Early Detection using Information Technology in Health (EDITH) (Previously known as ArITHMetIC)

Artificial intelligence (AI) for breast screening image analysis could address workforce shortages and improve cancer detection in the NHS Breast Screening Programme (NHSBSP). Studies to date show promise but critical evidence gaps remain. Non-interventional studies fail to measure downstream impact of AI on reader decisions. Interventional studies from other countries lack generalisability to the UK because of differing screening intervals, population demographics, qualification of readers and image acquisition systems. We propose to undertake the first pragmatic randomised multi-vendor mammography and AI algorithm trial to assess whether AI can be used safely, sustainably, and equitably in UK clinical practice with radiologist and radiographer readers. Research question What is the impact on cancer detection, recall to assessment, clinical and cost-effectiveness of integrating AI into the NHSBSP? Objectives 1. To measure the difference in clinical and cost-effectiveness between three study arms: 1) replacing the second reader with AI; 2) using AI for triage; and 3) the current double reading pathway (control). 2. To understand differences in clinical effectiveness by AI vendor and by population subgroup. 3. To monitor barriers to participation in women opting out and adverse psychological outcomes in those taking part. Design Multi centre, pragmatic, cluster, Randomised Controlled Trial allocated 1:1:1 ratio across control and intervention arms. Each site will be randomly allocated one of 5 AI systems, and randomise each clinic-day (cluster of ~36 women) to one of the 3 trial arms. Additional nested test accuracy study will compare AI systems’ accuracy by ethnicity, age, socioeconomic status and breast density. A cloud-based platform will host the images and facilitate integration with participating sites. A UK test set and use in shadow mode at each site will ensure safety before use in the trial. A data monitoring group will review ongoing real time safety results. 660000 women are required to detect a difference of 1 cancer per 1000 screened (5% significance 90% power). Women routinely invited for breast screening at participating centres in England, Scotland, Wales and Northern Ireland are informed about the study and can choose to opt out. Semi-structured qualitative interviews with women who opted out (maximum n=30) will explore barriers to acceptability of AI. Data analysis Primary outcomes of cancer detection and recall rates will be analysed separately using hierarchical models. Analysis will be intention to treat. In the nested test accuracy study, all AI systems will be tested on the same women. A micro-costing study will identify the resource use and total cost associated with each AI-driven intervention. A decision-analytic model will be created to estimate mean values for healthcare costs, life-years and quality-adjusted life years (QALYs) for each AI-driven intervention (standalone; triage) & AI-system; and compared to the current screening programme. Patient and Public Involvement (PPI) The trial benefits from an established PPI group and dedicated PPI researcher who have contributed to the trial design and carried out focus groups to explore the acceptance of AI in breast cancer screening. Timelines for delivery Completion within 4 years Dissemination We will publish our findings and work closely with the NHSBSP and the UK National Screening Committee who advise the four UK Governments on adoption of AI.