BACKGROUND
Within the trauma system, the emergency department (ED) is the hospital’s first contact and is vital for controlling and providing medical resources. However, ED mortality patients may have limited information. Therefore, we aimed to develop an artificial intelligence (AI) model to predict trauma mortality for all patients visiting the ED. Additionally, we aimed to analyze what information from trauma patients had a significant impact on mortality using that AI model.
OBJECTIVE
We aim to develop an artificial intelligent (AI) model to predict trauma mortality among emergency department patients using the international classification of disease (ICD)-10, triage scale, and other clinical features.
METHODS
We used the Korean National Emergency Department Information System (NEDIS) dataset (n=6,536,306), incorporating over 400 hospitals between 2016 to 2019. Next, we included the International Classification of Disease 10th Revision (ICD-10) and selected the following input features to predict ED mortality: patient’s age, sex, intentionality, injury mechanism, emergent symptom, Alert/Verbal/Painful/Unresponsive (AVPU) scale; Korean Triage and Acuity Scale (KTAS); and vital signs. For the AI input information, we compared three different feature set performances: all features (n=921), ICD-10 features (n=878), and features excluding the ICD-10 (n=43). We presented various machine learning models with Ensemble via five-fold cross-validation and compared each with traditional prediction models. Lastly, we investigated feature effects for explainable AI and deployed our final AI model on a public website (http://ai-wm.khu.ac.kr/mortality_visiting_ED), allowing access to the mortality prediction results among visiting ED patients
RESULTS
Our proposed AI model with the all feature set provided the highest area under the receiver operating curve (AUROC) of 0.9974 (AdaBoost, AdaBoost + LightGBM: Ensemble), outperforming other state-of-the-art machine learning and traditional prediction model AUROCs: XGBoost, 0.9972; LightGBM, 0.9973; ICD-based injury severity scores, 0.9328 (inclusive model) and 0.9567 (exclusive model); and KTAS, 0.9405. In addition, we compared our model’s prediction performance to the state-of-the-art AI model designed for in-hospital mortality prediction. The results indicated that our proposed AI model outperformed the in-hospital mortality AI model (0.7675 AUROC) for all ED visitors. Finally, from the AI model, we also found that age and unresponsiveness (coma) were the top two contributors in predicting mortality among visiting ED patients. Next were oxygen saturation, S224 (multiple rib fractures), painful response (stupor, semi-coma) and S320 (lumbar vertebra fracture).
CONCLUSIONS
Our proposed AI model for predicting ED mortality exhibits remarkable accuracy. Despite the necessity for external validation, a large nationwide dataset would provide a more accurate model and minimize overfitting.