Statistical analysis plan for the Strategy to Avoid Excessive Oxygen using Autonomous Oxygen Titration Intervention (SAVE-O2 AI) trial: protocol

Abstract

Introduction

Administering supplemental oxygen to prevent hypoxaemia is a fundamental treatment for patients hospitalised with acute injury or illness. However, the amount of oxygen administered frequently exceeds that needed to maintain normoxaemia, causing patients to experience hyperoxaemia and wasting supplemental oxygen. Closed-loop, autonomous oxygen titration systems are designed to optimise oxygen delivery by administering the lowest possible oxygen flow that maintains peripheral oxygen saturation (SpO₂) within a predefined range. For adults hospitalised with an acute injury or illness, it remains uncertain whether the use of a closed-loop, autonomous oxygen titration system safely increases the proportion of time spent in normoxaemia (SpO₂ 90%–96%) compared with usual care.

Methods and analysis

The Strategy to Avoid Excessive Oxygen using Autonomous Oxygen Titration Intervention trial is a multicentre, unblinded, parallel-group, randomised trial being conducted at four level 1 trauma centres in the USA. The trial compares an autonomous oxygen titration system versus usual care among 300 adults hospitalised for major trauma, burn, acute care surgery or acute respiratory illness. The primary outcome is the proportion of patient-time spent within the targeted normoxaemia range (SpO₂ 90%–96%) as measured by continuous non-invasive pulse oximetry, during the first 72 hours after randomisation. Secondary outcomes include the amount of supplemental oxygen administered and the proportion of time spent in hypoxaemia (SpO₂<88%) and hyperoxaemia (SpO₂ >96%). Specifying the protocol and statistical analysis plan before the conclusion of enrolment increases the rigour, reproducibility and interpretability of the trial. Enrolment began on 6 May 2024.

Ethics and dissemination

The trial protocol was approved by the single institutional review board at the University of Colorado School of Medicine and the Office of Human Research Oversight at the Department of Defense. We will present the results at scientific conferences and submit them for publication in a peer-reviewed journal.

Trial registration number

NCT06374225.

STRENGTHS AND LIMITATIONS OF THIS STUDY.

• This protocol describes in detail the design and methods for a multicentre, unblinded, randomised clinical trial comparing the use of a closed-loop, autonomous oxygen titration system to usual care among adults hospitalised with an acute injury or illness.

• Outcome assessments are performed by physicians and statisticians who are blinded to group allocation.

• The trial is conducted at multiple institutions in the USA, reflecting the characteristics of the healthcare system there.

• Open-label design is a potential limitation.

Introduction

Ensuring adequate oxygenation is a primary goal in surgical and medical patients to treat and prevent morbidity associated with hypoxaemia.^{1 2} However, the administration of excessive supplemental oxygen resulting in hyperoxaemia, defined as peripheral oxygen saturation (SpO₂) values greater than 96%, is common in hospitalised patients.3,6 Excessive oxygen administration often leads to the unnecessary use of supplemental oxygen, a limited resource in austere settings.⁷ Our previous Strategy to Avoid Excessive Oxygen (SAVE-O2) Clinical Trials demonstrated that targeting normoxaemia, defined as SpO₂ 90%–96%, safely reduced utilisation of supplemental oxygen without increasing hypoxaemia (SpO₂<88%), and maintained clinical outcomes among critically ill⁸ and major burn⁹ patients.

Recent publications have explored the use of a closed-loop, autonomous oxygen titration system for mechanically ventilated patients.10,12 Initial results indicate that an autonomous oxygen titration system is reliable and cost-effective.^{10 11 13} However, most patients who receive supplemental oxygen do not receive mechanical ventilation. They receive oxygen at lower flow rates through a nasal cannula or face mask.14,17 Limited evidence exists on the use of an autonomous oxygen titration system for non-mechanically ventilated acutely ill patients.

