Clinical Validation of an AI System for Pneumoconiosis Detection Using Chest X-rays.

Abstract

Abstract: Objective: To develop and evaluate "eTóraxLaboral," an intelligent platform for detecting signs of pneumoconiosis in chest radiographs, and to assess its predictive capacity. Methods: A retrospective analysis of 2300 randomly selected chest radiographs was performed. Sensitivity, specificity, false positive/negative rates, predictive values, likelihood ratios, efficiency, error rate, and area under the receiver operating characteristic curve (AUC) were evaluated. A Fagan nomogram and ROC curve analysis were included. Results: "eTóraxLaboral" demonstrated high sensitivity to signs of pneumoconiosis (LR+ 23, LR- 0.2). A slight tendency toward a higher number of false positives was observed, possibly due to the superposition of anatomical elements and increased lung markings. False negatives were less common, often misinterpreting pneumoconiotic opacities as consolidation-type findings. Conclusions: "eTóraxLaboral" facilitates early pneumoconiosis detection, providing crucial diagnostic support for healthcare workers in Chile and other developed or developing nations.