Nonnegative matrix factorization and random forest for classification of heart sound recordings in the spectral domain

Abstract

Stimulating the development of robust algorithms for the automated classification of phonocardiograms (PCGs) is the goal of the PhysioNet/CinC challenge 2016. In this paper, an approach to classify PCGs in the spectral domain is presented. First, the magnitude spectrogram is calculated. Next, the spectral shapes of four states of the cardiac cycle ('S1', 'Systole', 'S2', 'Diastole') are extracted using nonnegative matrix factorization, which is initialized with a time-domain segmentation algorithm. A Random Forest with 3000 trees is used for classification. Using 10-fold cross-validation on the unbalanced training data, a mean sensitivity of 0.92 at a specificity of 0.83 was achieved, resulting in an overall score of 0.88. On the complete hidden test data, a top score of 0.78 during phase II of the challenge with a sensitivity of 0.74 and a specificity of 0.83 was achieved.