Oxygenation thresholds for invasive ventilation in hypoxemic respiratory failure: a target trial emulation in two cohorts

Acute hypoxemic respiratory failure affects 20–50% of patients admitted to an intensive care unit (ICU) [1‐3]. Affected patients have a 20–40% mortality risk and survivors may experience decreased quality of life [1, 4‐6]. Invasive ventilation is a potentially lifesaving intervention that restores compensated gas exchange [7, 8]. However, invasive ventilation exposes patients to the risks of peri-intubation cardiac arrest, ventilator-induced lung injury, pneumonia, delirium, and ICU-acquired weakness [9‐13]. The best physiologic thresholds for the initiation of invasive ventilation are unknown [8].

Current practice varies. In qualitative research, clinicians factor multiple variables such as the degree of hypoxemia, work of breathing, and experience of the team into their decision for invasive ventilation [14, 15]. Randomized trials use multiple criteria incorporating hemodynamics, neurologic function, and respiratory status [16]. Observational cohorts show a low incidence of invasive ventilation after meeting various physiologic thresholds, including those used in trials [17], and profound inter-hospital variation in the use of invasive ventilation [18, 19]. Some of the observed practice patterns may cause harm through either overuse or delay in the initiation of invasive ventilation.

Relevant potential thresholds include the degree of hypoxemia, the criteria used in randomized trials, and thresholds that incorporate work of breathing, duration of respiratory failure, or clinical trajectory [20‐23]. A randomized controlled trial would be the most robust design to compare outcomes according to threshold, but this trial is not feasible at present due to disagreement on the region of equipoise and uncertainty about which thresholds to test [24]. A target trial emulation is an observational study design for causal inference which strives to mirror the eligibility criteria and interventions of the corresponding randomized trial, when that trial cannot be easily performed [25, 26]. Using the saturation-to-inspired oxygen ratio (SF), we performed a target trial emulation to compare the effect of using invasive ventilation initiation thresholds of SF < 110, < 98, and < 88 on 28-day mortality.

The primary analysis included 3,357 patients from MIMIC-IV (Additional file 1: Figure e9). The median age was 65 (interquartile range (IQR) 58 to 79) years and 45% (1,500) were women (Table 1). Most (63%) were admitted to a medical or surgical ICU. At eligibility, 16% (536) were using HFNC, 14% (483) NIV, and 70% (2,338) non-rebreather masks. The median baseline SF was 148 (IQR 136 to 174). Within 28 days, 896 patients (26.7%) received invasive ventilation and 745 patients (22.2%) died. Mortality was 17.7% in patients who did not receive invasive ventilation and 34.5% in patients who received invasive ventilation.

This target trial emulation of thresholds for initiating invasive ventilation in hypoxemic respiratory failure showed that using a threshold of SF < 110 as a trigger to initiate invasive ventilation resulted in more predicted 28-day invasive ventilation than thresholds of SF < 98 or SF < 88. In the primary analysis, the predicted 28-day mortality was lowest with a threshold of SF < 110. Across additional thresholds focused on respiratory rate, work of breathing, duration of hypoxemia, trajectory of hypoxemia, and multi-organ criteria from randomized trials, thresholds met at lower as opposed to higher severity led to lower predicted mortality. By contrast, in the secondary analysis, predicted mortality was highest with a threshold of SF < 110, and across the additional thresholds, those met at lower severity were associated with higher predicted mortality.

The different relationship between invasive ventilation thresholds and mortality comparing primary and secondary analyses might be explained by differences in internal validity or clinical context. Compared to the secondary analysis, the primary analysis had more patients, more measured confounding variables, better discrimination, better precision, and better calibration. The direction of residual bias was harder to predict because the target trial construction did not permit death before invasive ventilation during the 96-h target trial period, which favored higher severity thresholds, but the possible inclusion of patients with care limitations favored lower severity thresholds.

