September 1st, 2026, 9 pm EST

Kidney Int Rep,  2026 Jun 23;11(9):106673., doi: 10.1016/j.ekir.2026.106673.eCollection 2026 Sep.

Randomized Controlled Trial of Large Language Model–Assisted Diagnostic Accuracy in Nephrology 

Raphaël Bentegeac, Philippe Amouyel, Bastien Le Guellec, Wisit Cheungpasitport, Mehdi Maanaoui, Aghiles Hamroun

PMID: 42502668
DOI:  10.1016/j.ekir.2026.106673

Introduction

Modern Artificial Intelligence (AI) encompasses a spectrum of computational approaches, from traditional machine learning to deep neural networks capable of processing increasingly complex data (see NephMadness 2026, AI region). Among the most rapidly evolving applications of deep learning are large language models (LLMs) that are trained on vast amounts of information and can interpret, synthesize, and generate natural language.  Since the release of ChatGPT in late 2022,  medicine has been grappling with a collective identity crisis.  In the last few years, there has been an expanding body of research evaluating the performance of different LLMs across a wide range of medical tasks, from answering clinical questions and interpreting complex case vignettes to diagnostic reasoning and medical decision-making. A recent systematic review (Chen SF et al, Nat Med, 2026)  found that LLMs outperformed human comparators in 33% of studies, with better relative performance in knowledge-based assessments than in real-world clinical settings, highlighting the persistent gap between benchmark performance and clinical utility. Essentially, AI programs are very good at multiple-choice exam questions.

Nephrology has not been an exception to this trend. In an evaluation of 975 questions from the Nephrology Self-Assessment Program (NephSAP) and Kidney-Self-Assessment Program (KSAP), GPT-4 achieved 74% accuracy, although remaining below the 77% passing threshold and the average performance of nephrology examinees (Miao J et al, Clin J Am Soc Nephrol, 2024). Moreover, beyond knowledge-based assessments, the potential applications of LLMs in nephrology have rapidly increased to include patient management, clinical documentation, decision support, and medical training. Evidence supporting the safety and effectiveness of LLMs in real-world nephrology settings remains limited (Yongzheng Hu et al, Ren Fail, 2025). 

Against this vast background, the REFINe (Reasoning Enhancement With Feedback From a Generative AI in Nephrology) was conducted. Rather than asking whether a LLM could independently solve nephrology questions, REFINe addressed a more clinically relevant question: Can access to a high-reasoning LLM improve physicians diagnostic accuracy when faced with complex, open-ended nephrology cases?

REFINe tested whether AI can better diagnosticians of physicians- a more relevant question for the future of AI-assisted medicine. 

Methods

REFINe was a prospective, randomized, open-label, parallel-group superiority trial conducted entirely online between November 2025 and March 2026. English or French-speaking residents and board-certified physicians were randomized 1:1 to diagnose complex nephrology cases with or without LLM assistance. Each case evaluation was one unit of analysis. The review board of lille University Hospital approved the study. The study group planned to enroll 100 participants and recruited 97. They report no change to the criteria, the outcomes or the analyses after the trial started. 
The cases came from the Make Your Diagnosis series in Kidney International, published between August 2007 and January 2025. Two nephrologists (N.F. and A.H.) rewrote 245 of these cases. The aim was to stop the LLMs from recognizing cases used during training. Study designers changed every sentence and every number but kept the clinical logic and the final diagnosis. Ten cases were then chosen at random for the trial. 
It is worth noting what those ten cases detailed, because this shapes how we should interpret percentages in the paper. There were four glomerular and vascular cases, two tubular, interstitial and cystic disorders, two electrolyte and acid disorders, one case of acute kidney injury in critical care, and one transplant case. Seven of the ten included kidney or allograft pathology. One included urine microscopy. Two were text only. One was pediatric. This question inventory tells us more about the difficulty than any adjectives or descriptors could. 

Participants were residents or board-certified doctors who spoke English or French. They were recruited through professional societies and conferences. The main exclusion criteria was previous contact with the specific cases. The team used the Pocock and Simon minimization algorithm (80% fixed and 20% random) to balance five features: certification status, speciality, years of experience, how often the doctor used LLMs, and academic status. The process was automatic, and the allocation sequence was hidden from the study staff. This is a stronger allocation method than most similar AI studies have used. 

Each participant received the 10 cases, in random order, on a dedicated platform. They could not go back to review a previous case. In the control group, doctors gave up to three diagnoses and a confidence score, then moved directly to the next case- no AI input at any point. In the AI group, participants completed the same first step, and the platform showed them a fixed GPT-5 answer with short, structured reasoning: they could revise their diagnoses and confidence once before moving on. The AI answer came from a two step prompt with no examples given to the model. The first step returned exactly three ranked diagnoses. The second step turned them into case specific reasoning with probabilities and a short summary of the key differences. All GPT-5 answers were produced in October 2025 and then frozen (this was a fixed ranked list with probabilities, not a conversation). 

Then two board-certified nephrologists scored the main outcome. They worked independently and did not know from which group each answer was obtained. They agreed on about 97% of answers before any discussion, and the rest were solved by consensus. For emphasis, the whole result of this trial rested upon the “yes” or “no” judgments about free text answers by the adjudicators. 

The sample size came from a simulation. It assumed 50 participants per arm, 10 cases each, 50% accuracy in the control group, and 70% in the AI group. With 100 participants, the power was calculated to be above 99%. The main analysis used mixed-effects logistic regression, with a fixed effect for the AI study arm and random intercepts for the participant cases. It was adjusted for gender, country, speciality, academic status, experience, confidence in AI, how often the doctor used LLMs, and how many AI tools they used. Every randomized participant was analysed in the assigned group. The subgroup analyses were planned in advance. The analyses by number of cases completed were done afterwards. The R 4.3.2 software environment was used for data analysis, statistical modelling, and graphics.

REFINe Population

Results

The 97 participants produced 556 case evaluations. The median age was 38 yo (IQR 34 to 45). Three out of four said nephrology was their speciality, and 69% did not work in academic centres. Nine countries were represented, but 87% to 93% of participants were French, depending on the arm. The two groups were balanced at baseline. About four in ten participants had never used an LLM. 

Primary Outcome

Secondary Outcomes

Doctors only took half of the good advice!

There were 178 answers that were wrong at first where the AI suggestion prompted the correct answer. Of those, 85 (47.8%) were corrected, and just under half were not. In the AI arm, doctors changed their top-3 in 47.4% of answers (from 41.3% to 53.6%).

Harm was rare

Only 1 of 69 answers that were right at first became wrong after a bad AI suggestion (1.5%; 95% CI: 0.04 to 7.9). In this format, doctors were not simply following the machine. 

No subgroup effect reached significance (all interaction p > 0.05). The analyses by number of cases completed gave similar effect sizes, but they also show something the main paper does not discuss.

One pattern stands out despite the null result: monthly LLM users showed by far the largest AI benefit (OR 10.51, 95% CI 3.18-24.67), well above never users or daily users, and the LLM-usage interaction (p=0.008) was the closest to significance of any subgroup tested (supp figure 3).

The adjusted odds ratios were 4.3 (1.99 to 9.28) for doctors who completed at least five cases, and 3.94 (1.70 to 9.13) for those who completed all ten. 

Before the trial, the team tested eleven models on all 245 cases. This took about three hours and cost around EUR 300. GPT-5 was the best performer in the pre-trial benchmark (77.1%), so it was the model used in the AI arm.

DeepSeek V3.2 and MedGemma 27B were tested with text-only inputs because those setups could not accept images. Their scores therefore, mix model ability with a missing input. 

Funding: The authors received no specific funding for this work. 

Data availability statement: Participant consent did not include permission for public data sharing. The analysis code will be made publicly available in a Git repository.

Discussion

From AI versus nephrologists to AI with nephrologists

LLMs have already shown that they can readily solve board-style questions. This study has moved the field of study from multiple-choice questions to complex, integrated clinical vignettes requiring the synthesis of history, laboratory data, pathology, urine microscopy and images. This trial addresses a very practical and pressing question: Does giving clinicians access to an LLM actually improve diagnostic reasoning in nephrology? In this RCT of fixed scenarios, the answer appears to be yes. While the results seem like a slam dunk endorsement of our machine overlords (we will add your biological and technological distinctiveness to our own; resistance is futile), we need to read in between the lines. Top-3 and Top-1 diagnostic accuracy improved by around 20% with AI, and the effect was preserved among participants who completed at least 5 cases, but it could have been an even greater improvement if the physicians had trusted the AI more. 
Clinical nephrology is open-ended, and new clinical information is integrated in ways particular to our specialty. Our diagnoses rely on the incorporation of heterogeneous data types, including a urinary sediment, light microscopy and immunofluorescence pattern, blood gases, and a serology panel that half fits multiple etiologies (diseases rarely read textbooks and don’t deal in absolutes). There is rarely a diagnosis that relies solely on an individual, decisive test. Answers may depend on trajectory rather than on a point value, since the same creatinine of 2.4 mg/dL means one thing over three days and something entirely different over three years. Finally, nephrologists draw on a long tail of low-prevalence entities, precisely the region in which a probabilistic model may either shine or possibly exaggerate the exotic. 

The human - AI gap may be the real story

There were 178 occasions where the clinician's initial answer was wrong, but GPT-5 gave the correct diagnosis. Yet only 85 (47%) of those answers were eventually corrected. This is probably one of the most important findings of the study. The problem is not always whether AI can correctly determine the diagnosis, but rather whether physicians are willing and/or able to recognize their mistakes. When it comes to correct diagnosis, humans are subject to errors in logic, including anchoring, premature closing, and confirmation bias (Mutlack Z, et al. J Clin Med 2025). AI assistance brought clinicians substantially closer to the apparent performance ceiling of the model, but nowhere near all the way (maybe the real trial was about us, not LLMs). That is where this paper becomes less about how clever GPT-5 is and more about how we, as clinicians, respond when AI disagrees with us. Do we trust it, challenge it, ignore it, or simply refuse to change despite the evidence? 

Why does a tool capable of very high diagnostic accuracy not automatically transfer that accuracy to the clinician sitting in front of it? The study workflow and the AI interaction here were very artificial and may have contributed to the gap. GPT-5 produced exactly three ranked diagnoses with concise reasoning and probability estimates. The answer was pre-generated and frozen. There was no chat, no follow-up, no chance to ask, “Why not my diagnosis?” or “Which one feature in this case makes you think that?” That matters because most of us do not use AI like a static answer key. We question it, push back, ask for discriminating features, ask it to argue against itself, and sometimes keep asking until either we are convinced or the model starts hallucinating creatively enough that we close the window. So, the question is not simply whether GPT is accurate. It is how much of that accuracy can be transferred to a clinician in a safe and useful way.

What has been examined previously

Other studies with the use of LLMs and physicians diagnosing complex cases have had mixed results. Goh and colleagues found no benefit at all from giving doctors an LLM, while the model on its own beat both groups (physicians alone and AI assisted physicians) by 16 points, similar to the current study (Goh E, et al. JAMA Netw Open, 2024). Everett and colleagues did improve accuracy with collaborative workflows (Everett SS, et al. NPJ Digit Med, 2026). We know LLMs keep evolving (quickly), but human integration may take longer due to trust and fear of machine hallucinations. Presentation of source materials might help physicians be more accepting of AI prodding changes, especially if the logic is solid and the primary references could be reviewed. Again, a conversation versus a mandate seems to work better in altering a doctor’s clinical reasoning.

Many previous studies scored the quality of the reasoning against a rubric (Goh E et al, JAMA Netw Open, 2024| Everett SS et al, npj Digit Med, 2026| Qazi IA, et al. Nature Health, 2026). REFINe scored a binary “yes” or “no”: was the right diagnosis somewhere in the list. The nephrology study mixed a percentage with a completeness rating. 

