维持性血液透析患者贫血治疗中血红蛋白水平的机器学习预测OA

Objective To construct a prediction model for the next-month hemoglobin(Hb)level during anemia treatment in patients with maintenance hemodialysis(MHD).Methods A retrospective collection was performed on 7 139 anemia treatment cycles from 322 MHD patients with anemia admitted to Wenzhou People's Hospital from 2019 to 2025.Indicators collected included demographic characteristics,underlying medical history,physical examination findings,laboratory test results,and medication regimens.Feature selection was conducted using the Boruta algorithm combined with recursive feature elimination.The predictive performance for next-month Hb level was compared among 22 models covering three categories:non-temporal models,temporal models,and deep learning models.Model predictive performance and generalization ability were maximized through parameter optimization,feature interaction analysis,and a sliding window design.Finally,5-fold cross-validation was used to verify model stability.Results Among the 22 compared models,gradient boosting machine,extreme gradient boosting(XGBoost),light gradient boosting machine(LightGBM),and random forest showed favorable performance.After optimization,the LightGBM model achieved the best predictive performance for next-month Hb in the test set,with root mean square error(RMSE),mean absolute error(MAE),and coefficient of determination(R2)of 9.550 g/L,5.610 g/L and 0.658,respectively,which were superior to those of XGBoost(10.600 g/L,6.870 g/L,0.448,respectively),gradient boosting machine(11.719 g/L,8.954 g/L,0.431,respectively),and random forest(11.786 g/L,9.102 g/L,0.427,respectively).The optimized LightGBM model incorporated 18 indicators including previous Hb,serum ferritin,and C-reactive protein.In 5-fold cross-validation,the RMSE,MAE,and R2 were(9.03±0.54)g/L,(5.90±0.23)g/L,and 0.62±0.02,respectively.Conclusion The LightGBM model optimized by sliding window design and feature interaction can provide a reference for predicting Hb level based on routine monitoring indicators in clinical practice.