Weidman Aaron C, Sedor-Schiffhauer Zach, Zikmund Chase, Salcido David D, Guyette Francis X, Weiss Leonard S, Poropatich Ronald K, Pinsky Michael R
Department of Emergency Medicine, University of Pittsburgh School of Medicine, Pittsburgh, Pennsylvania, USA.
Center for Military Medicine Research, University of Pittsburgh School of Medicine, Pittsburgh, Pennsylvania, USA.
Acad Emerg Med. 2025 May;32(5):516-525. doi: 10.1111/acem.15067. Epub 2025 Jan 24.
Prehospital emergencies require providers to rapidly identify patients' medical condition and determine treatment needs. We tested whether medics' initial, written impressions of patient condition contain information that can help identify patients who require prehospital lifesaving interventions (LSI) prior to or during transport.
We analyzed free-text medic impressions of prehospital patients encountered at the scene of an accident or injury, using data from STAT MedEvac air medical transport service from 2012 to 2021. EMR records were used to identify LSIs performed for these patients in prehospital settings. Text was cleaned via natural language processing and transformed using term frequency-inverse document frequency. A gradient boosting machine learning (ML) model was used to predict individual patient need for prehospital LSI as well as seven LSI subcategories (e.g., airway interventions, blood transfusion, vasopressor medication).
A total of 12,913 prehospital patients were included in our sample (mean age = 52.3 years, 63% men). We observed good ML performance in predicting overall LSI (area under the receiver operating curve = 0.793, 95% confidence interval = [0.776-0.810]; average precision = 0.670, 95% confidence interval = [0.643-0.695] vs. LSI rate of 0.282) and equivalent-or-better performance in predicting each LSI subcategory except for crystalloid fluid administration. We identified individual words within medic impressions that portended high (e.g., unresponsive, hemorrhage) or low (e.g., droop, rib) LSI rates. Calibration analysis showed that models could prioritize correct LSI identification (i.e., high sensitivity) or accurate triage (i.e., low false-positive rate). Sensitivity analyses showed that model performance was robust when removing from medic impressions words that directly labeled an LSI.
ML based on free-text medic impressions can help identify patient need for prehospital LSI. We discuss future work, such as applying similar methods to 9-1-1 call requests, and potential applications, including voice-to-text translation of medic impressions.
院前急救要求急救人员迅速识别患者的病情并确定治疗需求。我们测试了急救人员对患者病情的初始书面印象中是否包含有助于识别在转运前或转运过程中需要院前救生干预（LSI）的患者的信息。
我们分析了从2012年到2021年STAT MedEvac空中医疗运输服务机构在事故或受伤现场遇到的院前患者的急救人员自由文本印象。电子病历记录用于识别这些患者在院前环境中接受的LSI。文本通过自然语言处理进行清理，并使用词频-逆文档频率进行转换。使用梯度提升机器学习（ML）模型预测个体患者对院前LSI的需求以及七个LSI子类别（例如，气道干预、输血、血管加压药物）。
我们的样本共纳入了12913名院前患者（平均年龄 = 52.3岁，63%为男性）。我们观察到在预测总体LSI方面ML表现良好（受试者工作特征曲线下面积 = 0.793，95%置信区间 = [0.776 - 0.810]；平均精度 = 0.670，95%置信区间 = [0.643 - 0.695]，而LSI发生率为0.282），并且在预测除晶体液输注外的每个LSI子类别方面表现相当或更好。我们在急救人员印象中识别出预示高LSI发生率（例如，无反应、出血）或低LSI发生率（例如，下垂、肋骨）的单个词汇。校准分析表明，模型可以优先进行正确的LSI识别（即高灵敏度）或准确的分诊（即低假阳性率）。敏感性分析表明，从急救人员印象中去除直接标记LSI的词汇时，模型性能稳健。
基于自由文本急救人员印象的ML可以帮助识别患者对院前LSI的需求。我们讨论了未来的工作，例如将类似方法应用于911呼叫请求，以及潜在应用，包括急救人员印象的语音转文本翻译。