Differences in the clinical context of the two cohorts may also explain the findings. The two sites differed in terms of time period, healthcare system, ICU beds, patient population, oxygen device availability, and clinical practice. The secondary analysis encompasses an older observation period where contemporary approaches to ventilation, weaning, and sedation may not have been employed, potentially increasing the harms associated with invasive ventilation. In the secondary analysis, no patients used high-flow nasal cannula; however, this would be expected to bias results toward finding invasive ventilation more beneficial, the opposite direction from our findings. At BIDMC patients rely on private medical insurance and ICU beds comprise 11% of total hospital beds, while at Amsterdam UMC there is universal coverage for hospital care and ICU beds comprise only 4.4% of hospital beds [55, 56]. The impact of ICU bed availability on the results is harder to predict, but one potential impact is that the decision for invasive ventilation may be more commonly made prior to ICU admission in Amsterdam UMC as compared to BIDMC.

The results support the hypothesis that the benefit of invasive ventilation is related to underlying disease severity. Mortality was lower in the AmsterdamUMCdb cohort, implying a lower disease severity for non-intubated patients in that database. Other research also supports this hypothesis. In an observational study, ARDS patients with arterial-to-inspired oxygen ratio less than 150 mmHg managed using invasive as opposed to noninvasive ventilation had lower mortality [21]. In patients with COVID-19, higher baseline sequential organ failure assessment scores were associated with better outcomes when patients were managed with invasive ventilation as opposed to noninvasive oxygen strategies [57]. For patients with higher predicted mortality, invasive ventilation thresholds triggered at a lower severity of illness could confer benefits by avoiding catastrophic deteriorations, emergency intubations, or patient self-inflicted lung injury [58, 59]. By contrast, for patients with lower predicted mortality, the benefits of avoiding iatrogenic complications associated with intubation and invasive ventilation may predominate.

However, not all research accords with this conclusion. Two small randomized trials from 1998 and 2005 suggested benefits with a higher severity threshold for invasive ventilation among a population with severe hypoxemic respiratory failure [60, 61]. Options for noninvasive oxygen support, and the use of contemporary best practices for ventilation, weaning, and sedation may have been reduced in those studies, highlighting that the balance of benefit and harm associated with invasive ventilation will depend on the use of noninvasive oxygen strategies and best practices during invasive ventilation [11, 62‐65].

This study has important limitations. Unmeasured confounding is present because the clinical decision for invasive ventilation incorporates information about the diagnosis, prognosis, and nuanced respiratory assessment that are not available in the data studied. Unavoidably, the methods involved many modeling decisions which may affect the results in unpredictable ways [66]. The study does not report functional outcomes, where the harms of excess invasive ventilation may be more evident [6, 67]. The choice of SF ratio for the primary thresholds is problematic for people with darker skin pigment in whom peripheral oximeters can overestimate arterial oxygen saturation; we recommend using arterial oxygen saturation when a discrepancy is possible [68]. The target trial duration of 96 h captures most but not all intubations for hypoxemic respiratory failure [20, 69]. Some of the data were gathered more than 10–15 years ago and may not reflect current clinical practice.

The study also has considerable strengths. The methods are novel, fully documented, and address many challenges in correlating treatment decisions with clinical outcomes from retrospective data, including immortal time bias, indication bias, and time-varying confounding [66, 67]. The predictive validity of the component models has been explicitly assessed and documented. The thresholds evaluated are simple and clinically applicable.

These results highlight many areas for future research. Optimal thresholds may additional physiologic data such as standardized dyspnea assessment, electrical impedance tomography, or esophageal manometry [22, 70, 71]. More complex thresholds could be found through reinforcement learning [72, 73]. More information is also needed to compare thresholds with respect to functional outcomes, cost-effectiveness, and patient preferences.

For patients with hypoxemic respiratory failure, initiating invasive ventilation at lower hypoxemia severity will increase the rate of invasive ventilation, but this can either increase or decrease the expected mortality, with the direction of effect likely depending on baseline mortality risk and clinical context.