A total of 1,073 chest X-rays were acquired within the specified period at the department. The data collection remained intact and unaffected throughout the testing phase, ensuring the integrity of the dataset. The collected sample accurately represents the prevalence of findings within the observed population. After excluding ineligible studies such as X-rays from patients under 18 years of age, lateral projection X-rays, and scans of insufficient quality, a total of 956 relevant CXRs were identified for further assessment.

The proposed DLAD (Carebot AI CXR) is a deep learning-based medical device designed to assist radiologists in interpreting chest X-ray images acquired in anteroposterior (AP) or posteroanterior (PA) projection. By employing advanced deep learning algorithms, this solution enables automatic detection of abnormal findings by analyzing visual patterns associated with specific conditions. The targeted abnormalities include atelectasis (ATE), consolidation (CON), pleural effusion (EFF), pulmonary lesion (LES), subcutaneous emphysema (SCE), cardiomegaly (CMG), and pneumothorax (PNO). The DLAD functions as a prediction algorithm complemented by various application peripherals, such as web-based communication tools, DICOM file conversion capabilities, and storage and reporting libraries supporting both DICOM Structured Report and DICOM Presentation State formats.

The primary objective is to evaluate the performance parameters of the proposed DLAD (Carebot AI CXR) in comparison to individual radiologists. The performance test includes sensitivity and specificity, positive and negative likelihood ratio, and positive and negative predictive value. The aforementioned parameters are statistically compared using confidence intervals (CI) and p-Values. The comparison procedure consists of two steps: a global hypothesis test is conducted to determine whether there are significant differences between DLAD and radiologists. If the global hypothesis test yields a significant result, individual hypothesis tests are performed. Additionally, multiple comparison methods, such as McNemar with continuity correction for Se and Sp, Holm method for LRs, and weighted generalized score statistics for PVs, are applied to control the overall error rate. All tests are performed as two-tailed tests at the 5% significance level.