For the detection of murmurs in heart sound recordings collected from multiple auscultation locations, our team SmartBeatIT developed a model based on a deep recurrent neural network architecture that classifies individual recordings. It is composed of two stacked bidirectional Long Short-Term Memory (LSTM) layers and robust to signals of variable length.
The signal features used at the input of the network were the homomorphic envelope, the Hilbert envelope, the power spectral density envelope, and the wavelet envelope, which were downsampled to 50 Hz. These envelopes form a multivariate time series, from which we extracted fixed-length segments of 200 samples for training. To deal with class imbalance, for the minority classes (Present and Unknown) these segments were extracted with 75\% overlap. Recordings from patients in the Present class without an audible murmur were excluded from training.
The final label and confidence scores for each patient were generated by selecting the recording with the highest probability for the Present class, since it is only necessary that the murmur is audible in one location to confirm its presence. Using the challenge scoring metric, we achieved a score of 1001 on the validation data. With 5-fold cross-validation on the training data, we achieved an average score of 822.78, an accuracy of 0.671, and a sensitivity of 0.618.
LSTMs can model long distance dependencies in the heart sounds sequential data. Based on these preliminary results, we believe that classification performance can be improved by supplementing the temporal envelopes with additional signal features, to better capture the relevant variability in the signals.