Publication

• Title: Autonomous Oxygen Titration for Maintaining Normoxemia in Acutely Ill Adults: The SAVE-O2 AI Randomized Clinical Trial
• Acronym: SAVE-O2 AI
• Year: 2026
• Journal published in: JAMA Internal Medicine
• Citation: Douin DJ, Rice JD, Xiao M, et al; SAVE-O2 AI Investigators. Autonomous oxygen titration for maintaining normoxemia in acutely ill adults: the SAVE-O2 AI randomized clinical trial. JAMA Intern Med. Published online August 3, 2026.

Context & Rationale

• Background

  • Supplemental oxygen is one of the most frequently administered acute-care treatments, but oxygen flow is usually adjusted intermittently by clinicians rather than continuously in response to changing oxygen saturation.
  • Intermittent manual titration can leave patients below the prescribed saturation range between observations and can also permit prolonged unnecessary oxygen administration and hyperoxaemia.
  • Closed-loop systems continuously use pulse-oximetry feedback to adjust oxygen flow. A 2024 systematic review of 21 studies involving 1,577 participants found that closed-loop control increased time within the prescribed saturation target by a mean 25.47 percentage points (95% CI 19.7 to 30.0), while evidence for reduced hypoxaemia and staff workload was less certain.¹
  • Most previous studies were small, used selected respiratory or perioperative populations, evaluated heterogeneous devices and targets, or focused on mechanical ventilation or high-flow oxygen. Evidence in a broad population receiving conventional low-flow oxygen was limited.
  • The investigators' earlier SAVE-O2 trials showed that clinician-led targeting of an SpO₂ range of 90%-96% reduced oxygen exposure in trauma and burn populations, but manual implementation remained dependent on staff recognition and repeated bedside adjustments.²³
• Research Question/Hypothesis

  • In adults hospitalised with acute illness or injury and receiving 1-10 L/min of supplemental oxygen, would autonomous closed-loop oxygen titration increase the proportion of observed time spent at SpO₂ 90%-96% compared with clinician-led manual titration?
  • The investigators also hypothesised that autonomous titration would reduce time spent in hypoxaemia and hyperoxaemia and reduce overall oxygen utilisation.⁴
• Why This Matters

  • Closed-loop oxygen delivery could convert a commonly intermittent treatment into a continuously controlled therapy, reducing unnoticed excursions outside the prescribed range.
  • Automation could reduce bedside workload and conserve oxygen, which may be valuable during staffing constraints, oxygen shortages, prolonged field care or other resource-limited conditions.
  • However, greater time within a pulse-oximetry target is a physiological process outcome. Whether this produces fewer episodes of respiratory failure, shorter hospital stay, lower mortality or meaningful cost savings requires separate demonstration.

