Home › Blog › Cardiology › DOAC Dosing in Renal Impairment

Quick Answer

DOAC dosing in renal impairment is one of the few everyday prescribing questions where the regulators themselves disagree. The US FDA permits apixaban and rivaroxaban in patients on haemodialysis; the European Medicines Agency does not. The pivotal trials used Cockcroft-Gault creatinine clearance, yet most laboratories report CKD-EPI eGFR. The randomised evidence in dialysis comes from three small trials, none adequately powered. When the sources disagree, a trustworthy medical AI should not quietly pick one answer. It should show the disagreement, name the source behind each position, and flag which parameter the dose depends on.

Key Takeaways

• FDA and EMA labels diverge most sharply below a creatinine clearance of 30 mL/min and in dialysis, where the same molecule can be permitted in one jurisdiction and not recommended in another.
• In older, low-weight patients, substituting CKD-EPI eGFR for Cockcroft-Gault changes the DOAC dose category in roughly half of cases, a finding with direct relevance to Indian practice.
• Off-label dosing is common and not benign: registry data associate both overdosing and underdosing with worse outcomes.
• Evidence on AI accuracy for anticoagulant questions is limited, but a 2026 evaluation found general-purpose chatbots matched guideline recommendations in under two-thirds of anticoagulant vignettes.
• The right design standard for medical AI is not "always give one answer." It is "show the sources, show where they differ, and show what the answer depends on."

Why Is DOAC Dosing a Useful Test for Medical AI?

Many clinical questions look simple and turn out not to be. DOAC dosing in renal impairment is a clear example, and it exposes something most discussions of medical AI miss.

The public conversation about AI in medicine tends to frame reliability as a binary: the model either gets the answer right or it hallucinates. That framing assumes there is a single right answer waiting to be retrieved. For a surprising number of real prescribing questions, there is not. There are several authoritative answers, issued by bodies that looked at the same pharmacokinetic data and reached different conclusions.

A cardiologist in Pune treating a 78-year-old woman with atrial fibrillation, a serum creatinine of 1.6 mg/dL and a body weight of 48 kg is not short of information. She is short of a clear picture of where the information conflicts. That is the problem this article examines, using DOAC dosing as the case study, because it shows what "evidence-based" should actually mean for medical AI.

Where Do FDA and EMA Disagree on DOAC Dosing in Renal Impairment?

They broadly agree on moderate CKD and diverge on severe CKD and dialysis. For creatinine clearance above roughly 30 mL/min, US and European labelling point in similar directions. Below that threshold, and especially in end-stage kidney disease, they separate.

The divergence is well documented. For apixaban, the FDA does not exclude patients with end-stage renal disease or those on dialysis, whereas the EMA does not approve apixaban for patients with a creatinine clearance below 15 mL/min. Rivaroxaban follows the same pattern: both agencies approve 20 mg daily above a CrCl of 50 mL/min and 15 mg daily between 15 and 50 mL/min, but the FDA has additionally approved use in dialysis. Apixaban and rivaroxaban are the only two DOACs with FDA-approved dosing recommendations for haemodialysis, and the EMA recommends against either drug in that population.

The disagreement extends into severe CKD for apixaban. The EMA summary of product characteristics reports that apixaban plasma concentrations rise by up to 44% in patients with a Cockcroft-Gault CrCl of 15 to 29 mL/min compared with normal renal function, and on that basis recommends 2.5 mg twice daily for this group irrespective of the other dose-reduction criteria. The US label, by contrast, applies its familiar "two of three" rule (age 80 or older, body weight 60 kg or less, serum creatinine 1.5 mg/dL or higher) without a separate CrCl trigger.

Professional societies add another layer. The EMA and the ESC advise against apixaban in patients on haemodialysis, while the ACC/AHA/HRS and the German AWMF take intermediate positions, stopping short of contraindication while acknowledging the moderate quality of the evidence.

The table below summarises the main points of divergence for stroke prevention in non-valvular AF. It is a comparison of source positions, not a dosing recommendation. Clinicians should confirm against the current approved prescribing information for the product in use.

Based on FDA and EMA product labelling and the ENGAGE AF-TIMI 48 trial. Always check the current prescribing information before use.