The O2matic PRO100 (Herlev, Denmark) is a closed-loop, autonomous oxygen titration system that is approved for clinical use in 32 countries, including in the European Union, Australia and New Zealand. However, clinical trials have not yet been conducted in the USA, and the device has not received US Food and Drug Administration (FDA) approval.¹⁸ The PRO100 system is specifically designed for patients receiving supplemental oxygen via nasal cannula or face mask. Using input on SpO2 values from a pulse oximeter, the device applies an algorithm to automatically adjust the oxygen flow rate between 0 and 15 L/min to maintain the patient’s SpO₂ within a provider-specified range. The device titrates oxygen flow to deliver the lowest flow that will maintain the oxygen saturation within the target range. Preliminary clinical evidence demonstrated that an autonomous oxygen titration system reduced hypoxaemia and hyperoxaemia in patients with COVID-19 pneumonia and hypoxaemic respiratory failure,¹⁷ as well as inpatients and outpatients with chronic obstructive pulmonary disease (COPD).^{19 20} However, the safety and efficacy of using an autonomous oxygen titration system for hospitalised adults has never been evaluated in the USA in a randomised trial. As a result, autonomous oxygen titration systems are not approved for use in current clinical care.

To address this uncertainty, we designed the Strategy to Avoid Excessive Oxygen using Autonomous Oxygen Titration Intervention (SAVE-O2 AI) trial, to determine the effectiveness of using an autonomous oxygen titration system to maintain normoxaemia (SpO₂ 90%–96%) and reduce the incidences of hypoxaemia and hyperoxaemia during the first 72 hours in adults hospitalised with an acute injury or illness who are receiving supplemental oxygen. We hypothesise that the autonomous oxygen titration system will increase the time spent in normoxaemia compared with usual care.

Methods and analysis

This article was written in accordance with the Standard Protocol Items: Recommendations for Interventional Trials guidelines (online supplemental appendix).²¹

Patient and public involvement

Materials used to communicate details of the study with patients and family members were developed with input from the Colorado Clinical and Translational Sciences Institute. Study authors will disseminate the results of this study online and via social media in forms suitable for public understanding.

Study design

SAVE-O2 AI is a multicentre, parallel-group, randomised controlled trial at four US level 1 trauma and tertiary care hospitals to determine the effectiveness of using an autonomous oxygen titration system to maintain normoxaemia during the first 72 hours of randomisation for acute injury or illness compared with usual care with manual oxygen titration (table 1). A 72-hour intervention duration is consistent with previous clinical studies of autonomous oxygen titration.²² The primary outcome is the proportion of time spent in normoxaemia (SpO₂ 90%–96%) during the first 72 hours after randomisation. The trial was registered prior to initiation of enrolment (ClinicalTrials.gov identifier: NCT06374225) and is being conducted under a US FDA Investigational Device Exemption (IDE) (G230325). The University of Colorado serves as the Clinical Coordinating Centre (CCC) for this trial.

Table 1. Schedule of enrolment, interventions and assessments in the SAVE-O2 AI trial.

Day	0	1	2	3	28 or hospital discharge
Eligibility and baseline data
Assess eligibility	X
Informed consent	X
Baseline data collection	X
Study intervention data
Randomisation	X
Implementation of study intervention	X	X	X	X
Device data	X	X	X	X
Hospital data
Oxygenation data	X	X	X	X	X
Secondary outcomes and safety assessments	X	X	X	X	X
Adverse event and unanticipated problem monitoring	X	X	X	X	X
Hospitalisation summary					X

Table entries without data mean that data elements were not collected at that time.

SAVE-O2 AI, Strategy to Avoid Excessive Oxygen using Autonomous Oxygen Titration Intervention.

