中国实用外科杂志

中国实用外科杂志 ›› 2025, Vol. 45 ›› Issue (07): 812-818.DOI: 10.19538/j.cjps.issn1005-2208.2025.07.16

• 论著 • 上一篇    下一篇

基于机器学习算法建立直肠癌低位前切除术后吻合口漏早期诊断模型及其效能评价

王大广1，杨少康1，武    平2，房立军3，刘志成1，陈羽佳1，何其通1，所    剑1   

  1. 1吉林大学白求恩第一医院普通外科中心胃结直肠外科，吉林长春 130000；2辽源市中心医院普二外科，吉林辽源 136200；3辽源市人民医院普外科，吉林辽源 136200
  • 出版日期:2025-07-01 发布日期:2025-07-27

  • Online:2025-07-01 Published:2025-07-27

摘要： 目的    筛选对直肠癌低位前切除术后吻合口漏（AL）有预测价值和诊断效能的病人临床特征、血液学指标及其复合指标，并构建早期诊断模型。方法    回顾性分析2019年1月至2024年6月吉林大学第一医院普通外科中心胃结直肠外科收治的行腹腔镜低位前切除术的1195例直肠癌病人的临床资料，其中训练集839例，验证集356例。收集病人临床特征指标以及术前、术后第1~3天血液学指标。根据有无AL发生分为AL组和非AL组，运用3种机器学习算法筛选差异有统计学意义的特征指标，通过多因素Logistic回归构建AL早期诊断模型，并验证模型效果。结果    1195例病人中83例（7.0%）发生AL。3种机器学习算法共筛选出8个差异性指标：术后第2天，白细胞（WBC）、C反应蛋白-白蛋白比值（CAR）；术后第3天，WBC、预后营养指数（PNI）、中性粒细胞-淋巴细胞比值（NLR）、衍生中性粒细胞-淋巴细胞比值（dNLR）、WBC与淋巴细胞比值（WLR）、CAR。多因素Logistic回归构建的模型由术后第3天的WBC、WLR及CAR构成，P值为0.008、0.004和<0.0001，OR值为1.2（95%CI 1.08-1.35）、1.05（95%CI 1.01-1.08）及1.61（95%CI 1.39-1.87）。训练组中模型ROC曲线下面积为0.851（95%CI 0.786-0.916），敏感度为75.9%，特异度为86.9%。在验证组中其ROC曲线下面积为0.808（95%CI 0.719-0.900），敏感度为86.2%，特异度为67.3%。结论    对于行腹腔镜前切除术的直肠癌病人，术后第3天WBC、WLR、CAR是术后发生AL的独立危险因素，基于此构建的模型可早期诊断AL，为早期干预提供临床依据。

关键词: 直肠肿瘤, 吻合口漏, 诊断指标, 生物标记物

Abstract: To explore the predictive value and diagnostic efficacy of clinical characteristics, hematological indicators and composite indicators for anastomotic leakage (AL) after laparoscopic anterior resection of the rectum in patients with colorectal cancer, and construct an early diagnosis model. Methods    The clinical data of 1195 rectal cancer patients who underwent laparoscopic anterior rectal resection at the Department of Gastric and Colorectal Surgery, General Surgery Center of the First Hospital of Jilin University between January 2019 and June 2024 were retrospective analyzed, with 839 cases in the training group and 356 cases in the validation group. Clinical characteristic indicators of patients and hematological parameters before and 1-3 days after surgery were collected. Patients were divided into the AL group and the non-AL group based on the occurrence of AL. 3 machine learning algorithms were employed to screen for differential characteristic indicators, and a multivariate Logistic regression was used to construct an early diagnosis model of AL, with the model effect verified in the validation group.  Results    A total of 83 of 1195 patients were diagnosed with AL, accounting for 7.0%. 3 machine learning algorithms identified 8 differential indicators (WBC, CAR on the second day after surgery and WBC, PNI, NLR, dNLR, WLR, CAR on the third day after surgery). The model constructed by multivariate Logistic regression was composed of WBC, WLR and CAR on the third day after surgery, with P values of 0.008, 0.004 and <0.0001, respectively, and OR values of 1.2 (95%CI 1.08-1.35), 1.05 (95%CI 1.01-1.08) and 1.61 (95%CI 1.39-1.87), respectively. In the training group, the area under the ROC curve of the model was 0.851 (95%CI 0.786-0.916), with a sensitivity of 75.9% and a specificity of 86.9%. In the validation group, the area under the ROC curve could also reach 0.808 (95%CI 0.719-0.900), with a sensitivity of 86.2% and a specificity of 67.3%. Conclusion    WBC, WLR and CAR on the third day after surgery are independent risk factors for AL after laparoscopic anterior resection of the rectum. The Logistic regression model constructed by these indicators can be used for early and accurate diagnosis of AL, providing a clinical basis for early intervention in AL patients.

Key words: rectal tumor, low anterior resection, anastomotic leakage, prediction index, biomarker