Abstract: Background　The incidence of postoperative fatigue（POF） after gastrointestinal tumor surgery is high，which brings negative effects on patients' prognosis and quality of life.However，available studies about factors associated with POF are limited to physiological and psychological aspects，and predicting POF preoperatively and screening POF in high-risk patients have not been achieved.Objective　To establish a risk screening scale based on predictive model for POF after gastrointestinal tumor surgery，to provide a basis for early identification and delivering interventions for high-risk patients.Methods　360 patients who underwent gastrointestinal tumor surgery in the First Affiliated Hospital with Nanjing Medical University from January to June 2018 were included for prospective analysis.Physiological，psychological，and sociological data were collected before surgery，and POF was measured after surgery.The independent risk factors of POF were screened by univariate and multivariate analyses，and were substituted into the binary logistic regression model，BP neural network model and decision tree model，respectively.The AUC and generalization ability of the model tester were compared to select the optimal model.On this basis，POF Risk Screening Scale was formed，then its reliability and validity were analyzed and the cut-off value was determined.Finally，105 patients with gastrointestinal tumor surgery fro the same hospital were included to test the scale.Results　The AUC for the binary logistic regression model，BP neural network model and decision tree model was 0.857，0.894，0.774，respectively，indicating that all of them had a good generalization ability，but BP neural network model was the optimal，with tumor stage，education level，personal monthly income，subjective support，age，preoperative anxiety and depression，and preoperative albumin included.The risk screening scale scored 0 to 15 points.The CVI for the scale was 0.90，and ranged from 0.80 to 1.00 for its items.Exploratory factor analysis indicated that the scale consisted of 3 factors which explained 66.04% of the total variance，and factor loading of all items ranged from 0.552 to 0.751.The Cronbach's α coefficients for the scale，and its 3 factors were 0.730，0.839，0.763，and 0.637，respectively.The AUC of the scale in predicting POF was 0.839，and the cut-off value was 8.In the stage of verification，the scale showed an accuracy of 90.49% in predicting POF.Conclusion　The POF Risk Screening Scale for Gastrointestinal Tumor Patients based on BP neural network model has good reliability and validity，which can be used to effectively predict POF，providing guidance for POF risk screening and delivery of targeted interventions.

Key words: Gastrointestinal neoplasms, Fatigue, Logistic models, Nerve net, Decision trees

摘要： 背景　消化道肿瘤患者术后疲劳（POF）发生率高，影响患者预后和生活质量。目前对POF的影响因素研究局限于生理、心理两方面，且现有研究无法实现术前预测和高危人群筛查。目的　构建基于预测模型且针对消化道肿瘤患者POF的风险筛查评分量表，为临床高危患者早期识别和干预提供依据。方法　前瞻性纳入2018年1—6月在南京医科大学第一附属医院行消化道肿瘤手术的360例患者进行统计分析，术前收集相关生理、心理、社会等资料，术后测量其疲劳程度。通过单因素分析、多因素分析筛选出POF独立危险因素，并将其代入至二元Logistic回归模型、神经网络模型、决策树模型三种模型中，对模型的受试者工作特征曲线下面积（AUC）、泛化能力进行比较以选出最优模型，在此基础上形成POF风险筛查评分量表，进行信效度分析和截断值确定。前瞻性纳入2018年8—9月在南京医科大学第一附属医院行消化道肿瘤手术的105例患者进行量表验证。结果　二元Logistic回归、神经网络、决策树模型AUC分别为0.857、0.894、0.774，均具有良好的泛化能力；神经网络模型为最优模型，纳入因素包括肿瘤分期、文化程度、个人月收入、主观支持、年龄、术前焦虑抑郁、术前清蛋白；形成的风险筛查评分量表最低分0分，最高分15分。量表各条目的内容效度指数（CVI）为0.80~1.00，总CVI为0.90。通过探索性因子分析可提取3个公因子，累积方差贡献率为66.04%，各条目的载荷值为0.552~0.751。量表Cronbach's α系数为0.730，提取的3个公因子的Cronbach's α系数分别为0.839、0.763和0.637。POF风险筛查评分量表预测POF的AUC为0.839，量表截断值为8分；量表验证阶段，总体正确率为90.49%。结论　本研究依据神经网络模型构建的消化道肿瘤POF风险筛查评分量表信效度良好，可有效预测POF，并为临床术后早期风险筛查和针对性干预提供依据。

关键词: 胃肠肿瘤, 疲劳, Logistic模型, 神经网络, 决策树