Study population and inclusion/exclusion criteria

Adults who are hospitalised from the emergency department (ED) with major trauma, burns, acute care surgery or acute respiratory illness and who are able to be randomised within 36 hours of hospitalisation are potentially eligible. Acute respiratory illness is defined as a primary or secondary lung insult resulting in hypoxaemia, such as respiratory infection, pulmonary embolism, pleural effusion, pulmonary oedema, COPD exacerbation, asthma exacerbation, pulmonary contusion, aspiration or other cause of acute hypoxaemic respiratory insufficiency or failure. Altered mental status or sedation as the primary cause of hypoxaemia is not considered to be an acute respiratory illness. Patients who were anticipated to be discharged from the hospital within 24 hours or with imminent plans to discontinue supplemental oxygen are excluded. Likewise, patients with imminent plans to escalate to high-flow nasal oxygen, non-invasive ventilation or invasive mechanical ventilation are also excluded. Complete lists of inclusion and exclusion criteria are provided in box 1.

Box 1. Inclusion and exclusion criteria.

Inclusion criteria

Age 18 years or older.

Hospitalised or will be hospitalised from emergency department for major trauma, burn, acute care surgery or acute respiratory illness.

Able to be randomised within 36 hours of hospital arrival.

Receiving supplemental oxygen 1–10 L/min for documented or presumed hypoxaemia (must be higher than baseline for those on chronic oxygen therapy).

Signed and dated informed consent from patient or legally authorised representative.

Exclusion criteria

Anticipated hospital discharge within 24 hours.

Imminent plans to discontinue supplemental oxygen.

Imminent plans to administer high-flow nasal oxygen, non-invasive ventilation or invasive mechanical ventilation.

The clinical team is unwilling or unable to follow the prescribed oxygen titration method in either randomised group.

Known prisoner.

Known pregnancy.

Known contraindicated conditions for use of the autonomous oxygen titration system:

• Carbon monoxide poisoning.

• Incapable of handling airway secretions.

• Increased methaemoglobin.

• Cyanide poisoning.

• Cluster headaches.

• Undrained pneumothorax.

• Sickle cell crisis.

• Paraquat poisoning.

• History of bleomycin poisoning.

• Patients for whom the peripheral oxygen saturation signal is not stable.

Assessment of eligibility

Study personnel screen potential patients for eligibility via the electronic health record (EHR) at each site. If the patient appears to meet all eligibility criteria, as confirmed by the site investigator, the site investigator or research coordinator approaches the treating clinical team either in person, by phone, or via EHR secure chat, to ask permission to approach the patient or legally authorised representative (LAR). The clinical team agrees to the protocol, specifically the targeted normoxaemia SpO₂ range. Patients are then approached by study personnel to obtain written informed consent before initiating any study activities.

Process of obtaining informed consent

The patient and/or their LAR are provided with a written informed consent form describing the study interventions, study procedures and risks in appropriate detail. Patients/LARs have the opportunity to carefully review the written informed consent form and ask questions before signing. All participants can discuss the study with their family or surrogates before agreeing to participate. We have fully translated consent forms for commonly encountered languages at enrolling sites, and short-form consents for less commonly encountered languages. A qualified interpreter is used for all non-English languages. Use of an interpreter and the interpreter’s identity are documented on the paper/electronic consent. The patient/LAR must sign the informed consent document before any study procedures are initiated. All enrolled patients receive a copy of the informed consent document for their records. Further details on the consent process can be found in the online supplemental appendix.

We are using both a traditional paper consent process and a 21 CFR Part 11 compliant electronic consent (eConsent) process in Research Electronic Data Capture (REDCap). The consent procedures outlined above can be completed either via eConsent or paper consent.

Randomisation and treatment allocation

Enrolled patients are randomised in a 1:1 ratio to receive intervention (autonomous oxygen titration system) vs usual care (manual oxygen titration) in permuted blocks of variable size, stratified by study site and patient subgroup (trauma/burn/surgical vs acute respiratory illness). Employing permuted block randomisation ensures balance across the intervention and control groups at regular enrolment intervals within each site, thereby reducing the risk of unequal allocation across groups within site and patient type.

Study interventions

Interventions in both groups

All enrolled participants in both trial groups have a target SpO₂ of 93% with an acceptable range of 90%–96%.⁸ All other aspects of clinical care, including frequency of vital sign assessments, level of care, choice of oxygen delivery device (eg, nasal cannula or face mask) and escalation to high-flow oxygen or mechanical ventilation, are determined by the primary clinical team, per usual care. The recommended maximum oxygen flow rate is 8 L/min for standard (non-humidified) nasal cannula and 15 L/min for humidified nasal cannula and face mask.