We don't want doctors to blindly accept whatever an AI says. If AI gives a diagnosis that doesn't fit the patient, it should be questioned, because currently AI has its own set of logistical errors (hallucinations, prejudice due to training errors, and context loss). There’s another side to this; even if the AI is actually right, our own clinical experience and preconceived notions can make us dismiss it. The trial results show that the ability to decide when to listen to AI may be just as important as the AI's accuracy. The authors make a similar point, suggesting that AI literacy and how the interface feels to us may be important steps in human/AI integration to achieve the best results.

Considering the reverse scenario, among the responses that were initially correct, very few became wrong after viewing the AI suggestion. This means the participating clinicians didn’t simply abandon a correct diagnosis whenever the AI disagreed with them. That is reassuring, yet we should be cautious, because the study was small and the AI was presented in a very controlled format somewhat dissimilar from real world clinical reasoning.

It is important to note, the control group in this study did not have access to the usual clinical resources. That is quite different from trials in which doctors could use Google, UpToDate, or other conventional resources. In addition, LLMs are evolving at an incredible rate, making what’s true in machine logic yesterday almost obsolete in just a few weeks. We cannot simply say that REFINe proves that LLMs are better than physicians (this sounds like a headline AI would write). It just shows that adding this particular AI workflow to unaided clinical reasoning improved diagnostic performance. In clinical practice, AI is being integrated in a somewhat haphazard way, slowing potential real world improvements. However, AI will be a useful tool for clinicians who choose to enhance diagnosis and treatment strategies with an eye on confirming machine-generated recommendations.

The number we are highlighting is the distance between 57.5% (physician enhanced with AI) and 77.1% (AI alone), and the fact that half of the correct suggestions were simply not used. That gap is not an AI model problem. It is a problem about how we read, trust, and wrestle with these tools. Qazi and colleagues showed just how much this can be moved: an AI-literacy curriculum alone lifted diagnostic assistance accuracy from 42.6% to 71.4%, a 27.5 point gain - arguably a bigger lever than the model itself (Qazi IA et al, Nature Health, 2026). It is becoming quite apparent that AI literacy is no longer optional for the modern physician. Mastery of AI (a potentially daily consulted assistant) sits closer than the urine sediment and biopsy interpretation in core kidney education than most of us are comfortable admitting. Similar to the acceptance of the electronic medical record, AI will be transformational once everyone gets on board.

Strengths

Methodologically, there is quite a lot to like. Group allocation used the Pocock-Simon minimization algorithm, with 80% deterministic and 20% random assignment. This balanced certification status, specialty, years of experience, LLM usage frequency and academic status. Randomization was automated and allocation remained concealed from study personnel during enrollment. The primary analysis used mixed-effects logistic regression with random intercepts for both participant and vignette, which is important because each clinician contributed multiple cases and each vignette had a different intrinsic level of difficulty. The model also adjusted for gender, country, specialty, academic status, experience and AI-use characteristics. Although the intervention itself could never be blinded, the diagnostic outcome adjudication was. This is a stronger allocation method than most similar AI studies have used. 

Weaknesses

The sensitivity analysis is also reassuring as the benefit did not disappear when the analysis was restricted to people who completed more of the task. At the same time, 97 participants generated only 556 completed evaluations, despite being assigned up to 10 cases each, and that’s before counting the 46 signups (32% of all registrants) who never completed a single case and were excluded from analysis all together. That raises a question: were difficult cases more likely to be abandoned or did participants simply decide that ten consecutive renal diagnostic puzzles were enough nephrology for one evening? Completion was voluntary and uncompensated, so informative non-completion remains worth thinking about even though the sensitivity analyses are reassuring.

Baseline confidence was similar between control and AI groups, but confidence increased significantly after exposure to AI (p<0.001). That makes intuitive sense if someone or something has just agreed with you, produced three polished diagnoses, and assigned probabilities to them. You naturally feel more certain. But there is an important caveat here. The confidence has to move in the correct direction, not just increase. If AI makes a correct clinician more confident, that is useful; if it makes a wrong clinician more confident, that may be dangerous.

The control arm also had no access to conventional resources. That makes the trial clean experimentally, but it is not quite how nephrology is practiced. If we come across a rare genetic tubulopathy, an unusual biopsy or a toxic alcohol case, or a lot many numbers, we may look something up. That is not cheating; that is how nephrology is currently practiced. The clinically relevant future comparison is therefore probably usual resources versus usual resources plus LLM. 

Conclusion

This is the first randomized trial in nephrology to show that a structured LLM workflow improves diagnostic accuracy, with very few errors introduced by the AI. Although this study may not currently have real world implications, the growing breadth and speed of LLM evolution will make it an essential tool for all future physicians.

One last detail, buried in the disclosures: the authors used Claude Sonnet 4.5 and GPT-5 to help write and edit this manuscript. A paper about human-AI diagnostic collaboration was itself a product of human-AI collaboration, which feels about right.   

Summary by

Anca Elena Stefan
Nephrology specialist, Assistant Lecturer
Romania

Assad S M

Assistant Professor, Nephrology

Christian Medical College, India

Srinivasavaradan Govindarajan
Assistant Professor (Ped Neph)
VMMC and Safdarjung Hospital, India

Luis Daniel Ramírez-Calvillo

Nephrology Fellow
Instituto Nacional de Cardiología Ignacio Chavez, México

Reviewed by

Akshaya Jayachandran, Brian Rifkin, Cristina Popa

Special thanks to guest reviewer & author

Wisit Cheungpasitporn (The AI guy)

Header created by AI and prompts from Assad SM

Check out the REFINe VA, and don’t forget to join us on #NephJC!

Revisa el resumen visual de REFINe y no olvides acompañarnos en #NephJC

Vasopressors or Fluids in Early Septic Shock

ARISE FLUIDS investigators, the ANZICS Clinical Trials Group, and the ACEM Clinical Trials Network; Sandra L Peake, Stephen P J Macdonald, Belinda D Howe, Elissa Milford, Peter Jones, Glenn Arendts, Michael Bailey, Jonathon Burcham, Anthony Delaney, Diana Egerton-Warburton, Daniel Fatovich, John F Fraser, Alisa M Higgins, Gerben Keijzers, Andrew A Udy, Patricia Williams, Paul J Young, Rinaldo Bellomo

PMID: 42274006

DOI: 10.1056/NEJMoa2516225

Why was the study done?

Fluid resuscitation has long been a pillar of septic shock treatment. However, the evidence so far has weakened the idea for treating 30 mL/kg as a rigid universal target. At the bedside, the decision is rarely based on metrics. Fluids may restore effective circulating volume, yet beyond the point of fluid responsiveness, another bolus may provide diminishing hemodynamic benefit while worsening tissue edema, venous congestion, pulmonary edema, cardiac dysfunction, and oxygenation (Prescott HC et at, CCM 2026 |  Monnet X et al, Intensive Care Med 2025).

So, when to stop giving fluids and transition to early vasopressor support?

The “fluid wars” have challenged the assumption that more is better, but so far, less has not clearly translated into better outcomes. The closest precedent is the CLOVERS (Shapiro NI et al, NEJM 2023) and the CLASSIC trial (Meyhoff TS et al, NEJM 2022 | NephJC Summary | Freely Filtered discussion), which tested different points along the resuscitation pathway without improving mortality.  Perhaps mortality alone does not capture the entire effect of an early hemodynamic strategy? 

ARISE FLUIDS (Peake SL et al, NEJM 2026) moved the question into an ED strategy comparison: after initial fluid resuscitation has begun, can restricting additional fluid and initiating vasopressors earlier improve patient-centered outcomes? Rather than selecting mortality alone, the investigators chose days alive and out of the hospital through day 90 (DAOH-90) as the primary outcome integrating survival, hospitalization, and readmission. The hope was that this new trial would help shape recommendations towards a more individualized approach to fluid resuscitation in sepsis (Prescott HC et at, CCM 2026 |  Monnet X et al, Intensive Care Med 2025).

What they did?

• Australian Resuscitation in Sepsis Evaluation: Fluid or Vasopressors in Emergency Department (ARISE FLUIDS).

• Investigator-initiated, multicenter, open-label, randomized superiority trial conducted at 51 hospitals in Australia, New Zealand, and Ireland.

• Randomization occurred in patients presenting to the ED with suspected sepsis and persistent hypotension (systolic blood pressure <90 mm Hg or mean arterial pressure <65 mm Hg) and lactate >2.0 mmol/L while receiving the first dose of an antimicrobial agent, despite at least 1L of IV fluid.

• Important exclusions included more than 2 L of fluid before enrollment, more than 6 hours since ED presentation, more than 2 hours since the final inclusion criterion was met, hypotension attributed to another cause, an immediate need for surgery, imminent death, or a clinician determination that either strategy was unsuitable.

 These criteria selected an early, still-modifiable phase of shock rather than established refractory ICU shock.

• Patients were randomized 1:1, consent was not obtained for 37 patients; the intention-to-treat population included 963 participants: 481 assigned to the vasopressor strategy and 482 to the fluid strategy. The assigned strategy continued for at least 6 hours, and up to 24-hours, while patients remained in a critical-care area.

• Vasopressor strategy: routine resuscitation fluid was stopped, vasopressors were started immediately and titrated to the blood-pressure target, and 250-mL rescue boluses were allowed when clinically indicated.

• Fluid strategy: up to 1L was given during the first hour, followed by 500-mL boluses when hypotension or hypoperfusion persisted; vasopressors were started after the fluid target was restored or when further fluid was not considered appropriate.

Outcomes 

• Primary: 

• Days alive and out of the hospital through day 90 (DAOH-90)

• Secondary: 

• Mortality; days alive at home; and days free of mechanical ventilation, vasopressors, and acute kidney-replacement therapy. 

• Safety included pulmonary edema, vascular-access complications, ischemic events, and serious adverse events.

What do the results say?

The groups were balanced at baseline. Median age was 68 years in the vasopressor group and 69 years in the fluid group, median APACHE II score was 18 in both groups, and median lactate was approximately 3.2–3.3 mmol/L. Both groups had already received a median of 1.5 L, approximately 18 mL/kg, prior randomization. 

Did the trial separate the two strategies?

By 6 hours, the vasopressor group had received a median of 500 mL of additional fluid versus 1,500 mL in the fluid group. By 24 hours, the corresponding volumes were 1,140 mL versus 2,248 mL, a median difference of −1,108 mL (95% CI, −1,395 to −850). Vasopressors were started a median of 1 hour earlier (0.4 vs 1.4 hours), and more patients received them during the first 24 hours (86.5% vs 67.6%; difference, 18.9 percentage points; 95% CI, 13.3 to 24.5). Peripheral administration was common (71.9% vs 55.4%). 

Primary outcome: 

The median number of days alive and out of hospital at day 90 was: 76 days vs 76 days, for early vasopressors versus more fluids, respectively.

Restricting additional fluid and starting vasopressors earlier did not improve the primary patient-centered outcome. The adjusted analysis showed no clear difference for the major secondary outcomes, including mortality. 

Ninety-day mortality was 16.4% with the vasopressor strategy and 14.4% with the fluid strategy (RR, 1.14; 95% CI, 0.85 to 1.54). Mortality at 28 days, days alive at home, and days free of mechanical ventilation or acute kidney-replacement therapy were also similar. Also, vasopressor-free days were slightly fewer in the early-vasopressor group, 

Pulmonary edema occurred in 0.6% of patients assigned to early vasopressors versus 5.0% assigned to more fluid (RR, 0.12; 95% CI, 0.03 to 0.39; P<0.001). Other adverse events were uncommon and broadly similar; no ischemic events or serious adverse events were reported.

