基于不同核函数构建的退行性颈椎病支持向量机高危人群筛查模型的比较

摘要

目的 评价基于不同核函数构建的退行性颈椎病支持向量机高危人群筛查模型的优劣,为退行性颈椎疾病高危人群的筛查提供工具支持.方法 利用北京地区社区人群骨科退行性疾病研究数据库,采用线性核、多项式核、Sigmoid核和高斯核函数构建支持向量机模型,并根据十折交叉验证率最大的标准确定核函数参数.根据约登指数最大化的标准选择切点值,计算模型相应的灵敏度、特异度和预测准确率.采用ROC曲线评价不同核函数构建的模型的性能.结果 在四种核函数计算的支持向量机模型中,多项式核函数计算ROC曲线下面积最大,为0.6928(95％ CI:0.6502～0.7355), 但不同核函数的ROC曲线下面积的95％ CI存在重叠,尚未发现不同核函数建立本模型的优势.结论 可利用该模型进行高危人群筛查,但未发现不同核函数构建的支持向量机模型性能的差别.%Objective To evaluate the support vector machine model for screening high risk population of cervical spondylosis based on different kernel functions. Methods According to the community-based spinal degenerative disease data, support vector machine model was maximization criterion, the cut-off value, sensitivity, specificity and predictive accuracy were determined. ROC curve was used to evaluate the model' s capability of different kernel functions. Constructed using linear kernel, polynomial kernel, Sigmoid kernel and gauss kernel function. The parameters were determined by ten-fold cross validation rate. According to Youden index. Results Among the support vector machine models constructed by four kernel functions, the area under ROC curve of polynomial kernel function was the biggest. It was 0.6928 (95％ CI:0.6502 ～ 0.7355). However, the AUC 95％ CI of ROC curves of different kernel functions overlapped and have not yet found the difference among four kernel functions. Conclusion The model can be used for screening high-risk groups. There was no capability difference among support vector machine modelsconstructed by different kernel functions.