Artificial Intelligence Versus Anesthesiology Residents in Chest X-Ray Interpretation

This single-center, cross-sectional, observational diagnostic accuracy study will be coordinated by the Department of Anesthesiology and Reanimation at Kayseri City Hospital. Volunteer resident assessments will be collected prospectively using retrospectively selected chest radiographs. The observational model is classified as Other because this is a comparative multi-reader, multi-case assessment of a fixed image set, without longitudinal follow-up.

A dataset of 150 posteroanterior chest radiographs will be selected from the hospital PACS to represent normal findings, common thoracic abnormalities, and a limited number of atypical or diagnostically challenging cases. The distribution of findings will be balanced to avoid marked overrepresentation or underrepresentation of particular abnormalities. All patient-identifying information will be removed before assessment.

Two radiologists will independently evaluate each radiograph. Disagreements will be resolved through joint review and consensus. Their final assessments will serve as the reference standard. The same radiographs will be analyzed by the hChestXR artificial intelligence system and independently interpreted by approximately 35 anesthesiology residents who provide informed consent. Residents will record suspected findings using a predefined multiple-choice assessment form and will be blinded to the radiologist and artificial intelligence assessments. Resident training year will also be recorded.

Diagnostic accuracy, sensitivity, specificity, positive predictive value, and negative predictive value will be calculated overall and by finding. Comparisons between resident and hChestXR performance will account for the paired image assessments and the multi-reader, multi-case structure of the data. McNemar testing will be used for appropriate paired comparisons, and analyses will account for variability across readers and radiographs. Agreement with the radiologist reference standard will be evaluated using Cohen's kappa. A multi-reader, multi-case approach will consider individual resident performance and variability across readers interpreting the same image set. Analyses will use SPSS version 31.0, with p < 0.05 considered statistically significant.

An a priori sample size calculation was performed to determine the number of resident readers. Assuming an absolute difference of 15 percentage points in diagnostic accuracy between anesthesiology residents and hChestXR, a standard deviation of 25 percentage points in reader-level accuracy, a two-sided significance level of 0.05, and 90% power, the required minimum sample size was 32 residents (standardized effect size, Cohen's d = 0.60). To allow for incomplete assessments or participant withdrawal, at least 35 anesthesiology residents are planned for enrollment. The 150 radiographs constitute the image dataset and are distinct from the enrolled resident participants.

The study involves no patient intervention. Resident assessment data will be coded and de-identified for analysis, and identifying information will remain confidential. The planned study period is October 15, 2026, through November 15, 2026. Ethics approval was granted by the Kayseri City Hospital Non-Interventional Clinical Research Ethics Committee on August 26, 2026 (decision number 1081).