基于贝叶斯网络的膝骨关节炎患病风险定量预测研究

Quantitative Prediction of the Risk of Knee Osteoarthritis Based on Bayesian Networks

目的:膝骨关节炎(Knee Osteoarthritis,KOA)初早期症状不明显,存在大量膝关节已然病变、但未发现病情的患者。通过简便、易测的体质指标,建立KOA患病风险的定量预测模型,有助于实现大范围的KOA筛查。方法:按年龄、性别比例在南京市2个街道分层抽取1045名居民,进行KOA诊断与体质测试,测试指标包括性别、年龄、身高、体重、BMI、大腿围、30 s坐站、膝关节屈曲度、闭眼单脚站、计时起立行走。在GeNle 2.3软件中,通过贝叶斯网络学习建立KOA与上述指标间的数学模型,建模步骤包括数据离散化处理、基于爬山算法与K2算法的结构学习、基于最大期望算法的参数学习、模型验证与敏感度分析等。结果:10个指标在“未患KOA”与“患KOA”之间的组间差异均存在统计学意义(P<0.01)。建立的数学模型包括11个节点和19条有向线段,确定任意一个或多个节点的状态都可以预测KOA的患病概率。模型中,性别、BMI、体重、30 s坐站、膝关节屈曲度是与KOA直接关联或敏感度较高的节点,这些指标的预测价值较高。模型的准确率为78.9%,ROC曲线下面积为0.722。结论:构建了定量预测KOA患病风险的贝叶斯网络模型,模型的预测性能良好,具有推广应用优势。

Objective:The early symptoms of knee osteoarthritis(KOA)are not obvious,therefore there are a large number of patients with knee joint lesions that have not been detected.In order to screen high-risk groups or earlystage patients with KOA on a large scale,this study established a quantitative prediction model for the risk of KOA through simple and easy-to-measure indicators.Methods:According to age and sex proportion,1045 residents from two streets in Nanjing were sampled for KOA diagnosis and physical fitness tests.The test indicators included gender,age,height,weight,BMI,thigh circumference,30-second sitting and standing,knee joint flexion,single-leg standing with eyes closed,and time-up-and-go test.In GeNle 2.3 software,a mathematical model between KOA and the above indicators was established through Bayesian network learning.The modeling steps include data discretization,structural learning using mountain climbing algorithm and K2 algorithm,parameter learning using Expectation-Maximization algorithm,model verification and sensitivity analysis.Results:Univari⁃ate analysis showed that there were statistically significant differences between the groups of“no KOA”and“KOA”in the 10 indicators(P<0.01).The established mathematical model included 11 nodes and 19 directed line segments.Determining the state of any one or more nodes can predict the probability of KOA disease.In the model,gender,BMI,body weight,30-second sitting and standing,and knee joint flexion were nodes that were di⁃rectly related to KOA or had higher sensitivity.These indicators had high predictive value.The accuracy of the model was 78.9%,and the area under the ROC curve was 0.722.Conclusion:This study has constructed a Bayes⁃ian network model for quantitatively predicting the risk of developing KOA,which exhibits good predictive perfor⁃mance and has advantages for widespread application.