Deep-Learning Approaches to Identify Critically Ill Patients at Emergency Department Triage Using Limited Information

ImportanceTriage quickly identifies critically ill patients, helping to facilitate timely interventions. Many emergency departments use the emergency severity index (ESI) or abnormal vital sign thresholds to identify critically ill patients. However, both rely on fixed thresholds, and false activations detract from efficient care. Prior research suggests that machine-learning approaches may improve triage accuracy, but have relied on information that is often unavailable during the triage process.ObjectiveWe examined whether deep-learning approaches could identify critically ill patients using data immediately available at triage with greater discriminative power than ESI or abnormal vital sign thresholds.DesignRetrospective, cross-sectional study.SettingAn urban tertiary care hospital in the Northeastern United States.ParticipantsAdult patients presenting to the emergency department from 1/1/2012 - 1/1/2020 were included. Deidentified triage information included structured data (age, sex, initial vital signs, ESI score, and clinical trigger activation due to abnormal vital signs), and textual data (chief complaint) with critical illness (defined as mortality or ICU admission within 24 hours) as the outcome.InterventionsThree progressively complex deep-learning models were trained (logistic regression on structured data, neural network on structured data, and neural network on structured and textual data), and applied to triage information from all patients.Main Outcomes and MeasuresThe primary outcome was the accuracy of the model for predicting whether patients were critically ill using area under the receiver-operator curve (AUC), as compared to ESI, utilizing a 10-fold cross-validation.Results445,925 patients were included, with 60,901 (13.7%) critically ill. Vital sign thresholds identified critically ill patients with AUC 0.521 (95% CI 0.519 -- 0.522), and ESI less than 3 demonstrated AUC 0.672 (95% CI 0.671 -- 0.674), logistic regression classified patients with AUC 0.803 (95% CI 0.802 -- 0.804), neural network with structured data with 0.811 (95% CI 0.807 - 0.815), and the neural network model with textual data with AUC 0.851 (95% CI 0.849 -- 0.852).Conclusions and RelevanceDeep-learning techniques represent a promising method of enhancing the triage process, even when working from limited information. Further research is needed to determine if improved predictions can be translated into meaningful clinical and operational benefits.