Machine Learning-Based Prediction of Hemoglobin Variability in Patients with Chronic Kidney Disease Receiving Erythropoiesis-Stimulating Agents: A Real-World Retrospective Study.

Background: Anemia is a common complication of chronic kidney disease (CKD) and is predominantly managed with erythropoiesis-stimulating agents (ESAs). Hemoglobin response to ESA therapy varies substantially among patients and remains difficult to predict owing to multiple interacting clinical and treatment-related factors. Machine learning approaches may help model hemoglobin dynamics using routinely collected clinical data. Purpose: To develop and evaluate machine learning models for predicting weekly hemoglobin changes in patients with CKD receiving ESA therapy. Patients and Methods: A retrospective cohort study was conducted using electronic medical records from a tertiary care center in Thailand (January 2012 to December 2022). Adult patients with CKD stages 3B-5 and an estimated glomerular filtration rate (eGFR) below 45 mL/min/1.73 m Results: A total of 834 patients contributing 10,335 clinical visits and 9,935 time-series observations were included. The median age was 71 years and 58.6% were female. RMSE values were 0.121, 0.115, 0.117, and 0.115 for Decision Tree, Random Forest, XGBoost, and Support Vector Machine, respectively. Pearson correlation coefficients ranged from 0.457 to 0.553 (all p < 0.001). Conclusion: Machine learning models trained on routinely collected clinical data showed feasibility for predicting short-term hemoglobin variability in patients with CKD receiving ESA therapy. Larger multicenter and external validation studies are needed to refine predictive accuracy and evaluate clinical utility.