The edoxaban row carries a detail that surprises many readers: the renal concern is not only at the low end. In ENGAGE AF-TIMI 48, efficacy relative to warfarin was reduced in patients with high creatinine clearance, which led the FDA to restrict use above a CrCl of 95 mL/min. A DOAC question in renal impairment is really a question about the entire renal function range.

Which Renal Function Estimate Should Guide DOAC Dosing?

The label thresholds were built on Cockcroft-Gault creatinine clearance, and substituting eGFR can change the dose category, most often in older, low-weight patients.

This is where the problem moves from policy to arithmetic. The randomised trials used the Cockcroft-Gault formula to calculate renal function for DOAC dose adjustment, whereas the MDRD and CKD-EPI formulas dominate daily practice. The two are not interchangeable. CKD-EPI eGFR is indexed to body surface area and was never designed as the dosing input for these molecules.

The size of the discrepancy depends heavily on who the patient is. A multicentre Taiwanese analysis of 39,239 patients with AF found that for patients younger than 75 years with body weight of 50 kg or more, agreement with Cockcroft-Gault in dose-relevant renal categories was about 92% for both MDRD and CKD-EPI. For patients aged 75 or older weighing 50 kg or less, agreement fell to 41.2% for MDRD and 49.3% for CKD-EPI. For the elderly, underweight patient, the choice of equation is close to a coin toss for the dose band.

A literature review of simulated dosing reached a similar conclusion: dose discordance ranged from 28.8% underdosing to 59.2% overdosing when MDRD or CKD-EPI was substituted for Cockcroft-Gault, with eGFR equations overestimating renal function at lower levels, particularly in elderly patients and women. A 2025 Thai outpatient study found that roughly 70% of DOAC prescriptions complied with the approved labels, overdosing occurred in around 16%, and dose discordance between eGFR and Cockcroft-Gault was approximately 40%.

The contrast with Western populations is instructive. A 2026 Michigan registry analysis, in a cohort with a mean body weight of 95.6 kg, found rivaroxaban dosing disagreement in only 4.8% of patients when the race-free CKD-EPI 2021 equation replaced Cockcroft-Gault. The equations converge in heavier, younger patients and diverge in lighter, older ones. Evidence generated in one population cannot be assumed to transfer to another.

For Indian practice this matters directly. The typical Indian AF patient is lighter than the populations in which most dosing studies were run, and lower body weight can reduce volume of distribution and slow clearance; one pharmacokinetic study found apixaban exposure roughly 27% higher in subjects weighing 50 kg or less. The equation question is not academic. It is a daily source of dose error for exactly the patients Indian clinicians see most.

The newest guidance adds one more wrinkle. The first dedicated ESC guideline on cardiovascular disease and CKD, published at ESC Congress 2026, states that the evidence is sufficient to recommend DOACs, particularly factor Xa inhibitors at low GFR, over vitamin K antagonists in patients with AF and an eGFR of 15 mL/min/1.73 m² or above, while calling for more randomised trials in patients on kidney replacement therapy. The guideline frames its threshold in eGFR; the product labels frame theirs in Cockcroft-Gault CrCl. Both are correct within their own terms, and a clinician needs to see which unit each source is using.

What Does the Evidence Say About DOACs in Dialysis?

Three small randomised trials suggest DOACs are no worse than vitamin K antagonists in haemodialysis, but none was powered to settle the question.

This is the area where source disagreement is most defensible, because the underlying evidence is genuinely thin. Three randomised trials have been completed in patients on long-term dialysis with AF: Valkyrie (warfarin versus rivaroxaban 10 mg once daily, n=132), RENAL-AF (warfarin versus dose-adjusted apixaban, n=154) and AXADIA-AFNET 8 (warfarin versus apixaban 2.5 mg twice daily, n=97). Results leaned toward DOACs, but the studies lacked the power to support definitive conclusions.

RENAL-AF used the US label logic: patients were randomised to apixaban 5 mg twice daily, reduced to 2.5 mg twice daily for those aged 80 or older or weighing 60 kg or less, versus dose-adjusted warfarin, with major or clinically relevant non-major bleeding as the primary outcome. AXADIA-AFNET 8 chose the lower dose throughout. A 2024 systematic review concluded that AXADIA-AFNET 8 found no differences in safety or effectiveness between apixaban and VKA in dialysis patients, while RENAL-AF was too small to support conclusions.