Patients in both groups may receive nocturnal non-invasive positive pressure ventilation if they receive this therapy at baseline. Oxygen is titrated by the clinical team during nocturnal positive pressure ventilation. Patients receiving nocturnal positive pressure ventilation remain connected to continuous pulse oximeters from both the hospital monitor and the autonomous oxygen titration system. In the morning, oxygen titration is resumed based on the assigned group.

Usual care (manual titration) group

Patients randomised to the usual care group receive usual care, in which supplemental oxygen is titrated by clinical staff per hospital protocol at each site. In the usual care group, the patient is connected to the hospital’s continuous pulse oximeter and the autonomous oxygen titration system’s pulse oximeter in observation mode (ie, autonomous titration is disabled and the autonomous system is not used for titration decisions) to allow for data collection. Since oxygen titration in the usual care group is at the discretion of the treating clinician, bedside nurse or respiratory therapist, we anticipate some variation between the four sites. The recommended SpO₂ target, or 93%, and the acceptable range of 90%–96% are communicated regularly (at least daily) to the clinical team during the 72-hour study period. This daily communication should improve the harmonisation of usual care between sites.

Intervention (autonomous oxygen titration) group

Patients randomised to the intervention group receive supplemental oxygen via nasal cannula or face mask that is titrated by a closed-loop, autonomous oxygen titration system for the first 72 hours after randomisation, or until hospital discharge, whichever occurs first. The study team initially programmes the autonomous oxygen titration system to provide a maximum oxygen flow rate of 8 L/min if a low-flow, non-humidified nasal cannula is being used. Humidification is used at the discretion of the clinical team if the oxygen flow rate is ≤8 L/min. If the patient receives >8 L/min, the autonomous oxygen titration system is programmed to provide a maximum oxygen flow rate of 15 L/min. The patient then receives oxygen via face mask or humidified oxygen through a large-bore nasal cannula in order to reduce unfavourable outcomes related to administration of non-humidified high-flow nasal oxygen. The patient is monitored and vital signs documented using routine clinical SpO₂ assessments; however, oxygen titration occurs automatically through the autonomous oxygen titration system during the intervention period.

The recommended SpO₂ target is 93% with a range programmed into the autonomous oxygen titration system at 92%–94%. Specifically, the system targets the lower programmed SpO₂ value +1 (ie, 93% in this case). The acceptable SpO₂ range for the SAVE-O2 AI trial is 90%–96%. If a patient escalates to receive >15 L/min or develops other clinical signs of advancing respiratory failure, they are taken off the autonomous oxygen titration system and transitioned to higher flow oxygen devices or mechanical ventilation per usual clinical care. If the patient is not receiving any supplemental oxygen for at least 12 hours per the device’s autonomous titration, the clinical team may remove the oxygen delivery device (eg, nasal cannula) from the patient, but the patient remains connected to the pulse oximeter of the autonomous oxygen titration system for the duration of the intervention period for data collection and to monitor for new supplemental oxygen needs.

Autonomous oxygen titration system information

The PRO100 is a trend-based, automated oxygen titration system that is lightweight (1.85 kg) and operates on AC power, with a battery life of up to 3 hours. The system is specifically designed for patients receiving supplemental oxygen via nasal cannula or face mask. It is compatible with most commercially available portable and standard oxygen systems. The autonomous oxygen titration system allows providers to set a specific SpO₂ range, with a minimum allowable range of 2% between minimum and maximum values. The system then continuously and automatically titrates the flow of supplemental oxygen, based on closed-loop principles, to achieve the targeted SpO₂. Safety alarms are set to notify providers based on specific SpO₂ and oxygen volume parameters. For additional system specifications, see the Study Protocol in the online supplemental appendix.

