Modeling Longitudinal Relationships Between CKD-MBD Biomarker Trajectories with Interpretable Machine Learning in a...

Background/Objectives: Chronic kidney disease-mineral and bone disorder (CKD-MBD) is characterized by complex interactions between parathyroid hormone (PTH), fibroblast growth factor 23 (FGF23), phosphate, calcium, and vitamin D, yet clinical decisions are often based on static individual values. A trajectory-based approach that considers how longitudinal changes in CKD-MBD biomarkers relate to each other may provide a more integrated understanding. Methods: In 1968 adults with non-dialysis CKD (mean age 68 ± 12 years; 34% female; mean follow-up duration 3.02 ± 0.35 years), annualized slopes of CKD-MBD biomarkers (PTH, FGF23, phosphate, calcium) were derived from mixed-effects models, alongside slopes and intra-individual variability of other key features of CKD-eGFR, hemoglobin, albumin, bicarbonate, 25-hydroxyvitamin D, uric acid, urine albumin creatinine ratio (UACR). These 36 features, including baseline levels, were evaluated using Extreme Gradient Boosting regression with interpretability via SHapley Additive exPlanations (SHAP). Results: Models explained substantial variance in biomarker slopes (R²: 0.86 for PTH, 0.72 for FGF23, 0.71 for phosphate, 0.67 for calcium) without evidence of overfitting. Baseline values and eGFR slopes showed the strongest associations with biomarker trajectories. PTH and FGF23 slopes were positively correlated, while declining hemoglobin and greater hemoglobin variability were associated with steeper FGF23 and phosphate slopes. Phosphate slopes were associated with UACR slopes and bicarbonate slopes. Declining 25(OH)D slopes were linked with rising PTH slopes. Conclusions: This analysis revealed consistent relationships between CKD-MBD biomarker trajectories. Longitudinal changes in kidney function, hemoglobin, bicarbonate, UACR, and vitamin D were also associated with biomarker trajectories. Our findings confirm the multi-system nature of CKD-MBD, and provide a framework for examining potentially modifiable pathways in its progression.