Toward the Future: A Race‐Agnostic, Bayesian Approach to Defining Glomerular Filtration for Personalized Drug Dosing

Abstract

Accurate assessment of renal function is fundamental to dosing renally eliminated drugs. For decades, serum creatinine–based estimating equations have been used to approximate glomerular filtration rate (GFR) for this purpose. While these equations perform reasonably well at the population level, they exhibit substantial individual‐level imprecision and inaccuracy. Simultaneously, recent efforts to remove race from estimating equations addressed important ethical concerns but did not resolve the core limitation of population‐based estimation for individual drug dosing. In this perspective, we argue that personalized dosing requires a shift away from demographic surrogates toward direct measurement of renal clearance and clarify when incremental therapeutic complexity is warranted. We review the limitations of endogenous biomarkers, discuss the role of cystatin C, and propose exogenously measured GFR using iohexol integrated within a Bayesian pharmacometric framework as a clinically actionable approach to individualized drug dosing. Practically, this framework can be implemented by administering a small iohexol dose, collecting sparse timed samples, estimating absolute mGFR and its uncertainty with Bayesian software, and using that posterior estimate to refine renal dose selection.

Drug dosing is inherently an individual‐level intervention but is often guided by population‐based strategies. Ideally, individualized dosing strategies are guided by direct measurement of drug concentrations in the biologic matrix where pharmacodynamic effects (i.e., efficacy or toxicity) are expected to occur. In most clinical scenarios, access to the relevant tissue compartment (e.g., brain) is not feasible, and blood is therefore used as a practical surrogate. Even then, direct measurement of drug concentrations in blood is often limited by assay availability, turnaround time, and cost. ¹ As a result, clinicians frequently rely on indirect physiologic surrogates (e.g., patient response) to guide dosing decisions rather than consider drug exposure.

When considering overall drug exposure, dosing decisions for renally eliminated drugs are commonly guided by population‐based estimates of glomerular filtration rate (eGFR) derived from endogenous biomarkers, such as serum creatinine (in order to attempt to match average patient exposures with average intended effects). Historically, creatinine‐based formulas (e.g., Cockcroft–Gault, ² MDRD, ³ and CKD‐EPI ⁴ , ⁵ ) have been used as surrogates of GFR, and drug‐labeling authorities, including the US Food and Drug Administration (FDA), have relied on these estimates to define renal‐function subgroups (e.g., normal vs. reduced renal function) as they relate to drug dosing and corresponding dose‐adjustment recommendations. However, these equations were developed to characterize population‐level central tendency for purposes, such as CKD diagnosis, public health surveillance, organ donor and recipient evaluation, referral, and prognostication, ⁶ , ⁷ not to provide the individual‐level accuracy and precision required to predict the renal elimination of specific drugs, particularly in patients with disparate renal function. Notably, the FDA does not endorse a single estimating methodology ⁸ ; rather, it recognizes multiple creatinine‐based equations as acceptable for regulatory purposes, and clinical practice mirrors this heterogeneity. ⁹

Precision and accuracy of drug exposures at the level of the individual can be highly important. For drugs with narrow therapeutic margins or exposure‐response relationships that drive efficacy and toxicity, such as vancomycin, aminoglycosides, and β‐lactam antibiotics, inaccuracy in renal‐function estimation directly translates into suboptimal exposures and poor patient outcomes. ¹⁰ The problem is magnified in critically ill patients, where renal function is frequently unstable, augmented, or recovering. In such settings, the use of population averages, derived from equations curated in stable outpatients to guide individual therapy, is fundamentally misaligned with the goals of personalized dosing.

In an effort to be more “personalized,” these equations typically include a select few patient variables, including age and sex, along with biomarker values. Historically the equations also included a race variable, but recent recognition of the adverse impact of this factor on health equity has led to development of “race‐free equations.” ⁵ , ¹¹ , ¹² , ¹³ , ¹⁴ Removal of race does not alter the underlying population‐based structure of these models, and race‐free equations remain population‐derived estimators that do not resolve the central dosing problem, that is, population‐level accuracy does not guarantee individual‐level accuracy.

