An artificial intelligence (AI) machine-learning model has been developed that can predict the risk of early death in trauma patients with high accuracy.
Trauma refers to a group of conditions caused by accidents or injuries that can be life-threatening, and it is one of the leading causes of death worldwide. It is especially a major cause of death among younger people. Because a trauma patient's condition can worsen rapidly in a short time, it is crucial to identify high-risk patients accurately at an early stage. That allows medical staff to respond quickly and allocate critical resources such as intensive care units, operating rooms, and blood transfusions appropriately.
However, predicting the prognosis of trauma patients is not easy. This is because accident circumstances, the location and severity of injuries, the patient's age and physical condition, prehospital emergency care, and treatment after arrival at the hospital all interact in complex ways. Many previous studies relied on small-scale or single-center data, and they lacked sufficient validation to confirm whether the models maintained the same performance over time or under different medical environments. In addition, it was often difficult to explain the basis on which the AI made its predictions.
To address this, a joint research team led by Professor Lee Jae-myung of the Trauma and Critical Care Surgery Department at Korea University Anam Hospital and Professor Baek Seung-min of the Trauma and Critical Care Surgery Department at Ewha Womans University Mokdong Hospital analyzed community-based severe trauma survey data released by the Korea Disease Control and Prevention Agency (KDCA) National Injury Information Portal from 2016 to 2020. Of 237,616 cases in total, the team excluded cases with missing or inconsistent key outcome values and ultimately included 207,012 cases in the study.
The team compared and validated the predictive performance of six machine-learning algorithms: Logistic Regression, k-NN, Decision Tree, Random Forest, MLP, and XGB. Machine learning is an AI technology in which computers learn from large amounts of data to find patterns and then use those patterns to predict the risk of new patients. The researchers developed the models using data from 2016 to 2018 and validated their performance with data from 2019 to 2020.
The analysis showed that the XGB model delivered the best performance.
The model recorded an AUROC of 0.985 and an AUPRC of 0.957 in predicting trauma-related death. AUROC and AUPRC are metrics used to evaluate how well an AI model distinguishes truly high-risk patients, with values closer to 1 indicating better performance. The Random Forest model also showed strong predictive performance, with an AUROC of 0.984 and an AUPRC of 0.956.
Notably, the model's performance remained stable even on data from 2020, during the COVID-19 pandemic. The XGB model still recorded an AUROC of 0.984 during that period. This suggests that the model could operate relatively robustly even in situations where the emergency medical system is under significant strain.
The research team also analyzed which factors the AI considered important in making its predictions. For this, it used the SHAP method. SHAP is an explainable AI technique that shows how much each variable influenced the AI's decision. The analysis identified prehospital cardiac arrest, the Injury Severity Score (ISS), age, and the time to first transfusion as major factors with a strong impact on mortality risk prediction. The Injury Severity Score is an index that measures how severe a patient's injuries are.
This study is significant because it was based not on a single institution but on nationally representative public data. It also considered interpretability, scalability, and generalizability alongside predictive performance.
Professor Lee said, "We hope this will serve as foundational data that can be used in the future to quickly identify patient risk in emergency medical systems and trauma care settings." Professor Baek said, "Using nationwide trauma registry data, we confirmed the possibility of systematically identifying early death risk at the system level." He added, "We will further develop this into research that integrates AI-based risk screening into emergency medical systems and trauma systems through future calibration and prospective validation."
The study, titled 'Machine learning models for early mortality prediction in trauma patients using public data: a nationwide retrospective study,' was published in the international journal World Journal of Emergency Surgery, which covers emergency and trauma surgery. The journal is a Q1 publication, ranking in the top 2.6% in surgery and emergency medicine, with an average JIF percentile of 97.4.
Jang Jong-ho bellho@sportschosun.com
