Association between arterial oxygen partial pressure and mortality in sepsis

To the editor,

The appropriate arterial oxygen partial pressure (PaO₂) in sepsis patients has been investigated in dozens of studies. However, no consensus has been reached. In a recent study [1], Dr. Hyun et al. investigated the association between PaO₂ and mortality in critically ill sepsis patients. Data on PaO₂ of 4147 sepsis patients from the Korea Sepsis Alliance Registry (KSA) during the first three days was extracted, and patients were divided into conservative PaO₂ group (< 80mmHg) and liberal PaO₂ group (≥ 80mmHg). Statistical methods, including propensity score matching, mixed linear model, and competitive risk models, were used to explore potential causal relationships. The results showed higher PaO₂ (≥ 80 mm Hg) during the first three ICU days was associated with a lower 28-day mortality than conservative PaO₂. Several issues should be noted when interpreting these findings.

First, dividing the entire cohort into groups based on the cut-off value of one continuous variable is a common strategy in observational studies. In the current research, a cut-off value of 80 mmHg of PaO₂ was used. However, the results may be biased when the correlation between PaO₂ and mortality is non-linear. For instance, previous studies [2, 3] have pointed out that there may be a U-shaped correlation between PaO₂ and death, which means either extremely high or low PaO₂ may be associated with increased mortality compared to normal PaO₂. In this case, whether there is a difference between the conservative and liberal PaO₂ groups can be significantly affected by the cut-off value. For instance, in the current study, the restrictive cubic spline between PaO₂ and mortality showed that the PaO₂ with the lowest mortality possibility was around 100mmHg. Thus, when using 80mmHg as the cut-off value, the comparison between conservative (PaO_2min < 80mmHg) and liberal (PaO_2min ≥ 80mmHg) PaO₂ groups was actually comparing the low PaO₂ group with the normal PaO₂ group combined with the high PaO₂ group (< 80mmHg vs. (80 – 110mmHg +  > 110mmHg)), which thus lead to a potentially biased result that liberal PaO₂ was associated with low mortality rate. Similarly, the difference in mortality rate between the two groups may also be affected by the proportion of patients with high PaO₂. In the current study, we noted that the PaO₂ was relatively low even in the liberal PaO₂ group (median values were 107, 110, and 106 during three days). In the case of a low proportion of patients with high PaO₂, using 80mmHg as the cut-off value may actually be the comparison between the low PaO₂ group and the normal PaO₂ group, which may also be one factor for the inconsistent results of previous studies.

Second, there may be some statistical issues that also need to be noted. For instance, we note that a total of 4147 patients were included. However, in the multivariable COX regression analysis, only patients (n = 2422) matched in the propensity score matching (PSM) were included. Theoretically, implementing regression analysis only in the matched cohorts may reduce the variability of the cohort, which may result in a biased result. For instance, the PaO₂ in the matched cohorts was relatively low (median values were 107, 110, and 106 during three days). Thus, whether higher PaO₂ (such as PaO₂ > 180mmHg) was associated with poor prognosis cannot be inferred. Another minor issue is that the author used the 5th percentiles of PaO₂ as the reference point in restricted cubic spline analysis. This seems a little inappropriate as the selected cut-off point of PaO₂ was around 100 mmHg.

Finally, we thank Dr. Hyun et al. for their great work, and we hope our opinions will help interpret these findings.