Design & Methods

• Research Question: Does autonomous oxygen titration with the O2matic PRO100 improve time spent within SpO₂ 90%-96%, while reducing hypoxaemia, hyperoxaemia and oxygen use, compared with manual clinician titration during the first 72 hours of hospital care?
• Study Type: Investigator-initiated, multicentre, open-label, parallel-group, patient-level randomised controlled device trial conducted at four geographically distinct US tertiary hospitals from 6 May 2024 to 17 November 2025. The trial operated under a US Food and Drug Administration investigational device exemption and was funded principally by the US Defense Health Agency through the Medical Technology Enterprise Consortium.
• Population:
  • Adults aged at least 18 years who were admitted from the emergency department with acute respiratory illness, major trauma, burns or an acute-care surgical condition.
  • Participants had to be receiving 1-10 L/min of supplemental oxygen for documented or presumed hypoxaemia and be randomised within 36 hours of hospital arrival.
  • Acute respiratory illness was defined as a primary or secondary pulmonary process causing hypoxaemia; hypoxaemia primarily attributable to sedation or altered consciousness did not qualify.
  • Key exclusions were anticipated discharge within 24 hours, imminent discontinuation of oxygen, imminent escalation to high-flow nasal oxygen, non-invasive ventilation or invasive mechanical ventilation, pregnancy, imprisonment, inability of the clinical team to accept the assigned titration strategy and device-specific contraindications.
  • The analysed cohort comprised 300 patients: median age 66 years (IQR 55-75), 162 (54%) women, 208 (69%) admitted with acute respiratory illness, 73 (24%) in an ICU at enrolment and 155 (52%) receiving at least 3 L/min of oxygen.
• Intervention:
  • Supplemental oxygen was delivered through a nasal cannula or face mask and automatically titrated by the O2matic PRO100, a trend-based closed-loop system that could deliver 0-15 L/min.
  • The system was programmed to an SpO₂ range of 92%-94%, thereby targeting 93%; the protocol-defined acceptable clinical and outcome range was 90%-96%.
  • A maximum of 8 L/min was recommended through a standard non-humidified nasal cannula. Flows up to 15 L/min required a face mask or humidified large-bore nasal cannula.
  • The device titrated from its own Nonin pulse oximeter. A second hospital pulse oximeter, usually on the opposite hand, remained connected to routine central monitoring because the trial device alarmed only within the patient's room.
  • Autonomous titration continued for up to 72 hours, hospital discharge or a protocol-defined concluding event. It was discontinued if oxygen requirements exceeded 15 L/min or the patient required high-flow nasal oxygen, non-invasive ventilation or invasive mechanical ventilation.
• Comparison:
  • Supplemental oxygen was manually titrated by nurses and respiratory therapists according to local hospital practice.
  • The same target SpO₂ of 93% and acceptable range of 90%-96% were communicated to the clinical team at least daily throughout the intervention period.
  • Participants wore both pulse oximeters, but the PRO100 remained in observation mode and did not influence oxygen delivery.
  • Typical assessment frequency was every 1-2 hours in the ICU and every 4-8 hours on general wards. All other clinical care, including the oxygen-delivery interface and escalation of respiratory support, remained at clinician discretion.
• Blinding: Clinicians and participants were not blinded. The primary outcome was derived electronically from continuous pulse-oximetry data, reducing subjective outcome assessment, although knowledge of allocation could affect manual titration, device tolerance, monitoring and decisions to discontinue study procedures. Blinding of the final analysts was not reported.
• Statistics: A total of 300 participants (150 per group) was planned to detect a 10-percentage-point absolute increase in time spent in normoxaemia, from 60% to 70%, at a two-sided significance level of 0.05. The SAP calculation gave 99.1% power with an outcome SD of 20 percentage points and 82.1% power with an SD of 30 percentage points. The primary analysis compared groups as randomised using a quasi-binomial generalised linear model with an identity link, adjusted for study site and admission indication, with cluster-robust standard errors for the two pulse-oximeter streams and weighting by observed monitoring time. Secondary outcomes were not adjusted for multiplicity. A single safety interim analysis after approximately 150 participants used a one-sided α of 0.01 to identify excess hypoxaemia; there was no efficacy or futility stopping rule.
• Follow-Up Period: The intervention and continuous oxygenation assessment continued for up to 72 hours. Clinical outcomes were collected to day 28 or hospital discharge, whichever occurred first; there was no post-discharge follow-up.

Key Results

This trial was not stopped early. The data and safety monitoring board reviewed the planned safety analysis after 150 participants and recommended continuation without modification to the planned enrolment of 300 patients.

• Autonomous titration produced a large and precise improvement in the physiological control outcome, with an adjusted 21-percentage-point increase in time within SpO₂ 90%-96%.
• The principal separation was avoidance of hyperoxaemia; the reduction in time below SpO₂ 88% was statistically significant but much smaller in absolute terms.
• The primary result was consistent when analysis was restricted to participants with at least 12, 24 or 48 hours of valid monitoring and across prespecified subgroups, including Monk skin-tone categories. The subgroup analyses were not powered to establish true effect modification.
• No patient-centred benefit was demonstrated for respiratory failure, time to room air, oxygen-free days, hospital-free days or mortality.

