We retrospectively analyzed 3,009 patients treated with thoracic surgery from the Thoracic Perioperative Database for Geriatrics in the First Medical Center of Chinese PLA General Hospital (Beijing), and we finally enrolled 1,585 elderly patients with lung cancer treated with video-assisted thoracoscopic surgery between January 2012 and December 2019. The Thoracic Perioperative Database for Geriatrics was established at the First Medical Center of Chinese PLA General Hospital using the WebService interface. It automatically extracted data from the electronic medical record system, clinical laboratory system, medical image management system, radiation information management system, blood transfusion management system, and nurse workstation. Patient’s diagnoses, examinations, prescriptions, laboratory tests, medical insurances, and operation information were extracted. Then, all data (100%) were rechecked and reconfirmed by two medical workers at least and contradictory data were reviewed and confirmed after discussions. Data standardization complied with the Clinical Data Interchange Standards Consortium (CDISC). In the present study, patients were included if they were elderly patients (age≧65 years) with tissue-proved diagnosis of lung cancer and treated with video-assisted thoracoscopic surgery. Patients were excluded in the analysis, if (1) patients had the same admission ID, (2) received thoracic surgery due to esophageal cancer, gastric antral cancer, bullae of lung, and other reasons, (3) treated with open thoracotomy, (4) preoperative pneumonia confirmed by computed tomography (CT), and (5) missing data. Figure 1 shows the patient’s flow chart.

The enrolled patients were randomly classified into two groups: namely, the training group (n = 793) and the validation group (n = 792). The randomization of patients was computer-based and achieved with the ratio of 1:1, which was consistent with other studies (17–19). Patients in the training group were conducted to develop the algorithm and patients in the validation group were used to internally validate the predictive performance of the algorithm. External validation of the algorithm was achieved in the external validation dataset after enrolling 165 patients from the Hainan Hospital of Chinese PLA General Hospital (Sanya, Hainan) and Xiangya Hospital of the Central South University (Changsha, Hunan) between January 2019 and September 2021 according to the inclusive and exclusive criteria.

The study was conducted in accordance with the Declaration of Helsinki. The study was approved by the Ethics Committee Board of the First Medical Center of Chinese PLA General Hospital (No. S2019-311-03) and patient’s written consents were waived because all data were anonymized and the study was retrospective in nature.

The main surgical indications were as follows: (1) lung cancer at stage I, II, and III; (2) no cancer involvement in the chest wall and mediastinum; (3) no recent myocardial infarction and tendency of severe bleeding; and (4) tolerable to single lung ventilation. Generally, patients received general anesthesia and then were placed in lateral position. Operation was conducted with video-assisted thoracoscopic surgery, which was performed with one access window and one operation port. After surgery, a closed drainage was placed at the eighth intercostal space and a vacuum sealing drainage was placed at the operation port.

Based on the literature and available variables in the dataset, we collected 30 potential characteristics and the 30 variables could be divided into (1) basic characteristics including age, sex (male vs. female), body mass index (BMI, kg/m²), smoking (no vs. yes), and alcohol drinker (no vs. yes); (2) comorbidities including hypertension (yes vs. no), diabetes (yes vs. no), coronary heart disease (yes vs. no), myocardial infarction (yes vs. no), cerebrovascular disease (yes vs. no), asthma (yes vs. no), chronic obstructive pulmonary disease (COPD) (yes vs. no), renal function insufficiency (yes vs. no), and peptic ulcer (yes vs. no); (3) laboratory testing including serum albumin (g/L), leukocyte count (×10⁹/L), red blood cell volume distribution width (RDW, %), serum glucose (mmol/L), blood urea nitrogen (mmol/L), blood potassium (mEq/L), blood sodium (mEq/L), serum creatinine (μmol/L), and total bilirubin (μmol/L); (4) preoperative medication history including angiotensin-converting enzyme inhibitor (ACEI) drug (yes vs. no), statin drug (yes vs. no), and calcium channel blocker (yes vs. no); and (5) variables immediately before or during surgery including preoperative mean artery pressure (MAP) (mmHg), surgery duration (minutes, min), intraoperative administration of colloid (ml) (0 vs. >0 and ≦500 vs. >500), and intraoperative administration of hormone (yes vs. no). All laboratory testing data were collected before surgery. MAP was evaluated just before surgery. Surgery duration was defined as the time interval between skin incision and suture. Hormone injection included dexamethasone and methylprednisolone. Besides, hospitalization expenses of all patients were also recorded in the study. Hospitalization expense indicated all costs that were spent to treat the disease during the hospitalization. Hospitalization expense was converted from renminbi to dollar according to the current exchange rate. We compared the patient’s basic and clinical demographics according to the presence of postoperative pneumonia.

Postoperative pneumonia refers to abnormal radiographic findings including new infiltrations being found in postoperative chest x-ray or CT and patients had the following one or more symptoms postoperatively during hospitalization (20): (1) patients had new or progressive respiratory symptoms, such as coughing and expectoration; (2) patients had a vital sign of fever >38°C or hypothermia; (3) patients had a physical examination of lung consolidation signs or moist rale; (4) patients had a white cell count of more than 10 × 10⁹/L or less than 4 × 10⁹/L or a new rise in the C-reactive; and (5) patients had pathogen isolation derived from blood culture or sputum.

