Early Clinical Evaluation of AI Triage of Chest Radiographs: Time to Diagnosis for Suspected Cancer and Number of Urgent CT Referrals.
Authors
Affiliations (8)
Affiliations (8)
- South West London AI Working Group, London, UK.
- St George's University Hospital NHS Foundation Trust, London, UK.
- Epsom and St. Helier Hospitals NHS Trust, Carshalton, UK.
- Chelsea and Westminster Hospitals NHS Trust, London, UK.
- Kingston NHS Foundation Trust, Kingston upon Thames, UK.
- Great Ormond Street Hospital for Children, London, UK.
- University College London Great Ormond Street Institute of Child Health, Great Ormond Street Hospital for Children, London, UK.
- National Institute for Health and Care Research Great Ormond Street Hospital Biomedical Research Centre, London, UK.
Abstract
AbstractLung cancer is the leading cause of cancer mortality in the United Kingdom. Early diagnosis and rapid access to treatment are associated with better outcomes but are dependent on prompt recognition of findings and the availability of a computed tomography (CT) scanner and reporting. In this clinical evaluation case study, we report the use of artificial intelligence (AI) to triage suspected cancer on chest x-rays to a "radiographer-led same-day CT pathway." We included 56,257 chest radiographs acquired before and after real-world AI deployment across a National Health Service trust comprising five hospitals in South West London over a 3-year study period between January 2022 and December 2024 (1 year preintervention and 2 years postintervention). The time from chest radiograph acquisition to CT report for cases with suspected lung cancer was reduced from 6.0 to 3.6 days, with the most significant reductions at centers with colocated CT and x-ray facilities and no significant increase in the proportion of CTs performed relative to chest x-rays. In detecting suspected cancer, the 14-feature AI triage model has an 82.5% sensitivity, a 90.5% specificity, a 14.6% positive predictive value, and a 99.6% negative predictive value for possible lung cancer on chest x-rays. Features such as focal airspace opacity and single pulmonary nodule returned the highest proportion of false-positive results, while pulmonary and mediastinal masses had a greater diagnostic yield. This work demonstrates how AI triage can be implemented for the rapid onward diagnosis of lung cancer and informs the deployment of triage systems for urgent cancer pathways in general.