Internal Validity

• Randomisation and Allocation: Allocation was generated in REDCap using variable permuted blocks and stratified by site and reason for admission. Electronic allocation after consent should have provided concealment before randomisation.
• Post-randomisation Exclusions: Of 303 randomised patients, three were excluded before study procedures began: one in the autonomous group and two in usual care. The final 300-patient analysis was therefore a modified rather than complete intention-to-treat analysis, although the exclusions were few, occurred before exposure and are unlikely to explain the primary effect.
• Performance and Detection Bias: The open-label design was unavoidable for clinicians. Repeated target reminders could improve manual titration in the control group, while awareness of receiving an autonomous device could affect tolerance, monitoring and discontinuation. The continuous electronically recorded primary outcome was considerably less susceptible to subjective observer bias than clinician-rated outcomes.
• Protocol Adherence: Median intervention exposure was similar at 42 hours (IQR 23-60) with autonomous titration and 41 hours (IQR 18-59) with usual care. However, the proportion of participants whose assigned titration strategy was followed, the number of clinician overrides and the duration of temporary protocol interruptions were not reported, despite a planned descriptive adherence analysis.
• Baseline Characteristics: Groups were broadly comparable for admission indication, ICU status, chronic lung disease, home oxygen and skin pigmentation. Chance imbalances included median age 68 versus 64 years, oxygen flow 5-10 L/min in 12% versus 18%, and current or previous smoking in 42% versus 53% for autonomous titration and usual care, respectively. The primary model adjusted for stratification factors, not these additional variables.
• Patient Severity: The cohort had meaningful acute oxygen requirements, but it was not a population with severe evolving respiratory failure: oxygen flow was limited to 1-10 L/min at entry, imminent escalation was excluded and only one-quarter were in an ICU at enrolment. This reduced event rates and the capacity to detect clinically important benefit or harm.
• Heterogeneity: The population included medical, trauma, burn and surgical patients across four centres. The adjusted improvement in normoxaemia was directionally consistent by admission diagnosis, baseline oxygen flow, chronic lung disease, home oxygen and skin pigmentation. Small subgroup sizes and wide interaction uncertainty mean that absence of demonstrated heterogeneity is not proof of identical effects in every subgroup.
• Timing and Follow-Up: Randomisation within 36 hours appropriately targeted the early phase of oxygen therapy. Only 143/300 participants completed the full 72-hour intervention; common concluding events were stable room air for at least 12 hours and hospital discharge. Restriction to participants with at least 48 hours of monitoring retained a similar adjusted RD of +23 percentage points (95% CI +18 to +28).
• Dose and Target: The algorithm was programmed to SpO₂ 92%-94% while the outcome range was 90%-96%. This narrow control band created a strong intervention capable of maximising time within the wider outcome range. The protocol allowed clinician-specified alternative targets, but the frequency and duration of such periods were not reported.
• Separation of the Variable of Interest: There was clear physiological separation: time in normoxaemia was 85% versus 63%, time in hyperoxaemia was 9.2% versus 29.1%, and time in hypoxaemia was 2.0% versus 3.6%. Oxygen volume was numerically lower at 2.31 versus 2.99 thousand litres per patient per day, but the IRR of 0.83 had a 95% CI of 0.67 to 1.02.
• Key Delivery Aspects: Both groups wore two pulse oximeters and received continuous monitoring, while control clinicians received at least daily reinforcement of the target range. This created an active, protocol-aware comparator rather than entirely unmodified routine care.
• Discontinuation and Crossover: Continuous SpO₂ monitoring was discontinued before a concluding event in 9/152 autonomous-titration patients (5.9%) and 3/148 usual-care patients (2.0%). One patient was removed from autonomous titration because of anxiety and discomfort. Formal crossover or override counts were not reported.
• Outcome Assessment: The primary endpoint was objective and minute-level, but it remained dependent on pulse-oximeter performance. Using two independent sensors was a major safeguard; the treatment estimate remained favourable with the hospital monitor alone, although it was smaller than with the device monitor.
• Statistical Rigour: The primary model appropriately estimated an absolute difference in proportions, adjusted for stratification and accounted for paired device streams. Complete primary-outcome data were available for all 300 analysed patients. Important concerns are the timing of the available SAP after the safety interim analysis, evolution from the protocol's simpler two-sample t-test to a quasi-binomial model, and the absence of multiplicity protection for secondary outcomes.

Conclusion on Internal Validity: Internal validity is moderate to strong for the direction of the effect on time spent within the prescribed SpO₂ range: randomisation, objective continuous measurement, complete primary data and consistent sensitivity analyses all support a genuine technical benefit. Confidence is lower in the exact magnitude and in any clinical-benefit claim because of the open-label design, modified intention-to-treat population, sensor-dependent effect size, incomplete adherence reporting and late finalisation of the available SAP.