Assessment of skin pigmentation

In 2024, the FDA recommended directly measuring skin pigmentation when evaluating pulse oximeters.²³ Therefore, the SAVE-O2 AI trial directly measures skin pigmentation for all enrolled patients in both groups. Skin pigmentation is independently assessed by two trained research staff at the bedside with input from the patient. Specifically, we employ two validated scales (Fitzpatrick Skin Scale²⁴ and Monk Skin Tone Scale²⁵ and an objective spectrophotometer (Nix Spectro 2 Nix Sensor, Hamilton, Ontario, Canada).^{26 27} Patients are directly involved in adjudicating their Fitzpatrick and Monk values. The Monk Skin Tone Scale is the FDA-preferred method for measuring skin pigmentation, based on guidance published in January 2025.²⁸

All Nix Spectro 2 measurements are standardised across sites. First, natural sunlight is used through a window (hospital lights off) whenever possible. Next, a trained investigator scans the midpoint of the patient’s inner forearm (unless an apparent injury or deformity exists in that area). Finally, the research staff performs a visual quality assessment to ensure the Nix Spectro 2 is producing a skin tone similar to the patient’s. If a gross discrepancy is noted, the scan is repeated until it appears accurate. We record Hex and CIELAB values²⁹ from the Nix Spectro 2, which are used to calculate skin pigmentation via the individual topography angle (ITA).³⁰ ITA was explicitly recommended by the FDA in 2025 as the preferred objective skin pigmentation measurement.²⁸ For additional Nix device specifications, see the Trial Protocol in the online supplemental appendix.

Primary outcome

The primary outcome is proportion of time spent within the targeted normoxaemia range, defined as SpO₂ 90%–96%, as measured by continuous non-invasive pulse oximetry during the first 72 hours after randomisation, with observation censored at hospital discharge, escalation to high flow nasal oxygen/mechanical ventilation, transition to room air for at least 12 hours, or death if prior to 72 hours. Periods of time spent in the operating room, on non-invasive ventilation for sleep disordered breathing, or under procedural sedation are excluded.

Secondary outcomes

Secondary outcomes include (1) total volume of supplemental oxygen administered during the first 72 hours after randomisation, (2) proportion of time spent in hypoxaemia (SpO₂<88%) during the first 72 hours after randomisation, (3) proportion of time spent in hyperoxaemia (SpO₂>96%) during the first 72 hours after randomisation and (4) time to room air. Detailed definitions for all secondary outcomes can be found in table 2.

Table 2. Study outcomes.

Primary outcome	Proportion of time spent within the targeted normoxaemia range, defined as SpO2 90%–96% (target 93%), as measured by continuous non-invasive pulse oximetry, during the first 72 hours after randomisation.
Secondary outcomes	Amount of supplemental oxygen administered, defined as total estimated oxygen volume during the first 72 hours after randomisation. Proportion of time spent in hypoxaemia (SpO2<88%) during the first 72 hours after randomisation. Proportion of time spent in hyperoxaemia (SpO2>96%) during the first 72 hours after randomisation. Time to room air, defined as the time from hospital presentation to the first episode of no supplemental oxygen (room air), censored at discharge or death.
Exploratory outcomes	Supplemental oxygen free days, defined as number of days alive and not receiving supplemental oxygen during the index hospitalisation, censored at hospital discharge. Patients who are discharged without supplemental oxygen or with prior volume of baseline (prehospitalisation) home oxygen are assumed to remain without additional supplemental oxygen to day 28. Patients who are discharged on supplemental oxygen, if above any prior baseline home oxygen therapy, are assumed to have additional supplemental oxygen to day 28. All-cause mortality to day 28, defined as a dichotomous vital status (survived or died) at hospital discharge or day 28, whichever is first. Respiratory failure, defined as receipt of high flow nasal oxygen, non-invasive ventilation or invasive mechanical ventilation, to day 28 or hospital discharge, whichever is first, excluding time in the operating room. Hospital-free days to day 28, defined as the number of days alive and outside the hospital between the initial ED visit and 28 days later, censored at hospital discharge. This outcome is a composite of in-hospital mortality and length of hospital stay and has a range of 0 days (worst outcome) to 28 days (best outcome). Patients discharged to any location prior to day 28 are assumed to survive to day 28 and will be assigned 28 minus hospital length of stay hospital-free days. Duration of time receiving no supplemental oxygen (room air) during the 72 hours after randomisation.

