Heart Murmur Detection and Clinical Outcome Prediction using Multilayer Perceptron Classifier

Kiarash Jalali, Mohammad Amin Saket, Saman Noorzadeh
Shahid Beheshti University


Aims: This study aims to apply a multilayer perceptron (MLP) to classify patients suspected of having heart murmurs into three categories of Present (class 0), Unknown (class 1), and Absent (class 2) using their metadata and Phonocardiogram (PCG) recordings. This algorithm was submitted to the 2022 Moody PhysioNet Challenge under the team name AKSJ_97BSc. Methods: Among the 943 data given by the challenge, 179 belonged to class 0, 68 belonged to class 1, and 695 belonged to class 2. This shows a clear imbalance in the data. In order to balance the training data, SVM-SMOTE, a borderline oversampling technique was used. At the moment, the features given to the classifier were patient demographic data (sex, age, height and weight) as well as statistical properties of the PCG recordings (mean, variance and skewness of the sound signal). The classifier itself was a 6-layer MLP with each hidden layer containing 256, 128, 64 and 32 nodes successively. Relu activation function was used in the hidden layers, softmax function in the output layer, and the Adam solver was used for weight optimization. The dataset was split into 70 and 30%, respectively, for training and testing. Results: Out of the 283 test data, 50, 22 and 211 belonged to class 0, class 1 and class 2 respectively. The results showed F1= 0.29 and recall= 0.62 for class 0, F1= 0.24 and recall= 0.45 for class 1 and F1= 0.3 and recall= 0.19 for class 2. The algorithm received a score of 934 from the challenge. Conclusion: While the obtained results appear to be promising, they are far from satisfactory. Further investigation into preprocessing, extracting more appropriate features (e.g. higher-order statistical, band power, frequency and time-frequency features) from the PCG data, and optimizing the classification algorithm and its parameters are needed.