Which kidney function estimate should guide anticoagulant dosing and risk assessment in atrial fibrillation?

Handling Editor: Magnus Bäck

This editorial refers to ‘Impact of differences in kidney function estimation by creatinine compared to cystatin C on clinical outcomes in patients with atrial fibrillation’, by Philipp Krisai et al. https://doi.org/10.1093/ehjopen/oeag145

Why the choice of filtration marker matters in atrial fibrillation

Renal function is simultaneously a prognostic covariate and a dosing variable in the management of atrial fibrillation (AF). The renally eliminated fraction differs substantially across the direct oral anticoagulants (DOACs), from approximately 80% for dabigatran and 50% for edoxaban to roughly one-third or less for rivaroxaban and apixaban, yet every label conditions dose reduction on a single estimate of filtration, so the choice of that estimate propagates directly into drug exposure.¹ An estimate, however, is not a measurement, and creatinine is an imperfect filtration marker. Its generation depends on skeletal muscle mass and protein intake, and its clearance includes a variable component of proximal tubular secretion that is augmented by neurohormonal activation. In sarcopenic, cachectic, or advanced heart failure (HF) phenotypes, creatinine-based estimates therefore systematically overstate the glomerular filtration rate (GFR), and any dose derived from them is correspondingly liberal. Cystatin C is synthesized by all nucleated cells, freely filtered by the glomerulus, and catabolized in the proximal tubule; its concentration is largely independent of muscle mass, although it is not inert: inflammation, adiposity, thyroid dysfunction, and corticosteroid exposure all raise it. Overall, it is the more robust marker in comorbid populations, and the 2024 Kidney Disease: Improving Global Outcomes (KDIGO) guideline recommends the combined creatinine-cystatin C equation when creatinine-based estimation is expected to be inaccurate, and when GFR informs clinical management, including dosing of narrow-therapeutic-index drugs.²

Discordance between the two markers is not analytical noise. Their intraindividual difference, eGFRdiff, defined as the cystatin C-based minus the creatinine-based estimated GFR (eGFR), behaves as a phenotype in its own right: in chronic kidney disease (CKD), a negative eGFRdiff predicted mortality and end-stage kidney disease,³ and in PARADIGM-HF, it segregated patients carrying the biomarker signature and prognosis of advanced HF.⁴ A preliminary signal in AF came from the Swiss-AF cohort.⁵

In this issue of European Heart Journal Open, Krisai and colleagues extend this concept to AF.⁶ Among 3804 patients enrolled in the prospective Swiss-AF and Beat-AF cohorts, nearly one in three had an eGFRdiff exceeding 10 mL/min/1.73 m² in either direction, and 10.6% had an eGFRdiff below −10 mL/min/1.73 m². Across the three prespecified eGFRdiff strata, the annual incidence rates rose from 1.69 to 6.78 per 100 patient-years for major adverse cardiovascular events and from 1.24 to 6.07 for HF hospitalization. After multivariable adjustment, an eGFRdiff below −10 mL/min/1.73 m² carried hazard ratios of 1.50 for major adverse cardiovascular events and 1.49 for HF hospitalization. The gradient was mirrored biologically: high-sensitivity troponin T and N-terminal pro-B-type natriuretic peptide rose monotonically as eGFRdiff fell, and magnetic resonance imaging showed a stepwise increase in silent brain infarcts and white matter hyperintensities. Importantly, the authors frame their findings not only as risk stratification but as a warning about dosing, since a substantial minority of patients would be assigned different anticoagulant doses depending on which marker is used.

Which estimate did the pivotal trials actually use?

Contemporary practice has drifted from the evidence that generated it. RE-LY randomized fixed dabigatran doses with no renal adjustment algorithm, excluding a Cockcroft–Gault (CG) creatinine clearance (CrCl) below 30 mL/min.⁷ ROCKET AF reduced rivaroxaban to 15 mg once daily at a CG CrCl of 30 to 49 mL/min.⁸ ARISTOTLE dispensed with an equation entirely, halving apixaban when two of three criteria were met, one being a serum creatinine of 1.5 mg/dL or higher.⁹ ENGAGE AF-TIMI 48 halved edoxaban for a CG CrCl of 30 to 50 mL/min, a weight of 60 kg or less, or concomitant P-glycoprotein inhibition.¹⁰

Every licensed dose-reduction rule therefore derives from CG or from serum creatinine itself; none has been validated against a cystatin C-based estimate, and the European Heart Rhythm Association practical guide continues to specify CG.¹ Laboratories, however, report the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) or Modification of Diet in Renal Disease (MDRD) equations by default, so the value the prescriber sees is frequently not the value the label presupposes.

