A Prolonged Air Leak Score for Lung Cancer Resection: An Analysis of The Society of Thoracic Surgeons General Thoracic Surgery Database

Abstract

Background.

The objective of this study was to create a simple preoperative tool to assess the risk of prolonged air leak (PAL) using The Society of Thoracic Surgeons General Thoracic Surgery Database (STS GTSD).

Methods.

The STS GTSD was queried for patients who underwent elective lung cancer resection between 2009 and 2016. Exclusion criteria included pneumonectomy, sleeve lobectomy, chest wall resection, bilateral procedures, and patients with incomplete data sets. The primary outcome was PAL exceeding 5 days. Multivariable logistic regression was used to identify risk factors for a PAL. Model coefficients were used to generate a PAL score (PALS). The approach was cross-validated in 100 replications of a training set consisting of two-thirds of the cohort that was randomly selected and a validation set of remaining patients.

Results.

A total of 52,198 patients from the STS GTSD met inclusion criteria, with an overall rate of PAL of 10.4% (n = 5453). Final variables incorporated into the PALS included body mass index of 25 kg/m² or less (7 points), lobectomy or bilobectomy (6 points), forced expiratory volume in 1 second of 70% predicted or less (5 points), male sex (4 points), and right upper lobe procedure (3 points). A cumulative PALS exceeding 17 points stratified patients as high-risk or low-risk for PAL (19.6% vs 9% rate of PAL) with a cross-validated mean negative predictive value of 91%, positive predictive value of 19%, sensitivity of 30%, specificity of 85%, and correctly classifies 79% of patients.

Conclusions.

The PALS is a simple preoperative clinical tool that can reliably risk-stratify patients for PAL who are undergoing lung cancer resection.

Lung cancer remains the most lethal form of cancer in North America.¹ Although surgery remains central to the treatment of lung cancer, a prolonged air leak (PAL) occurs in 7% to 18% of patients who undergo pulmonary resection.^2–12 PAL is associated with increased pain from chest tubes, higher rates of pulmonary complications, prolonged hospital length of stay, and increased health care costs.^5,13–17 Given the substantial health impact of PAL, a simple tool to help identify patients who are at risk before surgery could be clinically valuable. It would provide an opportunity to improve preoperative counseling and to develop more effective mitigation strategies by allowing investigators to focus their time and resources on patients at increased risk.

Algorithms to risk-stratify patients undergoing lung resection for PAL have been reported in North American and European populations.^{3,4,6,7,9,10,12} Unfortunately, the heterogeneous definition of PAL, lack of external validation, and complex scoring systems proposed have limited the clinical application of these reports.

A study in a single-institution cohort showed that a simple scoring system, based exclusively on preoperative data, can reliably classify patients who are at risk for PAL after pulmonary resection.⁶ Preoperative air leak predictor scores have also been proposed based on review and synthesis of the literature.¹⁸ A more recent study proposed the use of a nomogram to quantify PAL risk.⁷ However, nomograms can be complex to apply in the clinical setting, and as in other reports, generalizability of the results is limited by the absence of external validation.

The Society of Thoracic Surgeons General Thoracic Surgery Database (STS GTSD) is the world’s largest prospective, audited thoracic database containing sociodemographic, procedural, and short-term outcome data.¹⁹ In the current study, we used the STS GTSD to create and validate a simple PAL score (PALS) in patients who underwent lung cancer resection using readily available preoperative factors.

Patients and Methods

Study Population and Data Source

Through the STS Participant User File mechanism, the STS GTSD was queried for all patients who underwent elective lobectomy, bilobectomy, segmentectomy, or wedge resection for lung cancer between January 2009 and December 2016. Exclusion criteria included age younger than 18 years, pneumonectomy, sleeve lobectomy, chest wall or diaphragm resection, or bilateral procedures. Variables considered in the model were specified a priori and were chosen based on the existing literature and expert consensus.^7–11,20 Patients with missing data points in the a priori selected variables were excluded from analysis so that only complete data sets were examined. In the case of comorbidities, PAL, discharge with a chest tube, readmission, and mortality, failure to code the presence of a variable was considered a negative response, as previously described.²¹ All data were deidentified and stored in an encrypted database.

