Artificial intelligence for severity triage based on conversations in an emergency department in Korea

초록

In the fast-paced emergency departments, where crises unfold unpredictably, the systematic prioritization of critical patients based on a severity classification is vital for swift and effective treatment. This study aimed to enhance the quality of emergency services by automatically categorizing the severity levels of incoming patients using AI-powered natural language processing (NLP) algorithms to analyze conversations between medical staff and patients. The dataset comprised 1,028 transcripts of bedside conversations within emergency rooms. To verify the robustness of the models, we performed tenfold cross-validation. Based on the area under the receiver operating characteristic curve (AUROC) values, the support vector machine achieved the best performance among the term frequency-inverse document frequency-based conventional machine learning models with an AUROC of 0.764 (95% CI 0.019). Among the neural network models, multilayer perceptron performed with an AUROC of 0.759 (+/- 0.024). This research explored methods for automatically classifying patient severity using real-world conversations, including those with nonsensical and confused content. To achieve this, artificial intelligence algorithms that consider the frequency and order of words used in the conversation were employed alongside neural network models. Our findings have the potential to significantly contribute to alleviating overcrowding in emergency departments of hospitals, with future extensions involving highly efficient large language models. The results suggest that a fluid and immediate response to urgent situations, a reduction in patient waiting time, and effectively addressing the special circumstances of the emergency room environment can be achieved using this approach.

키워드