ED, emergency department; SpO2, peripheral oxygen saturation.

Exploratory outcomes

Table 2 reports the primary, secondary and exploratory outcomes. Patient-centred outcomes, such as hospital-free and ICU-free days through day 28 and all-cause mortality, will be reported as exploratory outcomes.

Protocol deviations

Protocol deviations are defined as any non-compliance with the clinical trial protocol, whether by the participant, the investigator or study site staff. We anticipate crossover between study groups or changes to the settings of the autonomous oxygen titration system by the clinical team may occur. The site investigator will identify and report protocol deviations within five working days of their identification.

During routine care, oxygen monitoring and administration devices may be intermittently removed by patients or the clinical team during transitions of care. Therefore, we will not consider the following to be protocol deviations: unintentional removal of the autonomous oxygen titration system or pulse oximeter(s) resulting in gaps in SpO₂ monitoring for <8 hours per episode and <24 hours cumulatively during the 72-hour intervention period. These episodes will be recorded but will not be reported as protocol deviations. Detailed procedures for handling protocol deviations are described in the Trial Protocol in online supplemental appendix.

Data management and statistical analysis plan

Data collection and management

The Data Coordinating Centre (DCC) at Vanderbilt University Medical Center automatically extracts variables directly from the EHR at each institution, including all recorded SpO₂, FiO₂ and oxygen flow rate measurements.⁸ Oxygen flow rate and SpO2 measurements for all patients are also extracted directly from the autonomous oxygen titration system. Variables that require manual chart extraction by site coordinators (ie, home supplemental oxygen use and method of hospital arrival) are minimised and obtained using standardised operating procedures with flow charts and detailed procedures for extraction.

The SAVE-O2 AI trial uses a specialised, 21 CFR Part 11-compliant REDCap database. REDCap is a secure, encrypted, HIPAA-compliant web application specifically designed for research data management and export of de-identified data for statistical analysis.³¹ Essential compliance features, such as logging, electronic signatures, eConsent and encrypted file storage, are validated and audit-ready. Data are entered and uploaded directly into the DCC’s compliant REDCap database. The DCC assists with automated feedback reports and shares these with the CCC biostatistical core. Identifiable information for each site is only accessible by the local sites, respectively, and the DCC. The CCC and biostatistics core can access a limited dataset from all sites, which includes dates. Sites are intermittently audited by the CCC without notice to ensure both protocol compliance and data accuracy.

Statistical analysis principles

We will report descriptive statistics for all baseline characteristics stratified by treatment group, including mean and SD for continuous variables and frequency and percent for categorical variables.

We will analyse primary and secondary endpoints using a regression modelling framework, with specific distributions chosen depending on the type of outcome (eg, binary, count, ordinal, time-to-event). The primary analysis will be as-randomised (ie, intention to treat), so the primary exposure variable in these regression models will be a binary indicator of randomised treatment group (equal to 1 for experimental and 0 for control). A threshold of $p=0.05$ will be considered statistically significant and will be reported for the primary endpoint. We will report other secondary and exploratory endpoints as effect estimates with 95% CIs.

Sample size estimation

Based on prior trials in the same settings,^{8 9 32} we estimated the mean proportion of time spent in the targeted normoxaemia range (primary outcome) in the usual care (control) group would be 60%. Our patient and clinician partners recommended that the trial should have adequate statistical power to detect an absolute difference of at least 10 percentage points between groups in time spent in normoxaemia. This difference is more conservative than the 32 percentage point difference (95% CI 26 to 37) observed in a previous trial of autonomous oxygen titration.³² We used a two-sample t-test with α=0.05 to calculate the predicted power with various estimations of effect size. We found that our estimated sample size of 300 patients (150 per group) allows us to detect a 10% difference in time spent in normoxaemia (70% in the intervention group and 60% in the control group) with >99% power. This sample size also provides >80% power to detect differences between subgroups as small as 60 patients (20% of the total sample size).