Registry evidence that the estimate changes the answer

Asian registry data quantify the consequence. Among 39 239 Taiwanese patients with AF, categorical agreement with CG was 91.7% for the MDRD equation and 92.0% for CKD-EPI in those younger than 75 years weighing 50 kg or more, but fell to 41.2% and 49.3% in those aged 75 years or older weighing 50 kg or less.¹¹ In a subsequent analysis of the same population, the discordance rates of DOAC dosing were 7.6% with MDRD and 6.5% with CKD-EPI, with the highest rates observed for rivaroxaban (11–13%) and the lowest for edoxaban (1%).¹² Most importantly, only doses truly on-label by CG were associated with less major bleeding events than warfarin (adjusted hazard ratio 0.34).

The European experience is concordant. In a prospective registry of 8506 patients with AF in whom eight equations were applied to identical data, CG agreed well only with its body surface area-adjusted variant and moderately with the remainder, yet CG and CG adjusted for body surface area afforded the best discrimination for death, coronary revascularization, and thromboembolism (area under the curve 0.661 and 0.660).¹³

The most recent Asian analysis applied eight equations to 21 318 DOAC-treated patients with non-valvular AF. The proportion classified as having preserved renal function varied from 25.7% to 65.3% according to the equation alone; agreement between CG and MDRD or CKD-EPI was at best moderate, and CG again discriminated best for thromboembolism, death, and major bleeding (area under the curve 0.707 vs. 0.641 and 0.664).¹⁴

These comparisons hold the marker constant and vary only the equation; the more fundamental question is what happens when the marker changes. An ARISTOTLE substudy of 14 980 anticoagulated patients compared CG, MDRD, CKD-EPI, and two cystatin C-based equations directly: CG correlated most poorly with the cystatin C-based estimate (correlation coefficient 0.49), the latter discriminated best for cardiovascular death (C-index 0.628), whereas CG remained among the best for major bleeding (C-index 0.612).¹⁵ The apparent paradox that the least physiologically accurate equation performs best for dosing-related endpoints reflects the fact that CG embeds age and weight, the covariates that drive both drug exposure and event risk. The present study extends the argument to a real-world AF population.⁶  Table 1 summarizes previous studies comparing renal function estimation formulas in AF.

Table 1.

Different renal function estimation formulas and markers for direct oral anticoagulant dosing and clinical outcomes in patients with atrial fibrillation

Study (year)	Population	Formulas/markers compared	Concordance	Association with clinical outcomes	Implication for practice
Hijazi et al., 202015	14 980 anticoagulated patients with AF; substudy of the ARISTOTLE randomized trial	CG, MDRD, CKD-EPI, CKD-EPI cystatin C, CKD-EPI creatinine plus cystatin C	Agreement between CG and CKD-EPI cystatin C was the poorest of all pairs (correlation coefficient 0.49)	Cystatin C-based eGFR discriminated best for cardiovascular death (C-index 0.628); CG was among the best for major bleeding (C-index 0.612)	Changing the filtration marker, not only the equation, alters risk prediction
Chan et al., 202011, 202212	39 239 patients with AF, Taiwan	CG vs. MDRD vs. CKD-EPI	Agreement with CG for the classifications (eGFR <15, 15–50, >50 mL/min)): 91.7% (MDRD) and 92.0% (CKD-EPI) in patients <75 y and ≥50 kg, but only 41.2% and 49.3% in patients ≥75 y and ≤50 kg	Only doses truly on-label by CG were associated with a lower risk of major bleeding vs. warfarin (aHR 0.34, 95% CI 0.26–0.45)	MDRD and CKD-EPI tend to overestimate renal function in elderly and underweight patients, leading to overdosing of DOACs
Boriani et al., 202513	8506 patients with AF, prospective European registry	BIS-1, CG, CG-BSA, CKD-EPI, EKFC, FAS, LMR, MDRD	CG showed good concordance only with CG-BSA and moderate concordance with all other formulas (weighted kappa vs. CKD-EPI 0.607, the lowest of all pairs)	Risk rose sharply below 60 mL/min for all formulas; best discrimination for CG (AUC 0.661) and CG-BSA (AUC 0.660)	The equation chosen affects both prescribing and risk prediction in a European population
Cheng et al., 202614	21 318 patients with non-valvular AF receiving DOACs, Taiwan	CG, CG-BSA, MDRD, CKD-EPI, JSN-CKDI, FAS, LMR, EKFC	Proportion classified as having normal renal function ranged from 25.7% to 65.3%; weighted kappa 0.512 (CG vs. MDRD) and 0.481 (CG vs. CKD-EPI)	CG showed the best discrimination for the composite of thromboembolism, death, and major bleeding (AUC 0.707) vs. MDRD (0.641) and CKD-EPI (0.664)	CG remains essential for risk prediction in Asian patients with AF
Krisai et al., 20266	3804 patients with AF, Swiss-AF and Beat-AF prospective cohorts	Creatinine-based eGFR vs. combined creatinine and cystatin C-based eGFR; eGFRdiff = cystatin C-based eGFR minus creatinine-based eGFR	28.3% had an absolute eGFRdiff >10 mL/min/1.73 m2 (10.6% <−10 and 17.7% >10 mL/min/1.73 m2)	eGFRdiff <−10 vs. −10 to 10 mL/min/1.73 m2: adjusted HR 1.50 for MACE and 1.49 for HF hospitalization; stepwise increase in hs-TnT, NT-proBNP, silent brain infarcts and white matter hyperintensities	Discordance between creatinine and cystatin C identifies a high-risk phenotype and raises concern about inappropriate DOAC dosing