External Validity

• Population Representativeness: The trial included adults with respiratory, traumatic, burn and surgical diagnoses and intentionally enrolled a broad range of skin pigmentation: 22% light, 56% medium and 22% dark by the Monk Skin Tone Scale. This is more diverse than much of the preceding device literature.
• Selection: Of 2,663 screened patients, 303 were randomised. Major exclusions included 1,455 expected to stop oxygen imminently, 252 expected to escalate respiratory support, 113 expected to leave hospital within 24 hours, 93 with device contraindications and 65 whose clinical teams could not accept the assigned strategy. The findings therefore apply to a selected, clinically stable low-flow-oxygen population rather than all hypoxaemic inpatients.
• Clinical Setting: All four hospitals were US tertiary centres. Typical nurse-to-patient ratios were 1:2 in ICU and 1:4 or 1:5 on wards, and all patients had continuous hospital pulse oximetry. Effects may differ in systems with intermittent monitoring, different staffing models or less protocolised respiratory care.
• Excluded Populations: The trial does not directly inform children, pregnant patients, prehospital care, patients requiring high-flow nasal oxygen, non-invasive or invasive ventilation, or those with rapidly increasing oxygen needs. Device contraindications also excluded several dyshemoglobinaemia and toxicological conditions.
• Target Generalisability: The common trial target of 90%-96% may not match all local policies. Widely used British Thoracic Society guidance available when SAVE-O2 AI was published recommended 94%-98% for many acutely ill adults and 88%-92% for patients at risk of hypercapnic respiratory failure, with patient-specific targets where appropriate.⁵
• Device Generalisability: Only the O2matic PRO100 was tested, and it was not cleared for routine US clinical use during the trial. It required mains electricity, had an internal battery life of up to three hours, used a proprietary sensor and could not send alarms directly to central hospital monitoring. Results cannot automatically be transferred to other algorithms, sensors or delivery systems.
• Resource-Limited Settings: Autonomous titration could conserve staff attention and oxygen, but implementation requires reliable power, oxygen supply, compatible interfaces, pulse-oximeter consumables, maintenance, training and a separate safety-monitoring pathway. The trial did not test cost-effectiveness, task burden or the number of additional patients who could be treated from a finite oxygen supply.

Conclusion on External Validity: External validity is good for clinically stable adults receiving conventional low-flow oxygen in well-resourced hospitals using a similar saturation target. Generalisability is limited for rapidly deteriorating or ventilated patients, alternative target ranges, other autonomous devices and austere environments in which the necessary monitoring and technical infrastructure may not be available.

Strengths & Limitations

• Strengths:
  • Randomised, multicentre evaluation across four geographically distinct hospitals.
  • Continuous minute-level physiological data rather than intermittent charted observations.
  • Two independent pulse-oximeter streams, permitting a clinically important sensor-specific sensitivity analysis.
  • Complete primary-outcome data for all 300 patients in the analysed trial population.
  • Broad medical and surgical case mix and direct measurement of skin pigmentation, with substantial medium- and dark-skin representation.
  • Consistent primary findings across monitoring-duration restrictions and prespecified clinical subgroups.
  • Investigator-initiated and publicly funded; devices were rented, and the manufacturer had no stated role in design, conduct, analysis or publication.
• Limitations:
  • The primary outcome was technical target attainment, not a patient-centred endpoint.
  • The sample was too small and event rates too low to assess mortality, respiratory failure or uncommon serious device harms.
  • The open-label intervention and repeated control-group target reminders created risks of performance and Hawthorne effects.
  • Only 300 of 2,663 screened patients entered the analysis, excluding patients likely to wean or deteriorate imminently.
  • The primary effect estimate varied materially by pulse-oximeter source.
  • Adherence, overrides, device downtime and protocol crossovers were incompletely reported.
  • The available SAP was dated after the safety interim analysis and altered important details of the primary analytic approach.
  • Secondary outcomes were not multiplicity-adjusted, and the status of hypoxaemia as the “key” secondary outcome was not clearly established in the earlier protocol hierarchy.
  • The findings are specific to one device, algorithm and target strategy.
  • No formal workload, economic, implementation or oxygen-supply capacity outcomes were assessed.