In the training group, we used the Least Absolute Shrinkage and Selection Operator (LASSO) logistic regression model to screen the above-mentioned 30 variables. The LASSO method was a penalized regression tool with the ability of selecting variables via minimizing the comparatively irrelevant variables’ coefficient to 0 and subsequently excluding unwarranted variables. The 10-fold cross-validation was conducted to adjust the parameter λ to further improve the accuracy of the LASSO regression model. The LASSO-identified variables were included to develop the algorithm. The estimates of the algorithm and corresponding intercept were obtained from the maximum likelihood estimates according to the multiple logistic regression analysis and then the algorithm was created based on the logistic regression equation: P(Y=1) = e^{intercept + ax1 + bx2 + cx3 …+ ixn} /(1 + e^{intercept + ax1 + bx2 + cx3 …+ ixn} ). In the algorithm, a to i were the maximum likelihood estimates and x1 to xn were the LASSO-identified significant variables. P(Y=1) refers to the predicted probability of developing postoperative pneumonia among elderly patients with lung cancer treated with video-assisted thoracoscopic surgery.

The internal validation of the predictive performance of the algorithm was conducted in the validation group and the training group, and the external validation was performed in the external validation dataset. The predictive ability of the algorithm was evaluated using the discrimination and calibration abilities.

The discrimination ability was the algorithm’s capability to identify patients with postoperative pneumonia and patients without it. The discrimination ability was measured by the AUROC and the discrimination slope. The discrimination slope was the difference between the mean predicted probabilities of developing postoperative pneumonia in patients with (positive) and without (negative) actual postoperative pneumonia. Besides, the algorithm’s correct classification rate, sensitivity, specificity, false-positive rate, and false-negative rate were calculated and presented in the study.

The calibration ability was the algorithm’s capability to confirm the consistency between the observed and the predicted probability of developing postoperative pneumonia. The calibration ability was measured by the calibration slope and the R ² which was calculated by the goodness-of-fit test. Calibration curve refers to the decile plot of the observed probability of developing postoperative pneumonia against the predicted probability of postoperative pneumonia in each decile. If the slope of the calibration curve is close to 1 and intercept is close to 0, it indicates good calibration ability. The R² of the fitted line in the calibration curve was also presented in the study based on the goodness-of-fit test.

We presented statistical descriptions of all patients. If the continuous data were normally distributed, they were shown as mean ± standard deviation (SD); if the continuous data were not normally distributed, they were shown as median and interquartile range (IQR). The categorical data were presented in proportion. The Wilcoxon test, Chi-square test, and continuity adjusted Chi-square test were used to compare differences of characteristics between patients with and without postoperative pneumonia. The Pearson correlation analysis was used to analyze the correlation between the seven variables and the main prognostic parameter (postoperative pneumonia), and the correlation matrix of the Pearson correlation coefficients was visualized by the heat map. Three different risk groups were achieved according to the cutoff of the algorithm and the Chi-square test was used to compare the difference between the actual probabilities of developing postoperative pneumonia among the three risk groups. Principal components analysis was performed to explore potential major contributors to postoperative pneumonia and principal equations were given according to eigenvectors. All analyses were performed using SAS 9.4 software (SAS Institute Inc., Cary, NC) and the heat map was visualized using R version 3.5.3 for Windows XP.

Of all the included patients in the Thoracic Perioperative Database for Geriatrics, 9.15% (145/1,585) of patients developed postoperative pneumonia. Subgroup analysis of patients was performed based on the presence of postoperative pneumonia ( Table 1 ). In the study, patients with postoperative pneumonia were more likely to be male (68.97% vs. 56.18%, p = 0.03) and smoker (46.21% vs. 37.78%, p = 0.047) as compared to patients without it. Besides, patients with postoperative pneumonia had a higher level of preoperative leukocyte count (×10⁹/L) (6.36 [5.20, 7.40] vs. 5.92 [4.97, 6.97], p = 0.009) and serum creatinine (μmol/L) (75.30 [63.60, 84.40] vs. 71.70 [61.30, 81.30], p = 0.041), a higher rate of preoperative calcium channel blocker usage (35.17% vs. 25.49%, p = 0.012), and a longer surgery duration (min) (175.00 [145.00, 223.00] vs. 150.00 [113.00, 187.00], p < 0.001) than those without it. Regarding intraoperative drug administration, hormones were administered more among the patients with postoperative pneumonia than that among the patients without postoperative pneumonia (86.21% vs. 74.44%, p = 0.002) and so was colloid (p < 0.001). Patients with postoperative pneumonia had a significantly higher hospitalization expense as compared with those without postoperative pneumonia ($) (11,537.93 [9,488.58, 14,088.16] vs. 10,069.85 [8,297.86, 11,968.01], p < 0.001).