Variables

All variables were defined in accordance with STS GTSD definitions.²² The independent variables were selected for analysis were sex, age, body mass index (BMI), smoking status (never vs ever), preoperative percentage predicted forced expiratory volume in 1 second (FEV₁%), preoperative percentage predicted diffusion capacity of the lung for carbon monoxide (DLCO%), prior cardiothoracic surgery, Zubrod score (performance status; 0–4), steroid use, diabetes mellitus, video-assisted or open approach, location of resection, pathologic tumor size, and N stage. BMI was dichotomized as 25 kg/m² or less vs more than 25 kg/m², FEV₁% and DLCO% were dichotomized as 70% or less vs more than 70% based on previous reports and distribution of the data.^6,7 The Zubrod score was dichotomized as 0 vs 1 to 4 based on its skewed distribution. Smoking status (ever vs never smoker) was chosen for analysis instead of pack-years for ease of clinical use and because pack-years has been shown to not correlate with PAL.⁹ Also examined were operative procedure, whether multiple concurrent lung resections were performed under the same anesthetic, hospital length of stay, PAL, discharge from hospital with an indwelling chest tube, readmission within 30 days, in-hospital mortality, and 30-day mortality. If multiple procedures were performed under the same anesthetic, the procedure was analyzed as a bilobectomy if performed, if not, then lobectomy if performed, if not, then segmentectomy if performed, and if not, then wedge resection if no other procedures were performed. Patients with extreme outliers of BMI (<18 kg/m² and >45 kg/m²), FEV₁% (>120% or <30% predicted), or DLCO% (>120% or <30% predicted) were excluded from the analysis to create a model most representative of the North American lung cancer resection patient population. The dependent variable, PAL, was defined as an air leak lasting more than 5 days in accordance with the STS GTSD definition.

Statistical Analysis and Validation

Univariate analysis was used to identify factors potentially prognostic for developing a PAL. Factors were then selected for inclusion in forward stepwise multivariate regression analysis until the predictive ability (or discriminant function) was optimized with a minimum set of covariates. Model coefficients were rounded to the nearest unit to develop a weighted scoring system. An air leak score was then calculated for each patient in the validation set. The approach was cross-validated in 100 replications of a training set consisting of two-thirds of the cohort that was randomly selected and a validation set of remaining patients. A receiver operating characteristic curve was created to identify the score threshold providing the best ability to risk-stratify patients for PAL. These analyses were performed using the entire cohort and repeated in the subset of patients who underwent lobectomy. Pearson and Spearman correlations were run to assess collinearity between variables. Statistical calculations were performed using R 3.5 statistical software (The R Foundation for Statistical Computing, Vienna, Austria).

Results

During the study period, 52,198 patients (46% male) from the STS GTSD met inclusion criteria, with an overall rate of PAL of 10.4% (n = 5453). The PAL rate did not vary by year. The cohort was a median age of 68 years (Supplemental Table 1). Most patients were previous smokers and had a Zubrod score of 0 or 1. Lobectomy was performed in 76% of patients, with the right upper lobe being the most common location of resection (35%). Video-assisted thoracoscopy (VATS) was used in 65% of cases, and multiple concurrent ipsilateral resections were performed in 10%. The rate of PAL by clinical characteristic is provided in Supplemental Table 2.

Nearly all of the original 14 candidate variables were associated with PAL in the univariate and multivariate logistic regression analyses (Table 1). By sequentially adding covariates to the model in order of their absolute z value, the predictive performance was improved by AUC 1% or higher with the addition of lobectomy/bilobectomy, FEV₁ of 70% or less, right upper lobe procedure, and male sex to BMI of 25 kg/m² or less (Supplemental Table 3). Pearson and Spearman correlations revealed no evidence of collinearity between right upper lobe procedure and lobectomy/bilobectomy (P = .06).

Table 1.