In this manuscript, we review the historical development of GFR estimating equations and explain why population‐based approaches fail to provide the individual accuracy and precision required for optimal drug dosing. We then propose pathways for dosing renally cleared drugs in settings where direct measurement of drug concentrations is not feasible. Central to this framework is the premise that truly accurate and precise drug dosing requires direct measurement of renal clearance rather than reliance on demographic surrogates. To that end, we present a Bayesian, individualized approach using iohexol‐measured GFR (mGFR) that obviates the need for demographic assumptions altogether. We argue that Bayesian modeling of exogenous renal clearance measurements can achieve substantially greater precision and accuracy at the individual patient level when compared to currently available population‐based estimating equations, providing a fit‐for‐purpose solution for personalized drug dosing.

HISTORICAL DEVELOPMENT OF CREATININE‐BASED GFR ESTIMATING EQUATIONS AND IMPLICATIONS FOR DRUG DOSING

The Cockcroft–Gault equation, introduced in 1976, was derived from a cohort of 249 predominantly Caucasian male patients and was designed to estimate creatinine clearance using age, weight, sex, and serum creatinine. ² The equation was based on a linear regression of weight‐normalized creatinine excretion vs. age, with the implicit assumptions of stable renal function and linear relationships between age and creatinine production. Rearrangement of this regression yielded the familiar form of the equation used in clinical practice (i.e., CrCL = [(140‐age) *weight]/[Scr *72] *0.85 if female). Notably, patient‐level factors such as body morphometrics (e.g., obesity, amputation, and cachexia), race, and sex‐related differences in creatinine generation were not explicitly evaluated in the original derivation. ² , ¹⁵ Despite subsequent attempts to refine or replace the Cockcroft–Gault equation in selected populations, ¹⁶ no universally accepted update exists. Furthermore, the Cockroft–Gault equation was derived using non‐standardized creatinine measurements; it has not been recalibrated for modern isotope dilution mass spectrometry (IDMS)–traceable assays, which yields lower serum creatinine values than the conventional methods used in the original study. ¹⁵

Over time, new estimating equations expanded upon this framework with the goal of improving population‐level performance. The Modification of Diet in Renal Disease (MDRD) equation, developed in 1999 from a cohort of 1,628 patients (88% White), incorporated additional covariates and included a 21% upward adjustment in GFR estimation for Black patients. ³ The Chronic Kidney Disease Epidemiology Collaboration (CKD‐EPI) equation, developed in 2009 from more than 8,000 individuals and externally validated in nearly 4,000 additional subjects, further refined non‐linear relationships between serum creatinine and mGFR and improved accuracy at higher levels of renal function. This equation similarly included a race‐based adjustment, increasing estimated GFR by approximately 18% for Black individuals. ⁵ In 2013, the Kidney Disease: Improving Global Outcomes (KDIGO) Work Group recommended use of the CKD‐EPI equation, including both creatinine‐ and cystatin C–based versions, for chronic kidney disease staging and risk stratification. ¹⁷ Of note, the equation used a race‐based correction coefficient for Black patients to attempt to adjust for differences in the mean serum creatinine. ¹⁷

GFR estimates were developed for CKD diagnosis, public health surveillance, prognostication, and transplantation consideration. ⁶ , ⁷ However, these estimates became central for drug dosing considerations ⁸ despite the fact that they were not developed for that purpose. Recently, inclusion of race‐based adjustment factors in traditional estimating equations raised concerns that these corrections led to downstream health equity implications, particularly by delaying kidney disease classification and transplant eligibility for Black patients. ⁶ , ⁷ In response, the National Kidney Foundation and American Society of Nephrology (NKF‐ASN) Task Force recommended adoption of race‐free equations for estimating renal function. ¹² As a result, contemporary clinical guidelines now promote race‐free population equations (e.g., 2021 CKD‐EPI eGFR equations with creatinine, cystatin C, or both biomarkers). ¹¹ , ¹⁸ While these equations are an important improvement toward health equity, their widespread adoption for drug dosing extended their use beyond the context for which they were originally designed (just as is the case with the Cockroft–Gault equation).

