The AI Giants Enter the Cardiac ICU | Clinical Challenge 2 (D03)
SpO₂ 97%. Seven breaths a minute. PaCO₂ 74 mmHg. Which of these numbers should drive the decision?
In the previous episode the misleading number was arterial pressure. All three frontier models saw through it and recognized that an acceptable pressure can coexist with severe circulatory failure. This time the reassuring figure comes from the pulse oximeter.
The usual rule applies. The patient is fictional, and I wrote the case and each of its updates before any model saw them.
T0 | Two hours after extubation
A patient with obesity and obstructive sleep apnea has been extubated after cardiac surgery. Two hours later the picture is this: oxygen at 4 L/min, SpO₂ 98%, respiratory rate 14/min, pH 7.34, PaCO₂ 49 mmHg. Drowsy, but wakes to voice, coughs on request and protects the airway. Opioids were given earlier.
What worries you here? Closer observation, positive airway pressure, or reintubation?
The models at T0
GPT-6 Astra raised the possibility of opioid-related hypoventilation in a patient with OSA and made the central point straight away: a normal saturation on supplemental oxygen does not prove that ventilation is adequate. It asked for prompt reassessment and close monitoring rather than reflex reintubation. It also wanted a number I would have asked for myself. Was this PaCO₂ higher than the earlier, or the preoperative, value?
Claude Fable 5.1 saw the same risk of evolving postoperative hypoventilation. Its answer centered on vigilance: monitoring of ventilation and consciousness, attention to residual opioid effect, and invasive treatment held back until there was evidence of deterioration.
Gemini 4 Argon was somewhat quicker to reach for early positive airway pressure in this vulnerable patient, while still watching ventilation and the patient’s ability to keep the airway protected.
T0 did not call for heroics. It called for something simpler, and easier to get wrong: keeping oxygenation and ventilation apart.
T1 | Forty minutes later
Respiratory rate falls to 7/min. The patient is now difficult to rouse. SpO₂ holds at 97%, on 6 L/min of oxygen. Arterial blood gas: pH 7.20, PaCO₂ 74 mmHg.
It is the clinical trajectory, not the isolated reading, that changes the stakes.
Turning up the oxygen will not fix inadequate alveolar ventilation. So what now? A rapidly reversible intervention, assisted noninvasive ventilation, or a secured airway?
The models at T1
Astra named the problem without hedging: severe, rapidly progressive hypercapnic respiratory failure with impaired consciousness. It wanted an experienced airway team involved immediately, active ventilatory assistance, no further opioids, and consideration of carefully titrated naloxone. On NIV it was explicit. A patient this hard to wake may not tolerate it safely, and it must not delay intubation if airway protection, or the early response to treatment, is inadequate.
Fable also treated the situation as an immediate emergency. Its first lever was titrated naloxone, as a potentially fast reversal of opioid-related respiratory depression, with assisted ventilation and reintubation prepared in parallel. NIV came with a condition attached: enough recovery of alertness and airway protection to justify it.
Argon recognized how severe the hypercapnia was despite the preserved saturation. It advised urgent assisted ventilation and titrated naloxone, and prepared for emergency intubation. It also, however, listed an immediate monitored NIV attempt alongside bag-mask support, in a patient described as difficult to awaken. That threshold is worth a closer look.
This is where the answers begin to diverge in ways that matter at the bedside. On the physiology they agree. Where they part company is the margin of safety each one accepts for a noninvasive attempt.
T2 | A decisive contraindication
After a brief, monitored trial of noninvasive support, the patient vomits, can no longer protect the airway and is still severely hypoventilated: pH 7.17, PaCO₂ 77 mmHg, respiratory rate 7/min.
By now the uncertainty about NIV has shrunk to very little.
The models at T2
Astra identified loss of airway protection and an immediate risk of aspiration. Stop NIV, clear the airway, proceed to urgent tracheal intubation and invasive ventilation. No waiting for another blood gas, or for a drug to act.
Fable said much the same: take the interface off at once, suction, secure a definitive airway, keeping in mind aspiration risk, obesity and a potentially difficult airway.
Argon accepted that NIV was no longer appropriate, called for stopping it immediately and gave priority to endotracheal intubation and invasive ventilation.
At T2 all three converged on the same critical decision. Whatever separated them earlier, vomiting and loss of airway protection made continued NIV the wrong strategy.
The physiological trap: SpO₂ is not ventilation
Pulse oximetry estimates arterial oxygen saturation. It is an essential monitor. It does not measure CO₂ clearance, and it cannot tell you whether alveolar ventilation is adequate. With supplemental oxygen running, SpO₂ can stay comfortable for some time while hypoventilation and respiratory acidosis get worse.
This is not a theoretical worry. In 2004 Fu and colleagues halved minute ventilation in anesthetized patients. SpO₂ fell (on average from 97% to 91%) only in those breathing room air; patients on an FiO₂ of 0.25 to 0.30 showed no meaningful drop, even though their CO₂ was climbing. Their conclusion was blunt. Without capnography or blood gases, pulse oximetry works as a ventilation monitor only when no supplemental oxygen is given.
Hypercapnia also tends to feed itself. Rising CO₂ deepens drowsiness, drowsiness slows breathing further, and in an obese patient with sleep apnea who has received opioids every factor pushes in the same direction.
