The AI Giants Enter the Cardiac ICU | Clinical Challenge 3 (D04)
When Oliguria Means Congestion: Three AI Models Face Postoperative Kidney Injury
Urine output is falling. Creatinine is rising. But the circulation is already volume overloaded. Would you give fluid, remove it, or prepare dialysis?
Oliguria after cardiac surgery can generate an almost reflexive urge to intervene. A fluid bolus seems attractive when urine output drops. A rising creatinine may encourage early consideration of renal replacement therapy. Yet neither reaction is justified without examining the underlying physiology.
This is the fourth installment of The AI Giants Enter the Cardiac ICU. As in the previous cases, we gave the same fictional patient to GPT-6 Astra, Claude Fable 5.1, and Gemini 4 Argon, allowing each model to revisit its decision as new information emerged.
The cases were established in advance. The previous clinical challenge addressed hypercapnia after extubation (Episode 03).
T0 — Oliguria without an established mechanism
On the first postoperative day after cardiac surgery, a 75-kg patient is producing 20 mL of urine per hour for four hours. MAP is 72 mmHg and lactate 1.8 mmol/L. Creatinine is 1.0 mg/dL, compared with a baseline of 0.9 mg/dL, and the postoperative fluid balance is +2.0 L. There is no known active bleeding, but neither the degree of venous congestion nor the effective circulating volume has yet been established.
What would you do now? Is a positive fluid balance enough to diagnose congestion? Is four hours of low urine output enough to diagnose acute kidney injury? Is fluid administration justified?
What the models said at T0
Astra identified significant oliguria, approximately 0.27 mL/kg/h, without prematurely labeling the current creatinine change as established AKI. It prioritized checking the measurement and urinary catheter, assessing perfusion and fluid status, and avoiding either automatic fluid boluses or diuretics based on urine output alone.
Fable focused on the early postoperative oliguria as a warning requiring prompt, targeted assessment. It did not assume that all oliguria reflected intravascular depletion, and advocated clarification of the patient’s hemodynamics and fluid status before choosing a treatment.
Argon likewise treated oliguria as clinically important but insufficient for a definitive mechanism. It emphasized reassessment of perfusion and congestion, rather than indiscriminate fluids.
The key point at T0 is epistemic: urine output tells us that something may be wrong; it does not yet tell us what the correction should be.
T1 — The missing mechanism becomes visible
Six hours later, urine output is 10 mL/h and creatinine has risen to 1.8 mg/dL. MAP is 73 mmHg, CVP 19 mmHg, and cumulative fluid balance +3.5 L. Echocardiography reveals a dilated right ventricle with moderate dysfunction, and there is clinical evidence of systemic congestion.
The patient now has evolving AKI in the context of marked venous congestion and right ventricular dysfunction.
What the models said at T1
Astra recognized the likely contribution of elevated venous pressure to renal dysfunction. It recommended urgent, monitored decongestion using a loop diuretic, while reassessing right ventricular function and other reversible causes. It explicitly discouraged empirical fluid administration.
Fable also identified congestion-related AKI and the importance of venous pressure in reducing effective renal perfusion. It proposed decongestion, optimization of right ventricular support, and nephrology involvement. Its response was notably more interventionist in one respect: it proposed preparing for renal replacement therapy and considered progression to ultrafiltration if the diuretic response remained inadequate after approximately two hours.
Argon recognized acute cardiorenal dysfunction driven by right ventricular failure and systemic venous congestion. It recommended stopping nonessential fluids, active but monitored decongestion, and consideration of hemodynamic support to improve forward circulation and perfusion.
All three avoided the most tempting inappropriate simplification: oliguria means low volume, therefore give more fluid. The difference lay in how early renal replacement therapy should enter the operational plan.
A related problem—right ventricular dysfunction despite apparently reassuring arterial pressure—was explored in Episode 02.