STRUCTURAL LIMITATIONS OF SERUM CREATININE

An ideal endogenous biomarker for GFR estimation should demonstrate constant production, exclusive glomerular filtration, no tubular secretion nor reabsorption, and no extrarenal elimination; yet, serum creatinine fails to meet these criteria. ¹⁹ Creatinine production varies with muscle mass, diet, age, sex, and critical illness‐associated catabolism (though there are attempts to control some of these factors in population‐based equations). ²⁰ , ²¹ Tubular secretion contributes variably to creatinine elimination, particularly as renal function declines, leading to overestimation of GFR in chronic kidney disease and acute kidney injury. ¹⁹

Most importantly for drug dosing, creatinine exhibits delayed kinetics when renal function is changing ¹⁹ , ²² , ²³ , ²⁴ , ²⁵ , ²⁶ , ²⁷ Most estimating equations rely on a single serum creatinine value and implicitly assume steady‐state conditions. ¹⁹ , ²⁸ In patients with acutely worsening or improving renal function, serum creatinine lags behind true changes in GFR. ²⁸ , ²⁹ As a result, creatinine‐based equations tend to overestimate true GFR in the early stages of acute kidney injury and underestimate GFR during renal recovery. ²⁸ , ²⁹ In patients with augmented renal clearance, serum creatinine often remains within the normal range, resulting in gross underestimation of actual GFR when creatinine‐based equations are used. ²⁸ , ³⁰ Serial‐creatinine approaches such as the Jelliffe method and related kinetic GFR concepts partially address this non‐steady‐state problem and, in our view, are more biologically plausible interim tools than static single‐point equations in patients with changing renal function. However, they remain endogenous‐marker models and therefore do not fully overcome uncertainty in creatinine generation, distribution, and secretion. ³¹ These predictable kinetic limitations explain why even well‐calibrated population equations systematically misrepresent renal clearance in many patients, including those with dynamic kidney function.

CYSTATIN C: IMPROVED SIGNAL, PERSISTENT CONSTRAINTS

Cystatin C is an alternative endogenous biomarker for GFR estimation. ³² , ³³ , ³⁴ , ³⁵ , ³⁶ Cystatin C has important non‐renal determinants (e.g., inflammation, cell turnover, obesity, medication exposures). Though, these liabilities may have less of a clinically significant impact on observed concentrations in acutely ill patients than non‐renal determinants of creatinine, ³² , ³⁷ , ³⁸ , ³⁹ making cystatin C a potentially attractive tool for assessment of renal function in acutely ill patients. ³³ , ⁴⁰ , ⁴¹ , ⁴² , ⁴³ The available clinical evidence suggests that prediction of GFR is best with cystatin C in addition to not as a replacement for creatinine. ⁴ , ¹¹ Cystatin C testing is increasingly available, ⁴⁴ and its use in chronic kidney disease staging continues to expand. However, as with creatinine‐based approaches, cystatin C–based equations remain population‐derived estimators. Substitution of cystatin C for creatinine does not materially reduce individual‐level bias nor improve accuracy when compared with measured GFR. Thus, cystatin C represents an important incremental improvement in signal responsiveness but does not overcome the fundamental limitations of endogenous, population equation‐based estimation for personalized drug dosing.

LIMITATIONS OF ALL eGFR EQUATIONS TO GUIDE DOSING OF RENALLY ELIMINATED DRUGS

Despite decades of refinement, all currently available eGFR equations, whether creatinine‐based, cystatin C–based, race‐inclusive, or race‐free, remain population‐derived estimators. Improving average bias at the population level does not guarantee adequate accuracy for individual patients. As a result, even the most contemporary equations remain ill‐suited to guide individualized dosing of renally eliminated drugs.

This limitation reflects the original purpose of these models. Estimating equations were developed to characterize population‐level central tendency for epidemiologic comparisons and chronic kidney disease classification, not to precisely estimate renal clearance for individual drug dosing. Regulatory labeling conventions subsequently adopted these equations to define renal impairment categories, and the FDA continues to recognize multiple estimating methodologies without endorsing a single preferred approach. ⁸ While this flexibility accommodates heterogeneity, it perpetuates reliance on surrogate estimates that assume stable renal function and homogeneous biomarker kinetics. Another practical limitation is that MDRD‐ and CKD‐EPI‐based values are generally reported as indexed mL/min/1.73 m², whereas drug dosing requires absolute clearance; failure to de‐index those values when appropriate, especially in patients at body‐size extremes, can introduce additional dosing error. ¹⁵ Of note, the FDA guidance document acknowledges the limited utility of creatinine‐based estimating equations in the setting of acutely changing renal function ⁸ but provides no specific recommendations for dosing in critically ill patients with dynamic renal dysfunction.