In the training group, the LASSO regression model (10-fold cross-validation) selected five variables, including sex, surgery duration, leukocyte count, intraoperative administration of colloid, and intraoperative administration of hormone, all of which were included to develop the algorithm. Besides, according to previous studies, we also added smoking and history of COPD into the algorithm. The Pearson correlation analysis further confirmed the relationship between the seven variables and the outcome variable, and the correlation matrix of the Pearson correlation coefficients was presented by heat map ( Figure 2 ). It further demonstrated that the five LASSO-selected variables were significantly and positively related to postoperative pneumonia, which meant that male sex, longer surgery duration, higher leukocyte count, intraoperative use of colloid, and intraoperative use of hormone were associated with a higher risk of developing postoperative pneumonia.

According to the estimates of each enrolled variable from the multiple logistic regression, the equation was shown as follows ( Table 2 ). The predicted probability of postoperative pneumonia was P (Y=1) = e^x/(1+ e^x), _X= -6.03 + 0.60*sex + 0.01*smoking + 0.57*history of COPD + 0.006*surgery duration + 0.16*leukocyte count + 0.31*intraoperative administration of colloid + 0.86*intraoperative administration of hormone. A calculator was created to promote the application of the algorithm ( Supplementary File 1 ). An example about how to use the algorithm was given as follows: If a male (1 point) elderly patient with lung cancer is treated with video-assisted thoracoscopic surgery, if he was a smoker (1 point) with a medical history of COPD (1 point), if the surgical duration was 120 min and his preoperative leukocyte count was 9.09×10⁹/L, and if he received an intraoperative injection of colloid (500 ml, 2 points) and hormone (1 point), his predicted probability of developing postoperative pneumonia was P (Y=1) = e^x /(1+e ^x) =e^x /(1+ e^x ) =23.23% (X=-6.03 + 0.60 * 1 + 0.01 * 1 + 0.57 * 1 + 0.006 * 120 + 0.16 * 9.09 + 0.31 * 2 + 0.86 * 1=-1.1956).

The predictive performance of the algorithm was validated using the discrimination and calibration ability in the training, internal validation, and external validation dataset. In the external validation dataset, 10.30% (17/165) of patients had postoperative pneumonia. The C-index from the AUROC was 0.70 in the training group ( Figure 3A ), 0.67 in the internal validation group ( Figure 3B ), and 0.71 in the external validation dataset ( Figure 3C ). The discrimination slopes were 0.06 (95% CI: 0.04-0.07) in the training group ( Figure 4A ), 0.05 (95% CI: 0.03-0.06) in the internal validation group ( Figure 4B ), and 0.08 (95% CI: 0.04-0.12) in the external validation dataset ( Figure 4C ). The corresponding correct classification rates were 64.10%, 61.50%, and 78.80%, respectively. More details are presented in Table 3 , which also showed that the algorithm had acceptable sensitivity and specificity and low false-negative rate.

Considering the calibration ability of the algorithm, the calibration slopes were 0.88 (95% CI: 0.37-1.39) in the training group ( Figure 5A ), 0.90 (95% CI: 0.46-1.34) in the internal validation group ( Figure 5B ), and 1.03 (95% CI: 0.24-1.83) in the external validation dataset ( Figure 5C and Table 4 ). The X-intercept and Y-intercept of the algorithm were all close to 0 in the three datasets. The R ² from the goodness-of-fit test showed the fitting of the line was good.

A predicted probability of developing postoperative pneumonia was capable of being calculated according to the algorithm in each patient. Then, three risk groups were created in the Thoracic Perioperative Database for Geriatrics according to the best cutoff values in the training group (cutoff = 8.00%), the internal validation group (cutoff=10.00%), and the external validation group (cutoff = 14.00%). The average cutoff value of the three groups was about 10.00%; thus, a predicted probability of less than 10.00% was classified into the low-risk group, a predicted probability of 10.00% to 20.00% belonged to the medium-risk group, and a predicted probability of more than 20.00% (twice the average prevalence) was the high-risk group ( Table 5 ). The corresponding mean predicted probabilities were 5.65%, 13.54%, and 27.69% in the three risk groups and the corresponding mean actual probabilities were 5.14% (53/1031), 15.07% (71/471), and 25.30% (21/83), respectively (p < 0.001, Chi-square test). The predicted probabilities were similar to the actual probabilities among the three groups, which further confirmed the consistency and reproducibility of the algorithm.

Principal components analysis was performed to explore major contributors to postoperative pneumonia and principal equations were given according to eigenvectors ( Supplementary File 2 ). In the training group, seven principal components were developed and the cumulative contribution rate was 59.08% for the first three principal components, which indicated that the first three principal components were responsible for 59.08% information of all seven variables. According to the coefficients of the first three principal components, we might speculate that sex (coefficient = 0.62), intraoperative administration of colloids (coefficient = 0.67), and intraoperative administration of hormones (coefficient = 0.71) might be the major contributors to postoperative pneumonia since the three variables had the largest coefficient in the corresponding principal equations. Similar trends were observed in the validation group.