P1032 - Use of telemedicine with an artificial intelligence-enabled stethoscope to support decision-making and reduce inappropriate use of echocardiography in children with heart murmurs

Background/Purpose: A pathologic murmur may be the first sign of heart disease in children and is an appropriate indication for echocardiography, while an innocent murmur is not. However, stethoscope skills are in decline among primary providers making distinction of innocent from pathologic murmurs increasingly difficult. This may lead to over-referral to the cardiologist and overuse of echocardiography, resulting in unnecessary emotional and financial burdens on families. The use of telemedicine and an artificial intelligence (AI)-enabled electronic stethoscope may aid decision-making regarding referral of children with a murmur.

Objective: Assess the feasibility of using an electronic stethoscope with AI and remote cardiologist interpretation of recorded heart sounds to screen for pathologic murmurs.

Design/

Methods: In this observational, non-intervention pilot study, children with heart murmurs and no prior cardiac evaluation were recruited from a community practice, hospital-affiliated resident clinic, newborn nursery, and pediatric cardiology clinic. Control patients with and without heart disease were recorded as well. A single study team member performed all heart sound recordings with an electronic stethoscope. Using the eMurmur smartphone app, recordings were uploaded to a cloud server for quality check and storage. AI and remote cardiologist interpretation of heart sound recordings were not shared with practitioners for decision making, but were compared to echocardiography, clinical cardiologist auscultation, or follow-up chart review. Feasibility, sensitivity, and specificity of AI and remote cardiologist interpretations were determined.

Results: Among 137 subjects, 96% were successfully recorded (80 male, average age 6.4 years, SD 6.1). 5 were not recorded due to network connectivity or failed quality check. Of the 132 recorded (98 study cases, 34 controls), 5 were uninterpretable (1 by AI and 4 by remote cardiologist, primarily due to noise). Net feasibility was therefore 93%. On chart review of subjects not initially referred (38 cases, average follow-up interval from date of recording 12.2 months, range 8.8-14.4 months) there was only one patient that had a subsequent echocardiogram: a newborn, recorded on day of life 1 but not referred until day 2. The echo showed a trivial patent ductus arteriosus. Sensitivity and specificity for remote auscultation by a cardiologist with AI-decision support was 91% (78-97% CI) and 80% (71-87% CI), respectively. Accuracy of AI alone was similar to remote cardiologist alone, though with slightly lower sensitivity.

Conclusion(s): In this pilot study, the paradigm of screening heart murmurs with an AI-enabled electronic stethoscope and remote listening by a cardiologist was feasible. Future studies are needed to assess the utility of this approach when actually used in clinical decision-making and its potential to improve referrals and reduce inappropriate use of echocardiography.