Construction Of Hypoproteinemia Risk Prediction Model For Maintenance Hemodialysis Patients Based On Machine Learning

Objective:To describe the occurrence of hypoproteinemia in maintenance hemodialysis patients and analy Ze the associated risk factors;A hypoproteinemia risk prediction model for maintenance hemodialysis patients was constructed based on four machine learning algorithms,and the performance of each model was compared by observing the difference between the prediction and actual classification of the model.After internal verification of the four prediction models,the optimal machine learning algorithm was selected to visuali Ze the quantitative relationship between risk factors and hypoproteinemia,providing a basis for clinical nursing workers to conduct early accurate screening and intervention.Methods:(1)Search the Chinese and English databases to obtain the risk factors related to the occurrence of hypoproteinemia in maintenance hemodialysis patients,and extract the related factors by discussing with mentors and experts and drawing up the Extraction Table of Risk Factors for hypoproteinemia in MHD Patients.(2)A total of 468 maintenance hemodialysis patients admitted to the nephrology department of a Grade-A hospital in Changchun,Jilin Province from January 1,2020 to December 1,2021 were selected as the research objects,and the relevant data of hypoproteinemia in maintenance hemodialysis patients were retrospectively collected.SPSS25.0 software was used for single factor analysis of the basic information of patients,corresponding disease data and biochemical examination data,and the variable P < 0.05 was obtained as the predictive variable of the model.(3)Taking the occurrence of hypoproteinemia(serum total protein < 60g/L or albumin level < 35g/L)as the outcome event,the optimal parameters were found using the 50 fold cross validation method and grid search method in Python3.10 software,and four models including random forest,support vector machine,logistic regression and BP neural network were constructed.Finally,multiple indexes such as area under ROC curve(AUC value),accuracy rate,accuracy rate,sensitivity,specificity and F1 balance score were used to evaluate model differentiation ability,and the optimal prediction model was selected.(4)Construct an importance matrix diagram for the optimal prediction model to describe the importance of each prediction variable.Results:(1)Among 468 maintenance hemodialysis patients included in this study,144 cases,accounting for 30.8%,developed hypoproteinemia,among which the incidence of hypoproteinemia was 31.6% in the training set and 27.7% in the test set.(2)18 risk factors including age,weight,dialysis age,dialysis frequency,high throughput dialysis,diabetes mellitus,hepatitis B,pulmonary infection,pulmonary tuberculosis,aspartate transferase,cholinesterase,prealbumin and globulin were obtained by single factor analysis of 468 patients with maintenance hemodialysis.(3)In accordance with 8: 2.After dividing the training set and the test set,the model was tested and verified.The average AUC value of the random forest was 0.924,95%CI(0.891-0.956),and the average accuracy rate was 0.924,95%CI(0.896-0.954).There were statistical differences in the prediction performance of the four machine learning models(P< 0.05).By comparing four models under different algorithms in the test set,it is found that the AUC values of the four models are all greater than 0.90,and the accuracy and specificity are above 0.85.The random forest model had higher AUC value(0.981),accuracy(0.955),accuracy(0.936),specificity(0.985)and F1 score(0.875)than the other three models.(4)In the feature importance ranking of the optimal random forest model,The top 10 were hypersensitive C-reactive protein(0.18),age(0.113),body weight(0.102),whether or not high throughput dialysis(0.086),dialysis age(0.079),whether or not lung infection(0.064),retinol-binding protein(0.055),whether or not diabetes mellitus(0.053),and hemoglobin(0.052),dialysis frequency(0.04),etc.Conclusion:(1)The incidence of hypoproteinemia in hospital patients undergoing maintenance hemodialysis is generally higher.(2)Hypersensitive C-reactive protein,age,body weight,whether it was used in high-throughput dialysis,dialysis age,dialysis frequency,diabetes mellitus,lung infection,retinol-binding protein,hemoglobin,etc.,were high risk predictors of hypoproteinemia.(3)Four novel machine learning algorithms built based on 18 risk factors for hypoproteinemia in maintenance hemodialysis patients showed excellent performance in predicting the occurrence of hypoproteinemia,among which the overall prediction performance and diagnostic efficiency of random forest model was better than the other three models,which could effectively distinguish the occurrence of hypoproteinemia in maintenance hemodialysis patients.More suitable for clinical work.(4)The predictive variables of the visuali Zed optimal model can help nurses identify changes in patient related indicators as soon as possible,screen out high-risk patients,enhance nurses’ risk perception,help clinical nurses do a good job in primary prevention,improve early accurate screening of hypoproteinemia,and provide basis for patients to carry out targeted health education and dietary guidance.