Classification tree analysis for the discrimination of pleural exudates and transudates

Background: Classification and regression tree (CART) analysis is a non-parametric technique suitable for the generation of clinical decision rules. We have studied the performance of CART analysis in the separation of pleural exudates and transudates. Methods: Basic demographic, radiologic and laboratory data were retrospectively evaluated in 1257 pleural effusions (204 transudates and 1053 exudates, according to standard clinical criteria) and submitted for CART analysis. The model's discriminative ability was compared with that of Light's criteria, in both the original formulation and an abbreviated version, i.e., deleting the pleural fluid (PF)/serum lactate dehydrogenase (LDH) ratio from the triad. Results: A first CART model built starting from all available data identified PF/serum protein ratio and PF LDH ratios as the two best discriminatory parameters. This algorithm achieved a sensitivity of 96.8%, slightly lower than that of classical Light's criteria (98.5%) and comparable to that of the abbreviated Light's criteria (97.0%), and significantly better specificity (85.3%) compared to both classical (74.0%) and abbreviated (79.4%) Light's criteria. A second CART model developed after excluding serum measurements selected PF protein and PF LDH as the most discriminatory variables, and correctly classified 97.2% of exudates and 77.0% of transudates. Conclusions: CART-based algorithms can efficiently discriminate between pleural exudates and transudates.