Nomogram for the prediction of postoperative hypoxemia in patients with acute aortic dissection

Acute aortic dissection (AAD) represents a life-threatening condition that can be encountered in emergency and critical care setting [1]. Many factors can influence the clinical outcomes of these patients such as the comorbidities, complications, organ dysfunction and site of dissection. Surgical operation is usually needed to avert catastrophic complications of aortic dissection [2]. Postoperative hypoxemia has long been noted in substantial proportion of patients with AAD and has been found to be associated with poor clinical outcomes such as prolonged mechanical ventilation, increased length of stay (LOS) in the intensive care unit (ICU) and hospital [3‐5]. Also, several studies have attempted to identify preoperative risk factor of hypoxemia [4, 5]. However, there is no report on training a model for early prediction of postoperative hypoxemia. Since early prediction of post-operative hypoxemia makes early intervention possible, it is of clinical utility to train and validate such a prediction model. The study aimed to develop a model for early prediction of postoperative hypoxemia. Discrimination and calibration of the model were validated in an independent dataset that was not used for model training. A nomogram was depicted for clinical use.

The study included patients with AAD who had undergone operation for the repair of the dissection. Risk factors for postoperative hypoxemia were identified via purposeful selection procedure. These factors included age, lactate, preoperative P/F ratio, WBC, BMI and CRP. Internal validation was performed for the model, which showed good discrimination and calibration. A nomogram was established for clinical utility.

The incidence of hypoxemia was lower than those reported in the literature (20% vs. 30%) [4, 14]. Different definitions of postoperative hypoxemia may explain its different incidences. For example, Wang’s study defined hypoxemia as P/F ratio less than 200 at 24 h after operation and they reported an incidence of 49.5% [5]. Several studies have been conducted to explore risk factors for postoperative hypoxemia. Consistent with our study, the study by Liu and colleagues also showed the preoperative P/F ratio and WBC were independently associated with postoperative hypoxemia [4]. It is not surprising that preoperative P/F ratio is independently associated with postoperative hypoxemia. Furthermore, WBC is a biomarker reflecting systemic inflammatory response, and higher responses may contribute to the respiratory dysfunction. There is empirical evidence that inflammatory response is associated with hypoxemia in patients with aortic dissection [15]. Furthermore, we also found CRP was associated with hypoxemia. CRP is a well-known biomarker of inflammatory response, which is more specific and sensitive than WBC [16, 17]. However, the effect of CRP disappeared after adjusting for covariates. Liu’s study also included time from symptom onset and deep hypothermic circulatory arrest time in their regression model. None of the intraoperative variables were associated with hypoxemia in univariate analysis, but the CPB time was associated with hypoxemia in multivariable regression model. BMI was identified as an independent predictor of postoperative hypoxemia in the study, consistent with the study by Sheng and colleagues [14]. Actually, the association of obesity and hypoxemia is not limited to aortic surgery [18]. In a large cohort of noncardiac surgery, Kendale SM and colleagues found that the odds of experiencing hypoxemia increased significantly with increasing categories of BMI [19]. Similar results were documented in other studies [20, 21]. BMI is an important determinant of respiratory function and studies show morbidly obese patients have a typical restrictive pattern with a reduction of forced vital capacity (FVC), forced residual capacity (FRC) and total lung capacity (TLC) with a decreased expiratory reserve volume (ERV) [22‐24]. Sex has been found to be related to postoperative hypoxemia in AAD (women were more likely to have hypoxemia than man), which was not replicated in our study. Most probably, the associated was confounded by other factors and the authors failed to adjust for these potential confounding factors [25].

An interesting finding in our study was that serum lactate was associated with postoperative hypoxemia in patients with AAD. Hyperlactatemia is an indicator of tissue ischemia [26]. In cardiac surgery patients, there is evidence that hyperlactatemia is associated with a compromised respiratory function and prolonged mechanical ventilation. Also, hyperlactatemia can explain organ dysfunction in our study [27‐29]. However, the effect of lactate disappeared in multivariable model, indicating that the effect could be explained by preoperative P/F ratio, as lactate was a biomarker of hypoxemia. Similarly, preoperative P/F ratio can explain the elevations of biomarkers of acute organ injury such as Scr, bilirubin and ALT. That was why the significant associations in univariate analysis disappeared in the multivariable regression model.

The potential utility of our prediction model is that interventions such as anti-inflammatory agents can be given to patients with high risk of postoperative hypoxemia. Furthermore, the model can be employed to design clinical trials to identify high risk patients who will benefit the most from treatment [30].

There were some limitations in the present study that must be acknowledged. First, the study was retrospective single center study. Although the prediction model was validated in a separate dataset that was not used for model training, its external validity was still unknown. In different cardiac centers, the healthcare process can be quite different and the predictive accuracy of a newly developed model needs to be tested [31, 32]. This is also our future work to perform a multicenter study to examine the external validity of the model. Second, postoperative hypoxemia reported in the study was defined by P/F ratio less than 200 for consecutive 2 days. While this definition was simple to perform, it suffers from case-mix. Some patients may have intensive respiratory support, and their P/F ratio is not comparable to those with spontaneous breathing. Thus, the definition of postoperative hypoxemia needs to be further explored. However, in order to make our results comparable to that in the literature [4], we opt to use this simple definition. Third, the retrospective design of the study suffers from its inherent limitation of selection bias. Some patients were excluded because they did not perform operation in our hospital. They might go to other hospital for further evaluation. Thus, the included patients may not well represent the whole target population of AAD, but they represent those who are willing to undergo operation in our hospital. Finally, the sample size of the study was small that only 43 patients had the event of interest. As a rule of thumb, 10–20 events per variable should be used [33]. However, the primary purpose of this rule is to prevent overfitting [34]. We addressed the problem of overfitting by validating the model in a dataset that was not used for model training (e.g. an overfitting model performs poorly in the validation dataset). The results showed that the developed model can predict accurately in the validation dataset with an AUC of 0.869. Further studies by employing multicenter data are mandatory to endure generalizability of the model.

In conclusion, the study developed a model for the prediction of postoperative hypoxemia in patients undergoing operation for AAD. Eight variables of age, BMI, WBC, HCT, CPB time, pH, Stanford type and P/F ratio were included in the model. The model showed good discrimination and calibration in an independent dataset that was not used for model training.