COVID-19-related ARDS: one disease, two trajectories, and several unanswered questions.

Since the early days of medicine, doctors have described the natural history of disease and its different forms, primarily based on personal interpretation or intuition, in contrast to modern evidence-based medicine. For example, leptospirosis has been described with icterohaemorrhagic or pulmonary subtypes, but the existence of these phenotypes has been confirmed only relatively recently. Recent improvements in analysis and comprehension have been made possible using modern statistical analysis. For example, a previous study showed that two distinct phenotypes of acute respiratory distress syndrome (ARDS) co-exist, but also—and more importantly for clinicians—that those phenotypes differed by their response to different treatment strategies. Unfortunately, these strategies have not been validated in prospective randomised trials.

This modern side of critical care has received increased publicity during the COVID-19 pandemic. In the early phase of the pandemic, a strong debate between experts focused on the possible existence of two phenotypes and, more importantly, on modifications of mechanical ventilation settings according to each phenotype. Previous studies found different numbers of phenotypes, but these had several problems, including a retrospective nature, taking place at a single centre only, or absence of external validation.

In The Lancet Respiratory Medicine, Lieuwe Bos and colleagues reported that advanced statistical analyses cannot identify different phenotypes of COVID-19-related ARDS at the time of invasive mechanical ventilation initiation, in contrast to the results of previous studies. Furthermore, COVID-19 appeared to have two distinct phenotypes in the early course of mechanical ventilation. Mechanical power and ventilatory ratio can help to identify these two phenotypes, supporting the results of a previous study. Bos and colleagues should be congratulated for doing such studies in the difficult context of the COVID-19 pandemic. Although an increasing number of papers are dedicated to machine learning, few have as many quality criteria, and even fewer are informative for clinicians. However, I would like to raise several points in relation to the study.

First, COVID-19-related ARDS is a homogenous syndrome at initiation of mechanical ventilation, but it evolved during the early phase of ventilation into two distinct phenotypes. However, these phenotypes could be related to treatment heterogeneity in intensive care units, as acknowledged by the authors.

Second, the study highlights the importance of measuring several respiratory parameters multiple times, including static respiratory measures (PaO2/FiO2, plateau pressure, driving pressure, and static compliance) and dynamic measures (mechanical power and ventilatory ratio). For example, concerning the high respiratory drive of patients with COVID-19, spontaneous breathing with a high respiratory rate will substantially influence mechanical power and could potentially artificially induce a more severe phenotype. A large proportion of guidelines advocate a neuromuscular blockade or prone session according to the level of PaO2/FiO2. However, superiority of one measure over another has not been proven, leading clinicians to try to integrate them into each patient scenario.

Third, unfortunately, the authors were unable to study biomarkers. Biomarkers are the main determinant of ARDS phenotypes that have previously been studied, and have value regradless of physician ability to perform bedside measures (ie, static and dynamic ventilation indicators).

Finally, although multiple randomised trials have been dedicated to antiviral or immunomodulation treatments, the results of this study highlight that large randomised trials can be done to better define the best way to deliver ventilation to patients with ARDS, and to define the best settings for positive end-expiratory pressure for patients with ARDS (related or unrelated to COVID-19), despite a moderate level of evidence in ARDS guidelines and COVID-19 panel opinion. The same can be said for prone positioning—despite a mean PaO2/FiO2 ratio of 148 mm Hg (SD 75), only 30% of patients received prone positioning during the first day of mechanical ventilation.

In conclusion, such promising results must be replicated in randomised trials. Currently, randomised trials only support the use of higher anticoagulation doses for patients with COVID-19 in hospital wards, in contrast to patients managed in critical care. However, such trials were not stratified for phenotypes. Identification of these phenotypes might be difficult at the bedside. Previously validated tools (including machine learning) could help to simplify this aspect of modern critical care research and guidelines are available to develop such an approach in other areas of critical care. Tailoring treatment to phenotypes could be a good balance between evidence-based medicine, which requires many patients, and clinical personalised medicine.

I declare no competing interests.