Predictive Model Of Alzheimer’s Disease Based On Integrated Biomarkers And Semi-supervised Learning

Objective:Based on the performance characteristics of the Alzheimer’s Disease(AD)population at different stages,we used the sMRI data of AD and normal control(NC)to perform feature selection independent of the classification process and constructed the early neuroimaging features of AD,provided a theoretical basis for early diagnosis and prediction of AD.So as to provide a theoretical basis for early diagnosis of AD.Semi-supervised learning methods may also be used in statistical modeling of other biomarkers with potential trait attributes.Methods:We enrolled data from the Alzheimer’s Disease Neuroimaging Initiative(ADNI)database.A total of 427 subjects at different stages were selected,including 125 cases of NC,90 cases of stable Mild Cognitive Impairment(sMCI),48 cases of progesive Mild Cognitive Impairment(p MCI),26 cases of unknown Mild Cognitive Impairment(u MCI),and 138 cases of AD.Linear regression regularization method was used to select features from gray matter voxel values of AD and NC sMRI data independent of the classification process.The screened differential sMRI features were used to construct individual sMRI biomarkers and predict u MCI labels by the Semi-supervised Support Vector Machine(S3VM)method.Then we combined the sMRI biomarkers of MCI subjects with the baseline cognitive assessment scale and clinical information to construct integrated biomarkers using the Random Forest(RF).Model evaluation indicators include accuracy rate(ACC)and ROC area under the curve(AUC)etc.Finally we selected the optimally integrated biomarkers and used survival analysis methods to predict the conversion risk of MCI to AD within 1-3 years.The feature extraction of sMRI images and machine learning method were construct using the SPM package in Matlab R2013 b software and Python3.8software,respectively.The significance level was defined as 0.05.Results:A total of 34 different gray matter volumes of sMRI were selected through the elastic net regression model,including the Precentral,Frontal Sup,Frontal Mid,Frontal Mid Orb,Rolandic Oper,Supp Motor Area,Frontal Sup Medial,Frontal Med Orb,Rectus,Insula,Cingulum Ant,Cingulum Mid,Cingulum Post,Hippocampus,Amygdala,Cuneus,Lingual,Occipital Mid,Postcentral,Parietal Inf,Supra Marginal,Precuneus,Caudate,Pallidum,Temporal Sup,and Temporal Pole Mid.The S3 VM classification model was used discriminating sMCI from p MCI,with ACC0.6667 and AUC 0.6923,respectively.The RF classification model integrated biomarkers based on sMRI data,cognitive assessment scale and clinical information discriminating sMCI from p MCI to get optimally integrated biomarkers(including ADAS13,ADASQ4,FAQ,Occipital Mid L,RAVLT＿i,Temporal Pole Mid R,Cingulum Post R,ADAS11,Cuneus R,Cingulum Post L,Pallidum R,Age,Precuneus R,CDRSB,RAVLT＿f,Precuneus L,Frontal Sup Medial L,Supra Marginal R,Hippocampus L,and Lingual L)with ACC 0.8621 and AUC 0.9266,respectively.Conclusion:This study used a combination of Semi-supervised and Supervised learning methods to predict the conversion of MCI to AD,indicating the important role of sMRI in the prediction,and the integrated biomarkers combining sMRI with cognitive assessment scales and clinical information could improve the accuracy of MCI to AD conversion prediction.This research results could provide personalized services for early identification of MCI individuals with AD conversion risk and alleviate the patients disease progression.