Validation and update of the thoracic surgery scoring system (Thoracoscore) risk model

Abstract

OBJECTIVES The performance of prediction models tends to deteriorate over time. The purpose of this study was to update the Thoracoscore risk prediction model with recent data from the Epithor nationwide thoracic surgery database. METHODS From January 2016 to December 2017, a total of 56 279 patients were operated on for mediastinal, pleural, chest wall or lung disease. We used 3 recommended methods to update the Thoracoscore prediction model and then proceeded to develop a new risk model. Thirty-day hospital mortality included patients who died within the first 30 days of the operation and those who died later during the same hospital stay. RESULTS We compared the baseline patient characteristics in the original data used to develop the Thoracoscore prediction model and the validation data. The age distribution was different, with specifically more patients older than 65 years in the validation group. Video-assisted thoracoscopy accounted for 47% of surgeries in the validation group compared but only 18% in the original data. The calibration curve used to update the Thoracoscore confirmed the overfitting of the 3 methods. The Hosmer–Lemeshow goodness-of-fit test was significant for the 3 updated models. Some coefficients were overfitted (American Society of Anesthesiologists score, performance status and procedure class) in the validation data. The new risk model has a correct calibration as indicated by the Hosmer–Lemeshow goodness-of-fit test, which was non-significant. The C-index was strong for the new risk model (0.84), confirming the ability of the new risk model to differentiate patients with and without the outcome. Internal validation shows no overfitting for the new model CONCLUSIONS The new Thoracoscore risk model has improved performance and good calibration, making it appropriate for use in current clinical practice.