Interpretation & Why It Matters

• Technical Efficacy

  Autonomous oxygen titration clearly controlled SpO₂ more precisely than intermittent clinician-led titration. The 21-percentage-point adjusted improvement is large, precise and consistent with the intended closed-loop mechanism.
• Mechanism of Benefit

  The system responded continuously to pulse-oximetry feedback rather than waiting for a scheduled bedside assessment. This primarily prevented prolonged oxygen overshoot, while also modestly reducing time in hypoxaemia.
• Patient-Centred Outcomes

  The trial does not establish that better SpO₂ control improves symptoms, prevents respiratory failure, shortens admission or saves lives. Clinical implementation should therefore be justified as improved treatment precision and potentially better resource stewardship, not as proven outcome-improving therapy.
• Oxygen Stewardship

  Oxygen volume was numerically lower, but the 72-hour IRR confidence interval included no difference. The system's capacity to conserve oxygen is plausible and supported by lower delivered flows, but it was not definitively established by the prespecified total-volume result.
• Safety and Oversight

  Automation does not remove the need for clinical monitoring. A poor sensor signal led the controller to increase flow to 15 L/min in one patient, illustrating that a closed-loop device can respond precisely to an inaccurate input and therefore requires signal-quality checks, alarms and clinician override.
• Equity

  Consistent target-range improvement across measured skin-tone groups is reassuring, but an autonomous system inherits the accuracy and bias of its pulse oximeter. It regulates to the displayed SpO₂; it cannot determine whether that display accurately reflects arterial saturation.

