Abstract P225: Improving Prehospital Stroke Identification Using Natural Language Processing of Paramedic Reports
Introduction: Early identification of stroke by emergency medical services (EMS) providers in the prehospital setting is associated with increased treatment rates, improved functional outcomes, and reduced mortality. We hypothesize that a predictive model utilizing machine learning and natural language processing (NLP) techniques can be developed to analyze EMS run reports to identify stroke patients accurately. Methods: We analyzed EMS data from the Chicago Fire Department matched with inpatient data on confirmed and suspected strokes from 17 Chicago hospitals in the Get With The Guidelines-Stroke (GWTG-Stroke) registry from 11/28/2018 to 5/31/2019. Using features derived from paramedic notes, we developed a support vector machine classifier to predict the following categories: any stroke, AIS-LVO, severe stroke (NIHSS>5), and CSC-eligible stroke (AIS-LVO or ICH/SAH). Individuals were randomly assigned into model derivation (70%) and validation cohorts (30%). C-statistics were used to evaluate discrimination of the classifier for stroke categories. Results: A total of 965 patients were included for analysis. In a validation cohort of 289 patients, the text-based model predicted stroke better than models trained using the Cincinnati Prehospital Stroke Scale (CPSS, c-statistic: 0.73 vs. 0.67, P=0.165) and the 3-Item Stroke Scale (3I-SS, c-statistic: 0.73 vs. 0.53, P <0.001) scores. The text-based model also demonstrated improved performance over the CPSS and 3I-SS models in discriminating patients with other stroke categories (Table 1). Conclusion: We derived a predictive model using clinical text from paramedic reports that has superior performance to existing prehospital clinical screening tools to identify stroke in the prehospital setting. Future studies can evaluate the implementation of an NLP-based decision tool to assist in prehospital stroke evaluation and destination decision-making.