Univariate and Multivariate Prognosticators of Prolonged Air Leak

Risk Factor	Univariate	Multivariate
	P Value	OR (95% CI)	P Value
BMI ≤25 kg/m2	<.001	1.95 (1.84–2.07)	<.001
Bilobectomy or lobectomy	<.001	1.95 (1.79–2.12)	<.001
Right upper lobe procedure	<.001	1.40 (1.32–1.49)	<.001
FEV1 ≤70%	<.001	1.43 (1.34–1.52)	<.001
Male sex	<.001	1.39 (1.31–1.47)	<.001
Ever smoker	<.001	1.51 (1.36–1.68)	<.001
DLCO ≤70%	<.001	1.25 (1.18–1.33)	<.001
Diabetes mellitus	<.001	1.30 (1.20–1.41)	<.001
VATS approach	<.001	0.82 (0.77–0.87)	<.001
Multiple resections	<.001	1.31 (1.20–1.42)	<.001
Age (y)	<.001	1.01 (1.01–1.01)	<.001
Prior cardiothoracic surgery	<.001	1.24 (1.15–1.34)	<.001
Tumor size <3 cm	<.001	0.90 (0.85–0.96)	.002
Zubrod score >0	<.001	1.07 (1.01–1.13)	.027
Steroid use	.022	1.16 (1.00–1.34)	.05
N stage (node negative)	.016	0.99 (0.90–1.08)	.766

BMI, body mass index; CI, confidence interval; D_LCO, preoperative percentage predicted diffusion capacity of the lung for carbon monoxide; FEV₁, preoperative percentage predicted forced expiratory volume in 1 second; OR, odds ratio; VATS, video-assisted thoracoscopic surgery.

Final covariates included in the PAL model included BMI of 25 kg/m² or less (odds ratio [OR], 1.95; 95% confidence interval [CI], 1.84–2.07; 7 points), lobectomy or bilobectomy (OR, 1.95; 95% CI, 1.79–2.12; 6 points), FEV₁ of 70% or less (OR, 1.43; 95% CI, 1.34–1.52; 5 points), male sex (OR, 1.39; 95% CI, 1.31–1.47; 4 points), and right upper lobe procedure (OR, 1.40; 95% CI, 1.32–1.49; 3 points).

To obtain a PALS that is usable in clinical practice, each model coefficient was rounded to the nearest unit to develop a weighted scoring system (Table 2). The rate of PAL by PALS is provided in Supplemental Table 4. Using a cumulative PALS exceeding 17 points, we found this tool stratified patients as high-risk or low-risk for PAL (19.6% vs 9% rate of PAL) with a cross-validated mean negative predictive value of 91%, positive predictive value of 19%, sensitivity of 30%, and specificity of 85%. Risk classification was correct in 79% of patients. The cross-validated receiver operating characteristic AUC was 0.644 (Figure 1).

Table 2.

Prolonged Air Leak Scoring Sheet for Patients Undergoing Lung Cancer Resection

Prolonged Air Leak Scoring Sheet	Points	PALS
Body mass index ≤25 kg/m2?	7
Lobectomy or bilobectomy?	6
FEV1 ≤70%?	5
Male?	4
Right upper lobe procedure?	3
Total PALS
If total PALS >17 the risk of PAL→	19.6%	(high risk)
If total PALS ≤17 the risk of PAL→	9%	(low risk)

FEV₁, preoperative percentage predicted forced expiratory volume in 1 second; PALS, prolonged air leak score.

Figure 1.

The cross-validated receiver operating characteristics curve based on the final prediction model for prolonged air leak in lung cancer resection patients. (AUC, area under the curve.)

When stratified by PALS, patients at high risk for a PAL (PALS >17 points) experienced a significantly longer median hospital length of stay (5 vs 4 days, P < .001), were more often discharged with a chest tube (14.2% vs 7.1%, P < .001), and had a significantly higher 30-day readmission rate (10% vs 7.5%, P < .001). In-hospital (1.5% vs 0.7%, P < .001) and 30-day mortality rates (1.9% vs 1%; P < .001) in patients with a PAL score exceeding 17 were nearly twice those of patients with a PALS of 17 points or less (Table 3).

Table 3.

