Bayesian Network Modeling: A Cohort Study Of Cognitive Function Evaluation Among The Elderly

Objective:Based on four states of Alzheimer's disease(AD),Bayesian networks have been chosen in order to produce an intuitive,graphical representation of the relationships between variables,and to make aSBN analysis two reasoning patterns: causal and evidential reasoning,so we can provide theory basis for different stages of AD prevention and early intervention among elderly people.Furthermore,the application of this kind of networks is of special interest,both from theoretical and practical point of view.Finally,we may extend the field of precise medicine to disease prediction.Methods:Our data came from preliminary longitudinal study,with information collected on May 2014 as baseline,and which finished on May 2016 as external validation data for the aim of testing Bayesian networks.First,Logistic models were made to find out factors which potential for cognitive impairment between two kinds of transitions,normal aging?Mild Cognitive Impairment(MCI),MCI?moderate to severe cognitive impairment,moderate to severe cognitive impairment?AD.Afterwards,all the potential factors were introduced to establish Bayesian networks,it allow us to capture the relationships between features through the relationships of dependency and conditional independency.Furthermore,the implementedSBN was validated using a 10-fold cross-validation for BN,and then used to make inferences i.e.,prediction and evidential reasoning.Results:1.Logistic regression results showed,age(OR: 1.794,95%CI: 1.200-2.682)?gender(OR: 4.125,95%CI: 2.017-8.436)?education(OR: 0.633,95%CI: 0.448-0.894)?depression(OR: 4.458,95%CI: 1.915-10.377)?hypertension(OR: 2.346,95%CI: 1.086-5.069)werestatistically significant for transition from normal aging to MCI;age(OR: 2.450,95%CI:1.212-4.953)?gender(OR: 0.118,95%CI: 0.031-0.442)?education(OR: 0.614,95%CI:0.375-1.004)? character(OR: 0.092,95%CI: 0.013-0.662)? family status(OR: 0.272,95%CI: 0.086-0.862)?family income(OR: 0.456,95%CI: 0.273-0.762)?physical exercise(OR: 0.631,95%CI: 0.407-0.980)and reading(OR: 0.432,95%CI: 0.188-0.992)?depression(OR: 97.144,95%CI: 21.452-439.909)? hypertension(OR: 0.304,95%CI:0.077-1.199)?history of brain trauma(OR: 0.188,95%CI: 0.037-0.959)were risk factors for transition from MCI to moderate to severe cognitive impairment;gender(OR: 0.328,95%CI: 0.087-1.234)?family status(OR: 0.102,95%CI: 0.043-0.243)?employment before retirement(OR: 7.799,95%CI: 1.242-48.955)?drinking(OR: 0.126,95%CI: 0.016-0.997)?depression(OR: 3.560,95%CI: 0.998-12.705)played an important role for transition from moderate to severe cognitive impairment to AD.2.The potential factors for cognitive impairment which consist of depression,gender,age,character,education,family status,family income,employment before retirement,reading,physical exercise,drinking,history of brain trauma,hypertension,were included in Bayesian networks model.TheSBN obtaining an expected loss of 10.28.We found the prediction efficiency ofSBN is superior to other classifiers on MoCA(the prediction accuracy is 77.14%,sensitivity is 0.869,specificity is 0.770)and GDS(the prediction accuracy is 80.07%,sensitivity is 0.801,specificity is 0.648),and hypertension,education,employment before retirement and depression were connected with cognition,they also serve as intermediate variables for character,physical exercise,gender,reading,family income,family status which have indirect effects on cognition.Conclusion:1.There were casual effects between hypertension,education,employment before retirement,depression and cognition,and character,physical exercise,gender,reading,family income,family status have indirect casual effects on cognition.Moreover,physical exercise,reading,cultivating outgoing personality and moderate alcohol consumption are available,also effective to slow the progression of cognitive decline.2.This study demonstrates the feasibility of Bayesian networks in epidemiological cohort studies,we can then visualize the relationships of probabilistic casual dependenciesbetween 14 features in the domain of elderly cognitive evaluation,and implement the causal inference,and disease risk prediction.The ability ofSBN makes it an powerful and adequate modeling tool in epidemiological studies.