Autonomous AI Oxygen Titration Keeps Hospitalized Patients in Target Oxygen Range Better Than Nurses Can

Oxygen is the most commonly administered therapy in hospitalized patients, but its delivery is surprisingly imprecise. The problem is not lack of awareness — it is the logistics of a busy ward. Nurses check oxygen saturations intermittently, adjust flow rates when they notice drift, then move on. Between checks, a patient may spend hours in mild hyperoxia (SpO2 98–100% when 92–96% is the target) or, less commonly, brief hypoxia. Both extremes carry risk: prolonged hyperoxia generates reactive oxygen species that injure the lung and airways, and in patients with COPD it can suppress hypoxic drive.[1]

The SAVE-O2 AI trial enrolled 300 acutely ill adults requiring supplemental oxygen at four U.S. hospitals (UCHealth University of Colorado, Vanderbilt, Oregon Health & Science University, and Wake Forest) and randomized them to either the O2matic PRO100 closed-loop device — which continuously reads pulse oximetry and adjusts oxygen flow in real time — or usual care with nurse-driven titration. The primary outcome was the proportion of time spent within the normoxemia target range of SpO₂ 90–96%.[1]

The autonomous system significantly outperformed usual care. Patients in the device arm spent 85% of monitored time within their prescribed oxygen range, compared with 63% in the standard care arm — a 22 percentage point absolute improvement (35% relative increase). Time in hypoxemia fell by 45% and time in hyperoxemia fell by 68%. Daily supplemental oxygen volume was approximately 23% lower in the device arm. The number of manual oxygen adjustments by clinical staff was also reduced. Serious adverse events were not increased.[1]

Clinical Context

The question of optimal oxygen targets in hospitalized patients has been actively studied since the early 2010s, when emerging evidence challenged the assumption that more oxygen is always better. The IOTA systematic review (2018, Lancet) found that liberal oxygen targets — SpO2 ≥96% — were associated with higher mortality than conservative targets in critically ill adults, with excess deaths attributable to reactive oxygen species-mediated lung and end-organ injury.[2] The HOT-ICU trial (2021, NEJM), which randomized 2,928 ICU patients to PaO2 60 vs 90 mmHg targets, found no mortality difference at 90 days but confirmed the theoretical physiologic basis for oxygen conservation in the critically ill.[3] Practical implementation of conservative oxygen targets, however, remained elusive: nurse-driven protocols require constant vigilance and frequent manual flow adjustments that compete with dozens of other care priorities.

Closed-loop oxygen delivery systems — which pair continuous SpO2 monitoring to automated FiO2 or flow adjustments via algorithm — have been evaluated in intensive care settings and preterm neonates, where they demonstrably reduce time spent outside the target SpO2 range. Extension to general medical wards is technically simpler (lower-acuity patients, simpler delivery via nasal cannula) but logistically broader: the patient population is far larger, nursing ratios are lower, and oxygen is often treated as a set-and-forget background therapy rather than a titrated drug. This RCT addresses that gap directly.[2]

Why It Matters Clinically

The connection to the broader literature on liberal versus conservative oxygen targets is important here. The LOCO2, HOT-ICU, ICU-ROX, and Oxygen-ICU trials all found that liberal oxygen (high SpO2 targets) in ICU patients offered no benefit and possible harm. In mechanically ventilated patients, targeting SpO2 of 90–94% rather than 97–100% is now guideline-supported. The new trial extends that logic to ward patients receiving supplemental oxygen: the target matters, and hitting it consistently is harder than it looks without automation.[2]

Limitations

This trial reports a process measure — time in target range — not a clinical outcome like hospital mortality, 30-day readmission, or length of stay. Whether reducing hyperoxia exposure on the ward translates into improved hard outcomes remains to be demonstrated in a larger, outcome-powered trial. The autonomous device requires integration with pulse oximetry hardware that is not universally available.