AF, atrial fibrillation; aHR, adjusted hazard ratio; AUC, area under the receiver operating characteristic curve; BIS-1, Berlin Initiative Study 1; CG, Cockcroft-Gault; CG-BSA, Cockcroft-Gault adjusted for body surface area; CI, confidence interval; CKD-EPI, Chronic Kidney Disease Epidemiology Collaboration; DOAC, direct oral anticoagulant; eGFR, estimated glomerular filtration rate; eGFRdiff, difference between cystatin C-based and creatinine-based eGFR; EKFC, European Kidney Function Consortium; FAS, Full Age Spectrum; HF, heart failure; HR, hazard ratio; hs-TnT, high-sensitivity troponin T; JSN-CKDI, Japanese Society of Nephrology-Chronic Kidney Disease Initiatives; LMR, Lund-Malmö Revised; MACE, major adverse cardiovascular events; MDRD, modification of diet in renal disease; NT-proBNP, N-terminal pro-B-type natriuretic peptide.

Translating discordance into practice

First, dose selection should remain anchored to the CG CrCl: it is the metric embedded in the labels and in the trials, and the one most consistently associated with thromboembolic and bleeding events across independent registries.^11–14 CKD-EPI and MDRD are not interchangeable surrogates. The discrepancy warrants explicit verification in patients aged 75 years or older and in those weighing 50 kg or less, in whom upward misclassification and consequent overdosing cluster.^11,12 Second, cystatin C is worth obtaining when creatinine is expected to mislead: sarcopenia, cachexia or malnutrition, prolonged immobility, amputation, cirrhosis, advanced HF, and extremes of body habitus.² A markedly negative eGFRdiff should then be interpreted as a marker of aggregate disease burden warranting intensified surveillance, not as a substitute input for dose calculation, since no dosing algorithm has been validated on it.⁶ Third, renal function requires serial reassessment, more frequently as it deteriorates and during intercurrent illness, since a single baseline value cannot represent a variable that tracks congestion and catabolic state.¹

Where the field must go next

First, whether eGFRdiff can be converted from a prognostic marker into a dosing instrument will require pooled individual participant data from the pivotal DOAC trials and, ultimately, a randomized comparison of marker-guided against label-based dosing. Second, whether two markers suffice is equally unsettled. Albuminuria, hs-TnT, natriuretic peptides, β-2-microglobulin, and β-trace protein each carry prognostic information attributable partly to filtration and partly to non-GFR biology such as endothelial injury, inflammation, and myocardial stress. The biomarker gradients reported here suggest that eGFRdiff partly indexes the cardiovascular–kidney–metabolic phenotype rather than filtration alone.⁶ Third, calibration and trajectory also merit attention. The Youden-derived optimal threshold approximated zero rather than the prespecified ±10 mL/min/1.73 m²; longitudinal change in eGFRdiff has predicted mortality in CKD,³ and the incremental discrimination conferred by eGFRdiff was modest once a combined creatinine-cystatin C estimate was already in the model, in which the 10-year time-dependent area under the curve reached 0.772 for MACE and 0.803 for HF hospitalization.⁶

Krisai and colleagues establish that discordance between two routinely obtainable filtration markers is common in AF, biologically coherent and prognostically meaningful. Two caveats apply: the observational design cannot exclude reverse causation, whereby advanced disease drives both biomarkers and discordance, and brain imaging was limited to the older, more comorbid Swiss-AF cohort, raising selection bias. Whether acting upon the discordance alters outcome therefore remains untested. Until it is, the defensible position is to keep dosing tethered to the equation the trials employed, while allowing the discordance itself to flag the patients who merit closer scrutiny.

Author contributions

Chih-Min Liu (Conceptualization, Data curation, Funding acquisition, Investigation, Writing—original draft [lead]), and Tze-Fan Chao (Conceptualization, Data curation, Funding acquisition, Investigation, Supervision, Writing—review & editing [lead]).

Funding

This work was supported by grants from Taipei Veterans General Hospital (V114C-069 and V115C-103), and National Science and Technology Council (113-2314-B-075-034-MY3).

Data availability

No new data were generated or analysed in support of this article.