The two apixaban trials did not even use the same dose, which reflects the pharmacokinetic uncertainty underneath. An earlier pharmacokinetic study in end-stage renal disease found that 5 mg twice daily produced supratherapeutic drug levels at steady state while 2.5 mg twice daily produced acceptable exposure, though stringent exclusion criteria limited generalisability.

Larger trials are underway. VISIONAIRE, comparing warfarin, edoxaban 30 mg once daily and no anticoagulation with a target of 1,500 patients, is among several larger randomised trials in progress. As of September 2026, the question of whether patients on haemodialysis with AF benefit from anticoagulation at all, and with which agent, remains open. Any answer to "Can apixaban be given to a patient on dialysis?" that does not say so is overstating the evidence.

Does Getting the Dose Wrong Actually Matter?

Yes. Registry data link both overdosing and underdosing to adverse outcomes, and off-label dosing is common.

In the US ORBIT-AF II registry of 5,738 patients, 9.4% of NOAC-treated patients were underdosed and 3.4% overdosed relative to FDA labelling. After adjustment, overdosing was associated with higher all-cause mortality (adjusted HR 1.91; 95% CI 1.02 to 3.60) and underdosing with more cardiovascular hospitalisation (adjusted HR 1.26; 95% CI 1.07 to 1.50). The associations deserve caution: off-label dosing was not linked to bleeding or stroke outcomes, which would be expected to mediate those signals, so residual confounding may explain part of the findings.

The renal-specific data are sharper. In a US analysis of 14,865 patients with AF, 43% of patients who qualified for renal dose reduction received standard dosing; this overdosed group had a higher rate of bleeding (HR 2.19; 95% CI 1.07 to 4.46), while underdosed patients had a higher rate of stroke (HR 4.87; 95% CI 1.30 to 18.26).

In Asian practice, the error usually runs the other way. A pan-Asian review found that non-recommended DOAC dosing is common in Asia, driven mainly by fear of bleeding, with the net clinical benefit of standard dosing often overlooked. A Korean nationwide cohort of AF patients weighing 60 kg or less found DOACs more effective and safer than warfarin, consistent even below 50 kg, with regular doses performing comparably to reduced doses. The instinct to reduce "to be safe" is understandable. The evidence does not support it as a default.

How Well Does General-Purpose AI Handle Anticoagulant Questions?

Evidence is limited, but what exists suggests general-purpose chatbots handle complex, patient-specific anticoagulant questions inconsistently and rarely show their sources.

Caution is needed here. Companies building medical AI have an obvious incentive to exaggerate the failures of general-purpose tools, the evidence base is small, and model versions change faster than papers are published.

What exists is still informative. A 2026 study tested ChatGPT and Gemini on 25 standardised vignettes against the 2025 ASRA guidelines on antithrombotic management before regional anaesthesia, with variables including renal function and specific DOACs. ChatGPT matched the guideline in 64% of evaluations and Gemini in 62%; ChatGPT produced errors classified as dangerous, recommending hold times shorter than required, in 20% of evaluations, and gave no specific timeframe in 16%. The authors noted that model outputs can be stated confidently even when wrong, and rarely include the reasoning or source behind a recommendation.

That last point matters most. A 64% concordance rate is a problem. A confidently worded answer with no visible source is a bigger one, because it removes the clinician's ability to catch the other 36%.

Human prescribing is not a high bar to clear either. A 2025 prospective study from a general medicine ward in Nepal found that among 1,461 drug orders, 348 required renal adjustment, and 63.5% of these were inappropriately dosed. Renal dose adjustment is hard for everyone. The question is what a tool should do to make it easier without adding a new, less visible failure mode.

What Should Medical AI Say When Guidelines Disagree?

It should report the disagreement, attribute each position to its source, and state what the answer depends on, rather than resolving the conflict silently.

Much of the discussion of medical AI reliability assumes the goal is a single correct answer. For questions like DOAC dosing in CKD, that goal is wrong. A tool that always produces one confident answer will, by construction, hide the jurisdictional and evidential disagreements that a clinician most needs to see.

Consider what a single-answer system does with the question "What dose of apixaban for an AF patient on haemodialysis?" If it was trained predominantly on US content, it may answer "5 mg twice daily, or 2.5 mg twice daily if age 80 or older or weight 60 kg or less," which reflects the FDA position. If it leans European, it may answer "not recommended." Both answers are defensible. Neither is complete. The clinician receiving either one has no way of knowing that the other exists.