Outcomes of Patients at High and Low Risk for Prolonged Air Leak

Outcome	Full Cohort (N = 52,198)	Risk for Prolonged Air Leaka
		High (PALS >17) (n = 9301)	Low (PALS ≤17) (n = 42,897)
Prolonged air leak	5453 (10.4)	1800 (19.3)	3653 (8.4)
LOS, median (IQR), d	4 (3)	5 (5)	4 (3)
Discharged with chest tube	4239 (8.1)	1303 (14)	2936 (6.8)
Readmission ≤30 days	4116 (7.9)	935 (10.1)	3181 (7.4)
In-hospital mortality	440 (0.8)	125 (1.3)	315 (0.7)
30-day mortality	599 (1.1)	159 (1.7)	449 (1)

^aAll P values for comparisons between patients at high and low risk for prolonged air leak were statistically significant (P < .001).

The data are presented as n (%), unless otherwise noted.

IQR, interquartile range; LOS, length of stay; PALS, prolonged air leak score.

A subset analysis of 39,660 patients who underwent lobectomy for lung cancer identified BMI of 25 kg/m² or less (OR, 1.92; 95% CI, 1.81–2.05; 7 points), FEV₁ of 70% or less (OR, 1.42; 95% CI, 1.32–1.52; 5 points), right upper lobectomy (OR, 1.45; 95% CI, 1.36–1.54; 4 points), and male sex (OR, 1.34; 95% CI, 1.26–1.43; 3 points) as predictors of PAL. The lobectomy model had approximately the same performance as that of the full cohort, with a cross-validated negative predictive value of 90%, positive predictive value of 21%, sensitivity of 19%, and a specificity of 91% (Supplemental Tables 5 and 6) and correctly predicted the presence or absence of a PAL in 83% of cases, with an AUC of 0.629 (Supplemental Figure 1).

Comment

The current study used the STS GTSD to create and validate a contemporary classification algorithm capable of risk-stratifying patients for PAL before lung cancer resection. The study population was restricted to patients with lung cancer—the most common indication for lung resection—thereby eliminating confounding populations that may have different risk factors for PAL, such as bullectomy and lung volume reduction surgery, among others . Upon cross-validation, the model demonstrated the ability to correctly prognosticate PAL 79% of the time using a PALS cutoff of more than 17 points. The current model differs from previous risk models in that it was developed and validated in a large, audited, multiinstitutional population of lung resection patients, it relies exclusively on preoperative data, and it is simple enough to be readily applied in the office by clinicians. In its current iteration, this PAL model divides patients in 2 risk categories (high-risk and low-risk) instead of 3 to 4 groups, as previously reported, to facilitate ease of use in clinical decision making.

The overall rate of PAL observed and preoperative variables found to be prognostic in the PALS model (BMI, lobectomy or bilobectomy, FEV₁ ≤70%, male sex, and right upper lobe procedure) are consistent with those previously reported.^{5–10,12,14,17,18} Patients with a PALS exceeding 17 points had not only more-than-double the rate of PAL than those with a PALS of 17 or less but also a longer hospital length of stay, duration of chest tube, rate of discharge with a chest tube, readmission rate, and mortality rate.

The concept of developing a prognostic model to quantify PAL risk is not new. However, no reported algorithm to date has gained widespread clinical adoption because of complexity of use, inclusion of intraoperative or postoperative variables, reliance on variables not readily collected, differing definitions of PAL, or lack of external validation. Our previously reported model had slightly superior performance (AUC = 0.8) but was derived from a relatively small, single-institution cohort, potentially introducing bias and reducing external validity.⁶ In addition, it relied on the Medical Research Council dyspnea score, a variable not used in many centers and not collected in the STS GTSD. The current model identified FEV₁ of 70% or less as a prognostic factor, which like the Medical Research Council dyspnea score, may serve as a proxy for parenchymal lung disease.

Brunelli and colleagues^9,12 have published separate risk models for VATS lobectomy and open lung resections. Whether European data can be applied to North American patients is unknown, given the previously demonstrated differences in pulmonary resection practice patterns between the European Society of Thoracic Surgeons Registry and the STS GTSD.²³ In addition, the European VATS algorithm uses BMI of 18.5 kg/m² or less as one of the predictors for PAL, resulting in only 2.3% of patients being classified into the highest risk category. When a risk model is created, it is important to ensure an adequate percentage of patients meet criteria for the high-risk group, otherwise, the number needed to screen when validating the findings in a prospective fashion will make accrual exceedingly difficult. Using the PALS model, we found nearly 14% of patients were classified as high risk, making the model easier to validate and potentially more applicable in real-world practice.