Critical view

So, the ARISE FLUIDS is not exactly a trial of “vasopressors versus fluids”. Both groups received fluids, and also received vasopressors. The trial compared the timing and relative emphasis of these two therapies, achieving a clear separation between treatment strategies in a controlled setting. This strategy did not improve the overall recovery, but it did reduce the fluid exposure. 

This finding fits into an increasingly consistent story from CLASSIC  and CLOVERS : changing fluid exposure is feasible, but simply giving less fluid has not produced a universal advantage. 

CLASSIC trial achieved 2L lower ICU fluid exposure, but 90-day mortality was essentially identical (42.3% vs 42.1%), as were days alive without life support and days alive out of hospital. CLOVERS moved the question and prioritized vasopressors over additional fluid. The restrictive group received about 2.1L less fluid, with earlier and longer pressor use, but death before discharge home by day 90 was similar (14.0% vs 14.9%; P=0.61). In ARISE FLUIDS, the fluid separation was smaller, about 1.1L less fluid, with 18.9% vasopressor use, and yet both groups showed no superiority at all.  

What about the apparent few patients with pulmonary edema in the vasopressor group? The infection source was heterogeneous, as expected in the sepsis population, and there was a modest imbalance in respiratory infections, with more cases in the fluid group. Could this influence safety outcomes? Still findings remain provocative, particularly given the lower fluid exposure. 

This may also be one of the trial's most interesting blind spots. Hemodynamic assessment was still largely driven by conventional clinical variables, and POCUS was not systematically incorporated to characterize fluid responsiveness as part of the strategy. Although CLOVERS allowed its use, it was not part of a standardized intervention. 

Conclusion

Where does this leave us? Unsurprisingly, there may not be one resuscitation strategy for every septic patient. The next step may be less about finding another universal fluid threshold and more about identifying the patient's hemodynamic phenotype in real time, integrating perfusion, fluid responsiveness, cardiac function, and venous/pulmonary congestion, to decide who needs another bolus and who needs vascular support instead.

Written by Milagros Flores

Reviewed by Brian Rifkin 

Randomized Trial of Dietary Nitrate Supplementation in CKD Pregnancy (ORCHARD-BEET)

Priscilla Smith, Danielle Ashworth, Rachel Hung, Yanzhong Wang, Andrew J. Webb, Kathryn Dalrymple, Katherine Clark, Lucy C. Chappell, Kate Wiles, Kate Bramham, on behalf of the ORCHARD Consortium

DOI: 10.1016/j.ekir.2026.106704

Why was the study needed?

CKD affects 0.2% to 6% of pregnancies (Williams D & Davison J, BMJ, 2008| Piccoli GB et al, J Clin Med, 2018), rising with diabetes, obesity, hypertension, and delayed childbearing (GBD 2015 Maternal Mortality Collaborators, Lancet, 2016). In stages G3-G5 CKD, ~46% of women lose over a quarter of pre-pregnancy eGFR or need renal replacement therapy within 12 months of delivery. (Wiles K et al, NDT, 2021). Outcomes haven’t improved in two decades (Smith PA et al, Nat Rev Nephrol, 2025), and no dedicated interventional trial has run since Kincaid-Smith’s 1995 heparin/ dipyridamole study (Nephrology, 1995)- a three-decade gap ORCHARD-BEET was designed to address (Smith P et al, KI reports, 2026). 

RAAS inhibitors and SGLT2 inhibitors, the strongest CKD progression-slowing agents, are contraindicated in pregnancy and lactation because of fetal risk and insufficient breastfeeding safety data (Wiles K et al, BMC Nephrol, 2019). Dialysis, meanwhile, is neither universally accessible nor scalable. KDIGO 2024 itself concedes that pregnancy-specific management evidence beyond monitoring remains unclear (KDIGO CKD Work Group, KI, 2024). 

Nitric oxide physiology: normal pregnancy, CKD, and preeclampsia

Normal pregnancy
Nitric oxide (NO) drives two relevant adaptations. Renally, NO-mediated afferent/efferent arteriolar dilation lowers vascular resistance and raises renal blood flow and GFR- the physiological hyperfiltration of pregnancy. In the placenta, NO drives trophoblast remodeling of spiral arteries from narrow, high-resistance vessels into wide, low-resistance ones able to meet fetal demand.

CKD
NO bioavailability falls via impaired synthesis (Wever R et al, Arterioscler Thromb Vasc Biol, 1999) and accumulated NO synthase (NOS) inhibitors- such as asymmetric dimethylarginine (ADMA) (Schmidt RJ, Kidney Int, 2000). A kidney entering pregnancy with impaired NO synthase activity cannot mount the normal hyperfiltration response, which is a key contributor to eGFR loss. 

Preeclampsia
Same deficit, different organ: reduced NO and endothelial dysfunction impair spiral artery remodeling, leaving the placenta underperfused, triggering anti-angiogenic factor release, maternal endothelial dysfunction, and fetal growth restriction- contributing to, but not strictly causing it. CKD and preeclampsia share this mechanism, explaining CKD pregnancies’ elevated preeclampsia risk.

Dietary nitrate as an NO synthase-independent bypass

Oral/ enteric bacteria reduce nitrate to nitrite, then to NO, a route Lundberg, Weitzberg and Gladwin showed works independently of the classical L-arginine-NOS route (Lundberg JO et al, Nat Rev Drug Discov, 2008), and one preferentially active in hypoxic, acidic tissue, exactly where NOS activity is the weakest. Nitrite-mediated vasorelaxation of human placental chorionic plate arteries is enhanced under hypoxia (Tropea T et al, Nitric Oxide, 2018), and low-dose nitrite restores renal tissue oxygenation and limits the creatinine rise following ischemia-reperfusion injury in animal models (Cantow K et al, Sci Rep, 2018). This gives the intervention a mechanistic rationale specific to the compromised, hypoxic tissue found in pregnancy with CKD and preeclampsia.

Beetroot juice had already shown plausible benefit and acceptable safety in hypertensive, non-CKD pregnancy (Ormesher L et al, Nitric Oxide Biol Chem, 2018), and in non-pregnant CKD cohorts (Kemmner S et al, Nitric Oxide Biol Chem, 2017| Webb AJ et al, Hypertens Dallas Tex, 2008). The effect of beetroot juice on blood pressure itself has been reported in several, albeit all small, RCTs (Gronroos et al, Nutr Metab CV, 2024), though whether this effect is mediated through nitrates or not may be debatable (Sagar et al, KI Reports 2024). 

This was a feasibility study, not an efficacy trial- designed to test whether recruitment was achievable and to gather preliminary safety/signal data for a future powered trial.

How was the trial done? 

Design

Trial-within-a-cohort (“modified Zelen”) design (Figure 1): participants first consented to the ORCHARD observational cohort and then were randomized 1:1 to dietary nitrate or standard care. Only those allocated to the nitrate were approached for a second consent to actually receive it; standard care participants were enrolled automatically. This structure, originally proposed by Zelen in 1979, reduces recruitment bias, but it weakens the crossover: even modest rates bias estimates and cut power (Homer CSE, J ADV Nurs, 2002).

Setting and population: 8 UK tertiary centers, February 2020-April 2024, with a 6-month pause for COVID-19. Eligible patients included singleton pregnancies, KDIGO-defined G2-G5 CKD, <25 weeks’ gestation, pre-pregnancy eGFR >90 ml/min/1.73 m² or pregnancy creatinine >70 μmol/l (0.79 mg/dl) without evidence of acute kidney injury.
Intervention vs comparator: Daily 70 mL “Beet-It-Sport” shot (approx. 400 mg nitrate) from consent or 12 weeks’ gestation (whichever was later) until delivery, versus standard care. The trial was open-label, without a placebo due to expense.
Follow-up and outcomes: Up to 5 visits through 6 months postpartum, assessing kidney function, proteinuria, blood pressure, and maternal/neonatal adverse events. The primary outcome was recruitment rate per site per month. 
Intention-to-treat: All standard care participants were analyzed. Only nitrate-arm participants who gave the second consent were analyzed. This departs from the classical Zelen analysis (which retains everyone originally randomized); a deviation the authors acknowledge as a limitation.
Of 119 eligible women, 108 were randomized, yet only 30/54 allocated to nitrate consented versus 53/54 retained in standard care (Figure 2)- crossover made concrete, and a reminder that the analyzed nitrate arm is self-selected, not the full allocated group.

Results

The primary outcome- recruitment feasibility- was achieved, with a mean recruitment rate of 1.01 participants per site per month (SD 0.90)-useful for planning, likely an underestimate in higher-burden settings.

The nitrate arm had numerically less chronic hypertension and markedly fewer Black participants to standard care, both established risk factors for CKD progression; despite comparable pre-pregnancy eGFR between arms, these imbalances were not statistically adjusted for in subsequent comparisons.

Kidney outcomes were directionally suggestive rather than statistically definitive, though biologically coherent. Among participants with pre-pregnancy eGFR below 45 ml/min/1.73 m², postpartum creatinine tended to be lower with dietary nitrate (176 vs 226 μmol/l at week 6, p= 0.23, 173 vs 216 at μmol/l 6 months, p=0.18), whereas in the milder CKD stratum (Figure 3c, eGFR > 45) the 2 arms’ trajectories were superimposable- consistent with the hypothesis that benefit concentrates where baseline NO deficiency is greatest.

Figure 3. (a) Median eGFR changes from prepregnancy to 6 months postpartum, by randomization eGFR group and treatment. (b) Median creatinine by intervention. (c) Median creatinine by intervention and randomization eGFR group, from Smith P et al, KI reports, 2026

There were no significant differences in delivery mode, gestational age, or birthweight centile but a numerically lower NICU/neonatal unit admission rate with nitrate use (23.3% vs 39.6%, p= 0.13) and a higher median birthweight. The authors note this magnitude of NICU admission reduction is comparable to that reported for antenatal corticosteroids in the updated Cochrane review of preterm birth (Roberts D et al, Cochrane Database Syst Rev, 2026), though this remains a descriptive comparison across studies rather than a direct statistical contrast.

Blood pressure approached but missed significance (Table 3): SBP 120.3 vs 123.2 mmHg (p= 0.064) and DBP 75.4 vs 77.7 mmHg (p= 0.051). A magnitude comparable to both the CHAP trial of mild chronic hypertension treatment in pregnancy (Tita AT et al, NEJM, 2022| NephJC summary) and non-pregnant dietary nitrate meta-analyses (Ashor AW et al, J Hypertens, 2017).

The single result reaching statistical significance appears in Table 4. Serious adverse events occurred in 23.3% of the nitrate arm vs 50.9% of standard care (OR 0.29, 95 CI 0.11-0.80), with no nitrate-attributable SAEs, no congenital abnormalities, and hyperkalemia above 6 mmol/L occurring in one participant per arm. A post hoc analysis found lower antihypertensive use with nitrate use (36.7% vs 66%, p= 0.0097). This analysis was explicitly designated post hoc by the authors and therefore is hypothesis-generating. 

Only 20.8% enjoyed the intervention-  taste, nausea, reflux, and beeturia (harmless urinary discoloration) were the main complaints- yet 62.5% would repeat it, and 70.8% would recommend it.

What are the implications?

The one significant finding- the SAE reduction- is also its most bias-prone result: open-label trials leave SAE ascertainment and decisions like antihypertensive initiation vulnerable to detection and performance bias without blinding. A nitrate-depleted placebo had worked before (Ormesher L et al, Nitric Oxide Biol Chem, 2018), so the comparator was feasible, just not funded here; its absence weakened every clinical outcome. Baseline imbalances compound this, as does the Zelen crossover itself: with nearly half the nitrate arm declining the second consent, the analyzed group is a self-selected, likely more engaged subgroup against a near-complete standard-care arm- exactly the bias methodological literature predicts, and one of the authors acknowledges. 

