Artificial intelligence based decision support system predicting prediagnosis using triage data

Objective: The aim of this study was to predict triage, preliminary diagnosis and admissions using a machine learning-based decision support system from complaint texts. Methods: Patients admitted to the emergency department of training and research hospital between Jan 2018 and Jul 2019 were included in the study. Triage and preliminary diagnosis predictions were made using XGBoost, Random Forest(RF), and Artificial Neural Networks(ANN) after Natural Language Processing of chief complaints text. Results: Out of the 54,867 patient data included in the study, yellow triage was found to be the most prevalent(76.4%). The three most frequently recorded diagnoses were acute abdomen(14%), upper respiratory tract infection(10.3%), and diarrhea(6.1%); within bi-grams, the most frequently used word pairs were abdominal pain, chest pain, and headache. For prediction of outcomes, the most successful method was ANN with an accuracy of 0.931. For triage prediction, the highest sensitivities were found in yellow and ESI 3 triages, specifically in XGB(0.771; 95% CI 0.763-0.777), random forest(0.765; 95% CI 0.760-0.773), and ANN(0.766; 95% CI 0.760-0.772). The highest specifity towards predicting pre-diagnoses, neurological ones was with XGB (0.936) and RF (0.936), and in the cardiovascular, specifity was again higher in XGB(0.865) and RF(0.865). Conclusion: While the sensitivities of the three methods in predicting triage and pre-diagnosis with unstructured complaints were found to be high for patients in yellow and ESI 3 categories, the highest accuracy in outcome prediction was detected in artificial neural networks. These models can aid clinicians in mortal pre-diagnosis and discharge decisions in emergency department.