prevalence cohort included arrivals via emergency medical services or public (e.g., LEO) vehicle; walk-in, private car, and air arrivals were excluded. LEO presence was identified by one discrete EHR data field (arrival method) and keyword search. The keyword search included 14 unique words chosen by pediatric ED clinicians. We manually reviewed encounters to determine LEO presence and assess algorithm performance. After the keyword algorithm was developed and validated against the reference set, an additional 255 charts from the general population were manually reviewed to ensure generalizability. Then a multivariable model including keyword algorithm result, arrival method, and chief complaint was developed. Lasso was used to perform controlled variable selection.

RESULTS: In the high prevalence population using the deterministic keyword algorithm, LEO presence was detected with 95 % sensitivity, 93 % specificity, and 93 % accuracy (positive predictive value 61 %). When applied to the general population, deterministic keyword algorithm detected LEO presence with 100 % sensitivity, 54 % specificity, and 66 % accuracy (PPV 44 %). The final multivariable probabilistic model included 21 of 29 variables after shrinkage and detected LEO presence in the entire cohort (n = 661) with 90 % sensitivity, 92 % specificity, 91 % accuracy (PPV 68 %) and AUC-PR of 0.85.

CONCLUSION: Retrospective analysis as well as surveillance and reporting of LEO presence during hospital encounters can be implemented with high accuracy using an EHR-based multivariable model. This will enable analysis of prevalence of LEO presence in treatments spaces and related impact on clinical care and health outcomes.

来源:Am J Emerg Med,查看原文。版权归原作者所有。

参考来源

  1. Abrams AH, Scarbro S, Gray C, et al. Development of an electronic health record model to predict law enforcement presence in pediatric emergency department encounters. Am J Emerg Med. 2025. PMID: 41411993. doi:10.1016/j.ajem.2025.12.010

供执业医师抢救与诊疗参考,不替代临床判断;剂量请结合体重、肝肾功能和本院药品规格复核,以指南原文为准。