Risk prediction of kidney function in long-term kidney transplant recipients

Abstract

Background: Limited tools exist for predicting kidney function in long-term kidney transplant recipients (KTRs). Elabela (ELA), apelin (APLN), and the APJ receptor constitute an axis that regulates vascular and cardiac physiology in opposition to the renin-angiotensin-aldosterone system.

Methods: Longitudinal, observational cohort of 102 KTRs who maintained graft function for at least 24 months, with no acute rejection history or active infection upon presentation. Serum APLN, ELA, fibroblast growth factor 23 (FGF-23) and α Klotho were tested using enzyme-linked immunoassay and compared with a control group of 32 healthy controls (HCs).

Results: When comparing with HCs, higher serum FGF-23, ELA and APLN, but lower ɑ Klotho concentrations were observed in long-term KTRs. Most KTRs had stable trajectories of renal function. Mean estimated glomerular filtration (eGFR) over 2-year follow-up was associated with significantly lower odds of graft loss (OR 0.04, 95% CI 0.01-0.15; p < 0.001). Baseline renal function was significantly correlated with mineral-bone markers (log[FGF-23]: r = -0.24, p = 0.02; log[α-Klotho]: r = 0.34, p < 0.001) but showed no significant association with aplnergic peptides (APLN: r = -0.07, p = 0.51; ELA: r = 0.17, p = 0.10). Univariable random forest regression indicated that baseline eGFR alone explained 87% of the variance in future 2-year eGFR, suggesting its overarching importance in late-term predictions. Incorporating both simple clinical characteristics and candidate serum biomarkers into a model predicting last available eGFR allowed for moderate predictive performance. In univariable Cox Proportion Hazard models, lower log(α-Klotho) (HR 0.26, 95% CI 0.12-0.58; p = 0.001) and higher log(FGF-23) (HR 2.14, 95% CI 1.49-3.09; p < 0.001) were significant predictors of death-censored allograft loss.

Conclusion: Both aplnergic and mineral-bone peptides appear as relevant candidate markers for future studies investigating their predictive performance regarding renal allograft outcomes.

Keywords: APLN; ELA; kidney; machine learning; risk prediction; transplantation.

Figure 1

Differences in serum concentrations of APLN, ELA, fibroblast growth factor type 23 (FGF-23) and alpha klotho compared between long-term kidney transplant recipients (LT-KTRs) and healthy controls. Point range is based on median with 1.5 * IQR for whiskers. *, **p value ≤ 0.05 and 0.001.

Figure 2

Temporal changes in renal function based on repeat eGFR and stratified by graft loss status (A). Predicted probabilities of renal allograft loss based on mean eGFR values calculated over time and age group based on quartiles (logistic regression—B).

Figure 3

Relationships between mean eGFR over 2-year follow-up and baseline serum concentrations of APLN (A, B), ELA (C), FGF-23 (D), alpha Klotho (E) and patients’ age (F). Analyses were conducted after excluding patients with initial eGFR <30 ml/min/1.73m².

Figure 4

Variable importance for “full” models predicting mean future 2-year eGFR minus baseline eGFR (A) and last available eGFR (B) based on random forest models.

Figure 5

Partial dependence plots for marginal effects of selected clinical features: mean eGFR over 2-year follow-up and baseline serum concentrations of alpha Klotho (A), FGF-23 (B), ELA (C), APLN (D), patients’ age (E) and body-mass index (BMI) (F).

Figure 6

Local prediction breakdown for last available eGFR over follow-up based on two real-life patient cases.