2.1 Study Design and Participants

BLCS is a longitudinal, ongoing cancer epidemiology cohort of lung cancer patients established in 1992 at Massachusetts General Hospital (MGH, Boston, MA, USA) [24]. The study enrolls patients newly diagnosed with histologically confirmed lung cancer at the time of recruitment, with no restrictions on stage or treatment type. All patients were 18 years or older at enrollment, and the majority were white (95.3%). Demographic information, including age, sex, body mass index (BMI), and smoking status and intensity, was collected at enrollment by trained staff using a standardized questionnaire. Clinical information, including tumor histology, clinical stage, and initial cancer treatment types, was retrieved retrospectively from electronic medical records. Between 1992 and 2020, among 5940 patients with NSCLC, 5826 patients received follow-up, and 2348 patients had lung volume and TLC tests performed before initiating lung cancer treatment and were included in our analytical cohort (Figure 1). This study was approved by the institutional review board of MGH (Partners Human Research Committee, Protocol No. 1999P004935/MGH), and informed consent was collected from each study participant or their surrogates.

2.2 Pulmonary Function Measurement

TLC and RV test data were obtained from pulmonary function test (PFT) reports or medical records in BLCS, which were entered upon patient enrollment into the electronic medical record system. The majority of PFTs were performed on-site at MGH during lung cancer diagnosis, following the standards of the American Thoracic Society (ATS)/ European Respiratory Society (ERS). Lung volume tests were primarily measured using body plethysmography or helium dilution techniques [25]. Quality control of collected PFT values was performed by trained staff during data extraction and independently by two pulmonary physicians (RB and DC) post-data extraction. Data collection and processing were conducted blind to patients' survival status. We excluded PFTs that were performed after the initiation of lung cancer treatment, as treatments could alter pulmonary function.

RV was calculated by subtracting slow vital capacity (SVC) from TLC. The predicted value and lower limit of normal (LLN) of each PFT were calculated through the online calculator of the Global Lung Function Initiative (GLI) (http://gli-calculator.ersnet.org/index.html) [26]. Tests for survival analysis included RV and percent predicted RV (RV%), TLC and percent predicted TLC (TLC%), RV/TLC ratio, and percent predicted RV/TLC (RV/TLC%). Analysis on percent predicted data was restricted to patients with age < 80 years based on the capability of GLI prediction for lung volumes [26]. Additionally, where available, we collected and calculated diffusing capacity for carbon monoxide (DLCO) and percent predicted DLCO (DLCO%) as our secondary variables of interest. Spirometry, including FEV₁, percent predicted FEV₁ (FEV₁%), FVC, percent predicted FVC (FVC%), and FEV₁/FVC ratio, was also collected and calculated.

2.3 Survival Data Collection

Our primary outcome was overall survival, i.e., the time lag in months between the date of pathological diagnosis of lung cancer and the date of death (event) or the last known date alive (censor), whichever came first. Data were collected from the following sources as previously described [24]: MGH patient and outpatient records, MGH tumor registry, social security death index, primary physician's office, and patient or family contact. As BLCS is an ongoing longitudinal cohort, there was no fixed cut-off date for follow-up. Patients who were alive at their last recorded contact were censored at that time, following standard survival analysis procedures. These patients remain under observation unless they withdraw from the study or become untraceable, resulting in a loss to follow-up rate of 1.9%.

2.4 Statistical Analysis

Cox proportional-hazards (PH) models were used to estimate hazard ratios (HRs) and 95% confidence intervals (CIs) for lung volume with overall survival, modeling time from diagnosis to death (event) or censoring (alive at last follow-up). To illustrate the shape of the relationship between each lung volume and overall survival, we first fitted penalized smoothing splines and visualized the HRs in response to the changes in each lung volume. This approach also served as our tool to examine the nonlinear associations, given the continuous nature of the lung volume values. In the main analyses, to reduce the impact of outlier data, we categorized values into three percentile groups (1%–25%, 25%–75%, and 75%–100%), and used the 25%–75% group as the reference group. We also modeled each measure on a continuous scale by examining one standard deviation (SD) increments of TLC, RV, and RV/TLC.