Together, these observations underscore that the central limitation of current renal dosing paradigms is not the inclusion or exclusion of race, nor the choice of endogenous biomarker(s), but rather the reliance on population‐based surrogates to guide individual therapy. Moving beyond these limitations will require direct measurement of renal clearance and modeling approaches capable of translating such measurements into actionable, patient‐specific dosing decisions.

ASSESSING PERFORMANCE AT THE INDIVIDUAL VS. THE POPULATION LEVEL

Large validation studies demonstrate that race‐free creatinine‐, cystatin C‐, or combined creatinine–cystatin C‐based equations can achieve acceptable bias at the population level (i.e., population‐level differences between measured GFR (mGFR, iohexol, and iothalamate) and race‐free estimated GFR (eGFR, creatinine)). In a landmark study, median predictions of eGFR were only off by 3.7 and 0.5 mL/min/1.73 m² in black patients and non‐black patients, though, at the individual level only 85% of eGFRs were within ±30% mGFR (i.e. proportion within ±30% of mGFR, P30). ¹¹ Other validation studies have found similar agreement at the population level (i.e., mGFR−eGFR of −0.6 mL/min/1.73 m² (95% CI, −1.2 to −0.2)), ⁴⁵ yet, differences at the individual level were marked (i.e., only about half of patients were within ±15% of eGFR across the continuum of GFRs from 15 to 90 mL/min/1.73m²). ⁴⁵ This study further demonstrated that the issue is variability between individuals and not that variability is driven by race. Very low population‐level racial differences between the mGFR and the eGFR (<5 mL/min/1.73 m²) were noted, but large individual‐level differences (i.e., a 95% prediction interval that spanned >50 mL/min/1.73 m²) existed. ⁴⁵ These findings underscore that average central tendency predictions are population‐level metrics that are not directly translatable to individual patients, regardless of the consideration of race. ⁴⁵ This pattern mirrors a well‐recognized phenomenon in population pharmacokinetic (PK) modeling. Well‐fit models can accurately estimate mean population parameters with modest between‐subject variability (i.e., population predictions show low average bias), yet still exhibit substantial between‐subject variability at the individual level, with CV%s of 50% and approaching or surpassing 100% in critically ill patients. ⁴⁶ When translated into drug exposure, such variability can correspond to orders‐of‐magnitude differences between individuals, ⁴⁷ often even among those receiving the same dose. From a drug‐dosing perspective, this degree of imprecision is clinically unacceptable for many individual patients receiving narrow therapeutic drugs (antimicrobials, chemotherapy agents, etc.). Importantly, limitations in population models persist regardless of whether race is included, excluded, or replaced by alternative covariates in population models. The problem is not race adjustment per se, but the reliance on demographic surrogates to infer renal clearance rather than directly measuring it. Once models are focused on an individual level, it is possible that race (or any other demographic feature for that matter) becomes irrelevant, not as a philosophical stance, but as a biologic reality. A tenet of Bayesian individualization is that patients are more likely to follow their own clearance trajectory than to regress toward a population mean. In this context, race‐agnosticism becomes a simple consequence of focusing on the individual (and starting with a model that does not include race). This argument persists for agnosticism in sex, age, and other demographic variables.

DIRECT MEASUREMENT OF RENAL CLEARANCE USING EXOGENOUS TRACERS

Exogenously administered tracers provide a direct and biologically grounded method for measuring glomerular filtration rate that does not rely on demographic assumptions. When freely filtered by the glomerulus, not secreted or reabsorbed by renal tubules, and eliminated exclusively by the kidney, such tracers serve as gold standard surrogates for measured GFR (mGFR). ⁴⁸ , ⁴⁹ , ⁵⁰ Additional desirable properties include reliable assay availability and rapid reflection of acute changes in renal function, allowing measurement of clearance under changing clinical conditions. ⁴⁸ , ⁴⁹ A key component of exogenous biomarkers is that since they are given as opposed to measured, the exact amount delivered to the body is known. This stands as different from endogenous markers that are measured, for which population methods are needed to estimate the amount delivered to the body.

