关键词: 轻度认知障碍, 阿尔茨海默病, 机器学习模型, 影像组学, 诊断

Abstract: Objective To explore the differential diagnostic efficacy of machine learning models for MRI mild cognitive impairment(MCI) and alzheimer's disease(AD). Methods 513 MCI patients who visited our hospital from April 2022 to April 2025 were selected as the research subjects,and whether the patients developed AD was used as the stratification factor.The stratified random sampling method was used to divide the patients into a training set(n=359 cases) and a testing set(n=154 cases) in a 7:3 ratio.Based on whether MCI has been converted to AD,the training set patients were divided into AD group(n=122 cases) and MCI group(n=237 cases).Compared and analyze the differences in general information and experimental indicators of patients,and used multiple logistic regression to analyze the influencing factors of AD occurrence in patients;Adopted the Least Absolute Shrinkage and Selection Operator(LASSO) regression method for screening omics features and constructed an imaging omics model;The machine of learning algorithms,logistic regression(LR),support vector machine(SVM),random forest(RF),and decision tree(DT) were used to construct clinical prediction models and radiomics models based on independent influencing factors and radiomics,respectively and evaluated the diagnostic performance of each model. Results There were significant differences among the two groups in years of education,age,hypertension,diabetes,cerebrovascular history,negative life events,less physical exercise and MMSE(P＜0.05);The results of multiple logistic regression model showed that age,hypertension,diabetes,cerebrovascular history,marital status and education level were independent risk factors for AD,and physical exercise was protective factor(P＜0.05); The optimal imaging features obtained through LASSO regression screening were:wavelet-HH glcm ClusterShade, log-sigma-4-0-mm-3D firstorder 10Percentile, log-sigma-5-0-mm-3D glrlm RunVariance,log-sigma-3-0-mm-3Dgldm DependenceNonUniformityNormalized, log-sigma-2-0-mm-3D glcm Imcl, wavelet-LH glcm Imcl,original shape Flatness, 1og-sigma-5-0-mm-3D glszm ZoneVariance; Among the clinical prediction models and radiomics models constructed by four machine learning algorithms,the RF model was the optimal model,with AUC values of 0.841 and 0.821 for predicting AD occurrence in the training and testing sets, respectively;The AUC of radiomics models were 0.864 and 0.832, respectively;The AUC of the combined model were 0.896 and 0.857,respectively.The AUC of the combined model was significantly higher than that of any single model(P＜0.05);The analysis of calibration curve and clinical decision curve showed that the joint model had the highest calibration degree and the highest clinical net benefit. Conclusion The RF model based on the combination of clinical and radiomics can locate and identify MCI and AD with high accuracy,which can significantly improve the accuracy of automatic imaging diagnosis.

Key words: Mild cognitive impairment, Alzheimer's disease, Machine learning models, Radiomics, Diagnosis