Validação externa de modelos de predição de pneumonia pós cirurgia cardíaca

Abstract

This study concerns the external validation of a prediction model for pneumonias after cardiac surgery. It also presents a review of methods and techniques for critical appraisal and performance assessment of clinical predictive models and nomogram applications. Methods: A consecutive series of 527 patients who underwent cardiac surgeries between June 2000 and August 2002 was used to develop a prognostic model.. Multiple logistic regression analysis was performed to predict the occurrence of pneumonia. Diverse simple and conventional preoperative risk factors were evaluated. The model was internal validated with bootstrap. A nomogram was developed to enhance clinical applicability. The performance was evaluated by calibration, discrimination and global measures. Prospective study was done to validate models predicting pneumonia after cardiac surgery with 333 adult patients who underwent cardiac surgery from October 2006 to May 2007. Previously constructed logistic regression (LRM) and classification and regression tree (CART) models were validated with external data. Results: a simple nomogram model was developed and showing low discrimination and good calibration (AUC 0.79, Brier score 0.064, discrimination slope 0.13, Hosmer-Lemeshow p=0.27). Pneumonia occurred in 7.5% of patients in the external validation set. LRM performed better with moderate discrimination (R2 7.1%, Brier=0.06, AROC=0.694) and calibration (Hosmer-Lemeshow P=0.08). Conclusions: Overall agreement between the predicted probabilities and observed frequencies was good in the development and the internal validation set. The nomogram predicts the probability of pneumonia for individual patients and may help in informing patients and surgeons before undergoing cardiac surgery. We validated a model that can identify which patients undergoing cardiac surgery are at high risk for pneumonia. CART performs well in derivation, and looses more discrimination and calibration than LRM in the validation set.