Data and safety monitoring board

The composition and responsibilities of the Data and Safety Monitoring Board (DSMB) are described in the online supplemental appendix.

Interim analysis

The DSMB will review a single formal interim analysis, prepared by the unblinded study biostatistician at the anticipated halfway point of the trial after enrolment of 150 patients. The purpose of this analysis is to evaluate whether the intervention group demonstrates evidence of an increased incidence of hypoxaemia, the primary safety outcome for the study. We are testing the null hypothesis that there is no difference in the proportion of time spent in hypoxaemia between the two treatment groups using a one-sided test with α=0.01. We are analysing the incidence of hypoxaemia for the interim analysis using the same approach as the final primary analysis (see Primary outcome analysis section). The interim analysis is conducted for safety monitoring only and is not intended to evaluate efficacy or futility. Early stopping for efficacy is not planned, as continued enrolment is needed to ensure adequate representation of prespecified subgroups and robust estimation of secondary outcomes. See the full statistical analysis plan (SAP) for additional details. The DSMB has the authority to recommend that the trial stop at any point, request additional data, request additional interim analyses or request modifications to the study protocol.

Primary outcome analysis

The primary outcome, proportion of time spent in normoxaemia (SpO₂ 90%–96%), will be analysed using an intention-to-treat comparison between randomised groups. We will fit a quasi-binomial generalised linear model with an identity link to estimate the absolute difference in mean proportions between the intervention and usual care groups. The model will include fixed effects for study site and patient subgroup (trauma/burn/surgical vs medical) and will use cluster-robust SEs to account for repeated measures.

Each pulse oximeter (hospital monitor and autonomous oxygen titration system) contributes a separate device-level observation. If both devices are active, both contribute; no direct averaging is performed. The interpretation of treatment coefficient, however, represents the marginal/average effect between devices. Because the outcome is modelled as a proportion, and the quasi-binomial framework includes a variance function that depends on the total time observed, this approach accommodates both overdispersion and variation in follow-up time. Participants with longer monitoring contribute more information to the model, effectively giving greater weight to more precise estimates. The identity link ensures the treatment effect is interpreted as a difference in mean proportions. We will report 95% CIs and a two-sided p value; statistical significance will be defined as p<0.05. A rationale for the primary outcome and its analysis is provided in the SAP. All analyses will follow reproducible research principles using R (R Foundation for Statistical Computing, Vienna, Austria).

Additional analyses of the primary outcome

Sensitivity analyses

We will assess the robustness of the findings of the primary analysis through a series of prespecified sensitivity analyses that explore alternative definitions and modelling assumptions. First, we will repeat the primary analysis after restricting to patients with increasing thresholds of valid SpO₂ data availability (eg, ≥12, ≥24 and ≥48 hours of monitoring). Second, we will repeat the primary analysis using only SpO₂ data from the hospital pulse oximeter. Third, we will repeat the primary analysis using only data from the autonomous oxygen titration system. Fourth, we will examine the sensitivity of the treatment effect estimate to assumptions about the mean–variance relationship in the quasi-binomial model, including potential alternative dispersion structures. Details on sensitivity analyses can be found in the SAP.

Analyses of heterogeneity of treatment effect

We will apply the three complementary approaches to analyse the heterogeneity of treatment effect described in the Predictive Approaches to Treatment Effect Heterogeneity (PATH) Statement:³³ (1) traditional one-variable-at-a-time subgroup analyses; (2) a risk-modelling approach and (3) an effect-modelling approach.

