Predicting anemia management in dialysis patients using open-source machine learning libraries

Abstract Background Managing anemia in patients with end-stage kidney disease (ESKD) undergoing hemodialysis (HD) is a major clinical challenge. It requires precise hemoglobin (Hb) control while minimizing erythropoiesis-stimulating agent (ESA) dosages to reduce cost and side effects. Despite recent advances, maintaining Hb levels within the recommended range (10–12 g/dL) remains difficult owing to inter- and intra-patient variability. Machine learning (ML) has shown potential in optimizing anemia management by predicting Hb levels and reducing ESA usage, though clinical implementation remains limited. Methods This study utilized ML to predict anemia management using data from 67 long-term HD patients treated at Tokyo Women’s Medical University. Key laboratory and treatment variables—such as blood counts, iron levels, and ESA dosages—were included in the analysis. ML models were developed using PyCaret, an open-source Python library. Performance metrics were compared across models, including XGBoost and LightGBM, to identify the most accurate algorithms. Results LightGBM and XGBoost outperformed logistic regression in predicting ESA and iron dosage changes, achieving high accuracy (e.g., area under the curve (AUC) = 0.86 for iron dosing). Feature importance analysis identified prior ESA and iron doses, Hb levels, and ferritin as critical predictors. These models closely mirrored actual prescribing patterns, suggesting feasibility for clinical integration. Conclusions ML models can accurately predict physician-prescribing behavior for anemia management in HD patients, indicating a promising role for artificial intelligence (AI) in improving treatment quality and operational efficiency.