An AI Oxygen System Kept Patients in Range 85% of the Time

Oxygen is one of medicine’s most familiar treatments. Delivering the right amount is less simple.

Too little oxygen can leave a patient hypoxemic. Too much can also cause harm, depending on the clinical setting. Yet supplemental oxygen is often adjusted manually after intermittent checks, while a patient’s needs can change between observations.

A randomized clinical trial published online in *JAMA Internal Medicine* on August 3, 2026, tested a more continuous approach. A closed-loop system monitored oxygen saturation and automatically adjusted supplemental-oxygen flow. Across 300 adults at four US hospitals, the autonomous-control group spent a mean 85% of monitored time in the trial’s target range, compared with 63% under usual care.

That adjusted 21-percentage-point difference is not proof that the system saves lives or shortens hospital stays. It is evidence of something more specific: under the conditions studied, a bounded automated system kept a measurable physiological signal within a clinician-defined range more consistently than manual adjustment.

Research references: [1]

The trial tested continuous adjustment instead of periodic checks

SAVE-O2 enrolled adults hospitalized for acute respiratory illness, trauma, burns, or acute care surgery who were receiving supplemental oxygen. Participants were recruited from May 6, 2024, through November 17, 2025. The registry required randomization within 36 hours of hospital arrival—not within 36 hours of starting oxygen—and excluded people expected to need high-flow oxygen or ventilation imminently.

The median age was 66 years, 54% of participants were women, and 69% were being treated for an acute respiratory illness. Researchers also reported investigator-assessed skin tone: 22% of participants were categorized as light, 56% as medium, and 22% as dark.

The intervention was O2matic PRO100. It used pulse-oximeter readings to adjust oxygen flow automatically, targeting 93% saturation within an acceptable range of 90% to 96%. Treatment continued for up to 72 hours, with clinical escalation or hospital discharge ending the intervention earlier.

This is not generative AI. It does not write, diagnose, or hold a conversation. It is a closed-loop control system: it takes in a physiological signal, applies a predefined algorithm, and changes one treatment variable within limits. That narrowness is part of the story. In health care, some of automation’s most defensible uses may be systems whose inputs, outputs, and escalation boundaries can be clearly described.

Research references: [1] [2]

Autonomous control kept oxygen levels closer to target

The primary result was straightforward. Mean time between 90% and 96% oxygen saturation was 85% with autonomous control and 63% with usual care. The adjusted difference was 21 percentage points, with a 95% confidence interval from 18 to 25 percentage points.

Time below 88% saturation was also lower: 2.0% with autonomous control versus 3.6% with usual care. The adjusted difference was minus 1.3 percentage points, with a 95% confidence interval from minus 2.0 to minus 0.5.

The larger change was on the high side of the range. Participants using autonomous control spent 9.2% of monitored time above 96% saturation, compared with 29.1% under usual care.

Together, those findings show that the system did not simply respond to low readings by supplying more oxygen. It held participants closer to a defined middle range and reduced time on both sides of it.

Median oxygen use was 2,310 liters per participant per day with autonomous control and 2,992 liters with usual care. But the incidence rate ratio was 0.83 with a 95% confidence interval from 0.67 to 1.02, so this trial did not establish a clear reduction in oxygen use.

Research references: [1]

This is bounded automation at its most understandable

Much of the debate around medical AI centers on systems whose reasoning can be difficult to inspect. Closed-loop oxygen control presents a different model.

The task is narrow. The target is explicit. The input is continuously measured. The output is limited to oxygen flow. Clinicians retain diagnosis, treatment goals, escalation decisions, and the ability to stop the system.

That does not make the system simple or risk-free. Pulse oximeters are imperfect, and their performance can vary across devices and skin pigmentation. The trial’s authors therefore reported skin-tone categories and examined results across prespecified subgroups; those analyses should not be read as proof of equal performance for every patient or sensor.

Still, the design offers a useful way to think about AI support in care. Rather than asking a model to replace broad clinical judgment, a hospital can ask whether automation can reliably manage one high-frequency adjustment while people remain responsible for context and escalation.

Research references: [1]

Better control did not yet mean better patient outcomes

The trial should not be stretched beyond what it measured. Its primary outcome was time in a target oxygen-saturation range during the first 72 hours, not survival, recovery, length of stay, staffing relief, or cost.

The paper reported no serious adverse events attributed to the intervention. Four nonserious events in the autonomous group—two episodes of nosebleed and two reports of discomfort from the oxygen interface—led to a protocol amendment requiring humidification when flow exceeded 8 liters per minute.

The study was unblinded, tested one commercial device, and depended on pulse-oximeter accuracy. It took place at four US hospitals, so its performance may not transfer automatically to other settings, populations, devices, or workflows.

Regulatory and commercial context matters too. The ClinicalTrials.gov record, updated October 14, 2026, lists the device as FDA-regulated and unapproved; it names the University of Colorado as lead sponsor and the US Department of Defense, O2matic, and IDTS Medical as collaborators. The article also reports author disclosures, including grants and fees outside or during the conduct of the study.

Research references: [1] [2]

The next test is whether precision changes recovery

SAVE-O2 shows that autonomous oxygen control can improve control of a measurable bedside target. That is a meaningful result. It answers the first practical question: can the system keep hospitalized adults in range more consistently?

The next questions are harder and more consequential. Does tighter control reduce complications? Does it help people recover sooner? Can it reduce repetitive work without adding new alarms or oversight costs? Does performance remain reliable across hospitals, skin tones, sensor types, and clinical conditions?

Those questions require larger and longer studies designed around patient outcomes and implementation.

For now, the strongest conclusion is also the most useful one. AI did not prove it could replace clinical judgment. It showed that, within a carefully bounded task, continuous automation could perform a routine adjustment more consistently than intermittent manual care. That is the kind of evidence on which responsible medical automation can be built.