TRICORDER

Triple Cardiovascular Disease Detection with an Artificial Intelligence-Enabled Stethoscope (TRICORDER) in the UK: a Cluster-Randomised Controlled Implementation Trial

Patient / Population Intervention / Exposure Comparison Outcome

In 1,553,175 patients registered at 205 UK primary care practices (701,933 at 96 intervention and 851,242 at 109 control practices), training and implementation of an AI-enabled stethoscope in routine care did not improve detection of newly diagnosed heart failure over 12 months compared with routine care.

N
1,553,175
Design
Pragmatic cluster-randomised implementation trial, UK primary care (205 practices)
Endpoint
Incidence of any newly coded heart-failure diagnosis per 1,000 patient-years (IRR); co-primary: detection stratified by community vs hospital diagnosis
Relevance
1
ResultHeart-failure detection IRR 0.94 (95% CI 0.86–1.02) in intention-to-treat analysis, with no difference in community-based or hospital-based diagnoses (P>0.05); intervention practices recorded 12,725 AI-stethoscope examinations across 972 clinical users.
Kelshiker MA, et al. Lancet. 2026;407(10529):704-715. 10.1016/S0140-6736(25)02156-7
Discussion & critique

The largest randomised implementation trial of an AI diagnostic device, and a negative one: despite regulatory-approved algorithms for reduced ejection fraction, AF and valvular disease, practice-level heart-failure detection did not rise in 12 months. Use was the problem — 12,725 examinations across a registered population of over 700,000 — and use itself was independently associated with higher detection of heart failure, AF and VHD, the adoption mechanism EAGLE and Rushlow had exposed. Open-label by necessity; it teaches that implementation, not accuracy, is the rate-limiting step for AI in primary care.