SAVE-O2 AI randomized trial: autonomous oxygen titration kept acutely ill adults in the target range 85% of the time vs 63% under usual care (JAMA Internal Medicine, 2026)

Verification status: IN CHECKING — not publish-ready, not pending, not verified.

The problem

Supplemental oxygen is one of the most common interventions in hospital care, yet clinicians cannot watch every patient’s oxygen saturation continuously, so patients drift out of the intended range into either too little oxygen (hypoxemia) or too much (hyperoxemia). The question this case file tests is narrow and measurable: when an autonomous closed-loop system manages the oxygen flow instead of the bedside clinician, does the share of time a patient actually spends in the target range rise on the independent public record, rather than in a device maker’s brochure (source).

What was built

The intervention tested was the O2matic PRO100, which the trial describes as “a closed-loop, trend-based, autonomous oxygen titration system” (source). The CU Anschutz institutional release describes the same device as continuously monitoring oxygen levels and adjusting oxygen flow in real time using pulse oximetry without requiring clinician intervention between assessments (source). To test it, investigators ran the SAVE-O2 AI trial: “300 adults…randomized…152 in the autonomous oxygen titration group and 148 in the usual care group” at “4 tertiary care hospitals”, enrolling from “May 6, 2024, to November 17, 2025” (source). The trial was academic-led and independent of the device maker: “This effort was sponsored by the US Department of War’s Defense Health Agency Research and Engineering Directorate, Combat Casualty Care Portfolio” (source).

The outcome

On its primary endpoint the trial reported that “the mean (SE) proportion of time spent in normoxemia (Spo2 90%-96%) was greater in the autonomous oxygen titration group than in the usual care group (85% [1%] vs 63% [2%]; adjusted RD, 21 percentage points [pp]; 95% CI, 18-25 pp; P < .001)” (source). The independent pulmonology outlet Healio reported the same result: patients receiving autonomous versus manual oxygen titration “spent a significantly higher proportion of time in normoxemia (mean, 85% vs. 63%; adjusted risk difference, 21 percentage points; 95% CI, 18-25 percentage points; P < .001)” (source), and a second independent outlet, HealthDay News (carried by Pulmonology Advisor), reported the same figures firsthand: “the mean proportion of time spent in normoxemia was greater in the autonomous oxygen titration group versus the usual care group (85 versus 63 percent)” (source). The CU Anschutz release carried the headline figures too, “85 percent of their time in the target oxygen range compared with 63 percent in the standard care group” (source), though as the lead academic center it is first-party to the investigators, not an independent voice.

The safety-relevant tail moved too. The trial reported that “the mean (SE) proportion of time spent in hypoxemia (Spo2 <88%), the key secondary outcome, was lower in the autonomous oxygen titration group than in the usual care group (2.0% [0.2%] vs 3.6% [0.4%]; adjusted RD, -1.3 pp; 95% CI, -2.0 to -0.5 pp; P = .002)” (source), a result Healio and HealthDay News (via Pulmonology Advisor) each reported independently in the same terms, the latter as “the mean proportion of time spent in hypoxemia was lower in the autonomous oxygen titration group versus the usual care group … 2.0 versus 3.6 percent” (source) (source).

A note on source strength. The figures are anchored in a Tier-1 peer-reviewed artifact and reported firsthand by two independent outlets, Healio and HealthDay News (via Pulmonology Advisor), so the corroboration is solid. The residual limit is the study itself, not the sourcing: the effect comes from a single trial, and it was “unblinded” (open-label), which is a real weakness for a subjective process measure even in a peer-reviewed journal (source). The primary outcome is a process measure, the share of time in a target SpO2 band, not a hard patient outcome such as mortality or ventilator-free days, and the independent outlets are reporting the same one study rather than re-measuring it (source). The 21 pp effect is real in this trial and independently reported; it has not yet been independently replicated by a second RCT.

