Research On Intradialytic Hypotension Prediction Models Based On Machine Learning Technology

Purpose:On the basis of evidence,to deeply explore the relevant factors affecting the occurrence of intradialysis hypotension(IDH)in maintenance hemodialysis(MHD)patients by machine learning techniques and construct a systematic and scientific model for predicting the risk of IDH,with the aim of providing a reference for optimizing and improving the blood pressure management in MHD patients.Materials and Methods:1.The original studies on factors influencing IDH in MHD patients in Pubmed,Web of Science,Embase,Cochrane Library,CNKI,CBM,Wanfang database and VIP database were systematically searched for the period from the establishment of the database to November 6,2021.Screening of the literature and extraction of data were performed independently by two investigators,and the Newcastle-Ottawa Scale(NOS)and the evaluation criteria recommended by the Agency for Health Care Quality and Research(AHRQ)were used to evaluate the quality of the literature for cohort studies and cross-sectional studies,respectively.Meanwhile,descriptive qualitative analysis of the included literature was conducted to summarize and summarize the current status of IDH occurrence in MHD patients and its main influencing factors.2.Based on the results of systematic evaluation,combined with the accessibility of clinical data and expert group meeting discussion of self-designed clinical information questionnaire on IDH in MHD patients,680 MHD patients who met the inclusion and exclusion criteria in a hemodialysis center in Chengdu from July 2018 to February 2022 were conveniently selected for a retrospective case-control study and divided into IDH group(n=170)and non-IDH group(n=510).The least absolute shrinkage and selection operator(LASSO)algorithm was used to screen key predictor variables affecting the occurrence of IDH in MHD patients to build a Nomogram model,a categorical regression tree(CART)model,and a extreme gradient boosting(XGboost)model,and the three models were internally validated and compared for predictive performance using self-sampling method.3.A prospective cohort study of MHD patients meeting the inclusion and exclusion criteria was recruited at 11 hemodialysis centers in Sichuan Province from March 2022 to July 2022 with a follow-up period of 1 month.The performance of the three pre-developed original prediction models in the external cohort was further explored and compared,and the coefficients of the best-performing original model was re-calibrated.In addition,new potential predictor variables(dialysis adequacy assessment metric Kt/V and body composition)were added to further update the models,model performance was assessed in terms of four dimensions:discrimination,calibration,reclassification,and net benefit,and a graphical score chart was drawn for visual presentation of the models.Results:1.This study included 15 studies,including 8 cohort studies and 7 cross-sectional studies.The influencing factors of the occurrence of IDH in MHD patients were summarized,mainly including the patient’s own factors,complications,laboratory tests and dialysis related factors,involving 35 subgroups of factors such as gender,age,diabetes,high-sensitivity C-reactive protein,albumin and ultrafiltration volume.2.Among the three risk prediction models,predialysis systolic blood pressure(Predialysis-SBP)was the most important predictor of IDH in MHD patients.In the internal validation,the area under the operating characteristic(ROC)curve(AUC)of subjects in Nomogram model,CART model and XGboost model were 0.979(95%CI:0.971-0.988),0.934(95%CI:0.914-0.954)and 0.992(95%CI:0.988-0.996),respectively.And the calibration plots showed that the Nomogram model and the XGboost model had better calibration than the CART model,and the predicted risk of IDH occurrence was in good agreement with the actual situation.3.A total of 2235 patients with MHD were included in this study,of which 327(14.6%)patients developed IDH during the 1-month follow-up period.External validation showed that the three original prediction models had poor prediction performance in the new patient cohort as a whole,and the original Nomogram model with the best performance was re-calibrated.For variables,Kt/V<1.2,body mass index(BMI)<18.5kg/m~2,standard percentage of mid-arm muscle circumference(%MAMC)<90%and possible osteoporosis,were added to the re-calibrated model for model updating.The updated model showed good discriminatory ability(AUC=0.817;95%CI,0.791-0.842),strong calibration ability(Brier score=0.081),the clinical net benefit within the threshold probability range of 15.2%to 87.1%,and the significant improvement of reclassification evaluated by the Net Weight Classification Improvement(NRI)(0.383;95%CI,0.271-0.503)and the Comprehensive Identification Improvement(IDI)(0.157;95%CI,0.092-0.285).In addition,the updated model was visualized as a graphical score chart,which can intuitively,conveniently and quickly obtain the risk of IDH.Conclusion:Based on the results of the system evaluation,this study used the regularized machine learning technology,through the combination of retrospective case-control study and prospective cohort study,qualitative analysis and quantitative analysis,finally developed a scientific,intuitive and low-cost IDH risk prediction model,which can be considered to guide the clinical management practice of MHD patients.