Evaluation of pulmonary tuberculosis diagnosis in chest radiographs using the AXIR-CX artificial intelligence program from confirmation results based on the Xpert MTB/RIF Ultra method in recipients with negative AFB sputum results

Abstract

Introduction: Tuberculosis is a problem for the public health system both globally and in Thailand. When screening for pulmonary tuberculosis using radiographic images, a shortage of radiologists to interpret the images was found. Therefore, the AXIR-CX artificial intelligence program was introduced for image interpretation. This study aims to evaluate the performance of the AXIR-CX artificial intelligence program in screening for pulmonary tuberculosis using digital chest radiographs, with the Xpert MTB/RIF Ultra method as the reference standard. Methods: A retrospective descriptive design was used, with digital chest radiographs randomly selected from samples at six primary hospitals in Chanthaburi province using the AeroDR 2 1417S DR detector. Data were collected from October 1, 2019, to September 30, 2023. The data were analyzed using diagnostic test statistics, ROC curve analysis, and Kappa statistics for agreement measurement. Results: The AXIR-CX artificial intelligence program analyzed 459 chest X-ray images and found that the sensitivity and specificity were 80.6% (95% CI: 69.5-88.9) and 59.2% (95% CI: 54.1-64.1), respectively. The accuracy was 62.5% (95% CI: 58.1-66.9). The positive likelihood ratio was 1.97 (95%CI: 1.67 to 2.33), and the negative likelihood ratio was 0.33 (95%CI: 0.20 to 0.53). The ROC curve analysis yielded a cutoff point of 0.5, with the area under the ROC curve being 69.9% (95% CI: 64.6-75.1). When setting the significance level at 0.05, there was no statistically significant difference in the area under the ROC curve between males and females (p-value = 0.12), and there was no statistically significant difference in the area under the ROC curve between the age groups of 15-54 years and those aged 55 years and older (p-value = 0.29). The kappa coefficient between the AXIR-CX artificial intelligence program and the Xpert MTB/RIF Ultra method was 0.22 (95% CI: 0.15-0.29). Conclusion: The AXIR-CX artificial intelligence program can be used with caution for screening pulmonary tuberculosis in primary hospitals, as the decision criteria for estimating the area under the ROC curve is relatively low, and the level of agreement strength is fair.