| With the increasing use of electronic medical records,intelligent diagnosis of medical data has great potential.By utilizing artificial intelligence technology,misdiagnosis caused by insufficient decision-making experience and pressure can be avoided,improving the quality of hospital treatment and patient survival rates.Pregnancy-induced hypertension is a common complication of pregnancy that occurs in the later stages,with symptoms including high blood pressure and proteinuria.The incidence of pregnancy-induced hypertension is approximately2.5% to 8% worldwide and can cause serious harm to maternal and infant health,including fetal death,premature birth,and birth defects.Therefore,effective diagnosis and treatment of pregnancy-induced hypertension are particularly important.This paper takes data-driven intelligent medical decision-making as the research background and non-balanced medical data as input,and establishes a classification model for non-balanced medical data to achieve intelligent diagnosis and decision support for pregnancy-induced hypertension.The main research is as follows:Firstly,to address the issue of class imbalance in medical clinical data,this paper proposed a hybrid oversampling algorithm SMOTE+Tomek links,which removed the noise data generated by SMOTE using Tomek links on the basis of oversampling,and improved the classification ability of the model compared to other sampling algorithms.Secondly,many existing medical evaluation models often focus only on obtaining high predictive accuracy,while ignoring the importance of preventing clinical patients’ own physical and functional features.Therefore,this paper proposed an improved F-Score feature selection algorithm,which calculated the F-Score value of each feature variable by computing the mean and variance of positive and negative samples and minority class samples in medical data,where a larger F-Score value indicates a stronger distinguishing ability of the feature.The improved F-Score algorithm not only screened out clinical patient features that perform better in the minority class classification and protected the authenticity of medical data,but also ranked the importance of features,which improved the classification ability of the model.The experiment results showed that the combination of multiple models with the proposed intelligent feature selection algorithm had improved predictive performance.The AUC of Bagging increased from 0.640 to 0.754,and the AUC of Ada Boost increased from 0.634 to0.695.Thirdly,to address the problem of poor generalization and predictive performance of a single model,this paper proposed a two-level Stacking ensemble model.The first level of the proposed model used 10 different base classifiers,including SVM,XGBoost,and Random Forest,and the predicted outputs of the 10 models after 10-fold cross-validation were passed on as input to the second-level LR model to obtain the final prediction result.Furthermore,to address the situation where a large number of base classifiers could result in the poor performance of the overall prediction performance in Stacking,this paper proposed a genetic algorithm-based ensemble strategy.Based on the multiple base classifiers in the first level of the Stacking model,multi-point crossover and mutation were conducted in the genetic algorithm,followed by the use of AUC as the fitness function for continuous iteration,to obtain the optimal model combination through natural selection.This method solved the problem of low efficiency in manual selection of classifiers and the impact of poor classifiers on the overall classification ability.Experimental results showed that the proposed Stacking model based on genetic algorithm achieved better performance compared to other classification models,with recall rate,F-measure,and AUC of 0.752,0.716,and 0.818,respectively.This research established a foundation for large-scale integration strategies,and made contributions to intelligent medical diagnosis in China. |
PDF Full Text Request