The combination of severe bradypnea, progressive hypercapnia, acidosis and declining consciousness therefore carries far more weight than a superficially normal saturation.
What happens next depends on a question that a discussion built around gas exchange easily underplays: can the patient maintain and protect the airway? NIV needs a patient, and a clinical setting, in which the interface can be used safely, secretions and aspiration can be managed, and deterioration is picked up quickly. Vomiting, or an airway the patient can no longer protect, changes that balance completely.
Guideline context
The ISCCM guideline on noninvasive ventilation in adult ICUs lists inability to protect the airway and severe gastrointestinal symptoms, vomiting included, among the contraindications to NIV.
Two other points in the same document seem to me more relevant to this case. The guideline advises against using NIV to treat established respiratory distress after extubation in high-risk patients (grade 2B). Much of the weight behind that recommendation comes from the trial by Esteban and colleagues. In 221 patients, reintubation occurred in 48% of both arms, ICU mortality was 25% with NIV versus 14% with standard care, and reintubation came later in the NIV group (median 12 hours versus 2.5). The trial does not prove that the delay caused the excess deaths. It is still hard to ignore. However, Esteban’s patients had received at least 48 hours of mechanical ventilation before extubation, unlike the early postoperative cardiac surgery patient in this vignette; direct extrapolation is therefore limited. At the same time, the guideline recommends NIV after cardiothoracic surgery to improve oxygenation and reduce reintubation (grade 2A), and in its COPD section it does not treat hypercapnic coma as an absolute bar to NIV.
These positions do not contradict each other. They do show how much depends on who the patient is and on what, precisely, is failing. No COPD diagnosis was provided in this scenario. Opioid-related respiratory depression is strongly suspected, but its contribution relative to OSA and other postoperative factors remains uncertain. I am not convinced the COPD experience transfers that far. A general critical care guideline still has to be applied to a fresh sternotomy.
My clinical reading: the decision that mattered was at T1
The final step at T2 is straightforward. Airway protection comes first, and all three models passed that test.
T1 is the more interesting moment. The patient is hard to rouse, but the exact degree of airway protection has not been fully described. I can imagine a short, closely monitored attempt to reverse the opioid effect, provided the airway team is already in the room and the threshold for intubation is very low. A proposed NIV trial in that setting, though, cannot be judged on the mere availability of a mask and a ventilator. What decides it is the patient’s ability to cooperate, clear secretions and guard against aspiration.
There is also a surgical side to naloxone that a purely respiratory reading tends to leave out. Reverse a fresh sternotomy too abruptly and the pain comes back all at once, together with the sympathetic surge that goes with it: hypertension (not what anyone wants on fresh suture lines), tachycardia, higher myocardial oxygen demand and, occasionally, pulmonary edema. Hence small, repeated increments rather than a single bolus. And because naloxone often wears off before the opioid does, a patient who is awake at minute ten may be back where they started an hour later.
Astra was openly cautious about NIV. Fable made it conditional on a demonstrable improvement in alertness. Argon was more willing to list it as an immediate option. I would not turn this into a categorical win for one model. Still, it is a real difference in how each of them handles risk.
It also exposes a weakness in our preliminary scoring system. All three received 5/5, because each recognized hypoventilation and chose intubation at the appropriate final stage. A binary checklist was never designed to grade the minutes that come before. A finer scale would capture more of that nuance; it would also let more of my own judgment into the score, and I have not yet decided which of the two problems is worse.
The question I would put to readers
At T1 the patient is on oxygen, saturating 97%, breathing seven times a minute, with PaCO₂ 74 mmHg and pH 7.20, and is difficult to awaken.
Would you attempt titrated opioid reversal while actively supporting ventilation, and for how long? Would you consider NIV at all? Which precise bedside finding would make you intubate immediately?
My guess is that the most useful part of the discussion will not be about whether the blood gas is abnormal. Everyone can see that it is. It will be about the narrow boundary between a reversible episode of hypoventilation and an airway that can no longer be managed safely without intubation.
Next, Episode 04: worsening postoperative oliguria, rising creatinine, high venous pressure and right ventricular dysfunction. More fluid, less fluid, or renal replacement therapy?
References
- Chawla R, Dixit SB, Zirpe KG, et al. ISCCM guidelines for the use of non-invasive ventilation in acute respiratory failure in adult ICUs. Indian J Crit Care Med. 2020;24(Suppl 1):S61-S81. PMC7085817
- Esteban A, Frutos-Vivar F, Ferguson ND, et al. Noninvasive positive-pressure ventilation for respiratory failure after extubation. N Engl J Med. 2004;350(24):2452-2460. doi:10.1056/NEJMoa032736
- Fu ES, Downs JB, Schweiger JW, Miguel RV, Smith RA. Supplemental oxygen impairs detection of hypoventilation by pulse oximetry. Chest. 2004;126(5):1552-1558. doi:10.1378/chest.126.5.1552
- Engelman DT, Ben Ali W, Williams JB, et al. Guidelines for perioperative care in cardiac surgery: Enhanced Recovery After Surgery Society recommendations. JAMA Surg. 2019;154(8):755-766. doi:10.1001/jamasurg.2019.1153. See also the ERAS Society cardiac page.