Multiple regression models were adjusted for known prognostic factors of lung cancer or potential confounders, including age, sex, BMI, smoking status (never vs. ever, current), NSCLC histological subtypes (adenocarcinoma vs. adenocarcinoma in situ [AIS], squamous cell, large cell, and unspecified), and clinical stages (I vs. II, III, and IV). These patient characteristics were correlated with the tests and believed to be confounders for the associations between RV, TLC, RV/TLC, and overall survival (Figure S1). Because all lung volume tests were performed before any cancer treatment, treatment was hypothesized to mediate the effects of lung volumes and other covariates on survival. To allow patients receiving different treatments to have distinct baseline hazards and to better isolate the association of lung volumes with overall survival, the Cox models were further stratified on lung cancer treatments (surgery only vs. surgery and chemo/radiation, and chemo/radiation only).

The PH assumption of the Cox model was evaluated for each variable by comparing its interaction with follow-up time, and violations were addressed by including their time interaction terms in each model. The deviance residuals were plotted to evaluate the goodness of fit for the Cox models. Model performance was evaluated by Harrell's C-index [27].

Complete-case analyses were conducted in the main analysis as there was no evidence for informative missingness, with a large sample size and a complete case rate of 71.8%. Missing values were imputed based on the mean, median, or mode of observed data [28], and the results based on imputed values were compared to those from complete-case analyses. We also explored the interactions between RV, TLC, and RV/TLC with spirometry by centering on the person with the median value of each test as the reference, and modeling one SD increment of each test in the Cox model. Further analysis was conducted by stratification on study subjects based on their spirometry status (within the normal range: FEV₁/FVC equal or above the LLN of GLI reference values, and with abnormal spirometry: FEV₁/FVC below the LLN) and cancer stage (stage I-II and stages III-IV) to evaluate the prognostic values of RV/TLC among patients classified on well-accepted prognostic factors. Kaplan–Meier curves were compared across the RV/TLC groups within these strata, and survival differences were assessed using log-rank tests.

To assess potential selection bias in our analytical cohort, we compared demographic and clinical characteristics as well as overall survival between NSCLC patients with and without available lung volume tests in the entire BLCS cohort. Because patient characteristics contributed to test availability (Figure S1), we further evaluated potential selection bias after covariate adjustment in survival analysis. For potential heterogeneity in pulmonary function tests conducted within and outside MGH, we indexed in-hospital and out-hospital lung volume tests to assess the effect of between-hospital variation by examining its interaction with each lung volume test variable.

All P-values were two-sided, and p < 0.05 was considered statistically significant. Statistical analyses were conducted using SAS (version 9.4; SAS Institute Inc., Cary, NC, USA) and R (version 4.0.2; R Foundation for Statistical Computing).

3.1 Patient Characteristics

Table 1 describes the demographic and clinical information of 2348 NSCLC patients in the analytic cohort (mean [SD] age: 67.49 [9.85] years, 49.74% male patients). The majority of patients were diagnosed at stage I (57.2%) and received surgery as their primary treatment (63.8%). A total of 1362 deaths were observed with a median survival of 66.9 months (95% confidence interval [CI]: 60.4–73.0 months). The 1-, 5-, and 10-year survival rates were 86.2%, 52.5%, and 34.6%, respectively, and the median follow-up length was 40.5 months (range, 0.07–281.54 months). From the univariate analyses, age, sex, BMI, smoking status, pack-years, NSCLC histological subtypes, clinical cancer stages, and post-diagnosis cancer treatments were all significantly associated with overall survival in lung cancer patients (Table 1).

TLC, RV, and diffusion capacity test results are summarized in Table 2. RV/TLC had a mean of 47.7% (SD: 11.2%) and was well-correlated with FEV₁ (Pearson correlation coefficient [r] = −0.70, p < 0.001), FEV₁% (r = −0.62, p < 0.001), FVC (r = −0.65, p < 0.001), and FVC% (r = −0.58, p < 0.001) (Figure S2).