Generalizability is also limited: women from lower socioeconomic groups and Black ethnicity - disproportionately affected by CKD and its pregnancy complications- were underrepresented in the intervention arm, partly from COVID-related recruitment disruption at sites serving them, limiting applicability to the population with the greatest disease burden. The mechanistic picture is incomplete too: planned salivary/ tongue sampling to characterize oral nitrate-reducing capacity was abandoned for infection-control reasons- a real gap since the pathway depends on bacterial nitrate reduction, and interindividual variation in this capacity (Willmott T et al, Hypertens Dallas Tex, 2023) plausibly explains some of the blood pressure heterogeneity (Lundberg JO et al, Nat Rev Drug Discov, 2008).

Conclusion

After thirty years without a randomized trial in CKD pregnancy and a pandemic working against it throughout, ORCHARD-BEET deserves credit for being a well-conducted, transparently reported multicenter RCT in a largely abandoned field. It met its stated goal: feasibility, not efficacy, was the primary outcome, and 1.01 participants per site per month with 76.8% completion gives solid planning data for what comes next.
It was not, and was never meant to, prove efficacy. The kidney, blood pressure, and neonatal trends are encouraging and biologically coherent, especially in more advanced CKD, but stem from a modified Zelen analysis that lost nearly half its intervention arm to a second consent in an open-label trial with no placebo.
Put simply, this is a solid proof of concept. Women with CKD will enroll in interventional pregnancy trials; the intervention raised no safety concerns, and the field now has the numbers it needs to design a properly powered trial that should follow. Closing a thirty-year gap is, on its own, quite an accomplishment.

                                                                                         By Cristina Popa

Reviewed by

Brian Rifkin, Swapnil Hiremath

Nondialyzable versus Dialyzable Beta-Blockers in Hemodialysis: A Target Trial Emulation Study

Ali Etemadi, Sai Liu, Wolfgang C Winkelmayer, Maria E Montez-Rath, Tara I Chang

PMID: 42160184

DOI: 10.1681/ASN.0000001147

Why was the study done?

Beta blocker use, although well established in cardiovascular disease, has been embroiled in controversy when it comes to their protective effects in patients in CKD (Han Hyung Seung et al, NDT 2026;  Yeh Tzu-Hsuan et al, PLOS One 2022).  Cardioselectivity and dialyzability are two major factors influencing physician choice in patients on dialysis. Non-cardioselective beta blockers, like bisoprolol, have proven mortality benefit in heart failure (Hans-Dirk Düngen  et al, Euro J Heart Fail, 2011) with some evidence describing their superiority in CKD (Shireman TI et al, BMC CVD 2016; Yu Albert et al, JASN supplement 2022). However, they may predispose to hyperkalemia, especially in patients on mineralocorticoid receptor blockers (Wu PH et al, NDT  2020).  Alternatively, non-dialyzable beta blockers may provide consistent cardioprotective benefit,  they may also predispose to bradycardia and intradialytic hypotension (Yeh Tzu-Hsuan et al, PLOS One 2022).  

Unfortunately, to date, observational data comparing dialyzable vs non dialyzable beta blockers use in dialysis populations has given mixed signals of cardiovascular benefits

(Weir et al JASN 2014; Assimon et al, AJKD 2019; Wu et al, NDT 2020). Importantly, apart from dialyzability, beta blockers have other pharmacokinetic and pharmacodynamic properties such as cardioselectivity, vasodilatory capacity, lipophilia which may affect outcomes. Cardioselective betablockers have been demonstrated to be more beneficial in dialysis patients compared to non-selective beta blockers (Shireman at al, BMC CV Dis 2016; Meta-analysis-Tao et al, PloS 2022).

Due to the inconsistent data regarding the benefit of one class over another, prescription practices vary widely and are often influenced by availability, insurance coverage, familiarity, and peer usage. Considering the clinical equipoise, the high variability in practices, and the difficulty in conducting RCTs in hemodialysis population, this target trial emulation study was done to demystify this controversial area. 

How was the study done and what did it report? 

Methods

This was a target trial emulation study (Hernán  M et al, JAMA 2022) using retrospective data from USRDS, Medicare and Davita to predict the outcomes of a randomized controlled trial of dialyzable versus non dialyzable beta blockers in CKD patients initiating hemodialysis.  Retrospective data from USRDS was obtained for patients with Medicare A, B, and D from 2008-2021. Elderly patients ≥ 67 years initiating hemodialysis while on beta blockers, with at least claim for hypertension in the past, were included in the study from 41 US states .

Inclusion and exclusion criteria are detailed below:

Adapted from Supplemental table 1 from Etemadi et al, JASN 2026

Nomenclature /definitions:

• Dialyzable beta blockers: metoprolol, atenolol, bisoprolol, acebutalol

• Non dialyzable beta blockers: carvedilol, propranolol, betaxalol

• Classification of patients by type of beta blockers was based on whether they had only one active prescription supply for either a dialyzable or a nondialyzable beta blocker on index date or within a 30-day grace period prior to index date.

Outcomes:

All-cause mortality and major adverse cardiac events  (MACE-defined as stroke, myocardial infarction, or all-cause mortality) were studied as per intention to treat analyses. Negative control outcomes (unrelated outcomes like conjunctivitis, contact dermatitis, onychomycosis, contact dermatitis) were analyzed both in the instrumental variable approach and the Cox model. Data were obtained from ICD codes used in Medicare.

Statistical procedures:

Quasi-randomized treatment groups were made based on an instrumental variable (IV) approach using regional prescribing preferences. Regions with a preferential prescription were included. 

RESULTS:

A total of 161,798 patient records were screened for eligibility, of which 80,353 patients were eligible for inclusion in the emulation study (Figure 1 below). Since only patients residing in areas with a preferential prescription (ATR>0.8) were included in the instrumental variable (IV) approach, the final sample size was 40,313 (n= 19,833 in the non-dialyzable beta blocker group, n=20,480 in the dialyzable beta-blocker group). 

Baseline characteristics (Table 1 below) were balanced between the two groups -patients residing in areas with preferential dialyzable vs preferential non-dialyzable beta blocker use (SMD <0.1 for all characteristics). Although p value should not be calculated for comparing baseline characteristics in randomized trials, a SMD in target trial emulation helps to determine if the quasi-randomization has genuinely been able to create balanced groups.  When actual group of patients receiving dialyzable versus non-dialyzable beta blockers was compared (supplemental table 3), the group of patients receiving non-dialyzable beta blockers were found to have a higher prevalence of congestive heart failure, ischemic heart disease, chronic pulmonary disease, complicated hypertension and diabetes,  higher mean Elixhauser mortality score, more frequent cardiology visits, and higher baseline diuretic use (probably reflecting sicker baseline status with higher comorbidity burden).

MACE and its individual component outcomes were significantly lower in the arm with preferential non-dialyzable beta blocker use compared to dialyzable beta blocker use.

(Figure 3 below). There was no significant association of any of the negative outcomes (conjunctivitis, cerumen accumulation, onychomycosis) with beta blocker dialyzability.

In the fully adherent cohort , the outcomes showed similar trends, shown in Figure 3, except for the outcome of stroke which was not significantly associated with beta blocker dialyzability at 6 months, 1 year or 3 years. 

Overall there were 25,720 MACE over the 3-year follow-up period, with 10,933 occurring in the first 6 months. The table below shows the incidence of MACE and individual outcomes at  6 months, 1 year and 3 years. The incidence of each outcome and of the composite outcome was highest during the first 6 months after hemodialysis initiation. 

Using a primary Cox model, hazard ratios for MACE, acute MI and stroke at 6 months, 1 year and 3 years showed a significant advantage of using non-dialyzable beta blockers in HD. However, negative outcomes (conjunctivitis, cerumen accumulation, contact dermatitis , onychomycosis) were found to be associated with beta blocker dialyzability at different time points after HD initiation. (A smart way to demonstrate the advantage of using target trial emulation over the Cox model!)

DISCUSSION:

Beta blockers are a frequent antihypertensive choice in patients on hemodialysis, especially in those with cardiac disease. The data regarding the use of dialyzable versus non-dialyzable beta blockers in the dialysis population has been fraught with controversy.

This study analysed data from >40,000 patients who were initiated on dialysis to compare outcomes in regions where there was prescription preference for dialyzable vs non-dialyzable beta blockers. The results demonstrated a significant benefit of non-dialyzable beta blockers in preventing MACE and its individual components.This target trial emulation study tried to bring an end to the controversy, but can anything but an RCT be the final answer ?

 Unfortunately, large scale RCTs on beta blocker use in hemodialysis have failed to reach completion due to various reasons (see table below).The first RCT of beta blockers in dialysis patients was published as early as 2001 (Cice G et al, JACC 2001) - it demonstrated that carvedilol decreased all cause mortality in dialysis patients with dilated cardiomyopathy. Subsequent to this small study (n=117), there was one RCT of atenolol vs lisinopril in HD patients (Agarwal et al, NDT 2014; HDPAL study)  which was terminated early due to more serious adverse events including CV events, all cause mortality , hyperkalemia and hypertensive crisis in the lisinopril arm. A feasibility trial of beta blocker vs placebo use in dialysis patients

(Roberts et al, AJKD 2016; BLOCADE) was unable to recruit patients due to high baseline use of beta blockers, apprehension of nephrologists to switch patients to a non-beta blocker, inability to get informed consent and a high drop out rate (with intradialytic hypotension and bradycardia). 

The study by Etemadi et al (JASN 2026)  had the strengths of using an instrumental variable approach and a target trial emulation design with a large sample size. It was a very well designed statistical approach with clear demonstration of techniques used to check independence, relevance and exclusion restrictions. They also demonstrated comparison with a Cox model to highlight the difference in results and flaws of using a Cox regression analysis in situations where RCTs are ideal (and trial emulations a distant second choice). Analysis was done both on ITT and the adherent population to bring out the differences. 

Limitations include the inherent limitations of a retrospective cohort and the use of medicare data rather than actual outcomes. The generalizability may be limited to older patients (age >67 years) and those with medicare coverage (adherence could also be influenced by that). Incidence of bradyarrhythmias and intradialytic hypotension episodes were not available to draw any mechanistic conclusions. Lastly PD patients were not included , similar to all previous studies quoted above of beta blockers in dialysis patients. 

Conclusion: 

In this target emulation trial, MACE and its individual component outcomes were significantly lower in the arm with preferential non-dialyzable beta blocker use compared to dialyzable beta blocker use.

This study also highlights the need for adequately powered RCTs in HD patients to study the effect of beta blockers accounting for dialyzability, cardioselectivity and other physicochemical factors unique to different drugs in this class.

                                                                                         Blog by Pallavi Prasad

Associate Professor Nephrology, VMMC and Safdarjung Hospital 

New Delhi, India

Reviewed by

Brian Rifkin, Cristina Popa

JAMA. 2026 Jul 30. doi: 10.1001/jama.2026.9268. Online ahead of print.

Dapagliflozin and Acute Kidney Injury Following Cardiac Surgery: A Randomized Clinical Trial

Maartina J P Oosterom-Eijmael, Abraham H Hulst, Nelson P Monteiro de Oliveira, Ed D Niesten, Nicobert E Wietsma, Bastiaan M Gerrits, Thierry V Scohy, Thijs C D Rettig, Ferdinand T F Snellen, Magiel F Voogd, Marc B Godfried, Rients N de Boer, Jeroen Wink, Lisa M M van der Werff, Christa M Cobbaert, L Renee Ruhaak, Susanne Eberl, Benedikt Preckel, Markus W Hollmann, Jimmy Schenk, Jeroen Hermanides, Daniel H van Raalte; MERCURI-2 Study Group

PMID: 42530910

DOI: 10.1001/jama.2026.9268

Introduction

SGLT2 inhibitors (SGLT2i or Flozins) are awesome! So much so that we at NephJC call them “The Undefeated”. A miracle class of drugs that works in chronic kidney disease (CKD) with or without diabetes, heart failure with or without reduced ejection fraction, proteinuric nephropathies; however, even Superman had his limits. 

