This was a pre-planned ancillary analysis of the Prev-HAP trial [12]. The PREV-HAP study was a multicenter, parallel-group, double-blind, randomized trial designed to investigate the effects of interferon gamma-1b in critically ill patients at risk of hospital-acquired pneumonia. Briefly, patients aged between 18 and 85 years receiving invasive mechanical ventilation were eligible if presenting with one or more acute organ failures at the time of inclusion. Participants received five subcutaneous injections of 100 μg of recombinant interferon gamma-1b (interferon-gamma group) or matching placebo (placebo group). Patients were followed until day 90, and the rates of the primary outcome components (all-cause mortality at day 28 and 90 and HAP at day 28), mechanical ventilation-free days on day 90 were recorded. We also collected responses to the EQ-5D-3L questionnaire on day 28 and day 90 to estimate Quality-Adjusted Life Years (QALYs) for this ancillary cost-effectiveness analysis.

The cost-effectiveness analysis adopted a collective perspective including “all of the people or institutions affected (in terms of health effects or cost) by the production of an intervention within the scope of the overall patient care” thus excluding production losses from the base case analysis, as recommended by the French guidelines for the economic evaluation of health care programmes [14], and a 90-day time horizon. The costs, effectiveness and incremental cost-effectiveness ratio (ICER) comparing interferon gamma‑1b use against placebo to prevent HAP were estimated from individual patient data collected during the PREV-HAP trial. The ICER is defined as follows:

where \(_, _,_,_\) correspond to estimates of total mean costs and total mean QALYs per patient over three months for the interferon gamma‑1b and the placebo arm, respectively.

The Ethics Committee of Ouest II Angers (France) approved the study protocol in March 2021 (N°2020–000620-18). This trial complied with the Declaration of Helsinki and was registered in March 2021 (number ClinicalTrial.gov NCT04793568). The patient’s legal surrogate provided written informed consent for participation.

The outcomes were expressed in terms of QALYs assessed from EQ-5D quality of life measures at baseline, 28 and 90 days. QALYs combine information on length and health-related quality of life (HRQol) in a single index. They are estimated by weighting each period of time by an HRQol weight, called a utility value or score, where 0 represents being dead and 1 represents the best imaginable health state and such that a higher score corresponds to a more preferred health state (negative utility values are allowed for states considered worse than being dead). A year lived in perfect health thus represents one QALY, and three months lived in perfect health 0.25 QALYs. Utility values were determined from patient responses to the generic HRQol EQ-5D-3L questionnaire. The EQ-5D asked five questions about five dimensions of HRQol (mobility, self-care, usual activities, pain/discomfort and anxiety/depression). They can be answered using three ordered items ranging from 'no problem' to 'extreme/severe problems' with the dimension considered. The patient's responses to the EQ-5D questionnaire were converted to utility values using published tariffs for France [15]. At the beginning of the study, all patients were sedated and assigned a unique negative utility value of -0.402, which corresponds to the state of 'being unconscious' in the original EQ-5D-3L UK study [16]. As in previous studies [17, 18], when patients could not respond to the questionnaire on day 28 and/or day 90 due to their health condition, a proxy response was questioned. The proxy respondent was either a relative or a healthcare professional. The total number of QALYs over the three-month period corresponds to the areas under the curves obtained by applying linear interpolation between each EQ-5D utility scores [19]. Due to the short time horizon, QALYs were not discounted.

Costs were estimated from a collective perspective. Resource use was documented using electronic case report forms (e-CRFs) for the initial hospitalisation and self-administered questionnaires for the follow-up period. The questionnaires asked participants retrospectively about their use of outpatient (GP and specialist consultations, antibiotic prescriptions, nurse visits) and hospital (rehospitalisation, emergency department visits and rehabilitation hospital stays) healthcare resources since their initial hospital discharge. They also reported on their use of some medical equipment (wheelchair, hospital bed), the help they received from relatives or professionals in carrying out their usual activities, and their absence from work. Resource units were valued monetarily using national health insurance tariffs, information from the national hospital cost database and wage information from the National Statistical Institute (Appendix 3). Total cost estimates over three months were obtained by multiplying resource quantities by their corresponding monetary value. All costs were expressed in 2019 euros. As with QALYs, no discounting was applied.

e-Table 1 indicates the percentage of missing observations per item. It ranges from 0% (initial hospital length of stay) to 24% (e.g. GP visits). The proportion of missing data is fairly balanced between the two arms with some exceptions such as the number of EQ-5D measures. We handled missing data for both costs and QALYs (derived from EQ-5D questionnaires) by using multiple imputation. Specifically, we used chained equations combined with predictive mean matching, which helps address the non-normal distribution of the data [20]. The regression models for imputation included the following baseline variables, which we assume are related to the missingness mechanism, costs, and EQ-5D scores: age, sex, whether patients were septic, whether they have trauma, a kidney, neurological, respiratory, or hemodynamic failure, respectively, SAPS (Simplified Acute Physiology Score) II score [21] and the study arm. We used 45 imputation sets. All analyses were performed using Stata version 15.0 (StataCorp, College Station, Texas 77,845 USA).

The base case analysis was performed according to an intention-to-treat (ITT) principle, used imputed missing information and estimated adjusted differences in costs and QALYs. Patients were kept in the arm to which they were assigned whether they received interferon gamma‑1b or not and whether they received the expected dose or not. Differences in costs and QALYs were estimated using seemingly unrelated regression which accounts for the correlation between costs and QALYs [20]. The two regression equations for costs and QALYs contained the following explanatory variables: age, sex, whether the patients were septic, admission for trauma, a kidney, a neurological, a respiratory, or a hemodynamic failure, respectively, and the study arm. QALYs were also adjusted for the SAPS II score. Because costs and QALYs have non-normal distributions, the confidence intervals around their respective differences between the two arms were estimated by bootstrapping the regressions (1,000 replications) using bias-corrected and accelerated bootstrapping [22]. Estimates were used to calculate the adjusted mean ICER [23] (the ratio of the difference in total mean costs divided by the difference in mean QALYs) and the corresponding adjusted incremental net monetary benefit (the difference between monetized QALYs and costs). To take into account sampling uncertainty, the results were presented as an acceptability curve plotting the probability that interferon gamma‑1b was cost effective compared to placebo for different threshold monetary values for a QALY: €20,000, €50,000, €100,000 and €150,000, respectively.

We conducted several sensitivity analyses to test the robustness of our results. First, we analyzed the complete case sample, which included only patients with no missing data. Second, we explored the possibility that data were not missing at random, an assumption required for multiple imputation. In this scenario, we assumed that patients with missing data had a lower quality of life (25% below the imputed EQ-5D utility value) and higher costs (25% above the imputed cost). Third, we ran an analysis using a baseline EQ-5D utility value of zero, instead of -0.402, to estimate QALYs. Fourth, we performed an analysis from a societal perspective that included production losses calculated using the human capital method. Finally, we conducted an analysis without adjusting for baseline covariates, except for the SAPS-II score in the equation for QALYs.