gms | German Medical Science

Introduction: Critically ill patients with hyperoxemia, a condition characterized by elevated levels of oxygen in the blood, have been shown to have increased all-cause mortality. However, little is known about the impact of hyperoxemia on neurosurgical patients, who are at an elevated risk for this condition. The aim of our study was to evaluate the extent of hyperoxemia during craniotomies, and to investigate its potential association with morbidity and mortality.

Methods: In this single-center cohort study, we extracted electronic perioperative data from adult craniotomy cases receiving general anesthesia with invasive mechanical ventilation and arterial blood gas (ABG) analyses. We tuned and trained two random forest algorithms for the estimation of perioperative arterial oxygen pressure (paO₂) at every time point of ABG, one based on training features excluding any ABG results (algorithm I), the other one based on training features including the first measured Horowitz quotient (algorithm II). These algorithms were then used for quasi-continuous (every 5 minutes) paO₂ estimations, where algorithm I was used before and algorithm II after the first ABG was available [1]. From these data, we calculated and normalized the integral values (AUC) for supraphysiological paO₂ values (>120 mmHg) for the peri- and intraoperative period using the composite trapezoidal rule. We then investigated the associations between hyperoxemia and in-hospital mortality, length of stay and complications using the Spearman rank-order correlation and the Kurskal-Wallis test, followed by Dunn’s test using Benjamini-Hochberg p value correction.

Results: The trained algorithms had adjusted R² values of 0.78 and 0.85 for algorithm I and II, respectively. The mean absolute percentage error was 14.6% for algorithm I and 11.7% for algorithm II. Our preliminary data consisted of 180,476 observations of 3,469 patients for the AUC calculation, of which 15,956 comprised actually measured paO₂ values. Among all patients, 12 patients did not receive supraphysiological oxygen at any time point, while 1,671 patients did not receive any intraoperatively.

The study found significantly higher normalized paO₂ AUC values in female patients, as well as in patients with ASA classes IV and V, with prolonged length of stay, in those who died, and in those with postoperative complications of pneumonia, cerebral embolism, and cerebral vasospasm.

Discussion: With hyperoxemia often being used as a suitable proxy to measure hyperoxia, we devised an effective approach for imputing paO₂ values, which outperformed several approximations for paO₂ such as the alveolar gas equation, Gadrey’s paO₂ or FiO₂ measurements alone [2]. The model fit as measured in R² as well as the mean absolute percentage error improved after including results from the first ABG, indicating that at least one ABG is advisable for good paO₂ estimations during craniotomy.

We found higher normalized paO₂ AUC values during the perioperative phase, likely due to intubation and extubation phases. Our findings of prolonged length of stay and higher mortality are consistent with the current literature [3]. However, the association between hyperoxia and secondary vascular complications is still controversial [4], [5].

Conclusion: Intraoperative paO₂ levels can be well estimated by random forest regression. Intraoperative hyperoxemia may be associated with higher mortality, increased rate of vascular complications, and prolonged length of stay. Therefore, a more cautious and sparing use of oxygen during craniotomies may be advisable.

The authors declare that they have no competing interests.

The authors declare that a positive ethics committee vote has been obtained.