(P013) PEDIATRIC HEART MURMUR CLASSIFICATION BY DEEP LEARNING ARTIFICIAL INTELLIGENCE

Background: Recognition and diagnosis of heart murmurs by auscultation is a declining skill. Studies show the ability of primary care providers to detect and classify innocent murmurs from pathology is unreliable. Echocardiography, a resource intensive and expensive technology is increasingly used to address this problem, burdening the healthcare system with unnecessary referrals and prolonging wait times. More effective and less expensive technology to assess murmurs will benefit patients and healthcare systems. We developed an artificial intelligence (AI) feature to classify innocent murmurs reliably and therefore all other murmurs are classified as pathological.

METHODS AND RESULTS: Heart sounds from 112 consecutive children average age 8 years (range 0-21) were recorded with a digital stethoscope. The cardiologist’s assessment made by auscultation was saved with each recording. The anonymized data and sounds were securely transferred to a bespoke computing-platform for analysis with a convolutional neural network (CNN) with or without long and short term memory (LSTM) learning, and with or without our innocent murmur recognition feature. The sounds were assessed for the presence of a murmur and murmurs were classified as innocent or pathological. Our innocent murmur recognition feature improved the specificity of murmur identification from 65% to 71% with CNN. Specificity increased to 85% with addition of LSTM. Sensitivity of murmur identification was 89% and did not change with our innocent murmur identification feature nor with adding LSTM to CNN. 60 patients had a murmur and 24 were pathological. Our innocent murmur feature and adding LSTM increased specificity of pathological murmur classification from 65% to 77% compared to CNN alone. Sensitivity of correct pathological murmur classification increased from 72% to 84%. These results suggest our AI feature in combination with CNN and LSTM can effectively identify and screen murmurs in children.

Conclusion: We propose our novel AI-feature used with CNN and LSTM should be further validated with a larger clinical study. We expect it will reduce unnecessary referrals of murmurs by primary care providers while decreasing the cost of managing murmurs and reducing wait times.