Operational characteristics of a clinical algorithm to optimize the detection of undiagnosed heart failure in primary care: insights from a prospective, multicenter study
S Stewart, D Wilson, J Smith, T Nguyen, A Garg, A R Stewart, R Perry, D Playford, Y K Chan, J BeilbyAbstract
Background
Although the genesis of heart failure (HF) largely occurs in the community, even among high-risk people, it remains chronically undetected in the primary care setting. Ideally, especially in rural settings, general practitioners (GPs) would have the capacity to proactively detect and manage HF much earlier without overwhelming scarce specialist resources.
Purpose
To examine the operational characteristics (with a specific focus on its diagnostic accuracy and ability to streamline specialist referral) of a pragmatic clinical algorithm designed to optimize the detection and care of previously undiagnosed HF in the primary care setting.
Methods
We are conducting a prospective, pragmatic, multicentre, HF surveillance study across metropolitan and rural-remote primary care clinics applying artificial intelligence (AI)-mediated point-of-care ultrasound (PoCUS) by Practice Nurses. This involves AI-assisted 2-dimensional imaging included parasternal long-axis and apical 4- and 2-chamber views. Manually acquired imaging included colour Doppler imaging of left heart valves, mitral inflow velocities and both septal and lateral mitral annular tissue Doppler velocities. Images then undergo AI-automated measurement for report generation to facilitate the detection of HF when combined with a clinical review of each person’s signs and symptoms (ESC HF criteria) and their NT-pro BNP level. Herein, we report on the initial performance and related referral outcomes of the clinical algorithm (FIGURE 1) we developed in the project's pilot phase.
Results
An initial 50 men and 50 women (aged 70.3±7.6 and 70.8±7.2 years) with a combination of hypertension and diabetes were screened for HF. Overall, 60 (60%) were found to have ESC-defined signs/symptoms of HF (27 and 9 NYHA II or III, respectively), whilst 43 (43%) had an elevated NT-proBNP level (median 289, IQR 172-684 pg/ml). Based on initial profiling, 34 (34%) cases had any form of HF ruled out – subsequently no cases were found to have an abnormality when still screened with PoCUS (100% sensitivity/positive likelihood ratio 2.00 [95%CI 1.58 to 2.54]). The rest spanned cases of 'heart stress of unknown origin’ to HF with a reduced ejection fraction. Based on their clinical judgement, GPs referred 29 cases (29%) for a full echo - resulting in 9 and 20 ‘false and true positive’ cases of HF, respectively (100% sensitivity/ 88.8% specificity). In comparison, the clinical algorithm recommended 20 cases be referred - resulting in 1 and 19 ‘false and true’ positive case (96.7% sensitivity/98.8% specificity and 77.3% [95% CI 11.0-543) positive likelihood ratio).
Conclusions
As designed, the clinical algorithm we’ve specifically developed for HF screening in the primary care setting shows promise, given its high sensitivity in detecting those without evidence of underlying cardiac dysfunction, whilst demonstrating high specificity when referring possible HF cases for specialist diagnosis/management.Fig. 1 Clinical AlgorithmFor image description, please refer to the figure legend and surrounding text.