Table 1.
Number of cases in the Training and Testing groups; pre-test probability for 1) local prevalence of tuberculous pleural effusion and 2) cases included in the machine learning analysis (that is, considering only the cases with exudative and lymphocytic effusion samples).
Table 2.
Comparative analysis of categorical variables age and ADA level in the reference (Training) and external validation (Testing) groups (considering p < 0.05 indicative of a statistically significant difference).
Fig 1.
Distribution of “Tuberculosis” and “others” samples in terms of ADA and age in Training (Gipuzkoa 2013-2022, left) and Testing (Bajo Deba 1996-2012, right) groups.
Fig 2.
Confusion matrices in cross-validation (Gipuzkoa, left) and in testing (Bajo Deba, right).
Table 3.
Accuracy, sensitivity, specificity, and positive predictive value for each dataset for the trained machine learning models.
Fig 3.
Positive and negative predictive values as a function of pre-test probability.
Table 4.
Real and estimated positive predictive values in the training and testing datasets.
Fig 4.
Confusion matrices for each classifier in Gipuzkoa (left) and Bajo Deba (right) (0: tuberculous, 1: malignant, 2: others; predicted values in columns, real values in rows).
Fig 5.
Tuberculous, malignant, and other cases as a function of adenosine deaminase level and age in Gipuzkoa (left) and in Bajo Deba (right).
False negatives produced by the support vector classifier in the Bajo Deba dataset are marked with red crosses.