随机森林模型在膝关节炎患者结构特征与症状定量分析中的应用

Quantitative analysis between knee’s structural features and symptoms for the patients with knee osteoarthritis using random forest model

目的探讨基于影像的膝关节定量结构特征是否可以预测膝关节的症状,并评估这些特征在几种不同的膝关节症状中的预测价值。材料与方法首先从骨性关节炎公用数据库(osteoarthritis initiative,OAI)的美国国立卫生研究院生物标志物联盟基金会项目(Foundation for the National Institutes of Health,FNIH)纳入了551个志愿者数据,并将他们分为训练集和测试集。然后提取其中5个影像特征数据集中153个结构影像特征以及西安大略大学和麦克马斯特大学骨性关节炎指数(Western Ontario and McMaster Universities,WOMAC),分别用于评估膝关节的结构特征和症状。接下来使用相关性分析和最小冗余最大相关性(minimum-redundancy maximum-relevance,mRMR)方法进行特征选择。最后,构建了基于随机森林(random forest,RF)回归的预测器模型,并评估了他们预测膝关节的症状评分的能力。结果影响膝关节不同症状(物理功能、疼痛、僵硬)的结构影像特征主要集中在股骨和胫骨的内侧位置。基于这些特征建立的预测器模型表现出了良好的可行性和准确性:4个预测器的R方值在训练和测试集中均高于0.9。疼痛和僵硬预测器的平均绝对误差(mean absolute errors,MAEs)和均方误差(mean squared errors,MSEs)在训练和测试集中均被限制在0.5以下,物理功能预测的MAEs和MSEs在训练/测试集分别为0.5296/2.2727、0.4449/7.8488,总分预测的MAEs和MSEs在训练/测试集分别为1.4167/3.3498、3.1651/16.3974。结论所建立的随机森林模型可以有效地用于预测和评估膝关节症状,筛选出来的结构特征可以在将来用作膝关节症状评估和指导治疗潜在的影像学标志物。

Objective:To investigate whether these imaging based on quantitative features can predict the symptoms of the knees and to evaluate which of these features show the highest predictive value in each distinct knee symptom.Materials and Methods:Five hundred and fifty-one subjects from Osteoarthritis Biomarkers Consortium of Foundation for the National Institutes of Health(FNIH)project in the osteoarthritis initiative(OAI)were included and divided into training and test sets.A total of 153 structural features from 5 quantitative structural feature sets and the Western Ontario and McMaster Universities(WOMAC)osteoarthritis index were included to access the structural characteristics and the symptom of the knees joints,respectively.Correlation analysis and minimum-redundancy maximumrelevance(mRMR)method were used to screen the structural features.Finally,four random forest(RF)regression model were constructed to predict the four WOMAC symptom scores of the knees joints based on the selected structural features,respectively.Results:The structural image features that affect the different symptoms(physical function,pain,stiffness)of the knee joint are manly concentrated at the medial part of the femur and tibia.The constructed predictors showed good feasibility and accuracy:the R-squared of the four predictors were all above 0.9 in both training and test sets.The mean absolute errors(MAEs)and mean squared errors(MSEs)of the pain and stiffness predictors were all restricted to below 0.5 in both training and test set,while the MAEs and MSEs of the physical function and total scores predictors were 0.5296/2.2727,0.4449/7.8488 in training/test sets,and 1.4167/3.3498,3.1651/16.3974 in training/test set.Conclusions:Constructed random forest model can be effectively used to predict and evaluate knee joint symptom,and the selected structural features can be future used as potential biomarkers in the knee joint symptom evaluation and treatment guiding.