Comparison of Machine Learning Models for Hemoglobin Prediction in Patients Undergoing Maintenance Hemodialysis

Ting Xie¹Xiaoyan Su^1,2Chen Yun^3,4Xiaohong Tang^1,5Xuejia Zheng⁶Jingjing Dong⁷Qi Guo⁸Shouping Zhu⁴Tang Donge^6*Yong Dai^6,9*Lianghong Yin^1*

• ¹The First Affiliated Hospital of Jinan University, Guangzhou, China
• ²Dongguan Tungwah Hospital, Dongguan, China
• ³Charite - Universitatsmedizin Berlin, Berlin, Germany
• ⁴Xidian University, Xi'an, China
• ⁵Guangzhou Medical University Second Affiliated Hospital, Guangzhou, China
• ⁶Shenzhen People's Hospital, Shenzhen, China
• ⁷Zhejiang Cancer Hospital, Hangzhou, China
• ⁸Southern University of Science and Technology, Shenzhen, China
• ⁹Anhui University of Science and Technology, Huainan, China

Objective: To estimate the next hemoglobin (Hb) levels in maintenance hemodialysis (MHD) patients, predictive models were developed using various Machine Learning (ML) algorithms. Methods: A total of 8,159 records from 2,104 MHD patients across 24 blood purification centers in Shenzhen were included. Eight ML algorithms were employed to develop prediction models: Linear Regression (LR), Least Absolute Shrinkage and Selection Operator (Lasso), Bayesian Ridge, Gradient Boosting (XGBoost), Random Forest (RF), Support Vector Machine (SVM), Multilayer Perceptron (MLP), and Long Short-Term Memory (LSTM). Subsequently, the performance of models was evaluated and compared. Results: Among all the models, the MLP performed the best performance, with an R² of 0.672, a mean absolute error (MAE) of 9.360 g/L, and a root mean square error (RMSE) of 12.438 g/L. The analysis indicated that the most recent Hb value (Hb(t-1)) was the strongest predictor. Conclusion: ML models based on demographic characteristics, dialysis records, and historical Hb data can effectively predict future Hb levels in MHD patients, which is helpful for early identification of anemia risk and timely clinical intervention.