3.2 RV, TLC Predicts Overall Survival in NSCLC

The crude and adjusted HRs of death by each static lung volume test are summarized in Figure 2. In the univariate models, compared to patients with average lung volume (having RV and TLC in the 25–75 percentile), patients with good lung volume (in the 0–25 percentile for RV, RV%, RV/TLC, and RV/TLC%) had lower mortality rates, and patients with resting pulmonary hyperinflation (in the 75–100 percentile for RV, RV%, RV/TLC, and RV/TLC%) had higher mortality rates. The associations of RV/TLC and RV/TLC% with survival remained significant after controlling for age, sex, BMI, smoking status, NSCLC histological subtypes, clinical stages, lung cancer treatments, and their time-varying effects. Similarly, patients with poor TLC and TLC% (in the 0–25 percentile) had higher mortality rates, while those in the 75–100 percentile had lower mortality rates.

Among all measures, RV/TLC showed the strongest dose-dependent association with overall survival, where higher values were associated with worse outcomes. In unadjusted models, patients in the 75–100th percentile for RV/TLC had a 34.3% higher mortality risk compared to those in the 25–75th percentile (HR = 1.343, 95% CI: [1.188, 1.518]), whereas those in the 0–25th percentile had a 28.2% lower mortality risk (HR = 0.718, 95% CI: [0.623, 0.829]). After covariate adjustment, patients in the 75–100th percentile had a 17.6% higher risk of death compared to those in the 25–75th percentile (adjusted HR = 1.176, 95% CI: [1.014, 1.364]), whereas patients in the 0–25th percentile had a 20.3% lower risk of death (adjusted HR = 0.797, 95% CI: [0.672, 0.946]). In contrast, after covariate adjustment, the associations of other lung function measures (RV, RV%, TLC, TLC%) with overall survival were no longer statistically significant (Figures 2 and S3).

Consistent with the survival analysis based on percentile groups, a 1 SD increment of RV, RV%, RV/TLC, and RV/TLC% was associated with higher mortality rates, and a 1 SD increment of TLC and TLC% was associated with lower mortality rates (Figure 2). The strongest association was observed for RV/TLC, where a 1 SD increase was associated with a 19.2% increase in mortality risk after covariate adjustment (adjusted HR = 1.192, 95% CI: [1.114, 1.277]).

3.3 RV/TLC Interacts With FEV₁% and FVC% on Mortality Risks

One SD increment of FEV₁, FEV₁%, FVC, and FVC% was all significantly associated with lower mortality rates in univariate and multiple regression models (Figure S4). Given the correlations between RV/TLC and spirometry tests and the strong association between RV/TLC and overall survival, we further explored how RV/TLC interacts with spirometry tests in predicting mortality risks in NSCLC patients.

Significant interactions between RV/TLC and FEV₁% (adjusted interaction HR = 1.087, 95% CI: [1.018, 1.160]) as well as FVC% (adjusted interaction HR = 1.096, 95% CI: [1.030, 1.168]) were found, suggesting a higher mortality risk with increased RV/TLC when the spirometry test remained consistent. Interestingly, better spirometry tests remained significant predictors of improved overall survival in the interaction models, where the mortality rates were compared per one SD increment of each spirometry test among patients with median values of RV/TLC (Figure S4).

3.4 Prognostic Values of RV/TLC Are Independent of Spirometry and Cancer Stage

Given the significant interactions found between spirometry and RV/TLC, we explored further the prognostic values of RV/TLC in patients with abnormal spirometry (FEV₁/FVC<LLN of GLI reference values) and in patients within the range of normal spirometry (FEV₁/FVC≥LLN) separately, especially among the latter who might be misclassified as at lower risks of death if relying solely on spirometry (Figure 3). Among patients with poor spirometry results, patients having less air trapping defined by RV/TLC (in the 0–25 percentile) had lower mortality rates and better overall survival (median survival = 70.4, 95% CI: [52.3, 98.6] months). And among patients within the range of predicted normal spirometry (FEV₁/FVC≥LLN), those with hyperinflation at RV defined by RV/TLC (in the 75–100 percentile) had higher mortality rates and worse overall survival (median survival = 55.8, 95% CI: [44.8, 67.0] months). Moreover, one SD increment of RV/TLC was significantly associated with worse overall survival among both patients with both good and poor spirometry (adjusted HR =1.189 [95% CI: 1.093, 1.293] and 1.241 [95% CI: 1.090, 1.413], respectively). These findings presented the consistent values of RV/TLC in predicting the risk of death regardless of spirometry test results.

