| 何文,雷益,赵滨.误差反向传播神经网络、随机森林预测结直肠癌根治术后并发症的效能对比[J].安徽医药,2025,29(7):1432-1437. |
| 误差反向传播神经网络、随机森林预测结直肠癌根治术后并发症的效能对比 |
| Comparison of the efficacy of backpropagation neural network and random forest in predicting complications after radical resection of colorectal cancer |
| |
| DOI：10.3969/j.issn.1009-6469.2025.07.034 |
| 中文关键词: 结直肠肿瘤 结肠直肠切除术 术后并发症 误差反向传播神经网络 随机森林 |
| 英文关键词: Colorectal neoplasms Coloproctectomy Postoperative complications BP neural network Random forest |
| 基金项目: |
|
| 摘要点击次数: 854 |
| 全文下载次数: 228 |
| 中文摘要: |
| 目的误差反向传播( BP)神经网络、随机森林预测结直肠癌根治术后并发症的效能对比分析。方法选取 2023年 1月至 2024年 1月眉山市人民医院收治的 128例行根治术治疗的结直肠癌病人为研究对象。根据受试者术后是否发生并发症分为并发症组( 90例)和无并发症组( 38例)。收集病人一般资料、病理学特征、围术期指标。采用 t检验或 χ2检验筛选出可能影响结直肠癌根治术病人术后发生并发症的因素。将受试者按照 7∶3的比例分为训练集( 90例)和测试集(38例)训练集用于构建 BP神经网络和随机森林模型，测试集用于评估上述两种模型的预测效能。结果 128例病人中， 38例病人术后，发生并发症，并发症发生率为 29.69%(38/128)。并发症组年龄( 67.54±10.36)岁、手术时间( 325.46±45.36)min高于无并发症组( 63.35± 10.42)岁、(268.36±46.52)min(P<0.05)贫血( 39.47%)、 aCCI评分 >4分( 52.63%)、 PNI<42.15(44.74%)占比高于无并发症组(12.22%、25.56%、12.22%)(P<0.05)。受试，者操作特征曲线( ROC曲线)分析结果显示， BP神经网络和随机森林模型预测测试集结直肠癌根治术病人术后并发症的曲线下面积( AUC)分别为 0.82［95%CI:(0.77，0.86)］、 0.92［95%CI:(0.89，0.98)］灵敏度分别为 85.70%、94.50%，特异度分别为 80.30%、90.50%，正确率分别为 79.00%、88.40%，召回率分别为 76.40%、90.20%，，精确率分别为 79.70%、90.20%。Delong检验结果显示，随机森林预测测试集结直肠癌根治术病人术后并发症的 AUC大于 BP神经网络，差异有统计学意义( D=.3.02，P=0.004)。结论 BP神经网络、随机森林模型对结直肠癌根治术病人术后并发症的预测效能较好，其中随机森林模型对该病病人术后发生并发症的预测效能更优，值得在临床推广应用。 |
| 英文摘要: |
| Objective To compare the efficacy of backpropagation(BP) neural network and random forest in predicting postoperativecomplications after radical resection of colorectal cancer.Methods One hundred and twenty-eight patients with colorectal cancer whounderwent radical surgery in Meishan People's Hospital from January 2023 to January 2024 were selected as the study objects. Thesubjects were assigned into complication group (90 cases) and non-complication group (38 cases) according to whether complicationsoccurred after surgery. General data, pathological features and perioperative indexes were collected. t test or χ2 test were used to screen out the factors that may affect the postoperative complications of patients with radical resection of colorectal cancer. The subjects weredivided into a training set and a test set according to a ratio of 7∶3. The training set was used to build the BP neural network and therandom forest model, and the test set was used to evaluate the prediction efficiency of the two models.Results Among the 128 pa-tients, 38 had postoperative complications, with a complication rate of 29.69% (38/128). The age of the complication group was (67.54±10.36) years and the operative time of (325.46±45.36) min was higher than that of the non-complication group (63.35±10.42) years and (268.36±46.52) min (P<0.05). The proportion of anemia (39.47%), aCCI score > 4 (52.63%) and PNI < 42.15 (44.74%) were higher than those in the non-complication group (12.22%, 25.56% and 12.22%) (P<0.05). ROC curve analysis results showed that the AUCs ofpostoperative complications in patients with radical resection of rectal cancer grouped by BP neural network and random forest modelprediction test were 0.82 [95%CI:(0.77, 0.86)] and 0.92 [95%CI:(0.89,0.98)], respectively, the sensitivities were 85.70% and 94.50%,the specificities were 80.30% and 90.50%, the accuracies were 79.00% and 88.40%, the recall rates were 76.40% and 90.20%, andthe precision rates were 79.70% and 90.20%, respectively. The results of Delong test showed that the AUC of postoperative complica-tions of patients undergoing radical resection of rectal cancer collected by random forest prediction test was greater than that of BP neu-ral network, with statistical significance (D=.3.02, P=0.004).Conclusion BP neural network and random forest model have good pre-dictive efficacies for postoperative complications of patients with radical resection of colorectal cancer, and random forest model has better predictive efficacy, which is worthy of clinical application. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
