Patients treated between January 2021 and August 2022 at the Department of Thoracic Surgery, Shandong Provincial Hospital Affiliated to Shandong University, and Qilu Hospital of Shandong University were screened if applicable. The primary outcomes were pulmonary complications within 30 days after lung resection. As a composite outcome measure, PPCs included prolonged air leak, pneumonia, pneumothorax, pleural effusion, atelectasis, acute respiratory distress syndrome, bronchopleural fistula, empyema, and respiratory failure (19). Pulmonary complications were assessed using the Clavien-Dindo classification, and the Clavien-Dindo grade ≥ 3 was considered a major complication (20, 21). The participants of this study were patients aged 60 years and older with stage I and II lung cancers. The histological classification criteria were based on the World Health Organization classification of lung tumors, and participants with adenocarcinoma or squamous cell carcinoma of the lungs were included (22). Patients were also selected if they intended to undergo video-assisted thoracoscopic surgery for lobectomy (sleeve lobectomy and bi-lobectomy) or sub-lobar resection (segmentectomy and wedge resection). Patients with contraindications to anesthesia or lung resection and those with severe cardiopulmonary impairment were excluded. Additionally, patients with a history of COVID-19 pneumonia were excluded. Patients with solid metastatic tumors, including in the lymph nodes, were also excluded from the study. Patients and outpatients with incomplete or missing data were not considered. Figure 1 presents an overview of the inclusion and exclusion criteria. This study conformed to the principles of the Declaration of Helsinki. Because of the anonymity of the data, all cases were exempt from the requirement for individual informed consent. The Institutional Review Board of the School of Nursing and Rehabilitation, Shandong University, approved this study (IRB no. 2022-R-107).

Eligible elderly patients with lung cancer were enrolled in this retrospective observational study. Data on the perioperative characteristics were collected. Pre- and intraoperative data include sex, age, smoking history, hypertension, chronic obstructive pulmonary disease (COPD), coronary artery disease, diabetes mellitus, histopathology, tumor location, clinical stage, albumin, hemoglobin, surgical procedures, time of operation, blood loss, forced vital capacity (FVC), forced expiratory volume in one second (FEV1), carbon monoxide diffusing capacity of the lung (DLCO), and so on. All required data were achieved manually from the electronic medical records system. Most variables had no missing data and albumin and hemoglobin levels had a small amount of missing data (<1%), except for FVC (2.2%), FEV1 (10.5%), and DLCO (11.5%). The series mean was used to substitute the missing values for albumin and hemoglobin. Multiple imputations were used to measure FVC, FEV1, and DLCO. Continuous variables were grouped according to actual clinical practice, and categorical variables were encoded prior to logistic regression ( Supplementary Table 1 ). Older adults with pulmonary complications were considered “cases” and, if not, “controls”.

The sample size was calculated based on ten events per variable (23), being widely recommended as the minimum standard sample size for logistic regression analysis. In a pre-survey of 50 elderly patients undergoing VATS, PPCs were present in 24% of the patients, comparable to previous studies (13%-33%). This study included 18 pre- and intraoperative variables; therefore, the final sample consisted of 750 elderly patients with lung cancer.

Statistical analyses were performed using R for Windows (version 4.2.1; TUNA Team, Tsinghua University, China) and IBM SPSS Statistics software version 27.0. Categorical variables were expressed as frequency (percentage). Continuous variables were expressed as the mean and standard deviation (SD) when distributed normally and as the median and interquartile range (IQR) when non-normally. Pearson’s chi-squared and Fisher’s exact tests (two-sided) were used for statistical tests of categorical variables, Student’s t-test for continuous variables after assessing the normality of data, and the Mann-Whitney U test for non-normal distributions.

Potential correlations between pre- and intraoperative variables were determined using Spearman’s rank correlation. Logistic regression was utilized to analyze the risk factors of pulmonary complications after surgery in 60-year-olds and older with lung cancer. Each variable was initially assessed using univariate analysis. Statistically significant variables (p<0.05) were then entered into further multivariate logistic regression analysis. Model construction refers to type 2b transparent reporting of a multivariable prediction model for individual prognosis or diagnosis statement (24). Data from 659 eligible elderly patients with lung cancer from Shandong Provincial Hospital Affiliated to Shandong University were collected for model building, and the data from 195 patients from Qilu Hospital of Shandong University were collected for validation to examine the fitting effect and extrapolation of the predictive model.

