The Establishment And Effect Evaluation Of Osteoporosis Prediction Model Based On BPNN

OBJECTIVE:This study based on the investigation data of osteoporosis,aimed to determine whether the back-propagation neural network(BPNN)model,established by correlation factors of bone mineral density(BMD),could be constructed to accurately in predicting osteoporosis(OP)for the elderly,and provided a method of a rapid economical and acceptable early screening for OP.Methods:1.A total of 789 subjects(599 females and 190 males)aged over 40 years were enrolled from Wuhan Union hospital and local communities.The lumbar l1-l4 and hip BMD of the subjects were examined by Dual Energy X-ray Absorptiometry(DEXA),and they were divided into three groups according to the"gold standard”:normal,bone loss and OP.According to the examination results and exclusion conditions,598 subjects(472 females with a average age of57.59±8.52 and 126 males with a average age 59.88±11.10)were finally included.The OP questionnaire was used to collected relevant information;enumeration data were expressed by rate(%)and compared using?~2 test;measurement data was presented as?±s and were compared by ANOVA.2.The network structure of BPNN was designed,and this model was established in pycharm software with python language.Statistically significant variables were selected for coding assignment and normalization as input variables.Cross entropy was used as loss function.The model was trained by the training set and its parameters were optimized.Finally,the network model after training is verified by the validation set.3.The subjects were randomly divided into the training group and the validation group,and their disease probability of OP was predicted through the BPNN model.OP assessment tools,such as OSTA(Osteoporosis Self-Assessment Tool for Asians)?ORAI(Osteoporosis Risk Assessment Instrument)?SCORE(Simple Calculated of Osteoporosis Risk Estimation)?OSIRIS(Osteoporosis Index of Risk)?OPERA(Osteoporosis prescreening risk assessment)and MORES(Male Osteoporosis Risk Estimation Score),were used to identify OP,and their scores were analyzed by ANOVA.Receiver Operating Characteristic(ROC)Curve was used to compare and evaluate the predictive power of BPNN and OP assessment tools.Result:1.The prevalence of OP in 126 male was 22.2%and that in 472 female was40.3%.the influencing factors of all subjects were gender,age,height,weight,BMI,education,fracture history,tea,coffee,carbonated drinks,milk,soy milk,calcium,vitamin D,steroids,and family members smoking(P<0.05).After analysis by gender,the influencing factors of female BMD were age,height,weight,education,menopause and postmenopausal years,awareness of OP,tea,coffee,carbonated drinks,soybean milk,vitamin D,steroids,and family members smoking(P<0.05).The influencing factors of male were weight BMI,milk,calcium and vitamin D(P<0.05).2.To establish the BPNN prediction model of OP,ReLU was selected as the activation function,30 hidden neurons and 0.01 learning rate were set.According to the prediction of the BPNN model,probability of OP in the training group was 0.42(female 0.40,male0.36);that in the validation group was 0.38(female 0.37,male 0.38).There was no statistical significance between the probability of OP of BPNN and the true prevalence(P>0.05),suggesting that the predicted value of BPNN can be used to estimate the true prevalence.3.The Areas Under the Curve(AUC)predicted by BPNN in training group were 0.970,0.984 and 0.859 respectively for the elderly,female and male(P<0.05),indicating that the model fitted the data well;the AUCs in validation group were 0.687,0.675 and0.677,respectively,and the male result was not statistically significant(P=0.1649),suggesting that the model could predict OP for the elderly and female.4.In training group,the AUCs of OSTA,ORAI,SCORE,OSIRIS,OPERA and MORES were 0.656(female 0.653),0.623,0.629,0.667,0.530 and 0.558,respectively.In validation group,the AUCs of OSTA,ORAI,SCORE,OSIRIS,OPERA and MORES were 0.726(male 0.677,female 0.742),0.749,0.653,0.718,0.549 and 0.573,respectively.Therefore,the above OP assessment tools could screen the OP in both groups with a moderate efficiency.Compared with OP assessment tools,the results of ROC in training group were higher(P<0.05).In validation group,there was no statistically significant difference between them(P>0.05).Conclusions:1.We found the prevalence of OP was high among all subjects.The influencing factors for the elderly were:gender,age,height,weight,BMI,education,tea,coffee,carbonated drinks,milk,soy milk,fracture history,calcium,vitamin D and steroids.After grouping by gender,there were two other factors,menopause and postmenopausal years,for female.The influencing factors for male were weight,BMI,milk,fracture history,calcium and vitamin D.2.A BPNN model based on OP related influencing factors was established,and the probability of OP was consistent with the true prevalence,which could be used to estimate the prevalence of population.This model could well fit the selected influencing factors and had certain ability to predict the individual disease probability.3.Compared with the OP assessment tool,this model had higher sensitivity,specificity,accuracy,and stronger prediction ability.According to the influencing factors,the software-based BPNN model could predict the incidence of OP more quickly.This model had high practical value in the screening and prevention of OP.