Acute kidney injury (AKI) is a significant health concern, especially among hospitalized cardiac surgery patients because it contributes to increased length of stay, morbidity, mortality, and costs (Fuhrman DY, Curr Opin Anaesthesiol 2017). The incidence of cardiac surgery-associated AKI occurs in up to 50% of cases, and 2%–5% of these patients require renal replacement therapy during an AKI episode (Fuhrman DY, Curr Opin Anaesthesiol 2017; Brown JK et al, J Cardiothorac Vasc Anesth 2023). Thus far, the prevention and management of AKI remains largely supportive - focusing on fluid balance, optimizing hemodynamics, and nephrotoxin avoidance. 

Flozins provide kidney-protective benefits theoretically by improving renal oxygenation, reducing oxidative stress, and decreasing tubular workload. In a recent meta-analysis, SGLT2i use was associated with a statistically significant decrease in both mortality (OR 0.73; CI 0.58-0.92) and acute kidney injury (OR 0.68; CI 0.61-0.76). However, like most meta-analyses, the outcomes and adverse effects data were muddied by considerable heterogeneity in data (Balbaa E et al, Endocrinol Diabetes Metab 2026). Hypothetically, SGLT2i should be beneficial in cardiac surgery patients, especially by lowering renal oxygen demands in the cardio-pulmonary bypass (CPB)  ischemia-reperfusion window. However, most current guidance recommends holding SGLT2i medications 3-4 days prior to any surgery because of the risk of euglycemic ketoacidosis, dehydration, and excess fluid loss. This study examines whether Flozins, given in the peri-operative period in cardiac surgery, can reduce the incidence of AKI.

Methods

Study design

MERCURI-2 was a multicenter, double-blind, placebo-controlled randomized clinical trial conducted in the Netherlands. Randomization (performed via Castor Electronic Data Capture) used computer-generated blocks of 4, 6, and 8, stratified by sex, type 2 diabetes, and study site. Allocation was visible only to the trial pharmacists, who prepared study medication but were otherwise uninvolved.

Study population

• Adults ≥18 years scheduled for any elective cardiac surgery (CABG or non-CABG) 

• Exclusion criteria included

• prior SGLT2 inhibitor or insulin use

• a history of diabetic ketoacidosis  

• baseline hypotension (defined by systolic blood pressure <100 mmHg) 

• eGFR <20 ml/min/1.73 m² 

• Pregnancy 

• Any contraindications based upon drug labeling

Intervention

Once randomized, patients received either placebo (a cellulose-containing excipient) or  AstraZeneca-dapagliflozin (encapsulated in-house by the academic trial pharmacy for blinding purposes - AstraZeneca supplied the source tables but was not a trial funder; more on that distinction below). 

The first dose was given the day before surgery, followed by daily doses until the second postoperative day (four doses total).

Outcomes
Primary Outcome

Incidence of AKI up to 7 days after cardiac surgery, defined by the 2012 KDIGO criteria:

• An increase in serum creatinine level of 0.3 mg/dL (26.5 μmol/L) or greater within 48 hours, or

• 1.5-fold or greater increase in creatinine level within 7 days of surgery, or

• Urine output less than 0.5 mL/kg/h for 6 to 12 hours after surgery.

They used the most recently obtained creatinine as the reference value, which mostly corresponded to the blood test taken the day before the surgery, though they don't actually report what percentage that applied to. Following the procedure, serum creatinine values were obtained daily during the first 7 days after the intervention. All patients had a urinary catheter in place for the entire 7-day period to measure urine output.

Secondary outcomes

• Stage 3 AKI as per KDIGO 2012 criteria 

• Postoperative maximum change in eGFR within the first 7 days

• De novo postoperative atrial fibrillation recorded on a 12-lead ECG; length of stay in the ICU  and hospital

• Incidence of major adverse kidney events (MAKE30): composite endpoint of death, new dialysis, or a persistent increase in serum creatinine concentration to 200% or more of the baseline, 30 days after the operation*

• Incidence of major adverse cardiac events (MACE30): composite endpoint of cardiovascular death, nonfatal myocardial infarction, nonfatal ischemic cerebrovascular accident, and hospitalization for heart failure, 30 days after the operation*

• World Health Organization Disability Assessment Schedule 2.0 (WHO-DAS 2.0) score within 30 days after the operation

• Days at home in the first 30 days (DAH30) after the operation

• EuroQol 5 dimensions 3 Levels (EQ-5D-3L) score measured at 30 days after surgery

Safety outcomes:  genital mycotic infections, ketoacidosis,  hypoglycemia

*NephJC writer’s note: The text of the article has the definitions for MACE30 and MAKE30 switched up. This text has been optimized for understanding purposes.

Funding

The study was supported by the Netherlands Organization for Health Research and Development (ZonMw), the European Society of Anaesthesiology and Intensive Care and BJA Charity, and the Netherlands Organization for Health Research and Development.

Dr. van Raalte (the last author) receives grants from AstraZeneca and is supported by a fellowship of the Dutch Diabetes Foundation and a Kolf fellowship from the Dutch Kidney Foundation.

Results

Out of 2551 considered, a total of 784 patients were randomized. Figure 1 lists 382 per arm as "included in primary analysis”- the full randomized cohort. But only 778 of 784 (99%) completed follow-up testing, and six patients never had surgery at all. 

Baseline characteristics were strikingly similar between groups; a median age of 68 years (IQR: 61–74) and a median BMI of 27 kg/m2. Male participants comprised over 75% of each arm, and white individuals accounted for more than 95% in both groups. Median eGFR was 80 mL/min/1.73 m² among both groups at baseline (Table 1). Surgical and clinical baseline characteristics as well as prior medication use were consistent across groups. Adherence to study protocol was high in both arms: 88% of the dapagliflozin group and 87% of the placebo group completed all four doses.

Table 1: Baseline and surgery characteristics from Oosterom-Eijmael MJP et al, JAMA 2026.

Primary Outcome

AKI occurred in 111 (28%) participants in the dapagliflozin group and in 205 (52%) participants in the placebo group (risk difference, −23.98% [−30.64% to −17.32%]; relative risk [RR], 0.54 [95% CI, 0.45-0.65]; P < 0.001).

Most primary-outcome events happened in the first 24-48 hours after surgery. That’s quick - closer to what you’d expect from transient postoperative oliguria (a normal, reversible dip in urine output from anesthesia, fluid shifts, or vasopressors) than from real structural kidney injury (see at-risk numbers under the curve), see discussion below.

Secondary Outcomes

Three (0.8%) in the dapagliflozin group and one (0.3%) in the placebo group had stage 3 AKI, however, this was not statistically significant (RR 3; 95% CI 0.31-28.71). Postoperative maximum change in eGFR compared with baseline was −7.4mL/min/1.73 m² in the dapagliflozin group and −1.75 mL/min/1.73 m² in the placebo group, likely reflecting the hemodynamic dip often seen with SGLT2i initiation. There were no between-group differences in MACE30 (5.6% vs 6.1%), MAKE30 (1.3% vs 0.5%), duration of hospital stay, or days alive and at home within 30 days after surgery (DAH30). 

Also unchanged was ECG-confirmed new-onset atrial fibrillation (32% vs 31%). Note that this is a separate variable from the “atrial fibrillation” adverse event in Table 3, which captures any AF, not just new-onset cases confirmed by ECG. Patient-reported quality of life (EQ-5D-3L) and functional status (WHO-DAS) were also unchanged. 

Exploratory Outcomes

In exploratory analyses, using a per-protocol method, the incidence of AKI was 27% in the dapagliflozin group and 51% in the placebo group, yielding a similar RR as the ITT analysis.

Post Hoc Outcomes

When AKI was defined exclusively by creatinine criteria, AKI occurred in 14% of the dapagliflozin group and 15% of the placebo group (RR, 0.93 CI, 0.66-1.32). However, when defined exclusively by urine output criteria, AKI incidence was 21% in the dapagliflozin group and 48% in the placebo group (RR, 0.44 0.35-0.54), indicating that the entire benefit lies in the AKI by urine output definition. 

Adverse Events

Recommended practice is to hold Flozins a few days before surgery to reduce the risk of ketoacidosis. However, only one episode of ketoacidosis was reported in the dapagliflozin group. Similarly, mycotic infections only occurred in one dapagliflozin treated patient. The most common post hoc adverse events were atrial fibrillation (45% vs 45%) and reoperation (11% vs 10%), which were nearly identical between groups.

Table 3. Postoperative adverse events, from Oosterom-Eijmael MJP et al, JAMA 2026.

Discussion

AKI after major surgeries, especially cardiac surgery, remains a significant issue contributing to high costs, prolonged hospital stays, and increased morbidity. Diuretics, CCBs, bicarbonate, and statins have all failed to create a significant difference in the incidence of AKI (Zanardo G et al, J Cardiothorac Vasc Anesth 1993; Tie HT et al, Crit Care 2014; Zhao BC, J Cardiothorac Vasc Anesth 2017; Bove T et al, PLoS One 2018; Takeichi T et al, J Extra Corpor Technol 2025). Here, we have flozins appearing to finally make a breakthrough where so many others have failed…or do we?

Mercury in Roman mythology was the messenger of news, but he was also considered a trickster god who used cunning and sleight of hand to confuse patrons. Upon closer examination, this trial seems more trick than treat. Unfortunately, the stunning headline of reduced AKI incidence does not hold up to closer inspection whatsoever. 

The creatinine-based signal is null, with the CI crossing 1 across all stages. The entire treatment effect lives in the urine-output criteria. The authors say this outright in the discussion: “the high placebo incidence is likely due to the urine-output criteria.”

That's the whole ballgame, right there! Dapagliflozin is a glucosuric osmotic diuretic. An agent that increases urine output will mechanistically make patients less likely to trip a low-urine-output threshold, independent of anything else happening in the nephron. The authors anticipated this and tried to defuse it by citing steady-state, fixed-sodium-intake, euvolemic, mildly hyperglycemic outpatient PK/PD studies (Scholtes RA et al, Diabetes Care 2021; Heise T et al, Clin Ther 2016) showing Flozins aren't meaningful diuretics beyond the first dose. 

But that literature isn’t appropriate here and highlights why indirectly comparing trial results is fraught with issues. Post-cardiac-surgery patients are not necessarily euvolemic, and usually hyperglycemic from surgical stress (catecholamines, cortisol), which means more filtered glucose, more glucosuria, more osmotic diuresis during the short 7 day window used to adjudicate the outcome. 

This is a testable hypothesis, and the authors had the data to test it. Perioperative glucose measurements and hemodynamics were prespecified secondary/exploratory outcomes but are explicitly flagged as "not analyzed at the time of this report." (BOOM goes the kryptonite) That's a significant thing to hold back given it bears directly on whether this is renoprotection or an osmotic-diuresis artifact. The same trick could be pulled off with hydrochlorothiazide dosing. 

The -7.4 ml/min/1.73 m² eGFR decline with dapagliflozin in MERCURI-2 in retrograde, in more ways than one - the filtration rate goes backward, and so does the “usual dip” explanation invoked to excuse it. CREDENCE’s (Perkovic V et al, NEJM 2019) eGFR dip averaged 3-4 ml/min/1.73 m² at 3 weeks (baseline 56); EMPA-REG’s (Zinman B et al, NEJM) was 2.7 ml/min/1.73 m² at 4 weeks (baseline 74)- Both drops were smaller and in stable outpatients over weeks. A “more normal” kidney startpoint doesn’t explain the size either: DECLARE-TIMI58 (Wiviott SD et al, NEJM 2018), the closest eGFR match (median 85), showed a smaller dip than EMPA-REG, not larger. The dip tracks baseline hyperfiltration, not baseline eGFR - a post hoc DINAMO analysis found a significant dip only in hyperfiltering patients (-11.7 ml/min/1.73 m²), not in normofiltering ones (Bjornstad P et al, CJASN 2025). MERCURI-e doesn’t report hyperfiltration status. The acute surgical/CPB context remains the likelier explanation. 