Next, we evaluated the prognostic role of RV/TLC in patients with NSCLC diagnosed at early stages and at advanced stages (Figure 4). The median survival was 86.2 months (95% CI: 78.0, 93.6) for stage I-II patients and 28.1 months (95% CI: 24.9, 34.2) for stage III-IV patients. Among both early-stage (log-rank p < 0.0001) and advanced-stage (log-rank p = 0.037) patients, survival outcomes differed significantly across RV/TLC percentile groups. In stage I-II patients, those in the 0–25th percentile had the best survival (median = 132.6, 95% CI: [106.2, 175.6] months), whereas those in the 75–100th percentile had the poorest survival (median = 58.7, 95% CI: [52.0, 69.9] months). A similar trend was observed in stage III–IV patients, with a median survival of 33.6 months (95% CI: 24.6, 52.3) in the 0–25th percentile group, compared to 24.9 months (95% CI: 17.2, 37.5) in the 75–100th percentile group. After covariate adjustment, absence of air trapping with RV/TLC in the 0–25 percentile of both groups was associated significantly with lower mortality rates. Specifically, stage I-II patients in the 0–25th percentile had a 32.6% lower mortality risk (adjusted HR = 0.674, 95% CI: [0.540, 0.842]), while stage III–IV patients in the same group had a 36.8% lower mortality risk (adjusted HR = 0.632, 95% CI: [0.455, 0.877]). One SD increment of RV/TLC was associated with a 17.9% increase in mortality risk for stage I–II patients (adjusted HR = 1.179, 95% CI: [1.085, 1.281]) and an 18.2% increase in mortality risk for stage III–IV patients (adjusted HR = 1.182, 95% CI: [1.046, 1.336]).

3.5 Additional Findings

Univariate analyses of both DLCO (HR = 1.709 [95% CI: 1.500, 1.947] and HR = 0.720 [95% CI: 0.619, 0.838], for 0–25 and 75–100 percentiles vs. 25–75 percentile, respectively) and DLCO% (HR = 1.840 [95% CI: 1.617, 2.093] and HR = 0.696 [95% CI: 0.598, 0.811], for 0–25 and 75–100 percentiles vs. 25–75 percentile, respectively) showed a dose-dependent association with overall survival, while the relationship of DLCO% became less clear after adjusting for covariates in multiple regression.

3.6 Sensitivity Analysis

Model diagnostics using the Deviance Residuals suggested adequate goodness-of-fit of these multiple regression models (Figure S5). The C-indexes were similar (range: 0.958–0.961) for all multiple regression models. The imputation-based results agreed with those of the complete-case analyses with slightly attenuated associations, which were expected as imputed values are prone to measurement errors and might thus reduce the predictive power of models [29].

Compared with NSCLC patients without reproducible lung function tests in the BLCS cohort, patients in our analytical cohort were older, with higher BMI, more likely to be smokers, less likely to have NSCLC of adenocarcinoma, and more likely to be in earlier stages and receive surgical treatment (Table S1). As a result, overall survival was lower in patients without reproducible tests in the BLCS cohort (median survival [95% CI]: 32.3 [29.9, 35.2] vs. 67.0 [60.4, 73.1] months), but the difference became insignificant after adjusting for covariates used in the multiple regression analyses (HR comparing patients with and without reproducible tests [95% CI] = 0.930 [0.852, 1.015]). Moreover, the source of lung function test records (inside vs. outside of MGH) was not associated with overall survival in the multiple regression models.