Detection of heart failure with an artificial intelligence-enabled stethoscope in primary care: a multicentre prospective external validation study

Background Heart failure (HF) is frequently diagnosed late, after an emergency hospital admission. The HF diagnostic pathway initiated in primary care requires echocardiography to confirm the existence and form of HF associated with an aabnormal NT-proBNP to guide initiation of treatment of major symptomatic and outcome benefit when started early, particularly in heart failure with reduced ejection fraction (HFrEF). Consequently, fewer than 10% of patients complete the HF diagnostic and treatment-initiation pathway to time and target. Determining the performance of an artificial intelligence-enabled stethoscope (AI stethoscope) for detection of HFrEF and abnormal NT-proBNP in primary care may inform a clinical pathway for early detection and treatment of HFrEF. Purpose This study measured the performance of an AI stethoscope for detection of left ventricular ejection fraction (LVEF) ≤40% (primary endpoint) and abnormal NT-proBNP (400pg/mL and 2000pg/mL) in patients attending primary care. Methods This was a prospective, multicentre, observational validation study conducted across 23 UK primary care practices. 423 patients being investigated for HF received a 15-second examination using the AI stethoscope. The AI stethoscope captures single-lead electrocardiogram and phonocardiogram waveforms that are examined by a cloud-based machine-learning algorithm trained to detect LVEF≤40%. Echocardiography-derived LVEF, and laboratory-reported NT-proBNP served as the ground truth for the primary and secondary endpoints, respectively. Results Between 27th July 2021 and 31st July 2023, 423 patients were recruited to the study and 41 patients had transthoracic echocardiographic LVEF <40%, with the corresponding AI stethoscope signal adequacy, the AI stethoscope achieved an AUC of 0.93 (95% CI 0.68-1.00), sensitivity of 100% (34-100), and specificity of 73% (57-84) for detection of LVEF ≤40%. The AUC for detection of abnormal NT-proBNP ≥400 pg/mL was 0.72 (95% CI 0.65-0.80) and 0.83 (95% CI 0.63-1.00) for ≥2000 pg/mL (Figure 1) For NT-proBNP ≥400 pg/mL, the AI stethoscope achieved a sensitivity of 43% (30-57) and specificity of 88% (83-92), with a positive predictive value of 0.45 and a negative predictive value of 0.87. Of 423 participants, 359 (84.8%) received NT-proBNP testing, with 74 (20.6%) showing abnormal results (≥400 pg/mL). No significant performance differences were observed based on age, sex, or ethnicity. (Figure 2) Conclusion The AI stethoscope can predict abnormal NT-proBNP in patients attending primary care for investigation of suspected HF, with high specificity and negative predictive value. These findings pose a compelling case for the AI stethoscope as a rule-in test to screen for HFrEF, and an indication for NT-proBNP testing in patients attending primary care.