In POST-CABGDM (Pitta FG et al, Diabetes Care 2025), using a creatinine-only AKI definition, the administration of empagliflozin in diabetic on-pump CABG patients - a population MERCURI-e excluded - showed a significant reduction (22.5% vs 29.1%, RR 0.57; CI 0.34-0.96), the inverse of MERCURI-2’s null creatinine finding. But it was a small, single-center, and open-label trial; an unblinded design is a real vulnerability for an outcome as sensitive to perioperative fluid and vasopressor management as AKI. Furthermore, the drug was stopped 72 hours before surgery, so its link to the acute injury window is indirect at best. 

A 2026 meta-analysis of 13 chronic dosing RCTs (84,581 patients; Chiang YC et al, Kidney Med 2026) found a 20% AKI reduction using largely creatinine/adjudicated definitions that exclude transient fluctuations - a different philosophy than MERCURI’s urine-output inclusive composite. Together, they suggest a real, more modest creatinine-based signal may exist that MERCURI-2 itself didn’t reproduce.

MERCURI’s associated editorial (Winkelmayer & Chertow, JAMA 2026) counters: in a 105-patient Amsterdam subset, patients on dapagliflozin got less IV fluid and norepinephrine, suggesting true kidney function rather than forced diuresis (not verifiable from the JAMA manuscript). It also cites lower KIM-1/TIMP-2xIGFBP7 from the earlier open-label pilot, a smaller, unblinded, non-MERCURI-2 cohort.

Limitations

OK take a deep breath…

The authors themselves describe up to 11 limitations in their study, some of them potentially questioning the entirety of their results. Let us start by stating that evaluating results just 7 days after starting a drug can potentially prove your intervention to be null across most meaningful endpoints. Using criteria with questionable clinical weight, such as transient oliguria without serum creatinine changes, could be disregarded as a mere hemodynamic change rather than a true kidney injury.

There was no mention of any other AKI prevention protocols implemented. Without data on pre-op discontinued medications (specifically RASi), fluid therapy, hemodynamic shifts, or even how old baseline creatinine values actually were, it's really hard to take those AKI numbers at face value. On top of that, there were no urine albumin measurements or differential diagnoses for the AKI cases they found. The use of  kidney damage biomarkers and/or MAKE30 in the primary outcome may have given us a better idea about the role of Flozins in preventing AKI. Similarly, the study does not state the need for post-op vasoactive drugs, fluid therapy, or initiation of additional drugs besides dapagliflozin.

The amount of secondary outcomes is difficult to square, and at the same time, there were endpoints that were registered, but ultimately not analyzed, including important end points like health care and productivity costs, hemodynamic vital signs and cardiac function.Why include them as endpoints if you’re not going to use the data in your published report? Given the results for the kidney outcomes, we could perhaps assume they just didn’t include those other endpoints because those results were also not statistically significant.

This study population was extremely homogeneous, with 97% white individuals, >75% males. While this may be explained away by geographical location of the study and higher prevalence / earlier age onset of cardiac disease in men vs women (though some studies suggest men are more aggressively treated for their cardiac disease than women), this heterogeneity does limit generalization despite the murky results. Furthermore, the medication provided was an SGLT2i drug from AstraZeneca, which is mentioned as one of the grant supporters for one of the main authors. Sponsorship bias alert. 

The authors go on to frame limitations as potential benefits? “Specific treatment protocols for AKI prevention were not implemented. The absence of study-specific protocols enhances the generalizability of the findings to routine clinical practice.” That second sentence should really give us food for thought. In fact, a better designed trial could have had multiple arms: one could imagine a trial with AKI prevention bundle +/- dapagliflozin versus placebo could perhaps produce more robust findings and elucidate if flozins are providing anything more than diuresis. 

To add fuel to the fire, the paper is littered with multiple spelling and grammatical errors.

Conclusion

In adults undergoing elective cardiac surgery, perioperative dapagliflozin did lower the incidence of postoperative AKI strictly according to the KDIGO AKI 2012 criteria at Day 7. However, the benefits were driven by urine output criteria, whereas serum creatinine remained completely unaltered. Although the study shows a large effect size, the inappropriate choice of the primary outcome measure marred the interpretation of this ambitious study. 

All in all, we have seen our superhero SGLT2i sullied by this poorly designed study. Fear not citizens of Nephropolis, SuperFlozinators will live to fight another day (in a much more logical and worthwhile study like The Renal Lifecycle Trial due to be completed in 2027). Until then, don’t hesitate…flozinate!

Summary by

David White FNP-C

Raquel Barba Teba Nephrology specialist, Spain

Reviewed by

Akshaya Jayachandran, Brian Rifkin, Cristina Popa,
Jade Teakell, Pallavi Prasad

Header Image created by AI, based on prompts by

Brian Rifkin

Permalink

]]>Too sweet to be true? Prophylactic SGLT2 inhibitors before cardiac surgeryVisual AbstractMilagros FloresMon, 17 Aug 2026 11:38:33 +0000http://www.nephjc.com/news/2026/8/17/mercuri-the-visual-abstract535bcb2fe4b05fe61b320c51:535bd92ae4b0a78001c0e260:6a828857daf64c7fc064d737Can perioperative dapagliflozin prevent AKI after cardiac surgery?

The MERCURI-2 trial offers an intriguing answer and raises an important question about what is driving the AKI benefit.

Check out the VA for next #NephJC by Andreea Manolea

El ensayo MERCURI-2 ofrece una respuesta interesante y plantea una pregunta clave: ¿qué está impulsando la reducción de la LRA?

Revisa el resumen visual para el siguiente #NephJC por Andreea Manolea

Iptacopan in IgA Nephropathy - Final 24-Month Data

Jonathan Barratt, Necmi Ere, Naoki Kashihara, Bart Maes, Dana V Rizk, Brad Rovin, Hernán Trimarchi, Hong Zhang, Weiming Wang, Ismail Kocyigit, Chuanming Hao, Vladimir Tesař, Kenan Turgutalp, Li Yang, Guangqun Xing , Valter Duro Garcia, Seung Hyeok Han, Wanhong Lu, Antonio Pisani, Julia Weinmann-Menke, Frank Eitner, Nicolas Guerard, Dmytro Butylin, Luca Monaco, Emil Scosyrev, Annabel Magirr, Ronny Renfurm, Thomas Hach, Vlado Perkovic; APPLAUSE-IgAN Study Group

Collaborators, Affiliations Expand

PMID: 41910396

DOI: 10.1056/NEJMoa2600743

Why was the study done?

And the hits keep on coming! Ever since the elucidation of the 4-hit model in IgA nephropathy (IgAN) there has been the introduction of a plethora (a plethora? Yes, a plethora!) of new therapeutics based upon the various pathways of destruction. NephJC reviewed the KDIGO IgAN guidelines in 2025, which didn’t even incorporate iptacopan use, as the preliminary APPLAUSE results had been contemporaneously released (Perkovic et al. New Engl J Med, 2025). The excellent work by the PARASOL project, and the use of proteinuria as a reliable clinical trial endpoint for rare kidney disease, has moved medications from the lab to human studies to market at a breakneck speed. However, we are often left with questions about how effective these new treatments will be in the real world. We all hold our collective breath and await the GFR data to hopefully prove what we “already know”. Iptacopan, a complement factor B binder (alternative pathway) that inhibits C3 convertase, has shown efficacy in Paroxysmal Nocturnal Hemoglobinuria (PNH), C3 nephropathy (C3GN), and IgAN. The 9-month APPLAUSE interim analysis of this phase 3 trial of iptacopan showed a significant reduction (38%) in the 24-hour urinary protein-to-creatinine ratio as compared with placebo. We now have the 24-month data of eGFR in IgAN…and hopefully no surprises.

How was the study done?

This was a phase 3, international, double blind, randomized, placebo-controlled trial. The trial included a screening period, a run-in period of up to 90 days, and a 24-month treatment period. At the start of the treatment period, eligible patients were randomly assigned, in a 1:1 ratio, to receive oral iptacopan at a dose of 200 mg twice daily or matching placebo, along with supportive care. After the treatment period, eligible patients could receive open label iptacopan in a rollover extension study.

Eligible patients were 18 years of age or older and had biopsy-confirmed IgA nephropathy, an eGFR of at least 30 ml per minute per 1.73 m², and a 24-hour urinary protein-to-creatinine ratio of 1 or higher, despite RASi use for at least 90 days before the receipt of iptacopan or placebo.

The primary end point for the final analysis was the annualized total eGFR slope estimated over a period of 24 months. Secondary end points a composite kidney-failure (defined as a sustained decline in the eGFR of ≥30% from baseline for ≥4 weeks, a sustained eGFR of <15 ml per minute per 1.73 m2 for ≥4 weeks, initiation of maintenance dialysis, receipt of a kidney transplant, or death from kidney failure), which was assessed in a time-to-event analysis. Safety end points, including adverse events, laboratory variables, and vital signs, were also assessed. 

What do the results say?

The trial population was representative of the wider population of patients with IgA nephropathy. Overall, 19% of the patients were taking an SGLT2 inhibitor at baseline. During the 24-month treatment period, alternative or rescue medication was taken because of disease progression by 4.6% of the patients in the iptacopan group and by 13.4% of those in the placebo group. Discontinuation was asymmetric (18.9% iptacopan vs 36.8%), mainly driven by the prespecified ≥ 30% eGFR decline stopping rule.

Iptacopan resulted in a significantly slower eGFR decline than placebo (whew, no surprises). The between-group difference in the eGFR slope was 3.02 mL per minute per 1.73 m² per year (95% CI, 2.02 to 4.01) in favor of iptacopan. Iptacopan resulted in a smaller decrease from baseline to month 24 in the eGFR than placebo. The effect of iptacopan on the eGFR decline appeared to be consistent across prespecified subgroups defined according to age, sex, geographic region, 24-hour urinary protein-to-creatinine ratio category, eGFR category, the presence or absence of hematuria at baseline, SGLT2 inhibitor use at baseline, and the MEST-C score at biopsy. The SGLT2i at baseline subgroup showed a numerically larger benefit (diff 2.97 vs 2.71 without), hypothesis-generating for combination synergy but not powered for interaction testing.

The iptacopan group had a significantly higher number of patients with a 24-hour urinary protein-to-creatinine ratio of less than 1 at month 9 (difference of 26.4 percentage points [95% CI, 18.7 to 34.0]). The risk of a composite kidney-failure end-point event was also significantly lower in the iptacopan group than in the placebo group (hazard ratio, 0.57; 95% CI, 0.40 to 0.81). FACIT-Fatigue change was not significantly different between arms- likely because baseline scores in both groups already approximated general population norms, leaving little room to detect improvement. 

Overall adverse event incidence was similar between groups (87% with iptacopan vs 89.1% with placebo). As was the serious AE incidence (12.2% vs 11.7%), with no deaths in either arm. However, serious infections were more common with iptacopan (6.7 vs 2.1%), predominantly pneumonia; 2 iptacopan-treated patients had confirmed S. pneumoniae infections despite prior vaccination (both resolved with antibiotics; both discontinued per protocol). No meningococcal or H. influenzae infections occurred. Serious adverse events pertaining to kidney and urinary disorders occurred in 5 patients (2.1%) receiving iptacopan and in 4 patients (1.7%) receiving placebo, and neoplasms were reported in 4 (1.7%) and 3 (1.3%), respectively. 