Multiple exogenous tracers have been used to measure GFR, including inulin, radiolabeled compounds (e.g., ⁵¹Cr‐EDTA), iohexol, iothalamate, and relmapirazin. ⁴⁸ , ⁴⁹ Other drugs that are given for therapeutic intent (e.g., cefepime) could potentially also be used. Inulin and radiotracers have largely fallen out of favor due to logistical complexity, regulatory constraints, and limited assay availability. In contrast, iohexol and iothalamate have emerged as contemporary tracers, supported by reproducible assays using high‐performance liquid chromatography or liquid chromatography–mass spectrometry and extensive clinical experience across diverse populations. ⁵¹ Both plasma‐only and plasma‐plus‐urine methodologies have been validated, with plasma‐based approaches offering greater feasibility for routine clinical application. ⁵¹ , ⁵² , ⁵³ , ⁵⁴ Importantly, Bayesian population PK estimation is not the only valid way to interpret iohexol data. One‐compartment slope‐intercept methods, Bröchner–Mortensen correction approaches, single‐sample formulas, and noncompartmental analyses remain well‐established and are endorsed by standardization groups, particularly for stable outpatient measurement with predefined sampling windows. ⁵¹ , ⁵³ , ⁵⁴ , ⁵⁵ , ⁵⁶ Our argument is therefore not that these methods should be abandoned, but that Bayesian approaches offer distinct advantages when sampling is sparse, exact times vary, or renal function is dynamic.

From a pharmacometric perspective, modeling the disposition of iohexol closely parallels the therapeutic drug monitoring of renally eliminated medications. Iothalamate is largely very similar to iohexol, though some have posited that a small amount of tubular secretion exists based on a slight positive bias when compared to inulin. ⁵⁰ Iohexol pharmacokinetics have been described using one‐, two‐, and three‐compartment models, with model selection driven primarily by sampling density and timing. ⁵⁴ As with therapeutic drugs, richer sampling supports estimation of more complex models, whereas sparse sampling necessitates greater reliance on robust prior information.

BAYESIAN INDIVIDUALIZATION AS THE SOLUTION FOR THE INDIVIDUAL

Precision drug dosing requires methods that explicitly account for interindividual variability and update estimates as patient‐specific data become available. Bayesian pharmacometric approaches provide a formal and clinically proven framework to achieve this goal for therapeutic drug monitoring, ⁵⁷ , ⁵⁸ and we propose that this can be adapted for exogenous tracers, such as iohexol for measuring GFR. Bayesian dosing integrates prior knowledge derived from population pharmacokinetic models with individual patient data to generate patient‐specific posterior estimates of drug exposure or clearance. ⁵⁷ Rather than relying on static demographic surrogates, Bayesian methods dynamically refine predictions as new information is observed. This approach has been successfully implemented for many drugs with narrow therapeutic windows, most notably vancomycin and aminoglycosides, where Bayesian‐guided dosing is now standard practice in many institutions.

At its core, Bayesian dosing requires three fundamental components: (1) a well‐characterized population PK model that describes typical drug behavior and variability across representative patients; (2) individual‐level dosing history, including timing and amount of administered doses; and (3) measured drug concentrations. Together, these inputs allow estimation of individual clearance and exposure with substantially greater precision than demographic‐derived population‐based dosing rules.

However, for many renally eliminated drugs, particularly β‐lactam antibiotics, routine measurement of drug concentrations is not widely available or practical at present. In settings where specific drug assays and Bayesian priors are unavailable, renal clearance itself becomes the next most salient variable for predicting drug exposure. Thus, an alternative pathway to Bayesian individualization for drugs that are highly renally cleared is to directly measure GFR rather than the drug concentration itself. This distinction reframes renal‐function assessment as a modeling problem rather than an estimation problem. If renal clearance can be measured directly and incorporated into a Bayesian framework, individual drug dosing can be optimized without reliance on demographic surrogates or population‐based estimating equations. Importantly, such an approach remains inherently race‐agnostic.

Bayesian approaches can forecast individual trajectories of exposure profiles for xenobiotics, endogenous substances, or tracers for the purpose of renal‐function estimation. Model structures and parameter estimates are retained as priors and updated using patient‐specific data to generate more accurate posterior predictions. ⁵⁹ Such approaches have been widespread in pharmacometrics since the late 1980s and have become integral to many aspects of medicine, including drug development and dose finding, therapeutic drug monitoring, and post‐marketing safety and efficacy studies. ⁴⁶ However, Bayesian approaches for forecasting endogenous and exogenous tracers for eGFR and mGFR, respectively, remain underused. Such methods have great potential, especially in the setting of patients with dynamic renal function.

