Lee Seungseok, Kim Do Wan, Oh Na-Eun, Lee Hayeon, Park Seoyoung, Yon Dong Keon, Kang Wu Seong, Lee Jinseok
Department of Biomedical Engineering, Kyung Hee University, 446-701 Electronic Information College Building, Kyunghee Univ, Global Campus, Seocheon-dong, Giheung-gu, Yongin-si, Gyeonggi, Republic of Korea.
Department of Thoracic and Cardiovascular Surgery, Chonnam National University Hospital, Chonnam National University Medical School, Gwangju, Republic of Korea.
Sci Rep. 2025 Jan 7;15(1):1100. doi: 10.1038/s41598-025-85420-5.
Artificial intelligence (AI) is being increasingly applied in healthcare to improve patient care and clinical outcomes. We previously developed an AI model using ICD-10 (International Classification of Diseases, Tenth Revision) codes with other clinical variables to predict in-hospital mortality among trauma patients from a nationwide database. This study aimed to externally validate the performance of the AI model. Validation was conducted using a multicenter retrospective cohort study design, analyzing patient data from January 2020 to December 2021. The study included trauma patients based on specific ICD-10 codes, with other clinical variables. The performance of the AI model was evaluated against conventional metrics, including the ISS, and the ICISS (ICD-based ISS), using sensitivity, specificity, accuracy, balanced accuracy, precision, F1-score, and area under the receiver operating characteristic curve (AUROC) analyses. Data from 4,439 patients were analyzed. The AI model demonstrated high overall performance, achieving an AUROC of 0.9448 and a balanced accuracy of 85.08%, thereby outperforming traditional scoring systems such as ISS, or ICISS. Furthermore, the model accurately predicted mortality across datasets from each hospital (AUROCs of 0.9234 and 0.9653, respectively) despite significant differences in hospital characteristics. In the subset of patients with ISS < 9, the model showed a robust AUROC of 0.9043, indicating its effectiveness in predicting mortality, even in cases with lower-severity injuries. For patients with ISSs ≥ 9, the model maintained high sensitivity (93.60%) and balanced accuracy (77.08%), proving its reliability in more severe injury cases. External validation demonstrated the AI model's high predictive accuracy and reliability in assessing in-hospital mortality risk among trauma patients across different injury severities and heterogeneous cohorts. These findings support the model's potential integration into emergency departments and offer a significant tool for enhancing patient triage and treatment protocols.
人工智能（AI）在医疗保健领域的应用日益广泛，旨在改善患者护理和临床结果。我们之前开发了一种人工智能模型，该模型使用国际疾病分类第十版（ICD - 10）编码及其他临床变量，从全国性数据库中预测创伤患者的院内死亡率。本研究旨在对该人工智能模型的性能进行外部验证。验证采用多中心回顾性队列研究设计，分析2020年1月至2021年12月期间的患者数据。该研究纳入了基于特定ICD - 10编码及其他临床变量的创伤患者。使用包括损伤严重度评分（ISS）和基于ICD的损伤严重度评分（ICISS）在内的传统指标，通过敏感性、特异性、准确性、平衡准确性、精确性、F1分数以及受试者工作特征曲线下面积（AUROC）分析来评估人工智能模型的性能。对4439例患者的数据进行了分析。该人工智能模型展现出较高的整体性能，AUROC为0.9448，平衡准确性为85.08%，从而优于ISS或ICISS等传统评分系统。此外，尽管各医院特征存在显著差异，但该模型仍能准确预测每家医院数据集中的死亡率（AUROC分别为0.9234和0.9653）。在ISS＜9的患者亚组中，该模型显示出稳健的AUROC为0.9043，表明其即使在损伤严重程度较低的情况下，在预测死亡率方面也具有有效性。对于ISS≥9的患者，该模型保持了较高的敏感性（93.60%）和平衡准确性（77.08%），证明了其在更严重损伤病例中的可靠性。外部验证表明，该人工智能模型在评估不同损伤严重程度和异质性队列的创伤患者院内死亡风险方面具有较高的预测准确性和可靠性。这些发现支持了该模型潜在地整合到急诊科，并为加强患者分诊和治疗方案提供了一个重要工具。
