基于MRI和机器学习对阿尔茨海默病的分类预测

Classification and Prediction of Alzheimer's Disease Based on MRI and Machine Learning

目的利用机器学习算法对分类模型进行优化,实现对阿尔茨海默病(AD)病程的分类预测。资料与方法纳入543例研究对象,根据疾病发展病程分为认知功能正常(CN)组、早期轻度认知功能下降(EMCI)组、晚期轻度认知功能下降(LMCI)组和AD组。收集结构磁共振成像(sMRI)和人口统计学指标共276项数据。将正则化惩罚项L1与逻辑回归模型(LR)和支持向量机(SVM)模型分别融合进行特征选择和学习器训练,获取表征性强的特征集合;同时对AD的病程进行两两分类,以十折交叉验证评估两种模型的分类准确率。采用受试者工作特征(ROC)曲线评价模型的分类性能。结果L1-LR和L1-SVM模型在AD病程转化时均表现出较好的分类效果。其中,L1-SVM模型区分CN组与LMCI组、CN组与AD组、EMCI组与AD组的准确率最高,分别为93.63%、100.00%和99.32%。L1-SVM模型区分CN组和AD组的准确率和特异度均为100.00%,ROC曲线下面积为1.00。L1-SVM整体平均准确率为91.49%,略高于L1-LR的90.81%。结论L1-SVM和L1-LR两种模型在不同的病程分类组中分类效果不同,均可作为AD的早期辅助诊断工具,其中L1-SVM模型的预测效能更好。

Purpose The classification model is optimized by machine learning algorithm to realize the classification and prediction of the progression of Alzheimer's disease(AD).Materials and Methods A total of 276 items of structural magnetic resonance imaging(sMRI)and demographic indicators were collected from 543 selected subjects.According to the course of disease,the patients were divided into four groups,including normal cognitive function(CN)group,early mild cognitive impairment(EMCI)group,late mild cognitive impairment(LMCI)group and AD group.The regularized punishment term L1 was fused with the Logistic regression model(LR)and the support vector machine model(SVM)respectively for feature selection and learning machine training,so as to obtain a characteristic set and classify the progression of Alzheimer's disease in pairs.The classification performance of the two models was evaluated by the 10-fold cross validation.Results Both L1-LR and L1-SVM models showed good classification effect in the transformation of AD course.L1-SVM model had the highest accuracy in distinguishing CN group and LMCI group,CN group and AD group,EMCI group and AD group,which were 93.63%,100.00%and 99.32%,respectively.Both the accuracy and specificity of L1-SVM in distinguishing CN group and AD group were 100.00%,and the area under the ROC curve was 1.00.The overall average accuracy of L1-SVM was 91.49%,which was slightly higher than that of L1-LR(90.81%).Conclusion Both L1-SVM and L1-LR have different effects in different course dichotomous groups,and can be used as an early auxiliary diagnostic tool for AD,but L1-SVM model is of higher predictive efficiency.