Population‐based renal estimating equations (i.e., eGFR) are severely limited for predicting GFR in patients with dynamic renal function. One reason for this limitation is that patients with unstable renal function are often systematically excluded from model development. Whereas such an approach may be useful for staging renal function in stable outpatients (i.e., intra‐patient variability is minimal and interpatient variation dominates) it breaks down in the dynamic setting of intensive care, where patients' renal function can change rapidly. Such challenges are well known when evaluating the behavior of therapeutic drugs in acutely ill patients. ⁶⁰ Several pharmacometric approaches exist to address dynamic drug clearance, including the use of time‐varying covariates (e.g., renal function) to describe inter‐occasion variability ⁴⁶ as well as multiple‐model approaches. ⁶¹ We therefore view renal‐function assessment as a continuum of tools selected according to clinical need: conventional creatinine‐based equations, de‐indexed eGFR when appropriate for body size, serial‐creatinine or kinetic methods for changing renal function, cystatin C‐augmented approaches, directly measured iohexol methods interpreted by slope‐intercept or Bayesian models, and ultimately drug‐specific therapeutic drug monitoring when available. Figure 1 conceptually situates these tools along that continuum.

Figure 1.

Conceptual continuum of renal function assessment approaches for drug dosing, ranging from conventional population‐based estimates when efficacy and toxicity risks are relatively low to direct individualized measurement when kidney function is uncertain and the clinical consequences of misclassification are greater. (Created and edited in ChatGPT and Grok with explicit input).

BAYESIAN MODELING OF IOHEXOL CLEARANCE‐ PRIOR WORK

Several groups have applied Bayesian pharmacokinetic frameworks to iohexol with the explicit goal of improving individual‐level precision in mGFR estimation. In one foundational analysis, richly sampled iohexol data from 176 patients (1–82 years; GFR 14–149 mL/min; 1,131 concentrations) were used to develop a two‐compartment, allometrically scaled model, which was subsequently validated in an independent cohort. ⁶² This study first analyzed a standard clinical 2‐sample monitoring strategy (i.e., 2 and 5 hours post‐dose) and demonstrated that this approach yielded limited individual‐level precision (i.e., ±10% accuracy in only a small proportion of patients). They then used Bayesian modeling and simulation to identify optimized early and late sampling times (e.g., 0, 30 minutes, 2 hours, and 5 hours) that substantially improved accuracy. Through simulation, they concluded that three strategically timed samples were sufficient to achieve clinically meaningful improvements in individual mGFR estimation. ⁶²

Subsequent studies extended this work to older adults and patients with reduced renal function. In a large analysis involving over 500 patients with external validation, Bayesian application of a three‐compartment iohexol model demonstrated P30 values of 98–99% using as few as two appropriately timed samples. ⁶³ Additional work will be needed for clinicians to balance feasibility of one vs. multiple samples vs. the desired degree of precision/accuracy. However by directly comparing the best P30 values achieved with population best estimates (i.e., 85%) ¹¹ to the iohexol based Bayesian modeling (i.e., P30 > 98%), it is clear that Bayesian modeling of exogenous tracers can provide more accurate and precise estimates of GFR.

Bayesian iohexol models have also been successfully applied in critically ill populations, including patients with unstable renal function, liver failure, and kidney transplantation. In these settings, sampling strategies involving three to four plasma samples, combined with Bayesian modeling achieved P30 values exceeding 95%, confirming that high individual‐level accuracy is attainable even in complex clinical scenarios. ⁶⁴ , ⁶⁵ Collectively, these data establish Bayesian modeling with iohexol as an important emerging method for measuring mGFR with high individual‐level precision.

Children also represent an important population for this framework. Pediatric patients are unique because creatinine generation changes with growth, body composition, and critical illness, while children similarly (to adults) require accurate dosing for oncology, transplant, and intensive care scenarios. Children across a continuum of ages are already represented in iohexol literature, including validation studies and studies in critical illness. ⁶² , ⁶⁶ , ⁶⁷ Taken together, this literature suggests that the general Bayesian iohexol paradigm is portable to pediatrics, although pediatric‐specific priors, reduced sample volumes, and microsampling approaches may be needed for implementation.