Controversies & Other Evidence

• Physiological Process Outcome: Time within a saturation range is objective and clinically relevant, but it remains a surrogate. The accompanying commentary emphasised that SAVE-O2 AI demonstrated more precise oxygen delivery without showing fewer episodes of respiratory failure, more oxygen-free days or other patient-centred benefit.⁶
• Dominant Reduction in Hyperoxaemia: The 18-percentage-point adjusted reduction in hyperoxaemia was much larger than the 1.3-percentage-point reduction in hypoxaemia. The biological and clinical importance of avoiding transient SpO₂ values above 96% during low-flow ward oxygen remains uncertain. In mechanically ventilated ICU populations, UK-ROX, HOT-ICU and ICU-ROX achieved lower oxygen exposure without demonstrating a mortality or other major clinical benefit from conservative oxygen targets.⁷⁸⁹
• Enhanced Usual Care: Control clinicians received daily target reminders and patients underwent continuous dual-sensor monitoring. Usual care therefore performed better than in many earlier studies: control patients spent 63% of observed time within target and only 3.6% below SpO₂ 88%. This may have reduced the apparent intervention effect, but it also means the comparator was not entirely naturalistic care.
• Selection Before Randomisation: Only 303 of 2,663 screened patients were randomised. The exclusion of patients expected either to stop oxygen or escalate respiratory support imminently selected a stable middle group. The commentary noted that the low control-group hypoxaemia rate could reflect enhanced care, preferential enrolment of less unstable patients, or both.⁶
• Measurement-Controller Coupling: The estimated primary effect was +28 percentage points using the device's own sensor but +14 percentage points using the independent hospital sensor. The latter result supports a real clinical separation, yet the twofold difference shows that the measured benefit depends partly on the sensor used by the feedback loop. This is a central issue for autonomous systems: a controller can only be as accurate as its input.
• Skin Pigmentation and Pulse-Oximetry Bias: The trial's deliberate inclusion of medium- and dark-skin participants is a major advance, and the primary effect was directionally consistent across Monk groups. Nevertheless, subgroup analyses were small, arterial saturation was not used as the reference standard and the device cannot correct occult hypoxaemia caused by pulse-oximeter bias.¹⁰
• Timing of the Statistical Analysis Plan: Protocol version 3.1, dated 11 August 2025, stated that the SAP would be finalised before the first formal DSMB interim analysis. The interim analysis had already occurred on 20 July 2025; the available SAP is dated 20 August 2025 and was signed on 30 August and 2 September. The final model was statistically reasonable, but its chronology weakens the claim that all analytic details were fixed before any comparative data were reviewed.⁴
• Evolution of the Primary Analysis: The protocol synopsis specified a two-sample t test of one participant-level proportion. The later SAP and final analysis used a quasi-binomial model in which each pulse-oximeter stream contributed a separate device-level proportion, with clustered standard errors and time weighting. This method better accommodates unequal monitoring and paired sensors, but the change should be recognised when judging prespecification.
• Secondary-Outcome Hierarchy: The protocol listed oxygen volume before hypoxaemia among secondary endpoints, and the SAP stated that secondary endpoints would be presented as estimates with 95% CIs without multiplicity adjustment. The publication designated hypoxaemia as the “key secondary outcome” and reported P=0.002. Its confidence interval excludes no effect, but it was not protected by a prespecified hierarchical testing procedure and should be interpreted as supportive rather than independently confirmatory.
• Informative Censoring: Only 143/300 patients completed 72 hours. Discharge occurred in 37/152 autonomous patients versus 24/148 controls, and early discontinuation of continuous monitoring occurred in 9/152 versus 3/148. The primary outcome therefore describes the proportion of observed, eligible time in range rather than a common fixed 72-hour interval. Monitoring-duration sensitivity analyses were reassuring but cannot remove every potential consequence of event-dependent follow-up.
• Unreported Adherence: The protocol planned to report the proportion of patients for whom the assigned treatment strategy was followed. The absence of this measure, clinician-override counts and device-downtime data limits assessment of how faithfully the autonomous and manual strategies were separated in practice.⁴
• Power-Calculation Inconsistency: The main publication and supplemental methods stated that 300 participants provided greater than 99% power to detect a 10-percentage-point difference, assuming an outcome SD up to 30 percentage points. The SAP table gives 99.1% power at an SD of 20 percentage points but only 82.1% at an SD of 30 percentage points. This does not threaten the highly significant primary result, but the stated power assumptions are internally inconsistent.
• Reporting Inconsistencies: The protocol described randomisation block sizes of 4-8, whereas the supplemental methods reported sizes of 2, 3 and 4. The main text and a post hoc supplemental table placed the ≥1 L/min daily titration counts in opposite treatment groups. The same post hoc table labelled a first-24-hour time-to-room-air result in “days” despite values of 9.9 and 13.4 and displayed a negative value labelled as an incidence rate ratio. These issues do not overturn the primary result but warrant correction.
• Consistency with Previous Closed-Loop Evidence: SAVE-O2 AI's adjusted +21-percentage-point target-range effect is close to the pooled +25.47-percentage-point effect in the 2024 meta-analysis.¹ Earlier perioperative work also favoured automated titration, and a 2026 emergency-department trial of automated high-flow oxygen found time in target of 96.4% versus 89.9%, an 8.0-percentage-point difference (95% CI 1.7 to 16.9; P=0.01).¹¹¹²
• Safety Ascertainment: Adverse events were collected only when serious or unexpected and considered definitely, probably, possibly or uncertainly related to trial procedures. Routine respiratory deterioration and death were recorded as clinical outcomes unless judged intervention-related. “No serious adverse events” should therefore be interpreted as absence of reported serious trial-related events, not as exhaustive capture of every clinical complication.
• Implementation and Economics: The trial did not measure nursing time, alarm burden, device failures, acquisition and maintenance costs, or the number of patients supportable from a fixed oxygen supply. These outcomes may ultimately determine whether the technology offers sufficient value for routine deployment, particularly outside high-resource hospitals.⁶
• Subsequent Evidence: SAVE-O2 AI was published only days before this summary. No independent confirmatory trial or post-publication guideline had yet incorporated its results; the immediate evidence base therefore consists of the index trial, its invited commentary and preceding closed-loop studies.

