Key Points:
- Eighty percent of HF cases are diagnosed only during emergency hospital admissions, which negatively impacts survival rates and increases treatment costs
- The TRICORDER trial was a cluster randomized controlled study of 200 primary care practices across the UK, one using the AI-stethoscope and the other continuing with standard diagnostic practices
- Though an intention-to-treat analysis was negative, the per-protocol analysis showed increased detection of HF, AF, and VHD with use of the stethoscope
- The trial found that, in a real-world analysis, utilization dropped to 60% over 12 months, but the pragmatic design of this trial might be useful for future AI implementation analyses
Heart failure (HF), atrial fibrillation (AF), and valvular heart disease (VHD) all suffer from the challenge of late diagnosis. For example, 80% of HF cases are diagnosed only during emergency hospital admissions, which negatively impacts survival rates and increases treatment costs. However, early diagnosis can significantly improve patient outcomes and reduce healthcare costs.
The TRICORDER trial investigated whether an artificial intelligence (AI)-enabled stethoscope (Eko DUO) can assist general practitioners (GPs) in diagnosing HF, AF, and VHD at an earlier stage, before patients require emergency care. The goal was to assess whether such a device can improve detection rates, reduce emergency hospital admissions, and lower healthcare costs.
This was a cluster randomized controlled implementation study involving 200 primary care practices in North West London and North Wales, UK. These practices were divided into two groups: one that used AI-enabled stethoscopes for use in routine patient assessments and one that continued with traditional diagnostic methods. The study evaluated the clinical and cost-effectiveness of the AI-stethoscope at 12 months. The trial investigators focused on novel methods in order to replicate real-world use of such a tool. For example, clinicians could opt to use this tool at their discretion.
The study’s primary outcomes were detection of HF per 1,000 patient-years (incidence rate ratio [IRR]) as well as routine diagnosis of HF at a community practice versus a hospital. Secondary outcomes included per protocol patient-level analyses (PSM), IRR for AF and VHD, utilization rates, implementation barriers, enablers, and behaviors, as well as AI algorithm performance. Out of 367 practices invited, 213 were randomized, 109 to control arm for a total of over 850,000 patients and 104 to intervention arm for a total of over 700,000 patients.
In the intention-to-treat analysis, the study found no difference in the primary outcome of IRR for HF between diagnosis in an emergency hospitalization versus a community practice (HR 0.940, 95% CI (0.860-1.020), p=0.14. Similarly, there was no difference of IRR for AF or VHD. In the patient-level per-protocol analysis, the study found a positive IRR for each condition: 2.33 for HF (95% CI 1.28-4.26, p=0.005), 3.45 for AF (95% CI 2.24-5.32, p=0.001), and 1.92 for VHD (95% CI 1.09-3.40, p=0.02), thereby favoring use of the stethoscope. Sensitivity analysis (looking at practices by low, medium, and high utilization) was non-significant for HF but with a possible dose-response trend toward increased utilization. Notably, 12-month utilization trends demonstrated drop in utilization to 60% at 12 months.
The TRICORDER trial was the largest diagnostic AI RCT to date with pragmatic design and real-world conditions. However, this real-world design resulted in limitations, including clinician discretion of use limiting utilization over time. In final analysis, the TRICORDER study investigators conclude that the use of AI-enhanced stethoscope increases the detection of multiple cardiac conditions, though the impact relies on clinician utilization. They suggest the randomized clinical implementation trial methodology for real-world evidence can characterize implementation gaps, ultimately serving as a blueprint for future analysis of AI implementation gaps.
