Predicting a prolonged air leak after video assisted thoracic surgery, is it really possible?

Validation of predictive risk models for prolonged air leak (PAL) is essential to understand if they can help to reduce its incidence and complications. This study aimed to evaluate both the clinical and statistical performances of 4 existing models. We selected 4 predictive PAL risk models based on their scientific relevance. We referred to these models as Chicago, Bordeaux, Leeds and Pittsburgh model, respectively, according to the affiliation place of the first author. These predicting risk models were retrospectively applied to patients recorded on the second edition of the Italian Video-Assisted Thoracoscopic Surgery Group registry. Predictions for each patient were calculated based on the logistic regression coefficient values provided in the original manuscripts. All models were tested for their overall performance, discrimination, and calibration. We recalibrated the original models with the re-estimation of the model intercept and slope. We used curve decision analysis to describe and compare the clinical effects of the studied risk mod els. Better statistical metrics characterize the models developed on larger populations (Chicago and Bordeaux models). However, no model has a valid benefit for threshold probability greater than 0.30. The Net benefit of the most performing model (Bordeaux model) at the threshold probability of 0.11 is 23 of 1000 patients, burdened by 333 false positive cases. One of 1000 is the Net benefit at the threshold probability of 0.3. The use of PAL scores based on preoperative predictive factors cannot be currently used in a clinical setting because of a high false positive rate and low positive pre dictive value