Original Article
Predictive models of malignant transudative pleural effusions
Abstract
Background: There are no firm recommendations when cytology should be performed in pleural transudates, since some malignant pleural effusions (MPEs) behave biochemically as transudates. The objective was to assess when would be justified to perform cytology on pleural transudates.
Methods: Consecutive patients with transudative pleural effusion (PE) were enrolled and divided in two groups: malignant and non-MPE. Logistic regression analysis was used to estimate the probability of malignancy. Two prognostic models were considered: (I) clinical-radiological variables; and (II) combination of clinical-radiological and analytical variables. Calibration and discrimination [receiver operating characteristics (ROC) curves and area under the curve (AUC)] were performed.
Results: A total of 281 pleural transudates were included: 26 malignant and 255 non-malignant. The AUC obtained with Model 1 (left PE, radiological images compatible with malignancy, absence of dyspnea, and serosanguinous appearance of the fluid), and Model 2 (the variables of Model 1 plus CEA) were 0.973 and 0.995, respectively. Although no false negatives are found in Models 1 and 2 to probabilities of 11% and 14%, respectively, by applying bootstrapping techniques to not find false negatives in 95% of other possible samples would require lowering the cut-off points for the aforementioned probabilities to 3% (Model 1) and 4% (Model 2), respectively. The false positive results are 32 (Model 1) and 18 (Model 2), with no false negatives.
Conclusions: The applied models have a high discriminative ability to predict when a transudative PE may be of neoplastic origin, being superior to adding an analytical variable to the clinic-radiological variables.
Methods: Consecutive patients with transudative pleural effusion (PE) were enrolled and divided in two groups: malignant and non-MPE. Logistic regression analysis was used to estimate the probability of malignancy. Two prognostic models were considered: (I) clinical-radiological variables; and (II) combination of clinical-radiological and analytical variables. Calibration and discrimination [receiver operating characteristics (ROC) curves and area under the curve (AUC)] were performed.
Results: A total of 281 pleural transudates were included: 26 malignant and 255 non-malignant. The AUC obtained with Model 1 (left PE, radiological images compatible with malignancy, absence of dyspnea, and serosanguinous appearance of the fluid), and Model 2 (the variables of Model 1 plus CEA) were 0.973 and 0.995, respectively. Although no false negatives are found in Models 1 and 2 to probabilities of 11% and 14%, respectively, by applying bootstrapping techniques to not find false negatives in 95% of other possible samples would require lowering the cut-off points for the aforementioned probabilities to 3% (Model 1) and 4% (Model 2), respectively. The false positive results are 32 (Model 1) and 18 (Model 2), with no false negatives.
Conclusions: The applied models have a high discriminative ability to predict when a transudative PE may be of neoplastic origin, being superior to adding an analytical variable to the clinic-radiological variables.
