Prolonged air leak after video-assisted thoracic anatomical pulmonary resections: a clinical predicting model based on data from the Italian VATS group registry, a machine learning approach

Abstract

Background: Prolonged air leak (PAL) is a frequent complication after lung resection surgery and has a high clinical and economic impact. A useful risk predictor model can help recognize those patients who might benefit from additional preventive procedures. Currently, no risk model has sufficient discriminatory capacity to be used in common clinical practice. The aim of this study is to identify predictive risk factors for PAL after video-assisted thoracoscopic surgery (VATS) anatomical resections in the Italian VATS group database and to evaluate their clinical and statistical performance.

Methods: We processed data collected in the second edition of the Italian VATS group registry. It includes patients that underwent a thoracoscopic anatomical resection for benign or malignant diseases, between November 2015 and December 2020. We used recursive feature elimination (RFE), using a backward selection process, to find the optimal combination of predictors. The study population was randomly split based on the outcome into a derivation (80%) and an internal validation cohort (20%). Discrimination of the model was measured using the area under the curve, or C-statistic. Calibration was displayed using a calibration plot and was measured using Emax and Eavg, the maximum and the average difference in predicted versus loess calibrated probabilities.

Results: A cohort of 6,236 patients was eligible for the study after application of the exclusion criteria. Five-day PAL rate in this patient cohort was 11.3%. For the construction of our predictive model, we used both preoperative and intraoperative variables, with a total of 320 variables. The presence of variables with missing values greater than 5% led to 120 remaining predictors. RFE algorithm recommended 8 features for the model that are relevant in predicting the target variable.

Conclusions: We confirmed significant prognostic risk factors for the prediction of PAL: decreased DLCO/VA ratio, longer duration of surgery, male sex, the need for adhesiolysis, COPD, and right side. We identified middle lobe resections and ground glass opacity as protective factors. After internal validation, a C statistic of 0.63 was revealed, which is too low to generate a reliable score in clinical practice.

Keywords: Prolonged air leak (PAL); risk factors; risk predictive model; video-assisted thoracoscopic surgery lobectomy (VATS lobectomy).

Figure 1

Patient’s algorithm. VATS, video-assisted thoracoscopic surgery.

Figure 2

Graphical representation of the performance profile of the different models across different subset sizes. RFE algorithm allows to examine variable importance for the selected features. The blue dot represents the optimal solution. ROC, receiver operating characteristic; RFE, recursive feature elimination.

Figure 3

Measures of importance. When using machine learning models, it is important to understand which predictors are more influential on the outcome variable. The area under the ROC curve is conducted on each predictor and used as the measure of variable importance. All measures of importance are scaled to have a maximum value of 1 and expressed as percentages. DLCO/VA, diffusing capacity of the lung for carbon monoxide divided by alveolar volume; COPD, chronic obstructive pulmonary disease; ROC, receiver operating characteristic.

Figure 4

Forest plot of the logistic regression coefficients of the optimal model selected by RFE. The variables are sorted from top to bottom in order of importance and coefficient values greater than one indicate increased risk, while values less than one indicate a protective effect. DLCO/VA, diffusing capacity of the lung for carbon monoxide divided by alveolar volume; COPD, chronic obstructive pulmonary disease; RFE, recursive feature elimination.

Figure 5

Calibration and distribution plot of our models. Black dots and their line ranges denote the observed probability for each decile of predicted risk, with their associated 95% CIs. 45-degree solid black line indicates perfect calibration. Dashed black line indicates the best-fitting straight line through the estimates (linear regression). Solid red line indicates the best-fitting curved line through the estimates (loess regression).