For the traditional subgroup analyses, we will assess whether prespecified baseline variables modify the effect of treatment on the primary outcome using formal interaction terms in regression models. Subgroups include race/ethnicity (non-Hispanic White, non-Hispanic Black, Hispanic and other), skin pigmentation (Monk Skin Tone Scale, categorised as light, medium and dark), home supplemental oxygen use (Yes, No), oxygen flow rate at randomisation (1–2.9 L/min, 3–4.9 L/min and 5–10 L/min), chronic pulmonary diseases (defined as the presence of chronic pulmonary disease on the Elixhauser comorbidity index)³⁴ and primary indication for hospitalisation (categorised as trauma/burn/surgical or acute respiratory illness). For categorical variables, we will report the absolute difference in means and associated 95% CIs within each subgroup. Details on subgroup definitions and hypothesised effect modification directions are provided in the SAP.

Analyses of secondary and exploratory outcomes

Continuous and proportional secondary outcomes will be analysed using methods appropriate to the scale and distribution of each variable. Proportional outcomes (including the proportion of time spent in hyperoxaemia, hypoxaemia and without supplemental oxygen) will be analysed using quasi-binomial generalised linear models with an identity link, consistent with the primary outcome analysis. This approach accounts for overdispersion and differing follow-up durations across patients. Skewed continuous outcomes defined over the positive real line, such as the total amount of supplemental oxygen administered, will be analysed using a quasi-Poisson regression model with a log link, including an offset for follow-up time (capped at 72 hours) to account for unequal observation periods. Event-based binary outcomes (eg, all-cause 28-day mortality, respiratory failure) will be analysed as time-to-event outcomes using a competing risks framework, with death and discharge treated as competing events, as appropriate. We will use cause-specific Cox proportional hazards models to estimate treatment effects, with time zero defined as the time of randomisation. Kaplan-Meier plots and cumulative incidence functions will be used to visualise differences between groups. For ordinal outcomes such as supplemental oxygen-free days to day 28, we will use a proportional odds logistic regression model to estimate treatment effects on the cumulative probability scale.

Handling of missing data

Missing primary outcome data are anticipated due to interruptions in device monitoring, variability in data transmission or other operational factors. All participants with any valid SpO₂ data from at least one pulse oximeter (autonomous oxygen titration system or hospital monitor) will be included in the analysis. The quasi-binomial model accounts for differing amounts of data by incorporating the total number of observed minutes per device in its variance function. We will not perform imputation for missing outcome data. The primary model uses a quasi-likelihood estimation framework, and inference will rely on cluster-robust SEs, which remain valid under standard regularity conditions even in the presence of unmodelled heteroskedasticity or within-participant correlation.³⁵ Participants with no valid SpO₂ data from either device during the first 72 hours will be excluded from the primary outcome analysis.

Trial status

The SAVE-O2 AI trial is a prospective, multicentre, unblinded, parallel-group, randomised trial comparing the use of a closed-loop, autonomous oxygen titration system with usual care among adults hospitalised for major trauma, burn, acute care surgery or acute respiratory illness. Patient enrolment began on 6 May 2024 and is being conducted at four level 1 trauma centres in the USA.

Ethics and dissemination

The SAVE-O2 AI trial was approved by the Colorado Multiple Institutional Review Board (COMIRB), which serves as the sIRB for the trial (COMIRB Number: 23-0866), and the US FDA (IDE G230325). The trial also received second-level human subjects review from the Office of Human Research Oversight with the Department of Defense, Defense Health Agency.

The study is being conducted in compliance with ethical principles from the Declaration of Helsinki, in accordance with the International Council for Harmonisation (ICH) Harmonised Guideline for Good Clinical Practice. The study fully conforms with Regulations for the Protection of Human Subjects of Research (45 CFR Part 46, 21 CFR Part 50, 21 CFR Part 56, and ICH E6).

Dissemination plan

Trial results will be submitted to a peer-reviewed journal and will be presented at one or more scientific conferences.

This trial was investigator-initiated. PRO100 devices were rented from O2matic (Herlev, Denmark) for the trial. O2matic was not involved in the funding, design, data collection, analysis or decision to publish. The views and conclusions contained here are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the US Government.