Quick and easy multi-pathology cardiovascular disease detection with an artificial intelligence-enabled stethoscope: multi-centre prospective validation study in secondary care

Abstract Introduction Heart failure with reduced ejection fraction (HFrEF), atrial fibrillation (AF), and valvular heart disease (VHD) are all frequently diagnosed late, often following emergency hospital admissions. Early detection is crucial for timely intervention and improved outcomes. Purpose This study determined the performance of an artificial intelligence (AI)-enabled stethoscope applied to a single precordial position for detection of HFrEF, AF, and VHD during routine clinical examination. Methods This was a prospective, observational, multicentre study conducted across three centres in the UK. We prospectively recruited 1,378 adult patients (18 years) attending for routine transthoracic echocardiography (TTE). Each patient received a 15-second examination using the AI stethoscope, which captured single-lead ECG and phonocardiogram (PCG) waveforms from four standard cardiac auscultation positions (aortic, pulmonary, tricuspid, and mitral). These waveforms served as input to three AI algorithms trained to detect HFrEF, AF and structural murmurs respectively. TTE-derived left ventricular ejection fraction (LVEF), detection of VHD, and presence of AF were used as the reference standard. The primary outcome was the area under the receiver operating characteristic curve (AUC) for detection of HFrEF (defined as reduced left ventricular ejection fraction ≤40%) using a combination of single-lead ECG and PCG as input. Secondary outcomes included feasibility (percentage of examinations with adequate signal quality); and performance in detecting AF (using single-lead ECG as input) and VHD (using PCG as input). Sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were also reported. This study is registered with ClinicalTrials.gov (NCT04601415). Results The greatest signal quality adequacy was achieved at the upper left sternal border (pulmonary position) at 92.0%. For detecting HFrEF (LVEF ≤40%), the AI-stethoscope yielded an AUC of 0.87 (95% CI 0.84-0.90), with sensitivity of 83% (70-89) and specificity of 76% (71-87) at the pulmonary auscultation position. For detection of moderate of severe aortic stenosis at the pulmonary position, the AUC was 0.81 (95% CI 0.76-0.86), with sensitivity of 61% and specificity of 85%. The AI-stethoscope demonstrated high specificity (93%) for detecting at least moderate VHD at the pulmonic position. For AF detection, the sensitivity was 84% (77-91) and specificity was 93% (92-95). Conclusion The AI-enabled stethoscope demonstrated accurate detection of HFrEF, AF, and clinically significant aortic stenosis, indicating potential utility in routine clinical practice facilitating early diagnosis in primary care settings, improving workflows, diagnostic rates and earlier initiation of treatment of these conditions.