Critical Review

So, we expected that the preliminary 9-month results of iptacopan in IgAN would show decreased proteinuria and a favorable safety profile. Now we have the eGFR data at 24-months that we also anticipated, with an improvement in eGFR slope of 3 mL per minute per 1.73 m² per year versus placebo. We can be at least a little cynical that the comparison group was placebo and additionally note that SGLT2 inhibitor use was still fairly uncommon during the study period (approximately 20%). Also, the standards for IgAN treatment are increasing, and the goal is slowing eGFR loss to background aging (0.5-1 ml per minute per 1.73 m² per year). So, we have another useful tool for IgAN, though one that still leaves a residual chronic eGFR loss of 3 mL/min/year and carries a real (if manageable) infection cost. 

The bigger issue for most nephrologists was not going to be if iptacopan decreases proteinuria and slows eGFR decline, but rather where does iptacopan fit in the grand scheme of IgAN treatments. It will definitely not be a “stand-alone” medication, but will it be enough when used with maximal non-immunosuppressive medications (i.e. RASi, DERA, SGLT2i)? Now that we have definitive eGFR data (and soon to be released sibeprenlimab eGFR data that was previewed at GlomCon 2026), there will be a push for studies with data with combined therapeutics (early versus late versus simultaneous initiation), just like GDMT in diabetic nephropathy. It is a new age, and I for one am happy to see KDIGO IgAN guidelines already out of date and badly in need of revising, less than 12 months after their release. It is a wonderful time to be a nephrologist with interest in rare glomerular diseases, because innovation and hard data have been a long time coming. A round of APPLAUSE for iptacopan the study group, and its proven benefits to a wide assortment of patients with varying severity of IgAN.

Summary by

Brian Rifkin

Reviewed by

Cristina Popa

The creation of our core content takes dozens of hours a week, with input from many team members. In a time when all the world’s knowledge is at one's fingertips, a simple web search by a conference presenter of their subject often yields NephJC content toward the top. We are prolific and have presented NephJC material every two weeks for more than 12 years (that’s at least 280 blogs, plus shorts, visual abstracts, editorials, podcasts, and social media threads). In fact, AI often uses our blogs as a reliable source of information for inquiries related to nephrology.

We at NephJC have one simple request of conference presenters and conference organizers. We strongly believe our efforts deserve credit. Honor our commitment to providing outstanding, critical, peer-reviewed material. Feel free to use our visual abstracts during conference presentations, but do not remove the NephJC logo or the creators’ names. We understand that sometimes images are cropped for the purpose of creating space, but failing to acknowledge the time and efforts of NephJC content creators feels insulting. By removing attributions, you are limiting the reach of our content that curious conference attendees might choose to engage with. We know coordinating national and international conferences is a large administrative burden, but we hope you will earnestly try to be more aware and considerate of attributing ownership of content. 

Thanks for your consideration,

The NephJC Leadership

#NephJC Ten POsts Chat

Tuesday, August 4th 2026, 9 pm Eastern

N Engl J Med. 2026 Jun 12. doi: 10.1056/NEJMoa2600526. Online ahead of print.

Sodium Bicarbonate for Critically Ill Adults with Metabolic Acidosis and Shock

SODa-BIC Investigators and the Australian and New Zealand Intensive Care Society Clinical Trials Group; Ary Serpa Neto, Mairead McNamara, Kyle White, D Jamie Cooper, Tomoko Fujii, Alisa M Higgins, Carol L Hodgson, Leanlove Navarra, Alistair Nichol, Sandra Peake, Alvaro Réa-Neto, Paul Secombe, Fiona Osborne, Hugh Burrill, Emily See, Meredith Young, Fernando G Zampieri, Mark Plummer, Connor Christie, Leo Nunnink, Josephine Mackay, Shay McGuinness, Joradee Marmol, Ravindranath Tiruvoipati, Michael Bailey, Paul J Young, Rinaldo Bellomo, Andrew Udy

PMID: 42283370

DOI: 10.1056/NEJMoa2600526

Introduction

Metabolic acidosis, a common manifestation in critically ill patients with shock, is independently associated with increased mortality, poor myocardial contractility, and impaired catecholamine response. Although identifying and correcting the underlying cause is the primary goal, there is still debate about whether correcting metabolic acidosis leads to improved clinical outcomes.

Intravenous sodium bicarbonate has been widely used for the correction of acidosis. Early trials (Cooper et al, Ann Int Med 1990; Mathieu et al, Crit Care Med 1991) corrected the acidosis without moving cardiac output to survival. Consequently, international guidelines recommend against sodium bicarbonate for hypoperfusion-induced lactic acidemia with pH ≥ 7.15, while conceding the evidence is weak (Surviving Sepsis Guidelines, Rhodes et al, Crit Care Med, 2016).
BICAR-ICU trial (Jaber et al, Lancet, 2018| NephJC summary)- a multicentric open-label randomized controlled trial- reopened the question: no day-28 mortality benefit overall, but lower mortality and RRT need in the AKI subgroup. To note the open-label design: renal replacement therapy (RRT) initiation is a clinician's judgment call, and an unblinded clinician who knows the patient is on bicarbonate has every incentive to defer dialysis a little longer. A subgroup RRT signal generated this way can be considered at least as consistent with detection bias as with a true renal effect- the trial cannot completely distinguish the two. A subsequent meta-analysis nonetheless reinforced a mortality benefit in AKI and shock (Ghauri et al, Cureus 2019), prompting the 2021 Surviving Sepsis Guidelines to endorse bicarbonate in septic shock with AKI (Evans et al, Crit Care Med, 2021)- guidance built on the same unblinded foundation- and setting up BICAR-ICU2. BICAR-ICU2 (Jung et al, JAMA 2025| NephJC Short) restricted enrollment to severe acidosis plus severe AKI, remained open-label, and still found no 90-day mortality benefit; its RRT reduction is open to the same criticism. A target-trial emulation in 1764 patients reported an association between bicarbonate and reduced mortality- observational, and therefore unable to settle what 2 unblinded RCTs already could not (Blank et al. Intensive Care Med, 2025). An individual-patient-data meta-analysis of BICAR-ICU trials concluded bicarbonate functions, at best, as an RRT-sparing strategy, with a mortality signal confined to pH ≤7.10—a conclusion built entirely on open-label data and inheriting all of its confounding (Fosset et al, Crit Care, 2026).
Taken together, this evidence points to a narrow niche- severe acidosis with concurrent AKI, RRT-sparing rather than survival-improving—and leaves patient selection, timing, target pH, and etiology dependence (lactic, septic, renal) unresolved. A more recent meta-analysis found no overall mortality benefit, but a 31% reduction in dialysis need in patients with acidosis and AKI (Chen et al, Crit Care Med, 2026).
Against this backdrop, SODa-BIC’s rationale rests on three unaddressed gaps. First, every positive trial to date was open-label, and any renal signal from an open-label trials is inseparable from performance bias in the RRT-initiation decision. Second, prior trials intervened only after severe acidemia was already established, leaving early correction untested. Third, only BICAR-ICU2 used a kidney-specific endpoint; the rest measured mortality or global organ dysfunction, diluting any renal-specific signal early correction might produce. 
After assessing feasibility in a smaller cohort (Serpa Neto et al, Crit Care Med, 2023), the SODA-BIC trial was designed as a double-blind, placebo-controlled trial to evaluate the effect of sodium bicarbonate treatment on MAKE30 (Major Adverse Kidney Events within 30 days), a composite kidney-specific outcome in critically ill patients with shock and moderate metabolic acidosis (pH <7.30) in. The central question SODA-BIC intended to answer was whether early correction of metabolic acidosis improves kidney-specific patient outcomes.

The Study

Methods

Population

All patients (>18 years) admitted to the ICU with metabolic acidosis and receiving vasopressors were screened for eligibility. Metabolic acidosis was defined as a pH < 7.30, a base excess ≤ -4 mEq/L, and a PaCO₂ ≤ 45 mmHg (non-intubated) or ≤ 50 mmHg (intubated) within the 2 hours prior to randomization. These 3 criteria together ensure the acidosis is truly metabolic, excluding primary respiratory acidosis, and is clinically significant (pH). The higher PaCO₂ threshold for intubated patients accounts for permissive hypercapnia from lung-protective ventilation. All 3 criteria (pH, PaCO₂, and base excess) needed to be met within the 2 hours preceding randomization. Patients with shock and metabolic acidosis for >48 hours, suspected secondary cause for metabolic acidosis (GI, urinary losses or DKA), eGFR <30, dyselectrolytemia (dysnatremia, serum K <2.5 mEq/L, iCa <0.8 mmol/L), or currently receiving RRT/ planned to start within 3 hours were excluded. Patients deemed at inevitable high risk of cerebral edema and clinicians’ decision for enrollment (pragmatic approach) were also excluded.

Study design

This pragmatic, adaptive, placebo-controlled, double-blind, phase III randomized clinical trial (NCT05697770) was conducted in 55 centers across seven countries between April 2023 to December 2025. Randomization was performed by variable block method and stratified for centres, pH (<7.25 or ≥7.25), and creatinine level (<150 or ≥150 𝞵mol/L). At each site, the investigator who randomized the patient and prepared the study drug was distinct from the (blinded) investigator collecting outcome data, to preserve the blinding.

The adaptive design had 2 components: 
1. A blinded sample-size re-estimation after the first 100 patients had outcome data, performed by an independent statistician unaware of treatment allocation, using the observed pooled event rate with the originally assumed relative effect held constant- bounded such that the sample size would be capped at 700 if recalculation exceeded this, and would not fall below the original 500 even if the observed event rate were higher than assumed 
2. A single formal Data and Safety Monitoring Board (DSMB) interim safety review at 125 patients (25% of target, with 90-day follow-up complete), conducted June 2025, after which the DSMB recommended continuation without protocol changes. Informed consent (or waiver, where locally permitted) was obtained per jurisdiction; if a patient died before consent, data were included where local regulation and ethics approval allowed.

Intervention

Trial participants were randomized to sodium bicarbonate or placebo groups and received infusion as per the algorithm given in Fig 1. The intervention group received either sodium bicarbonate infusion (300 ml sodium bicarbonate (8.4%) added to 500 ml 5% dextrose after removing 300 ml dextrose; final concentration 600 mEq/L) or placebo infusion (5% dextrose). The two solutions were macroscopically indistinguishable, as validated in a prior pilot trial (Serpa Neto et al, Crit Care Med, 2023). The study drug was permanently withheld if a contraindication developed (hypernatremia, hypocalcemia, refractory severe hypokalemia, clinically significant fluid overload, or per treating clinician judgment). Open-label bicarbonate after the infusion window was discouraged but permitted at clinician discretion, guided by non-mandatory suggested triggers (pH < 7.10 with PaCO₂ < 40; HCO₃⁻ < 8 mmol/L, BE < -15; or K > 7 with a pH <7.10)

Outcomes

Baseline serum creatinine was determined hierarchically using the following order: the most recent outpatient value obtained 7 days to 3 months before ICU admission; an in-hospital value measured within 7 days before admission; or, if no measured value was available, an estimated creatinine calculated using the formula: 0.74 − 0.20 (if female) + 0.08 (if Black) + 0.003 × age.
The primary outcome was evaluated across eight prespecified subgroups: baseline pH (<7.25 vs ≥7.25), AKI stage (KDIGO stage 2–3 vs none/stage 1), PaCO₂ (≤45 vs 46–50 mmHg), acidosis type (high anion gap >12 vs non-anion gap ≤12 mmol/L), illness severity (APACHE II score above vs at or below the median), pre-randomization bicarbonate use, presence of septic shock, and admission type (medical vs surgical).

