AI-assistent tolkning av hjerte-CT på akuttmottaket

This study is a prospective, pragmatic, randomized controlled trial designed to evaluate the impact of an artificial intelligence (AI)-based coronary computed tomography angiography (CCTA) interpretation tool (Angiomics) on the diagnostic accuracy, confidence, and decision-making of emergency physicians in patients presenting with acute chest pain.

Background and Rationale CCTA is widely used in the emergency department (ED) for patients presenting with acute chest pain or suspected acute coronary syndrome (ACS). It provides rapid and precise visualization of the coronary arteries and assists emergency physicians in making time-sensitive clinical decisions. However, interpretation of CCTA requires specialized training and experience, which may not always be available, particularly during nights or in resource-limited settings. This limitation can delay diagnosis and treatment, potentially leading to adverse patient outcomes.

Recent advances in AI technology have enabled the development of automated tools capable of analyzing CCTA images and identifying clinically significant findings such as coronary artery stenosis and myocardial ischemia. These tools have the potential to support emergency physicians by improving diagnostic accuracy, reducing interpretation time, enhancing physician confidence, and optimizing the use of healthcare resources. Despite these advantages, the influence of AI-based interpretation on real-world ED clinical workflows and physician decision-making remains insufficiently studied.

The present trial is designed to address this knowledge gap by evaluating whether the integration of an AI-based CCTA interpretation tool improves emergency physician diagnostic performance and clinical confidence, and whether such improvements translate into more reliable decision-making in the ED.

Study Design This study will employ a prospective, pragmatic, randomized controlled design. Eligible participants will be adults aged 18 years or older presenting to the ED with chest pain and classified as intermediate risk according to the HEART score (4-6 points). Participants are randomised after informed consent and before CCTA acquisition. Allocation is 1:1 between the AI-assisted interpretation group and the standard interpretation group.

1. AI-assisted group (Experimental Arm): Emergency physicians interpret CCTA scans with access to AI results provided by the Angiomics software.
2. Standard interpretation group (Active Comparator Arm): Emergency physicians interpret CCTA scans without access to AI results.

All physicians will evaluate stenosis in the four major coronary arteries (left main, left anterior descending [LAD], left circumflex [LCX], and right coronary artery [RCA]) and record their findings using a standardized format (Yes/No/Non-diagnostic/Other findings). Emergency physicians record diagnostic confidence using four categories: very high, high, moderate and low.

Reference Standard The reference standard for diagnostic performance will be the blinded consensus interpretation of board-certified radiologists specialized in thoracic imaging. Previous studies have demonstrated concordance of 98-99% between expert radiologist consensus and invasive coronary angiography, supporting the validity of this approach.

Outcomes The primary outcome is the negative predictive value (NPV) of CCTA interpretation at the patient level, comparing AI-assisted versus standard interpretations. Patient-level outcomes will be derived by aggregating findings across the four major coronary arteries.

Secondary outcomes include:

• Per-patient diagnostic performance: sensitivity, specificity, PPV and AUROC
• Per-vessel diagnostic performance for LM, LAD, LCx and RCA: sensitivity, specificity, PPV, NPV and AUROC
• Emergency physician diagnostic confidence (four categories)
• Protocol deviation rate Other outcomes include cardiology consultation rate and time to disposition decision.

Reader Experience and Image Quality Emergency physician imaging training, prior CCTA interpretation experience and years of clinical experience are collected. Image quality is assessed by a radiologist blinded to AI outputs, emergency physician interpretations and allocation, and classified as Poor/Fair or Good/Excellent. These variables are considered in exploratory adjusted analyses.

Analysis Population and Preliminary Interpretation Exposure The primary analysis uses a modified intention-to-treat population, excluding participants whose emergency physician was exposed to a preliminary radiology interpretation before completing the study interpretation. AI processing failures remain in the assigned group. A sensitivity analysis includes participants with prior preliminary radiology exposure when both index and reference interpretations are available.

Sample Size and Statistical Analysis The original enrolment target was 552 participants, with 276 participants per group, allowing for 10% dropout. Following consultation with a statistician, the target was revised using a pooled negative diagnosis rate of 64.9% to obtain 181 negative interpretations per group. Allowing for 10% dropout, the required sample size was 181/(0.649 × 0.90), rounded up to 310 participants per group, for a total of 620 participants. The IRB approved the sample-size amendment on 20 July 2026. The persons responsible for the sample-size decision did not have access to group-specific interim results.

An interim analysis was conducted on 8 September 2026. Type I error is controlled using the prespecified O'Brien-Fleming alpha-spending approach, accounting for the actual information fraction.

Diagnostic performance (sensitivity, specificity, PPV, NPV, accuracy) will be compared using chi-square or Fisher's exact tests as appropriate. Agreement between physician and radiologist consensus readings will be assessed using Cohen's Kappa. All statistical tests will be two-sided.

Data Collection and Confidentiality Data will be collected within the routine clinical workflow and stored in a secure, de-identified format according to institutional policies and national privacy regulations. Only authorized investigators will have access to the data. Data will be stored for three years after study completion before secure disposal.

Safety and Monitoring This trial poses minimal risk to participants, as it involves no procedures outside of routine clinical care. No additional radiation, medications, or invasive interventions are introduced by study participation. The study principal investigator will monitor protocol adherence and data quality on an annual basis. The study is subject to continuing IRB review and institutional monitoring or audit in accordance with institutional procedures. The principal investigator is responsible for protocol adherence, data integrity and safety reporting.

Significance This study will provide evidence regarding the clinical utility of an AI-based CCTA interpretation tool in the ED. Findings are expected to clarify whether AI support improves diagnostic accuracy and physician confidence, and whether it can be effectively integrated into real-world emergency care. The results may inform best practices for the implementation of AI tools to optimize patient outcomes and resource utilization in acute care settings.