Wang Xue, Yao Kuanda, Huang Zhicheng, Zhao Wanqing, Fu Jin, Lou Pei, Liu Yan, Hu Jiahui, Li Yansheng, Fang An, Chen Wei
Department of Clinical Nutrition, Department of Health Medicine, Peking Union Medical College Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, 100730, China; Department of Neurology, Zhongshan Hospital, Fudan University, Shanghai, 200032, China.
Institute of Medical Information, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, 100020, China.
Clin Nutr ESPEN. 2025 Aug;68:153-159. doi: 10.1016/j.clnesp.2025.03.178. Epub 2025 Apr 29.
BACKGROUND & AIMS: Nutrition screening is a fundamental step to ensure appropriate intervention in patients with malnutrition. An automatic tool of nutritional risk screening based on electronic health records will improve efficiency and elevate the malnutrition diagnosis rate. We aimed to develop an artificial intelligence (AI) malnutrition screening tool based on electronic medical records and compare it with the patient interview-based tool.
We conducted a cross-sectional study at a comprehensive tertiary hospital in China. Data of malnutrition information were extracted from electronic health records (EHR) and were used to train and test an AI tool for the malnutritional risk screening. We adopted the GLIM framework as a reference standard for assessing malnutrition. Six widely used machine learning algorithms for auxiliary diagnosis prediction, including Support Vector Machine, Random Forest, extreme gradient boosting (XGBoost), Logistic Regression, AdaBoost, and Gradient Boosting were compared and visualized using SHapley Additive exPlanations (SHAP). After feature screening, simplified algoritms were cross validated at an independent data set.
495 inpatients enrolled were randomly divided into training and validation groups for algorithm development. 10 features annotated manually from free texts and 32 features selected from structured EHRs entered the models. XGBoost had the highest area under the receiver operating characteristic curve (AUC) and the top six features were weight loss, decreased food intake, prealbmine, white cell, BMI group, and percent of neutrophils. In simplified models, Random Forest acquired the highest AUC of 0.97 based on first sources data from interviews and 0.87 based on EHR data.
Inpatients' EHR data could be integrated by AI to detect the risk of malnutrition. This AI-enabled tool may hold promise for timely and efficient nutrition screening in newly admitted inpatients.
营养筛查是确保对营养不良患者进行适当干预的基本步骤。基于电子健康记录的营养风险自动筛查工具将提高效率并提升营养不良诊断率。我们旨在开发一种基于电子病历的人工智能（AI）营养不良筛查工具，并将其与基于患者访谈的工具进行比较。
我们在中国一家综合性三级医院进行了一项横断面研究。从电子健康记录（EHR）中提取营养不良信息数据，并用于训练和测试用于营养不良风险筛查的AI工具。我们采用全球营养不良领导倡议（GLIM）框架作为评估营养不良的参考标准。比较了六种广泛使用的用于辅助诊断预测的机器学习算法，包括支持向量机、随机森林、极端梯度提升（XGBoost）、逻辑回归、AdaBoost和梯度提升，并使用夏普利值加法解释（SHAP）进行可视化。经过特征筛选后，在独立数据集上对简化算法进行交叉验证。
纳入的495名住院患者被随机分为训练组和验证组用于算法开发。从自由文本中手动标注的10个特征和从结构化EHR中选择的32个特征进入模型。XGBoost在受试者工作特征曲线（AUC）下面积最高，前六个特征是体重减轻、食物摄入量减少、前白蛋白、白细胞、BMI组和中性粒细胞百分比。在简化模型中，随机森林基于访谈的第一来源数据获得了最高AUC为0.97，基于EHR数据为0.87。
AI可以整合住院患者的EHR数据以检测营养不良风险。这种基于AI的工具可能有望为新入院患者进行及时有效的营养筛查。
Zotero 插件