Enhanced Phonocardiogram Classification Performance through Outlier Detection

Deep learning has been utilized in binary classi-fication of phonocardiogram (PCG) signals. Current learning techniques, segment PCG records into cycles for training. As the cycles are unlabelled, the cycle labels are inherited from the record labels, leading to wrong labelling of some of the abnormal cycles. Moreover, the heterogeneity within the training dataset arising from the use of different sensors, location of data acquisition, and different pathology further negatively impact the learning and subsequently the binary classification of the record as either normal or abnormal. Therefore, this work proposes a novel individual record-based method, instead of the whole dataset-based method to mitigate mislabelling of abnormal cycles during the training process and to simultaneously addressing heterogeneity in the dataset. Specifically, each abnormal PCG record undergoes individual outlier detection (due to hetero-geneity), enabling the detection and removal of outlier cycles before the PCG record is utilized for training. Thus, training is conducted using properly labelled abnormal and normal cycles. Comprehensive evaluations comparing with baseline methods across various data subsets is carried out. A range of performance metrics is considered to evaluate the proposed method. Results emphasize the effectiveness of integrating outlier detection into the proposed methodology, thereby significantly enhancing the accuracy of PCG classification.