文章摘要

文章信息/Info

Title:: Classification predicting length of stay for critically ill patients via XGBoost and exploring risk factors

Author(s):: LIN Zhiyun; CHENG Yunzhang; GENG Xiaobin; （Shanghai Interventional Medical Device Engineering Technology Research Center, University of Shanghai for Science and Technology, Shanghai 200093, China）

Keywords:: Intensive care unit; Machine learning; Time prediction; SHAP model; Feature analysis; Clinical decision support

摘要:: 为预测危重症患者在重症监护病房的住院时间（length of stay in intensive care unit，ICU LOS）,并探索实验室指标对ICU LOS的影响，本研究基于危重症患者的25个临床指标构建XGBoost模型,对患者是否发生超过3 d的ICU LOS进行预测，并基于SHAP模型对最佳性能模型进行解释性评估。结果显示，XGBoost模型准确率为87.9%。相比于其他预测模型，XGBoost模型在准确率、敏感度和区分度上均有明显优势。同时，SHAP模型增加了集成模型的可解释性和可靠性。研究表明， XGBoost模型可有效识别ICU LOS较长的患者，辅助医生优化临床治疗方案，改善患者预后状况。

Abstract:: To predict length of stay in intensive care unit (ICU LOS) of critically ill patients and explore the influence of laboratory indicators on ICU LOS. We constructed an XGBoost model based on 25 clinical indicators of critically ill patients, to categorically predict whether patients have ICU LOS exceeding 3 d, then interpreted and evaluated the optimal performance model based on the SHAP model. The results showed that the accuracy of XGBoost was 87.9%. Compared with other prediction models, XGBoost showed obvious advantages in accuracy, sensitivity and differentiation. At the same time, SHAP increased the interpretability and reliability of integration model. The research showes that XGBoost model can effectively identify patients with long ICU LOS, and assist doctors to optimize clinical treatment and improve the prognosis of patients.