T2 — The renal trajectory begins to improve
After team-guided hemodynamic interventions and reduction of congestion, CVP falls to 11 mmHg and urine output improves to 70 mL/h. Creatinine is 1.7 mg/dL. Potassium is 4.5 mmol/L, pH 7.38, and there is no refractory pulmonary edema.
The kidney injury has not disappeared. But the immediate physiology is changing in the right direction.
What the models said at T2
Astra interpreted the renewed diuresis and reduction in creatinine as early improvement, without claiming full renal recovery. It recommended close reassessment of fluid balance and hemodynamics, and did not identify a current indication for renal replacement therapy from the supplied findings.
Fable similarly stepped back from escalation. It regarded RRT as a contingency rather than an immediate intervention, and emphasized avoiding excessive decongestion or an abrupt drop in right ventricular preload.
Argon also supported continuing a monitored conservative strategy, avoiding recurrent congestion and iatrogenic hypovolemia. It did not recommend RRT in the absence of a clinical indication.
All three recognized that management should change with the patient’s response. A therapy considered at T1 should not automatically proceed at T2 simply because the creatinine remains abnormal.
The guideline comparison: treat the patient, not a number
Acute kidney injury is not managed by a single creatinine threshold. The NICE guideline on AKI recommends basing renal replacement therapy decisions on the patient’s overall condition, rather than an isolated creatinine, urea, or potassium result. It identifies indications for urgent consideration of RRT when serious problems such as hyperkalemia, acidosis, uremic complications, fluid overload, or pulmonary edema are not responding to medical treatment.
See NICE NG148: Acute kidney injury, especially recommendations 1.5.6–1.5.9.
The physiology is also important. Elevated central venous pressure in right ventricular failure can impair renal drainage and contribute to a reduced gradient for renal perfusion despite an apparently acceptable MAP. Venous congestion deserves active recognition; at the same time, excessive fluid removal can compromise preload and cardiac output. Neither “always give fluid” nor “always remove fluid” is a defensible rule.
The European Society of Cardiology’s discussion of acute right ventricular failure also highlights venous congestion as a driver of organ dysfunction. Read the consensus document.
My clinical reading: a good answer can still be too prescriptive
In this vignette, I do not see a major failure to identify the primary physiological mechanism. All three models recognized congestion, avoided unselected fluid loading, and backed away from RRT when the patient improved.
Fable’s early plan to prepare access and consider ultrafiltration after a very short period without diuretic response, however, deserves scrutiny. A poorly responsive, severely congested patient may indeed require RRT or ultrafiltration. But a fixed two-hour threshold is not itself a guideline-based universal indication. We would need to know the severity of volume overload, whether it is truly refractory, the clinical trajectory, the expected benefit of additional pharmacological treatment, and the wider hemodynamic picture.
This is a useful distinction between anticipating a possible need and creating an overly rigid trigger for action.
In our preliminary checklist-based exercise, all three models received 5/5. The more subtle question of proportionality was not captured by that simple score.
What would you do at T1?
The patient has a creatinine of 1.8 mg/dL, a CVP of 19 mmHg, a positive fluid balance of 3.5 L, right ventricular dysfunction, and severe oliguria. MAP remains 73 mmHg.
What is your first intervention? How would you judge whether decongestion is working? Under what additional conditions would you initiate renal replacement therapy?
I would be interested in how different clinical teams approach the relationship between right ventricular failure, congestion, renal injury, and the limits of medical treatment.
This completes the first three detailed clinical challenges on the blog. Further shorter challenges in the same numbered series will explore postoperative stroke and anticoagulation during active bleeding on LinkedIn, while the blog continues with its wider range of medical and AI topics.
Editorial note: Synthetic case only. The three model responses are summarized in English from original Italian outputs. The commentary is educational and does not replace guideline interpretation or individual clinical assessment.
Further reading on this site
- Dataset Shift in Medical AI: Why Validated Models Lose Reliability — on why clinical context and the patient population matter when evaluating AI systems.
- Machine Learning Model Evaluation — on the limits of summary metrics and the importance of evaluating decisions at clinically meaningful thresholds.