基于结构MRI和机器学习的阿尔茨海默病病程分类研究

A machine learning model for early diagnosis of Alzheimer’s disease

目的利用机器学习算法对阿尔茨海默病(Alzheimer’s disease,AD)病程进行分类,为临床早期诊断AD提供辅助工具。材料与方法将AD病程分为正常认知者、早期轻度认知障碍、晚期轻度认知障碍和AD 4组,收集这些研究对象的结构磁共振成像(structure magnetic resonance imaging,sMRI)数据,在此基础上加入年龄、性别、教育水平和简易智力状态检查(Mini-Mental State Examination,MMSE)评分,然后分别基于两个数据集用L1正则化支持向量机(L1 regularized support vector machine,L1-SVM)算法进行特征选择得到对分类组贡献最大的特征,将提取到的特征子集放入误差逆传播(back propagation,BP)神经网络模型中进行分类,并且与逻辑回归、随机森林、支持向量机3种机器学习模型作对比。用十折交叉验证法比较4种模型的准确率并给出最优组合模型的特异度、敏感度和AUC值。结果加入3项人口统计学指标和MMSE评分的特征集优于仅具有sMRI的特征集,此外,BP神经网络算法与L1-SVM特征选择算法结合的分类准确率优于其他机器学习模型,尤其是在从正常认知功能向AD转化的过程中,BP神经网络的准确率高达98.90%,敏感度98.75%,AUC值1.00。不同分类组之间略有差异。结论L1-SVM和BP神经网络组合模型可以用于AD早期诊断,并且AD进展转化的每一阶段的相关特征数据为临床基础研究和病理变化提供了依据。

Objective:Machine learning algorithm was used to classify the progression of Alzheimer’s disease(AD)and provide an auxiliary tool for clinical early diagnosis of AD.Materials and Methods:The progression of AD was divided into four groups,including normal cognitive subjects,early mild cognitive impairment,late mild cognitive impairment and AD.Structure magnetic resonance imaging(sMRI)datas were collected from these subjects.In addition,Age,sex,education level and Mini-Mental State Examination(MMSE)scores were also collected.Then,L1-regularized support vector machine(L1-SVM)algorithm was used to select the features that contributed the most to the classification group based on the two datasets respectively.The extracted feature subsets were classified in the back propagation(BP)neural network model,and compared with Logistic regression,random forest and support vector machine(SVM).The accuracy of the four models was compared with the ten fold cross validation method.The specificity,sensitivity and area under receiver operating characteristic curve(AUC)values of the optimal combination model were given.Results:The feature sets with three demographic indicators and MMSE score were better than that with only sMRI feature set.The classification accuracy of BP neural network algorithm combined with L1-SVM feature selection algorithm was better than other machine learning models,especially in the process of transforming from normal cognitive function to AD.The accuracy of BP neural network was as high as 98.90%,sensitivity was 98.75%,AUC was 1.00.There were slight differences among different classification groups.Conclusions:The combined model of L1-SVM and BP neural network can be used for the early diagnosis of AD,and the relevant characteristic data of each stage of AD progressive transformation provide the basis for clinical basic research and pathological changes.