That example points to four behaviours that should define evidence-based medical AI whenever the sources conflict:

1. Surface the conflict explicitly. The first sentence of a good answer to a contested question should say that it is contested. "US and European labelling differ for this patient" is more useful than any dose.
2. Attribute every position. Each recommendation should carry its source and date: the FDA label version, the EMA summary of product characteristics, the 2024 ESC AF guideline, the 2026 ESC CVD and CKD guideline, the 2023 ACC/AHA/ACCP/HRS guideline. Attribution is what lets a clinician apply local judgement.
3. Name the parameter the answer depends on. For DOAC dosing, the answer often turns on whether renal function was estimated with Cockcroft-Gault or eGFR, and on body weight. A good system asks for, or flags, the inputs that change the answer.
4. Grade the evidence behind each position. "Permitted by the FDA based on pharmacokinetic data" and "supported by a phase 3 outcomes trial" are very different statements. Dialysis dosing rests largely on the former, plus three underpowered randomised trials, and the answer should say so.

None of these behaviours requires the AI to be a clinical authority. They require it to be an honest reporter of the evidence, which is a more modest and more useful job.

What Does This Look Like in Practice?

Return to the patient from earlier: a 78-year-old woman with non-valvular AF, serum creatinine 1.6 mg/dL, body weight 48 kg, seen in a cardiology OPD in India. Her laboratory report shows a CKD-EPI eGFR of about 32 mL/min/1.73 m². Her Cockcroft-Gault creatinine clearance, calculated with actual body weight, is roughly 23 mL/min.

That single difference moves her across dose thresholds. On the eGFR figure, a clinician might treat her as having moderate CKD. On Cockcroft-Gault, which is the input the labels were built on, she falls into the 15 to 29 band, where the EMA recommends apixaban 2.5 mg twice daily, dabigatran is contraindicated in Europe, and she also meets two of the three US apixaban reduction criteria through her weight and creatinine. A system that takes the eGFR from the lab report at face value may place her in the wrong dose band for three of the four molecules.

This is the design principle behind OneRX, an AI clinical decision support platform for doctors, which is built to retrieve guideline and label sources, cite each one, and show where they differ rather than collapse them into a single unattributed answer. The point of such a system is not to make the decision. The point is to make sure the clinician sees the Cockcroft-Gault calculation, the FDA and EMA positions side by side, and the evidence grade behind each, in the time it takes to see the next patient.

How OneRX Supports Cardiologists With Questions Like This

OneRX began in oncology, where treatment evidence changes almost every week, and is now extending the same approach to cardiology. The principle does not change with the specialty: the doctor asks, OneRX summarises the published evidence, and the doctor decides. For a DOAC question in a patient with CKD, that works in five ways.

• Ask the question the way it comes up in OPD. A cardiologist can type or speak the question in plain language, for example "apixaban dose for a 78-year-old woman, 48 kg, creatinine 1.6", without building a search query.
• See the source behind every statement. Each part of the answer carries a reference, so the doctor can check which label, guideline or trial it came from before acting on it.
• Compare options side by side. Treatment options, contraindications and monitoring points are laid out in structured tables rather than long paragraphs.
• Check the newest trials. A clinical trials view helps find ongoing and recently reported studies, which matters in fast-moving areas such as anticoagulation in dialysis.
• Keep the decision with the doctor. OneRX organises and summarises published evidence. It does not diagnose, prescribe or write patient-specific treatment orders.

For hospitals, OneRX can also place the hospital's own approved protocols next to the global evidence, so the local anticoagulation policy and the international labels appear together on the same screen.

Why Is the Indian Context Different?

Indian patients differ from trial populations in body size and CKD burden, Indian guidance is consensus-based rather than trial-based, and no published study has benchmarked AI tools against Indian guidance.

First, the renal burden is substantial. The SEEK-India study estimated CKD prevalence at approximately 17% in its screened adult population, and CKD clusters with the same risk factors that drive AF: hypertension, diabetes and age.

Second, Indian guidance on this question is expert consensus rather than trial-derived. The 2015 Indian consensus guidance on stroke prevention in AF formulated expert opinion based on available trials and Asian and Indian subpopulation analyses, including dose recommendations across clinical scenarios and in the elderly. A 2022 consensus developed by Indian experts addressed frequently raised clinical questions on NOAC use and was positioned as a platform for future guidelines, policies, training and education. These documents are valuable, and they are also a reminder that Indian clinicians are often reconciling international labels with local expert opinion rather than following a single national standard.

