Artificial Intelligence to Improve Physicians' Interpretation of Chest X-Rays in Breathless Patients (XRAI)

Artificial Intelligence to Improve Chest X-ray Reading in Acute Dyspnoeic Patients: A Randomized Controlled Trial

Identifying the cause of breathlessness in acute patients in the emergency department is critical and challenging. The chest X-ray is central but challenging to read for non-radiologist physicians. Often the physicians read the CXR alone due to off-hours and shortage of radiology specialists. Artificial Intelligence (AI) has the potential to aid the reading of chest X-rays. The hypothesis is that AI applied to chest X-rays improves emergency physicians' diagnostic accuracy in acute breathless patients.

Background: Acute dyspnoea is a common symptom in the emergency department (ED) but possible differential diagnoses are numerous. The chest X-ray (CXR) is of great importance in distinguishing between these diagnoses and initiating proper treatment but is challenging to interpret for non-radiologist physicians. Radiology departments are confronted with a demand to read a constantly increasing number of acutely performed CXRs, which exceeds the necessary resources. Therefore, in the acute setting, emergency physicians must often read and diagnose the CXR alone. Altogether, there is an unmet need for help with the CXR interpretation in the ED. Artificial intelligence (AI) software for interpreting CXR has been developed for the detection of pathological findings. In this study, the primary aim is to investigate if AI improves the diagnosis on CXR by non-radiologist physicians in consecutive dyspnoeic patients in the emergency department. The investigators hypothesize, that AI applied to chest X-rays improves the emergency physicians' diagnostic accuracy in acute dyspnoeic patients. The study has the potential to impact the implementation of AI in clinical practice. Method: In a randomized, controlled cross-over study and multi-reader multi-case study, a total of 33 emergency physicians will review CXRs from 231 prospectively collected patients including vital patient information. Each physician will review data from 46 patients. In random order, and on two different days, each CXR is reviewed once with and once without AI-support. Each physician is asked to assess a diagnosis of heart failure, a diagnosis of pneumonia, and whether the CXR is with or without acute remarkable findings. The reference standard is the radiological diagnoses obtained by two independent thorax radiologists blinded to all clinical data. The physicians report their diagnoses in an online questionnaire based on REDCap®. Information that may affect diagnostic accuracy are also collected, such as level of education and experience with CXR reading, along with questions about how sure the physician feels of their tentative diagnosis. The physicians are asked about their interest in, former experience with and expectations to AI, along with an evaluation of these qualities afterwards.

Dyspnea, Dyspnea; Cardiac, Artificial Intelligence, Deep Learning, Emergency Department, Diagnostic, Physicians, Emergency Service, Hospital, X-Rays, Pneumonia, Heart Failure Acute, Diagnostic Accuracy, Multi-reader multi-case (MRMC), Chest X-ray, Randomized

In a crossover and multi-reader multi-case study, physicians read CXRs from acute dyspnoic patients. Each physician retrospectively interprets each image twice in two sessions - once with and once without AI-support in random order.The wash-out period was a minimum four weeks. The images were randomly allocated to the physicians via block randomization. Each image was viewed by at least one physician once with and once without AI on trial day 1.

Allocation of images was performed before inclusion of participants began. Allocation process ensured that is was unnecessary for the investigator to assess the randomization.

Images were allocated to participants. In randomized allocation, one half of the images for each participant are viewed with AI support and the other half is viewed without AI support on the first trial day. On the second trial day the same images are viewed without versus with AI, respectively. This ensures that all images are read twice by the same participant both with and without AI support.

The primary outcome is the difference in diagnostic accuracy of the non-radiologist physicians' diagnosis of ADHF on acute CXR compared with the gold standard. Odds of correct diagnosis are compared using an odds ratio with 95% confidence interval estimated using conditional logistic regression stratified by each image with and without AI. Thus, the improvement in the odds of correct classification after versus before AI support is reported. The significance level is 0.025.

The primary outcome is the difference in diagnostic accuracy of the non-radiologist physicians' diagnosis of pneumonia on acute CXR compared with the gold standard. Odds of correct diagnosis are compared using an odds ratio with 95% confidence interval estimated using conditional logistic regression stratified by each image with and without AI. Thus, the improvement in the odds of correct classification after versus before AI support is reported. The significance level is 0.025.

Inclusion Criteria: Medical Doctor (MD) Working experience with emergency patients Exclusion Criteria: Current or former employment as a radiologist Unwillingness to consent