目的： 构建预测老年胰腺癌患者在肿瘤根治术后发生术后谵妄（postoperative delirium，POD）的模型，并验证其预测效能，为早期识别和预防POD发生提供实践依据。方法： 根据纳入、排除标准，回顾性连续收集2025年1月至2025年6月在上海瑞金医院接受胰腺癌根治术的202例老年患者的资料，纳入建模组。术后7 d内，由经培训的护理人员每日使用意识错乱评估法（confusion assessment method，CAM）评估患者POD发生情况，将建模组分为谵妄组与非谵妄组，比较2组的临床特征，采用单因素分析法明确与POD相关的危险因素，运用多因素Logistic回归分析确定回归方程中各独立危险因素的偏回归系数，建立预测模型，并使用Hosmer-Lemeshow（H-L）检验及受试者操作特征（receiver operating characteristic，ROC）的曲线下面积（area under curve，AUC）评估模型效能。使用相同入组标准选取101例同期同类型患者纳入验证组，包括10例谵妄患者和91例非谵妄患者，以验证模型的预测效果。结果： POD相关危险因素的分析结果显示，谵妄组与非谵妄组患者在年龄、急性生理与慢性健康（acute physiology and chronic health evaluation-Ⅱ，APACHE-Ⅱ）评分、术中出血量、糖尿病、术前睡眠障碍等方面差异有统计学意义（P<0.05）。Logistic回归分析结果显示，年龄、APACHE-Ⅱ评分、术中出血量、糖尿病、术前睡眠障碍均为老年胰腺癌患者根治术后发生POD的独立危险因素（P<0.05）。根据回归分析得到各独立危险因素乘以对应回归系数构建预测模型P=e^x/（1+e^x）×100％，其中e为自然对数的底；x=-13.959+0.106×年龄+0.357×APACHE-Ⅱ评分+0.003×术中出血量+1.306×糖尿病赋值+1.785×术前睡眠障碍赋值。预测模型的AUC为0.893（95%CI为0.836~0.949），取约登指数最大值0.661时，灵敏度为81.0%，特异度为85.1%（P<0.001）。将验证组的病历资料代入预测模型验证，结果显示，模型的预测灵敏度为60.0%，特异度为94.5%，准确率为91.1%。结论： 年龄、APACHE-Ⅱ评分、术中出血量、糖尿病、术前睡眠障碍是老年胰腺癌患者根治术后发生POD的独立危险因素，根据上述因素构建的预测模型效能良好，可初步排除POD低风险患者，优化资源分配并指导早期干预。

Objective To develop a predictive model for postoperative delirium (POD) after radical surgery in elderly patients with pancreatic cancer, thereby providing a practical basis for the early identification and prevention of POD occurrence. Methods Data were retrospectively and consecutively collected from 202 elderly patients who underwent radical surgery for pancreatic cancer at Ruijin Hospital, Shanghai, from January 2025 to June 2025. These patients were included in the modeling group based on inclusion and exclusion criteria. Within 7 days after surgery, trained nursing staff used the confusion assessment method (CAM) daily to assess the occurrence of POD. The modeling group was further divided into delirium and non-delirium subgroups according to the occurrence of POD. Clinical characteristics of both groups were compared, and univariate analysis was used to identify risk factors for POD. Multivariable logistic regression analysis was performed to determine the partial regression coefficients of independent risk factors and construct a predictive model. Model performance was evaluated using the Hosmer-Lemeshow (H-L) test and the area under the curve (AUC) of the receiver operating characteristic (ROC). Additionally, 101 patients who underwent the same type of surgery during the same period and met the same inclusion criteria, including 10 with delirium and 91 without, were selected as the validation group to assess the predictive performance of the model. Results The analysis of risk factors associated with POD revealed significant differences in age, APACHE-Ⅱ score, intraoperative blood loss, diabetes, and preoperative sleep disturbance between the delirium and non-delirium groups (P<0.05). Logistic regression analysis identified age, APACHE-Ⅱ score, intraoperative blood loss, diabetes, and preoperative sleep disturbance as independent risk factors for POD (P<0.05). Based on the regression analysis, the predictive model was constructed by multiplying each independent risk factor by its corresponding regression coefficient: P = e^×/(1 + e^×) × 100%, where e is the base of the natural logarithm and x = -13.959 + 0.106 × age + 0.357 × APACHE-Ⅱ score + 0.003 × intraoperative blood loss + 1.306 × diabetes status + 1.785 × preoperative sleep disturbance status. The AUC of the predictive model was 0.893 (95% CI: 0.836-0.949). At the maximum value of the Youden index (0.661), the model achieved a sensitivity of 81.0% and a specificity of 85.1% (P<0.001). When the clinical data from the validation group were entered into the predictive model, the model demonstrated a sensitivity of 60.0%, a specificity of 94.5%, and an accuracy of 91.1%. Conclusions Age, APACHE-Ⅱ score, intraoperative blood loss, diabetes, and preoperative sleep disturbance are independent risk factors for POD after radical surgery in elderly patients with pancreatic cancer. The predictive model based on these factors performs well and can help identify patients at low risk of POD, optimize resource allocation, and guide early intervention.