At the same time, strong performance with limited sampling is not automatic. When only one or two iohexol concentrations are available, the posterior mGFR may be dominated by the prior, especially if the underlying model spans a very wide renal‐function distribution or if sampling times are weakly informative. In that circumstance, predictions can regress toward the population mean or have wide individual uncertainty, ⁵⁵ , ⁶⁸ recreating some of the same practical limitations of population modeling. Fit‐for‐purpose implementation will therefore require curated libraries of validated models and limited‐sampling schemes for select use cases (e.g., severe renal failure, obesity, transplantation, pediatrics, and critical illness), ⁶⁴ , ⁶⁵ , ⁶⁹ , ⁷⁰ rather than omnibus use of a single prior across all settings.

The approach to move from population models is not limited to a singular methodology. As machine learning is becoming increasingly used, hybrid approaches that combine pharmacometrics with machine learning may help aid individualized predictions in the future. Early studies suggest that machine learning can improve iohexol clearance prediction, assist model selection, or adapt the effective strength of priors so that predictions rely more on population information when little posterior data are available but increasingly weight individual data as additional samples are observed. ⁷¹ , ⁷² , ⁷³ , ⁷⁴ Such approaches are attractive for repeated iohexol studies in critically ill patients, but they will require calibration, validation, and workflow simplification before routine clinical use.

A PRAGMATIC PATHWAY FOR CLINICAL IMPLEMENTATION

We do not envision the transition from eGFR‐based renal dosing categories to iohexol‐informed dosing as an all‐or‐none replacement. In the near term, measured absolute GFR could serve as a better physiologic input for existing renal dosing categories (e.g., package insert dosing or institutional dose‐adjustment tables). However, the ultimate preferred end point is direct refitting of drug‐specific pharmacokinetic models and labeling paradigms so that clearance is linked to measured absolute GFR rather than indexed or estimated surrogates.

Consider the following patient case. A 42‐year‐old woman presents with fever, aphasia, focal seizures, and cerebral spinal fluid PCR positive for HSV‐1 encephalitis. She is started on high‐dose IV acyclovir. On Day 2 she develops nonoliguric AKI, and acyclovir crystal nephropathy is suspected. Her creatinine rises from 0.8 to 1.6 mg/dL, but urine output remains preserved and neurologic status is worsening. Direct acyclovir concentrations are not available at the hospital. In this setting, reducing the acyclovir dose too aggressively will risk undertreating encephalitis which could be catastrophic, but failing to adjust the dose correctly risks additional kidney injury. Measured iohexol clearance could help understand the true GFR and provide the best possible clearance‐based guidance. This example is intentionally pragmatic, though similar examples with acyclovir are identifiable in the literature. ⁷⁵ The key implementation principle is not that every patient requires exogenous tracer measurement, but that patients with the greatest combination of dosing consequence and renal‐function uncertainty should be able to move further along the continuum from convenient population estimates toward direct individualized measurement.

Operationally, this approach could be implemented in a stepwise manner. First, the clinician identifies a high‐consequence dosing scenario in which renal function is uncertain and drug concentration monitoring is unavailable or impractical. Second, a small intravenous iohexol dose is administered with exact dose and administration time recorded. Third, one to two plasma samples are obtained at prespecified times, with later sampling added when low GFR is suspected or when greater precision is needed. Fourth, iohexol concentrations and sampling times are entered into a validated Bayesian platform to estimate posterior absolute mGFR and its uncertainty. Fifth, that posterior estimate can be used either to assign the patient to the most appropriate existing renal dose category or, where available, to inform an mGFR‐linked pharmacokinetic model for the drug of interest. Finally, if the patient's renal trajectory changes materially during therapy, the measurement can be repeated to update dosing decisions.

PRACTICAL LIMITATIONS AND IMPLEMENTATION CHALLENGES

Despite the conceptual and empirical advantages of Bayesian modeling with exogenous tracer–measured GFR, several practical limitations must be acknowledged. First, widespread clinical implementation would require access to timely and reliable iohexol assays. At present, iohexol measurement is not supported by a universally available, FDA‐cleared commercial assay, and turnaround times may be prolonged in institutions without on‐site capabilities. Real‐time (or near–real‐time availability) of mGFR estimates is essential for dose adjustment in acutely ill patients, and delays in assay reporting could limit clinical utility. However, we must not let the capabilities of the present cloud our vision of a potential future. Just as chemistry‐based assays were critical translational developments for drugs like vancomycin ⁷⁶ , ⁷⁷ , ⁷⁸ and aminoglycosides ⁷⁸ , ⁷⁹ (and have functionally been implemented in widespread fashion because of this flexibility), it is anticipated that iohexol use (and hence a market for developers) will drive innovation in assay/platform space.