Sample size estimation

Based on the data from the BICAR-ICU trial (Fujii T et al, Crit Care 2021 and Mochizuki et al, Crit Care and Resusc 2023), assuming 40% incidence of MAKE30 in the control group with 90% power, 5% alpha-error and attrition considered, 500 patients were required to detect an absolute difference of 14%. Sample size was re-estimated using the event rate observed after completing the study of 100 patients and capped between 500-700 patients. 

Statistical analysis

A modified intention-to-treat analysis was performed with a two-sided superiority hypothesis (α=0.05). A mixed-effects generalised linear regression with a binomial distribution and an identity link was used to estimate the absolute risk difference, with clustering and stratification variables as random effects. Kaplan-Meier curves were used to display survival at days 30 and 90, with adjusted hazard ratios and 95% CIs estimated using shared-frailty Cox proportional hazards models (site as the frailty term). Heterogeneity of treatment effects across the 8 prespecified subgroups was tested via a treatment-by-subgroup interaction term. Sensitivity analysis was also performed, accounting for age, sex, pre-randomization open-label bicarbonate use, APACHE II and SOFA score. No correction for multiplicity was applied to secondary or exploratory outcomes; these are reported as point estimates with 95% CIs that should not be used to infer definitive between-group differences.

Funding 

The trial was funded by the National Health and Medical Research Council of Australia and coordinated by the Australian and New Zealand Intensive Care Research Centre at Monash University. The funding body had no role in design, trial conduct, analysis, or reporting of the data.

Results

Baseline demographics

In 33 months (April 2023-Dec 2025), 500 of 3060 patients were enrolled, and 498 were included in the analysis (Fig 1). Patients were recruited from 7 countries, predominantly from Australia and New Zealand. Enrolled patients were elderly, with a median age of 66 years. The majority were admitted to the ICU from the operating theatre and the emergency department, accounting for  76% of enrolled patients (Table S3). The median APACHE II and SOFA scores were 21 and 7, respectively, suggesting an association with higher mortality risk.

Approximately 50% patients were admitted with a medical diagnosis, and septic shock was the most common cause for metabolic acidosis (48%). Prevalence of severe AKI was slightly higher in the sodium bicarbonate group (46.9% vs 40%) than in the placebo group (Table S3). The median pH at enrollment was 7.26 with a median base excess of -9 mmol/L; approximately 40% had pH <7.25.

Notably, trial participants were substantially sicker than the general ICU population at the same centres over the same period (APACHE III 77.2 vs 55.4, hospital mortality 23.1% vs 10.4%, Table S2).

Intervention and outcomes

498 patients received the intervention, of whom 244 received sodium bicarbonate infusion (median dose 300 mmol or mEq). Infusion-rate adjustment per protocol was required in half of the patients in the bicarbonate group (49.8%) versus only 23.2% of the placebo group—expected since the algorithm responds to biochemical correction, which happened more often and faster with the active drug. Open-label sodium bicarbonate was given more often in the placebo arm (15.2% vs 7% in Fig S8), consistent with a lower need for rescue therapy once the study drug had already corrected the pH.  Protocol deviations (Table S5) were also more frequent in the bicarbonate arm, driven predominantly by infusion rate non-adjustment (13.5 vs 2.8% in the placebo group)- expected, since the titration algorithm fires more often with faster correction, not a conduct failure. Randomization of ineligible patients was balanced between arms (5.4% overall), and shouldn’t be counted as part of this pattern.  

Protocol deviations (Table S5) were more common in the bicarbonate group, driven largely by “infusion rate not adjusted per protocol” (13.5% vs 2.8%); randomization of ineligible patients, by contrast, was similar between arms (4.9% vs 5.9%; combined 5.4% overall) and was not a bicarbonate-arm-specific problem.

Primary endpoint 

At 30 days there was no significant difference in the incidence of MAKE30 between the two groups (40.2% vs 39.4%; P=0.78). 

There was no significant difference in persistence of kidney dysfunction in the prespecified subgroup analysis (Fig S7).

Sensitivity analyses were concordant: further adjustment for age, sex, pre-randomization bicarbonate use, APACHE II and SOFA gave an adjusted difference of 0.7 pp (-7.3 to 8.6), and an alternative eGFR-based definition of persistent renal dysfunction gave 1.3 pp (-7.2 to 9.8) (Table S9). A confidence distribution analysis of the primary result (Fig S4) estimated only 38.9% confidence that bicarbonate reduces MAKE 30 to any degree- i.e., the point estimate leans, non-significantly, toward no benefit or harm rather than benefit.

Secondary and additional endpoints

One fourth of the enrolled patients died within 30 days in both groups (25.4% bicarbonate vs 24% placebo; HR 1.09, 0.76-1.56- Fig 2A), same as death by day 90 (fig S9). RRT use within 30 days was numerically lower with bicarbonate, but not significantly so (16.8% vs 20.9%- Fig 2B), as was persistent renal dysfunction (14% vs 18.3%- subgroup analysis Fig S7). RRT dependence and ICU death by day 30 were also similar (Table 3).

AKI incidence at 7 days was similar between groups (about 64%), but recurrence of metabolic acidosis was lower with bicarbonate (32 vs 55.7%). Though reaching the pH target was earlier in the bicarbonate group, patients in both groups had similar pH after day 7 of randomization (Fig S2).

The number of vasopressor-free days, renal-replacement therapy–free days, and ICU-free days on day 30, and the number of hospital-free days on day 90, were also similar between the two trial groups (Table 3, distributions in fig S10).
Adverse effects were uncommon: 1.6% with bicarbonate vs none with placebo; driven by hypokalemia needing correction, with one patient also developing hypernatremia. No serious adverse effects occurred in either group.

Discussion

The SODa-BIC trial was a pragmatic trial that randomized 500 patients from 55 ICUs to receive sodium bicarbonate or dextrose infusions. They included critically ill patients dependent on vasopressors with a pH < 7.30, BE ≤ -4 mEq/L, and PaCO2 ≤ 45 mmHg (≤ 50 mmHg if intubated), representing a very common phenotype seen in the ICU. In this trial, sodium bicarbonate (an infusion of 8.4% sodium bicarbonate solution, 1000 mEq/L) did not reduce major adverse kidney events within 30 days (MAKE30) of randomization compared with an infusion of 5% dextrose.
MAKE30 is a well-validated, patient-centered composite: it captures both injury severity and survival. Here, the composite was driven almost entirely by death of any cause (25.4% vs 24%), while every renal component moved in a numerically favorable direction without reaching significance:  renal replacement therapy (16.8% versus 20.9% in placebo), persistent renal dysfunction (14% versus 18.3%), and dialysis dependence at day 30 (5.7% versus 9.1%). Three consistent point estimates favouring bicarbonate and none of them being significant is evidence of an underpowered trial, a genuinely absent effect, or both, and this trial cannot distinguish between these issues.
A  comparator is the PRESERVE trial - IV bicarbonate versus saline for the prevention of contrast-associated AKI, using a composite of death, dialysis, or persistent kidney impairment, and was stopped early for futility (Weisbord et al, NEJM, 2018). Two bicarbonate trials, two patient-centered composites, and two null results. That’s not a coincidence worth explaining away with endpoint choice. Perhaps the question is not whether MAKE30 was the right endpoint—it almost certainly was. Rather, isn’t it naive to think that correcting metabolic acidosis would somehow reverse such a complex multisystemic pathophysiology? 
Interestingly, despite protocolized treatment, clinicians retained the ability to administer open-label sodium bicarbonate whenever deemed clinically necessary. Although crossover was relatively limited, 14.3% of patients in the bicarbonate arm and 9.1% in the placebo arm had already received bicarbonate before randomization, while additional rescue bicarbonate administration occurred after randomization in both groups. This pragmatic design reflects real-world practice, but may also have attenuated any treatment separation. For what it is worth, the prespecified subgroup analyses were similarly unremarkable, with no evidence of heterogeneity across the evaluated subgroups. Neither the presence of septic shock, illness severity, baseline bicarbonate administration, nor admission type appeared to modify the treatment effect. 
BICAR-ICU2’s finding- lower RRT with bicarbonate- deserves more scrutiny than it usually gets (Jung et al, JAMA 2025| NephJC short). The primary endpoint, 90-day mortality, was neutral; RRT initiation, unlike death, is not a hard endpoint; it is a clinician’s judgment call, made in real time, by an unblinded physician. Faced with worsening acidosis in a patient already on bicarbonate, the path of least resistance is more bicarbonate—cheaper, non-invasive, and reassuringly proactive-feeling—rather than escalating to dialysis. This is evidence that open-label trials let treatment knowledge steer secondary endpoints. SODa-BIC, being double-blind, removes exactly this confounder, and the RRT signal shrank accordingly: 16.8% vs 20.9%, overall not significant. In the pH <7.25 subgroup, an 11-point difference persisted (17.2% vs 28.3%, 95% CI -22.3 to 0.4)—but with blinding intact, this is the honest version of the question BICAR-ICU2 never answered.

Old treatment, new trial

SODa-BIC differs from its predecessors in phenotype: vasopressor-dependent metabolic acidosis rather than severe acidemia and AKI. While BICAR-ICU suggested benefit in the subgroup with severe AKI and BICARICU-2 confirmed a reduction in kidney replacement therapy without improving death (dissociation more plausibly explained by unblinded clinicians than by a kidney-specific drug effect), SODa-BIC, blinded, found neither. Perhaps most consistent was the finding of no change in mortality.

Strengths

SODA-BIC used MAKE30, a patient-centered composite outcome, while accounting for competing risks and variation in baseline kidney function. Second, the trial incorporated a prespecified blinded sample-size re-estimation, improving confidence that the final sample size was adequate for the planned effect. Most importantly, the trial was also pragmatic and multicenter, increasing its relevance to real-world ICU practice across different health-care systems

Limitations

Several limitations are also important. Open-label bicarbonate was permitted when clinically indicated, and other sodium- or buffer-containing therapies were not fully controlled, potentially reducing treatment separation. Although the intervention corrected acidemia rapidly, it was brief, raising uncertainty about whether it was sustained long enough to influence a 30-day outcome. In addition, although blinding was carefully maintained, changes in pH and bicarbonate may have allowed clinicians to infer treatment allocation in some cases. The study did not collect physiological data at the time of initiation of renal replacement therapy, limiting interpretation of whether dialysis decisions differed between groups. Finally, the trial was powered to detect a moderate-to-large treatment effect. Therefore, while it convincingly excludes a major benefit, smaller but potentially clinically relevant effects, and here we are talking about kidney-specific outcomes, cannot be ruled out.

So, where does the SODA-BIC trial leave us? Sodium bicarbonate corrects the numbers, yes, but it still fails to correct patient-oriented outcomes and prevent death. Routine administration to prevent MAKE30 is difficult to justify. However, let’s imagine a real-world scenario. In a deteriorating patient with metabolic acidosis, vasopressor dependency and no clear contraindication, many clinicians may still use sodium bicarbonate as a “wait-it-out” measure - not because it improves survival, but because transient correction of acidemia may provide physiological support while the underlying cause is addressed (and it makes us clinicians feel better, like we’ve done something). Considering these aspects, maybe the takeaway message isn't that “bicarbonate isn’t useful”, but rather “don’t expect bicarbonate to change the prognosis”.  

Conclusion

In the SODa-BIC trial, sodium bicarbonate infusion in critically ill patients in the ICU with metabolic acidosis did not improve mortality or MAKE30 outcome.

Summary by

Srinivasavaradan Govindarajan
Assistant Professor (Ped Neph)
VMMC and Safdarjung Hospital, India

Anca Elena Stefan
Nephrology specialist
Romania

Reviewed by

Brian Rifkin, Cristina Popa, Sai Vani, Akshaya Jayachandran

Header image created by AI and prompts from Brian Rifkin