Diagnostic Efficiency Analysis Of Predicting Bipolar Disorder Based On Artificial Neural Network

Objective: Attempted to apply artificial neural network technology to construct the prediction model of bipolar depression,investigated and evaluated the efficiency of artificial neural network in predicting bipolar depression,to provide evidence for artificial neural network in early identification of bipolar disorder in clinical work.Methods:1 SubjectPatients with unipolar depression and bipolar depression diagnosed in the Sixth People’s Hospital of Hebei Province.There were 120 cases in total,unipolar depression group(n=60)and bipolar depression group(n=60).2 Admission criteriaThe current depressive episode in patients of bipolar depression groupconsistent with diagnostic criteria of "diagnostic and Statistical Manual ofmental disorders Fifth Edition"(DSM-5)for bipolar I disorder,currentmajor depressive episode(F31.31-F31.5)or bipolar II disorder,current majordepressive episode(F31.81).The current depressive episode in patients ofunipolar depression group is consistent with the diagnostic criteria forDSM-5 major depressive disorder,recurrent episodes(F33.0-F33.3).3 Data collection(1)The general information of the patients was recorded by self-made questionnaire,including demographic(gender,age,occupation,marital status,education level,personality,and family history),and clinical characteristics(accompanying characteristics,onset form,predisposing factors,total disease course,the duration of the onset,age of onset,type of onset,age of the first episode of depression,the number of depressive episodes,duration of depression and attempted suicide).(2)The application of Hamilton Depression Scale(HAMD),Montgomery’s and Asberg Depression Rating Scale(MADRS),Hamilton Anxiety Scale(HAMA),32-item hypomania light manic symptoms list checklist(HCL-32)to evaluate the clinical symptoms.Golbal Assessment Function(GAF)was used to evaluate the social function.Eysenck Personality Questionnaire(EPQ)was used to assess personality characteristics.The Defense Style Questionnaire(DSQ)was used to evaluate the defense mechanism.(3)Serum cortisol,thyroid function and sex hormones were detected by chemiluminescence.4 Statistical methodUsing SPSS18.0 statistical software and Medcalc statistical software for data entry and statistical analysis.(1)The data(120 cases)were randomly divided into the total set of training(108 cases)and test set(12 cases)according to the ratio of 9:1,respectively used to set up and test.In order to prevent over fitting the total set of training were randomly divided into train set(98 cases)and check set(10 cases)according to the ratio of 9:1.Using check set to check the results of training.(3)Application of Logistic regression analysis and neural network technology to construct the prediction model of bipolar depression.(4)The the receiver operating characteristic curve(ROC curve)of prediction models were plotted by using Medcalc software.The prediction performance of models were compared by the area under the ROC curve(AUC).Results:1 Comparison of the 36 main statistical variables between the two groupsThere were significant differences between the 13 variables in the two groups,include: age of onset(z=-3.276,P=0.001),duration of depression(z=-5.23,P<0.001),age of the first episode of depression(z=-2.596,P=0.009),free triiodothyronine(z=-2.559,P=0.010),EPQ E score(z=-2.447,P=0.014),average score of immature defense mechanism(z=-2.855,P=0.004),average score of mature defense mechanism(z=-2.208,P=0.027),with melancholy features(χ2=4.483,P=0.034),with atypical features(χ2=11.368,P=0.001),with psychotic features(χ2=7.728,P=0.005),predisposing factors(χ2=7.517,P=0.006),EPQ extroversion-unstable personality(χ2=6.530,P=0.011),extroversion impulse character(χ2=3.896,P=0.048).2 Results of univariate Logistic regressive analysiswithout melancholy features(OR=2.286,P=0.036),with atypical features(OR=4.600,P=0.001),with psychotic features(OR=4.808,P=0.009),no predisposing factors(OR=3.016,P=0.007),age of onset≤24 years old(OR=5.500,P<0.001),the number of depressive episodes≥4 times(OR=2.597,P=0.026),duration of depression≤28 weeks(OR=9.284,P<0.001),extroversion impulse character(OR=2.375,P=0.032),free triiodothyronine≥3.296 pg/ml(OR=2.890,P=0.008),EPQ E score≥46(OR=2.721,P=0.013),EPQ extroversion-unstable personality(OR=2.800,P=0.012),average score of immature defense mechanism≥5.6(OR=2.679,P=0.013),average score of mature defense mechanism≤5.1(OR=2.696,P=0.013),13 variables were associated with bipolar depression disorder.3 Construction of Logistic regression prediction model and prediction resultswith atypical features(OR=8.846,P=0.001),no predisposing factors(OR=5.624,P=0.003),age of onset≤24 years old(OR=6.045,P=0.001)and duration of depression≤28 weeks(OR=11.047,P<0.001)were used as predictor variables to construct the prediction model,and Logistic regression prediction model was established.The prediction model was: P=1/[1+exp(11.799-2.180×with atypical features-1.727×predisposing factors-1.799×age of onset-2.402×duration of depression)].The prediction accuracy of the total set of training was 81.5%,specificity was 77.8%,sensitivity 85.2%.The prediction accuracy of the test set was 75.0%,specificity was 66.7%,sensitivity was 83.3%.The total prediction accuracy of multivariate Logistic regression was 80.8%,specificity was 76.7%,sensitivity was 85.0%.4 Construction of neural network prediction model and prediction resultsThe standardized importance of the variables that construct the model ranked from large to small:duration of depression(100.00%),age of onset(72.30%),EPQ extroversion-unstable personality(56.20%),without melancholy features(46.60%),with psychotic features(29.30%),free triiodothyronine(26.30%),extroversion impulse character(25.60%),EPQ E score(24.40%),average score of mature defense mechanism(21.10%),predisposing factors(19.30%),with atypical features(15.90%),the number of depressive episodes(11.70%),average score of immature defense mechanism(10.40%).The prediction accuracy of the train set was 85.7%,specificity was 91.8%,sensitivity 79.6%.The prediction accuracy of the check set was 80.0%,specificity was 80.0%,sensitivity was 80.0%.The prediction accuracy of the test set was 91.7%,specificity was 100.0%,sensitivity was 83.3%.The total prediction accuracy of neural network was 85.8%,specificity was 91.7%,sensitivity was 80.0%.5 Comparison of prediction efficiency between Logistic regression model and neural network modelthe totall prediction accuracy of the neural network model was 85.8%,and the Logistic regression model was 80.8%.The total prediction accuracy of neural network was higher than that of Logistic regression model.The area under the ROC curve of neural network model was 0.902(95%CI:0.835~0.949),Logistic regression model was 0.881(95%CI:0.809~0.933).The neural network model had higher prediction performance.The prediction performance of Logistic regression model was moderate.Compared the area under ROC curve of the two prediction models.The results showed that the area under the ROC curve of neural network model was higher than Logistic regression model.There were no significant difference between the two models in the area under the ROC curve(z=0.961,P=0.336).Conclusions:It is feasible to construct prediction model of bipolar depression disorder with artificial neural network.The result show that the artificial neural network is better than the Logistic regression model in the prediction effect,but Logistic regression analysis has advantages in explaining the meaning of variables.Therefore,we can combine Logistic regression analysis and artificial neural network in clinical applications,to provide a predictive method for early identification of bipolar disorder.