Second, exogenous tracer–based approaches require blood sampling, and the number and timing of samples necessary to achieve acceptable precision remain important considerations. While existing studies demonstrate that high individual‐level accuracy can be achieved with sparse‐sampling strategies, the optimal balance between sampling burden and precision has not been fully defined across patient populations or clinical contexts. Novel capillary microsampling, dried blood spot testing, and volumetric devices may partially mitigate this burden, particularly in outpatient and pediatric settings. ⁷⁰ , ⁸⁰ However, bridging will be required when transitioning from serum or plasma to whole blood assays. Conversely, limited‐sampling protocols become more sensitive to timing errors as the number of samples decreases, making exact timestamp capture essential, especially when early samples are used. ⁶⁸ With sufficient model refinement and validation, a similar paradigm may ultimately be achievable from a single, strategically timed concentration in selected scenarios, substantially reducing sampling burden and improving clinical feasibility.

Third, implementation of mGFR‐informed dosing would require careful integration into existing pharmacokinetic models. Many current population PK models for renally eliminated drugs incorporate estimated GFR as a covariate on clearance. Substituting measured GFR for estimated GFR is not a simple plug‐and‐play exercise; rather, the functional relationship between renal drug clearance and mGFR would need to be refit and validated for individual drugs or drug classes. In the near term, measured absolute GFR could also be used to improve assignment to existing renal dose categories, but the longer‐term opportunity is to replace eGFR‐based bins with dosing models explicitly anchored to measured GFR.

Additional logistical considerations, including tracer procurement, nursing workflow, sample handling, and integration of Bayesian modeling into precision‐dosing software with bidirectional electronic health record connectivity, represent nontrivial barriers to adoption. These challenges are particularly relevant in intensive care settings, where competing clinical priorities may limit feasibility, even when the potential benefit to dosing precision is substantial. Additionally, more work will be needed to understand how to best estimate rapidly changing renal function, but real‐time monitoring would be able to document that instability is present.

While these are limitations that warrant consideration, it does not negate the importance of individualized renal clearance assessment for a subset of patients. Future work will be needed to understand where endogenous based methods are sufficient and how rapidly changing kidney function can be captured (e.g., kinetic GFR methods, dynamic adaptations of existing estimating equations, and development of new ICU‐specific models that better account for non‐steady‐state renal physiology). Ultimately, each of these strategies should be viewed as possible tools in the continuum of individualized GFR measurement.

CONCLUSION

Renal drug dosing remains fundamentally misaligned with its clinical objective: delivering precise, patient‐specific drug exposures in the setting of substantial biologic variability. Despite decades of refinement, creatinine‐ and cystatin C–based estimating equations, whether race‐inclusive or race‐free, were developed to describe population‐level central tendency and are not optimal for individual dosing decisions. The persistent reliance on these surrogates reflects historical convenience rather than biologic accuracy. We have argued that truly precise dosing of renally eliminated drugs requires a shift from demographic inference to direct measurement of renal clearance. Exogenous tracer–based measurement of GFR, coupled with Bayesian pharmacometric modeling, offers a race‐agnostic, biologically grounded framework capable of delivering substantially improved individual‐level precision. Existing data demonstrate that clinically feasible, sparse‐sampling strategies can achieve levels of accuracy unattainable with population‐based estimating equations, even in complex and critically ill populations.

While important practical and implementation challenges remain, including assay availability, sampling logistics, and integration into existing pharmacokinetic models, these barriers are navigable and should be viewed as opportunities for advancement rather than as arguments for the status quo.

Incremental advances in assay technology, model refinement, and workflow design can make individualized mGFR‐informed dosing increasingly achievable. Ultimately, the question is not whether population‐based estimates can be further optimized, but whether population averages should continue to anchor dosing for patients in whom exposure precision matters most.

FUNDING

No funding was received for this work.

CONFLICT OF INTEREST

M.H.S. reports serving as a consultant for DoseMe (Chair, Clinical Advisory Board) and is a co‐founder and equity holder of Safegate Therapeutics LLC, a company developing nephroprotective strategies for drug‐induced kidney injury. The remaining authors declare no competing interests. All other authors declared no competing interests for this work.