Identifying tuberculous pleural effusion using artificial intelligence machine learning algorithms

Abstract Background The differential diagnosis of tuberculous pleural effusion (TPE) is challenging. In recent years, artificial intelligence (AI) machine learning algorithms have started being used to an increasing extent in disease diagnosis due to the high level of efficiency, objectivity, and accuracy that they offer. Methods Data samples on 192 patients with TPE, 54 patients with parapneumonic pleural effusion (PPE), and 197 patients with malignant pleural effusion (MPE) were retrospectively collected. Based on 28 different features obtained via statistical analysis, TPE diagnostic models using four machine learning algorithms (MLAs), namely logistic regression, k-nearest neighbors (KNN), support vector machine (SVM) and random forest (RF) were established and their respective diagnostic performances were calculated. The respective diagnostic performances of each of the four algorithmic models were compared with that of pleural fluid adenosine deaminase (pfADA). Based on 12 features with the most significant impacts on the accuracy of the RF model, a new RF model was designed for clinical application. To demonstrate its external validity, a prospective study was conducted and the diagnostic performance of the RF model was calculated. Results The respective sensitivity and specificity of each of the four TPE diagnostic models were as follows: logistic regression – 80.5 and 84.8%; KNN– 78.6 and 86.6%; SVM – 83.2 and 85.9%; and RF – 89.1 and 93.6%. The sensitivity and specificity of pfADA were 85.4 and 84.1%, respectively, at the best cut-off value of 17.5 U/L. RF was the superior method among the four MLAs, and was also superior to pfADA. The newly designed RF model (based on 12 out of 28 features) exhibited an acceptable performance rate for the diagnosis of TPE with a sensitivity and specificity of 90.6 and 92.3%, respectively. In the prospective study, its sensitivity and specificity were 100.0 and 90.0%, respectively. Conclusions Establishing a model for the diagnosis of TPE using RF resulted in a more effective, economical, and faster diagnostic method. This method could enable clinicians to diagnose and treat TPE more effectively.