We adopted four criteria to evaluate the model performance. First, the receiver operating characteristic (ROC) curve was applied for model discrimination by calculating the area under the ROC curve (AUC) of the test group. Second, a calibration curve was utilized to assess the model calibration, with an outstanding prediction appearing at 45 degrees. Third, the Hosmer-Lemeshow test was employed for goodness-of-fit assessment, indicating an overall fit of the model (p>0.05). Fourth, decision curve analysis (DCA) was used to characterize the clinical usefulness and threshold probabilities of the predictive model.

The model was validated utilizing two methods to describe the generalization capability of the predictive model. In the first method, cross-validation was carried out with ten-fold cross-validation. Data from the model building were partitioned into ten equal-sized subsets, and each fold was used as the validation set, and the remaining folds were used as the derived set. Ten-fold cross-validation was performed to avoid overestimating the incremental values because the reported values were the average of running ten-fold cross-validation. In the second method, external data verification was conducted to prevent overfitting and ensure robustness. Data from one center were used for model construction, and data from the other center were used for model validation to test the fit and extrapolation of the predictive model.

Among the 854 eligible older adults included in the analysis ( Supplementary Table 2 ), data from the Shandong Provincial Hospital Affiliated to Shandong University (n=659) were availed to train the model, and data from the Qilu Hospital of Shandong University (n=195) were availed to test the model. The training and validation sets were further subdivided into controls and cases, and the comparison data for all four sets are shown in Table 1 . The characteristics of the two groups are presented in Table 2 . The preoperative and intraoperative characteristics were similar between the training and the test groups, and none of these were significantly different. As shown in Table 3 , 157 patients had a Clavien-Dindo grade <3, and 142 patients had a Clavien-Dindo grade of 3-5. The most severe pulmonary complications occurred in the participants with multiple complications.

Before establishing the nomogram, the preoperative and intraoperative variables were tested for correlation using Spearman’s coefficient. All variables had a low correlation (Spearman’s rho coefficient<0.5; p<0.001) and were retained in the subsequent analysis ( Figure 2 ).

As shown in Table 4 , univariate logistic regression was performed to compare 508 controls and 151 cases in the training set. Age, sex, smoking status, coronary artery disease, diabetes mellitus, COPD, histology, surgical procedure, operative time, blood loss, albumin, FVC, FEV1, and DLCO correlated significantly with PPCs (p<0.05). Statistically significant variables were included for further analyses.

Multivariate regression analysis showed that age, COPD, surgical procedure, operative time, FEV1, and DLCO were independent predictors of PPCs. Patients over or equal to 70 years of age had a 264.5% increased risk of PPCs compared to patients between 60 and 69 years. Patients undergoing lobectomy had a 175.0% greater risk of developing PPCs compared with those undergoing sub-lobectomy. The odds ratio (OR) of operative time equal to or greater than 120 minutes to those less than 120 minutes is 3.704, and the odds ratio of COPD diagnosis to those without is 6.443. Besides, FEV1 (<80% vs.≥80%) was the risk factor of PPCs (OR=2.836), and DLCO (<80% vs.≥80%) was also the risk factor of PPCs (OR=3.753).

A nomogram was derived to predict the probability of PPCs in elderly patients with lung cancer ( Figure 3A ). After performing univariate and multivariate logistic regression analyses to eliminate potential redundancy, the predicted probability was calculated by summing the scores for each item using a nomogram.

Our dynamic nomogram is available online at https://pulmonary-disease-predictor.shinyapps.io/dynnomapp/as a clinically easy-to-use tool to assist surgeons ( Figure 3B ). Whether or not pulmonary complications have occurred can be expediently judged with little effort by reading information related to the clinical characteristics, and the network server generates the results.

Four methods are available for evaluating prediction performance. The first method used all data to perform ten-fold cross-validation to characterize the generalizability of the nomogram. An AUC value of 0.839 (sensitivity=0.949, specificity=0.450) was the average value assessed by ten-fold cross-validation. The second method employs the results of the multivariate regression analysis to calculate the area under the ROC curve. The AUC and 95% confidence interval (CI) of the training set was 0.844 (95%CI: 0.806-0.882, p<0.001, Figure 4A ), and the AUC value of the test set was 0.796 (95% CI: 0.722-0.870, p<0.001, Figure 4B ), both of which fully embody the high discrimination of the nomogram. The third method combines the calibration curve and the Hosmer-Lemeshow test to evaluate the accuracy of the nomogram. The calibration curve visualizes the Hosmer-Lemeshow goodness-of-fit test (p>0.05). Calibration curves were constructed to compare the predicted and observed probabilities ( Figures 4C, D ). The Hosmer-Lemeshow goodness-of-fit for the training set was 0.099, and the result for the test set was 0.112. The fourth method is DCA, which showed that the nomogram had superior net benefits, with threshold probabilities ranging from 4%-97% and 6%-71% in Figures 4E, F , respectively.