Rivera and colleagues¹⁰ recently reported a PAL risk assessment model derived from a relatively large single-institution cohort. Unfortunately, application of the model requires the use of a complex nomogram with multiple variables and interaction effects. Although using a nomogram resulted in marginally improved classification performance, it makes day-to-day clinical use less practical. One of the primary goals of the current study was to create a model that is simple enough that it will gain widespread clinical application in risk-stratifying patients for PAL.

A similarity noted among all models attempting to risk-stratify patients for PAL, including the current PAL score, is modest prognostic ability. This is likely a reflection of the small degree of association between the clinical variables examined and PAL. No individual risk factor in the current analysis yielded an OR of more than 2 for PAL. Clinical risk factors associated with less than a 3-fold risk for an outcome have been shown to be unreliable for prognosticating individual patient events.²⁴ The limited performance of these models emphasizes the problem with using population-based data sets to predict individual patient outcomes.²⁵ Invariably, there is “noise” from unintended, unrecognized, and unexamined sources of bias that likely affect the performance of prognostic algorithms.

Although this is a large study examining risk factors for PAL, it shares limitations inherent to many national database studies.²⁶ Specifically, the STS GTSD was not designed as a research tool but as a quality improvement initiative. Therefore, certain variables that may be relevant to an analysis of PAL, such as use of pulmonary sealants, pleural tents, chest tube management strategies, use of digital drainage devices, and rate of subsequent reoperation, were not available for analysis. Similarly, when performing anatomic lung resections, some surgeons approach the vessels through the fissure and others use a “fissure-less” approach where the fissure is stapled last.

Finally, although this PAL classification tool is anticipated to be simple to apply in the clinic, it has not yet been clinically tested for ease of use. Feasibility testing is required to formally assess its clinical applicability.

The modest positive predictive value and high false-positive rate of this model could be perceived as a limitation. It is understood that efforts to improve sensitivity may come at the cost of potential overtreatment of patients incorrectly deemed to be at higher risk. Given that nearly all interventions aimed at minimizing PAL are associated with a minimal risk of adverse events, that patients will be harmed as a result appears to be unlikely.

Another concern may be increased health care costs as a result of overtreatment. It is reasonable to expect that expenditures related to air leak prevention or treatment may be offset by improvements in surgical quality and safety resulting in decreased hospital length of stay, complications, and mortality related to the prevention of PAL.

A systematic review of mitigation strategies, such as pulmonary sealants, has yielded conflicting results.²⁷ We think this observation is likely a reflection of our previously limited ability to appropriately select subgroups of patients who could benefit the most from such interventions. As a corollary, the high negative predictive value of the PALS model allows patients to be reassured and avoid unnecessary costs associated with interventions in populations unlikely to have a PAL. Lastly, if a patient is expected to have a PAL, perhaps the surgeon would be more likely to discharge them earlier with a chest tube rather than keeping the patient in the hospital waiting for the air leak to stop. A formal cost analysis performed in the setting of a prospective study is required for verification of the true therapeutic and financial implications associated with the use of the PALS model.

The clinical application of the PALS model may also have additional benefits that are more difficult to quantify. Specifically, the PALS classification system has the ability to provide surgeons with data to drive individualized preoperative discussions with patients. If realistic expectations are set, patients will be better prepared for complications if they arise, potentially improving patient satisfaction. In addition, because this model is derived from and validated in a multiinstitutional North American population, it may facilitate risk-stratified comparisons of PAL rates across surgeons and institutions. Most importantly, clinical application of the PALS has the potential to facilitate development of new and improved strategies for minimizing the rate of PAL after lung cancer resections. We hope that the proposed PALS will serve as a basis for further prospective research focusing on the effectiveness of existing and future risk-reduction strategies.