| 吴维妙,徐东丽,李小强.基于人工神经网络的结直肠癌预测模型研究[J].中国肿瘤,2019,28(8):621-628. |
| 基于人工神经网络的结直肠癌预测模型研究 |
| A Predictive Model for Colorectal Cancer Based on Artificial Neural Network |
| 投稿时间：2018-11-08 |
| DOI：10.11735/j.issn.1004-0242.2019.08.A011 |
|
 |
| 中文关键词: 结直肠肿瘤 预测模型 人工神经网络 贡献分析 |
| 英文关键词:colorectal cancer predictive model artificial neural network contribution analysis |
| 基金项目:上海市第四轮公共卫生行动计划重点学科建设课题（15GWZK0801） |
|
| 摘要点击次数: 1991 |
| 全文下载次数: 434 |
| 中文摘要: |
| 摘 要：［目的］ 筛选结直肠癌危险因素，建立人工神经网络（artificial neural network，ANN）模型，为筛查人群风险分层与优化筛查方案提供科学依据，提高筛查成本效果。［方法］ 2012年上海市闵行区163 240名50~80岁社区居民参加结直肠癌筛查。初筛采用调查表、粪便隐血试验（FOBT）和肛门指检，阳性者进一步做结肠镜和病理检查，诊断有无结直肠癌，同肿瘤登记链接补充筛查后2年内结直肠癌诊断信息。采用单因素和多因素分析筛选变量，研究对象按1∶1的比例随机分为训练集和验证集，分别用于建立和验证ANN模型。采用灵敏度、特异性、ROC曲线下面积（AUC）等指标评价模型。［结果］ 163 240名研究对象中新确诊结直肠癌病例363例。多因素Logistic回归筛选出年龄、性别、便秘、里急后重、血便、进行性消瘦与FOBT阳性7个变量作为ANN输入变量，其中，血便（20.8%）、年龄（18.1%）和FOBT阳性（17.1%）对结直肠癌的影响最大。训练集和验证集灵敏度分别为65.93%(95%CI：58.78%~72.43%)、60.22%(95%CI：52.95%~67.07%)，特异性分别为62.07%(95%CI：61.74%~62.40%)、61.92%(95%CI：61.59%~62.26%)，AUC分别为0.68(95%CI：0.64~0.72)、0.67(95%CI：0.63~0.70)，验证集符合率为61.92%(95%CI：61.59%~62.26%)，模型内部验证效果较好。［结论］ 本研究所建结直肠癌ANN预测模型总体区分度较好、诊断价值较高，但有待进一步优化和外部验证以提高其预测准确性和泛化能力。 |
| 英文摘要: |
| Abstract：［Purpose］ To establish an artificial neural network(ANN) model for prediction of colo-rectal cancer risk. ［Methods］ Total 163 240 residents aged 50~80 years in Minhang District of Shanghai participated in colorectal cancer screening in 2012. Questionnaires，fecal occult blood tests(FOBT) and anal examinations were conducted for initial screening of colorectal cancer. Further colonoscopy and biopsy were performed for subjects with a positive initial screening result. The data of identified colorectal cancer patients in 2 years after the screening were collected from cancer registry. Univariate and multivariate analysis were used to select variables for establishing ANN prediction model. The participants were randomly classified as training set and validation set by 1∶1 ratio for establishing and verifying the model. Sensitivity，specificity，AUC(area under the receiver operating characteristic curve) were calculated for model evaluation. ［Results］ In 163 240 screening participants，363 colorectal cancer cases were diagnosed. Seven variables including age，gender，constipation，rectal tenesmus，bloody stool，progressive emaciation and FOBT results were selected as the input variables of the ANN model by the multivariate logistic regression. Bloody stools(20.8%)，age(18.1%) and positive FOBT(17.1%) were the most important predictive variables for colorectal cancer. Sensitivity of the model in the training set and validation set were 65.93%(95%CI：58.78%~72.43%) and 60.22%(95%CI：52.95%~67.07%)，respectively. Specificity were 62.07%(95%CI：61.74% ~62.40%) and 61.92%(95%CI：61.59%~62.26%)，respectively. AUC were 0.68(95% CI：0.64~0.72) and 0.67(95% CI：0.63~0.70)，respectively. The agreement rate was 61.92%(95%CI：61.59%~62.26%) in the validation set. The internal validation effect of the model was performed well. ［Conclusion］ The ANN predictive model for colorectal cancer established in this study is performed well in the risk stratification，and is of good diagnostic value. However，further optimization and external validation are needed to improve its prediction accuracy and generalization ability. |
|
在线阅读
查看全文 查看/发表评论 下载PDF阅读器 |