Third, the product label question is not trivial. The prescribing information that accompanies a DOAC marketed in India may follow US wording, European wording or neither exactly, and clinicians should check the pack insert for the product actually dispensed rather than assume. Generic availability also means the same molecule may reach the patient from several manufacturers.

Finally, there is a gap that is one of the most important and least discussed facts in Indian medical AI. No published study identified for this article has benchmarked general-purpose AI chatbots against Indian clinical guidance, whether CSI, ICMR or the Indian NOAC consensus documents. Every accuracy figure in the literature, including the ones cited above, measures concordance with Western guidelines. For an Indian clinician, the reliability of these tools on the questions that matter locally is simply unknown.

What Should Clinicians Check Before Trusting an AI Answer on DOAC Dosing?

• Which renal estimate did the answer use? If the tool used eGFR, or did not ask, recalculate Cockcroft-Gault with actual body weight, particularly for patients aged 75 or older or weighing under 60 kg.
• Which jurisdiction does the answer reflect? If the answer concerns CrCl below 30 mL/min or dialysis and does not mention that FDA and EMA positions differ, it is incomplete.
• Is a source cited, with a date? An unattributed dose is not evidence. The 2024 ESC AF guideline, the 2026 ESC CVD and CKD guideline and the 2021 EHRA practical guide supersede older versions that many models may have absorbed.
• Is the evidence grade stated? Dialysis dosing rests on pharmacokinetic data and three small randomised trials. An answer that sounds as certain as a phase 3 result is overstating.
• Does it account for body weight and interacting drugs? P-glycoprotein inhibitors, age and weight all interact with renal thresholds.

Frequently Asked Clinical Questions

Should Cockcroft-Gault or eGFR be used for DOAC dosing in renal impairment?

Cockcroft-Gault creatinine clearance with actual body weight is the method used in the pivotal trials and in label thresholds, and the 2021 EHRA practical guide recommends it for DOAC dosing. Reviewers have recommended using Cockcroft-Gault in practice until large studies validate MDRD or CKD-EPI-based dosing. The discrepancy matters most in older and low-weight patients.

Can apixaban be used in patients on haemodialysis?

Positions differ by source. The FDA has approved apixaban for patients on haemodialysis, the EMA and ESC advise against it, and the ACC/AHA/HRS take an intermediate position. Randomised evidence is limited to RENAL-AF and AXADIA-AFNET 8, both small, and they used different doses. Larger trials are ongoing.

Is dabigatran an option below a CrCl of 30 mL/min?

In Europe, dabigatran is contraindicated below a CrCl of 30 mL/min. The US label provides a 75 mg twice-daily dose for CrCl 15 to 30 mL/min, based on pharmacokinetic modelling rather than outcome data. Availability of specific strengths varies by market.

Is reducing the DOAC dose to be safe reasonable in frail, low-weight patients?

Registry data suggest caution. In ORBIT-AF II, underdosing was associated with more cardiovascular hospitalisation, and Korean data in patients weighing 60 kg or less found regular doses performed comparably to reduced doses. Dose reduction should follow label criteria rather than general concern about bleeding.

How accurate are AI chatbots for anticoagulant dosing questions?

Published evidence is limited and model-specific. One recent evaluation found ChatGPT and Gemini matched guideline recommendations in 64% and 62% of anticoagulant vignettes respectively, with ChatGPT producing dangerous errors in 20%. No published study has benchmarked these tools against Indian guidance.

Conclusion

DOAC dosing in renal impairment is a small question with a large lesson. The FDA and EMA disagree on severe CKD and dialysis. The equation printed on a lab report may not be the equation the label assumes. The dialysis evidence rests on three trials with fewer than 400 patients combined. Dosing errors in both directions carry measurable harm.

A medical AI that answers this question with a single, confident, unattributed dose is not being helpful. It is hiding the part of the evidence the clinician most needs. The standard these tools should be held to is simple to state and hard to build: show the sources, show where they disagree, show what the answer depends on, and leave the decision with the doctor. The companion piece, Doctors Do Not Need AI to Choose GDMT, applies the same standard to heart failure.