How this was verified

• Method: The primary JAMA Internal Medicine article, Autonomous Oxygen Titration for Maintaining Normoxemia in Acutely Ill Adults: The SAVE-O2 AI Randomized Clinical Trial (Douin DJ, Rice JD, Xiao M, et al.; published online August 3, 2026; doi:10.1001/jamainternmed.2026.4023), was fetched live and its primary and secondary endpoints quoted verbatim for the time-in-normoxemia and time-in-hypoxemia figures, the sample size, the sites, the enrollment dates, the device and the funder. The 85% vs 63% and 2.0% vs 3.6% figures were then corroborated against two genuinely independent outlets, Healio (pulmonology press) and HealthDay News (carried by Pulmonology Advisor), each reporting the figures firsthand, plus the CU Anschutz institutional release (first-party to the investigators, not counted as independent).
• Date verified: 2026-08-27.
• Not claimed: that autonomous titration reduced mortality, length of stay or any hard patient outcome (the primary endpoint is time in the target SpO2 range, a process measure); that the effect has been independently replicated (only the single SAVE-O2 AI trial exists as of the research date); or any blinding of the intervention (the trial was unblinded / open-label).

Sources

1. JAMA Internal Medicine (Douin DJ, Rice JD, Xiao M, et al.) — “Autonomous Oxygen Titration for Maintaining Normoxemia in Acutely Ill Adults: The SAVE-O2 AI Randomized Clinical Trial” · 2026-08-03 · https://jamanetwork.com/journals/jamainternalmedicine/fullarticle/2852401 (archived: http://web.archive.org/web/20260825133436/https://jamanetwork.com/journals/jamainternalmedicine/fullarticle/2852401) — Tier 1 (primary peer-reviewed randomized-trial artifact; academic-led, sponsored by the Defense Health Agency, independent of the device maker; doi:10.1001/jamainternmed.2026.4023).
2. Healio (pulmonology) — “Autonomous vs. manual oxygen titration yields more time in normoxemia” · 2026-08-13 · https://www.healio.com/news/pulmonology/20260813/autonomous-vs-manual-oxygen-titration-yields-more-time-in-normoxemia (archived: https://web.archive.org/web/20260826225811/https://www.healio.com/news/pulmonology/20260813/autonomous-vs-manual-oxygen-titration-yields-more-time-in-normoxemia) — Tier 2 (independent specialty press; corroborates the 85% vs 63% primary result and the hypoxemia secondary result verbatim).
3. University of Colorado Anschutz Medical Campus — “CU Anschutz-Led Trial Finds AI System Improves Oxygen Delivery in Hospital Patients” · 2026-08-03 · https://news.cuanschutz.edu/news-stories/cu-anschutz-led-trial-finds-ai-system-improves-oxygen-delivery-in-hospital-patients (archived: http://web.archive.org/web/20260804002033/https://news.cuanschutz.edu/news-stories/cu-anschutz-led-trial-finds-ai-system-improves-oxygen-delivery-in-hospital-patients) — Tier 2 (institutional release from the lead academic center; corroborates the headline figures and the device; first-party to the investigators, so it is not counted as independent of the study itself).
4. Pulmonology Advisor (HealthDay News) — “Autonomous Oxygen Titration Increases Proportion of Time in Normoxemia” · 2026-08-03 · https://www.pulmonologyadvisor.com/news/autonomous-oxygen-titration-increases-proportion-of-time-in-normoxemia/ (archived: http://web.archive.org/web/20260827023838/https://www.pulmonologyadvisor.com/news/autonomous-oxygen-titration-increases-proportion-of-time-in-normoxemia/) — Tier 2 (independent HealthDay News wire report, distinct publisher from Healio; restates the 85% vs 63% and 2.0% vs 3.6% figures firsthand).
5. PubMed (National Library of Medicine) — record for the SAVE-O2 AI trial · 2026 · https://pubmed.ncbi.nlm.nih.gov/42546017/ (archived: https://web.archive.org/web/20260826225845/https://pubmed.ncbi.nlm.nih.gov/42546017/) — Tier 1 (indexing record confirming the citation, authors and journal).

Related case files

• Penda Health’s AI Consult clinical copilot in Kenya — another clinical-AI outcome where a strong process/documentation effect sits next to a null on the harder patient outcome, the same tension between a process win and a patient-outcome win recorded here.
• Boston Children’s and OpenAI o3 for rare-disease reanalysis — the peer-reviewed clinical-benchmark counterpart, useful for contrasting a diagnostic-accuracy result with this real-time control-system result.
• METR’s 2025 randomized controlled trial of AI coding tools — the evidence-standard reference: an independent RCT is the bar this file is measured against, and the reason a single unreplicated trial caps below green.