Heart Murmur Detection and Clinical Outcome Prediction with a Hybrid Deep Learning Model using Phonocardiogram Recordings

Cardiovascular diseases (CVDs) are a leading cause of global mortality, with millions of deaths annually, particularly in low- and middle-income countries. Early diagnosis is crucial for effective management and improved outcomes. Heart murmurs, caused by turbulent blood flow in the heart, are key indicators of certain CVDs, such as valve defects and congenital heart disorders. We developed a hybrid model with a combination of CNN and LSTM layers for murmur detection and clinical outcome prediction on a dataset including PCG recordings and demographic information of 942 patients from George B. Moody PhysioNet Challenge 2022 dataset. Our murmur detection model uses only PCG recordings, but the model trained for clinical outcome prediction, uses demographic information as well. We used focal categorical cross entropy for training phase of murmur detection, while trained clinical outcome prediction model using categorical cross entropy. Our hybrid CNN-LSTM model could reach weighted accuracy of 0.82 for murmur detection and clinical cost metric of 10786 for outcome prediction, outperforming leading models developed in Physionet challenge.