Summary

• SAVE-O2 AI randomised 303 adults receiving 1-10 L/min of supplemental oxygen at four US hospitals; 300 were analysed after three pre-intervention exclusions.
• Autonomous titration increased mean time at SpO₂ 90%-96% from 63% to 85%: adjusted RD +21 percentage points; 95% CI +18 to +25; P<0.001.
• Time below SpO₂ 88% decreased from 3.6% to 2.0%: adjusted RD -1.3 percentage points; 95% CI -2.0 to -0.5; P=0.002, while time above 96% decreased from 29.1% to 9.2%.
• The trial showed no clear difference in total oxygen volume, time to room air, oxygen-free days, respiratory failure, hospital-free days or mortality and was underpowered for these clinical outcomes.
• The primary technical result is convincing, but the exact effect size was sensor-dependent, the available SAP post-dated the safety interim analysis, adherence reporting was incomplete and applicability is limited to stable low-flow-oxygen patients.

Overall Takeaway

SAVE-O2 AI provides strong proof that autonomous closed-loop titration can deliver low-flow hospital oxygen far more precisely than intermittent manual adjustment, chiefly by preventing hyperoxaemia and with a smaller reduction in hypoxaemia. It is an important device-efficacy trial rather than definitive evidence of patient benefit: routine adoption should depend on independent replication, robust sensor and alarm safeguards, workload and cost evaluation, and trials powered for clinical outcomes.

Bibliography

• 1.Mól CG, Vieira AGS, Garcia BMSP, dos Santos Pereira E, Eid RAC, Schultz MJ, et al. Closed-loop oxygen control for critically ill patients: a systematic review and meta-analysis. PLoS One. 2024;19(6):e0304745.
• 2.Douin DJ, Rice JD, Anderson EL, et al; Strategy to Avoid Excessive Oxygen (SAVE-O2) Investigators. Targeted normoxemia and supplemental oxygen-free days in critically injured adults: a stepped-wedge cluster randomized clinical trial. JAMA Netw Open. 2025;8(3):e252093.
• 3.Douin DJ, Rice JD, Xiao M, et al; Strategy to Avoid Excessive Oxygen (SAVE-O2) Investigators. Effect of targeting normoxemia on supplemental oxygen-free days for adults with acute thermal burns: a stepped wedge cluster randomized clinical trial. J Trauma Acute Care Surg. 2025;99(4):619-627.
• 4.Douin DJ, Rice JD, Xiao M, Ginde AA; SAVE-O2 AI Investigators. Statistical analysis plan for the Strategy to Avoid Excessive Oxygen using Autonomous Oxygen Titration Intervention (SAVE-O2 AI) trial: protocol. BMJ Open. 2026;16(1):e110739.
• 5.O'Driscoll BR, Howard LS, Earis J, Mak V; British Thoracic Society Emergency Oxygen Guideline Development Group. British Thoracic Society guideline for oxygen use in adults in healthcare and emergency settings. BMJ Open Respir Res. 2017;4(1):e000170.
• 6.Weingart MF, Matthay MA. Potential transformative impact of autonomous oxygen delivery in hospitalized patients. JAMA Intern Med. Published online August 3, 2026.
• 7.Martin DS, Gould DW, Shahid T, et al; UK-ROX Investigators. Conservative oxygen therapy in mechanically ventilated critically ill adult patients: the UK-ROX randomized clinical trial. JAMA. 2025;334(5):398-408.
• 8.Schjørring OL, Klitgaard TL, Perner A, et al; HOT-ICU Investigators. Lower or higher oxygenation targets for acute hypoxemic respiratory failure. N Engl J Med. 2021;384(14):1301-1311.
• 9.Mackle D, Bellomo R, Bailey M, et al; ICU-ROX Investigators and the Australian and New Zealand Intensive Care Society Clinical Trials Group. Conservative oxygen therapy during mechanical ventilation in the ICU. N Engl J Med. 2020;382(11):989-998.
• 10.Shachar C, Drabo EF, Iwashyna TJ, Ferryman K. Addressing racial and ethnic bias in pulse oximeters: a wicked problem. JAMA. 2025;333(7):563-564.
• 11.L'Her E, Jaber S, Verzilli D, et al. Automated closed-loop versus standard manual oxygen administration after major abdominal or thoracic surgery: an international multicentre randomised controlled study. Eur Respir J. 2021;57(1):2000182.
• 12.Kirton L, Cook J, Hamill R, et al. Automated titration of nasal high flow oxygen in the emergency department: a randomised controlled trial. Emerg Med J. 2026;43(4):215-222.