Can electrocardiogram classification be applied to phonocardiogram data? — An analysis using recurrent neural networks

Both a Phonocardiogram (PCG) and an Electrocardiogram (ECG) are sequential measurements of heart activity used to distinguish normal from abnormal heart function. Although they measure different physical quantities, we show that training a long short-term memory network on the Physionet challenge using only the ECG data available for the MIT heart sounds database still yields a score of 0.74 compared to the reference score of 0.82 for a similar net trained on the PCG data. This finding suggests that it may be valuable to train a transformational neural network to produce an artificial ECG from a PCG. Such a transformational net would allow to harness the know-how of decades of research on ECG classification to improve PCG classification. Unfortunately, this task seems too hard for current state-of-the art architectures for neural networks given the data of the Physionet challenge 2016. However, it may be worthwhile to further pursue this approach using data with less variance in the ECG signals or a specialized network architecture.