基于多维临床特征与机器学习构建急性胰腺炎复发(RAP)的预测模型及临床验证

Construction and Clinical Validation of a Prediction Model for Recurrent Acute Pancreatitis (RAP) Based on Multidimensional Clinical Features and Machine Learning

摘要: 目的：探讨急性胰腺炎(AP)患者发生复发性急性胰腺炎(RAP)的独立危险因素，并结合多维临床特征构建列线图(Nomogram)预测模型。方法：回顾性分析2023年1月至2025年1月期间蚌埠医科大学附属合肥市第二人民医院收治的380例AP患者。通过LASSO回归与Boruta算法进行特征筛选，多因素Logistic回归确定预测变量，并结合血钙分类化、稳健性分析及病因分层分析对关键变量进行解释。结果：380例患者中位随访时间为18个月(IQR: 12~24个月)，RAP发生率为31.8%。多因素分析显示血钙、吸烟史、饮酒史、甘油三酯(TG)、尿酸(UA)、丙氨酸氨基转移酶(ALT)、入院年龄及白蛋白(ALB)是RAP发生的预测因素。血钙分类化及稳健性分析未改变其与RAP风险升高相关的总体方向；年龄和ALT的负向关联主要结合病因构成及病因去除干预进行解释。验证集AUC为0.822 (95%CI: 0.730~0.903)，校准曲线显示预测概率与实际发生率具有良好一致性。结论：本研究构建的Nomogram模型有助于临床早期识别RAP高危患者，辅助个体化随访和二级预防决策。

Abstract: Objective: To investigate the predictive factors for recurrent acute pancreatitis (RAP) in patients with acute pancreatitis (AP) and to construct a nomogram prediction model based on multidimensional clinical features. Methods: A retrospective analysis was conducted on 380 AP patients admitted to the Second People’s Hospital of Hefei, Affiliated to Bengbu Medical University, from January 2023 to January 2025. Feature selection was performed using LASSO regression and the Boruta algorithm, and predictive variables were determined by multivariate logistic regression. Categorized serum calcium analysis, robustness assessment, and etiology-based stratified analyses were further used to interpret key variables. Results: During a median follow-up of 18 months (IQR: 12~24 months), the incidence of RAP among the 380 patients was 31.8%. Multivariate analysis showed that serum calcium, smoking history, drinking history, triglycerides (TG), uric acid (UA), alanine aminotransferase (ALT), admission age, and albumin (ALB) were predictive factors for RAP. Categorized and robust analyses of serum calcium showed a consistent direction of association with increased RAP risk, whereas the inverse associations of age and ALT were mainly interpreted in relation to etiological composition and etiological interventions. The AUC of the validation set was 0.822 (95%CI: 0.730~0.903), and the calibration curve showed good consistency between the predicted probability and the actual incidence. Conclusion: The nomogram model constructed in this study helps identify high-risk RAP patients and supports individualized follow-up and secondary prevention decision-making.