Multiview Outlier Filtered Pediatric Heart Sound Classification

Abstract

The advancements in deep learning has generated a large-scale interest in development of black-box models for various use cases in different domains such as healthcare, in both at-home and critical setting for diagnosis and monitoring of various health conditions. The use of audio signals as a view for diagnosis is nascent and the success of deep learning models in ingesting multimedia data provides an opportunity for use as a diagnostic medium. For the widespread use of these decision support systems, it is prudent to develop high performing systems which require large quantities of data for training and low-cost method of data collection making it more accessible for developing regions of the world and general population. Data collected from low-cost collection especially wireless devices are prone to outliers and anomalies. The presence of outliers skews the hypothesis space of the model and leads to model drift on deployment. In this paper, we propose a multiview pipeline through interpretable outlier filtering on the small Mendeley Children Heart Sound dataset collected using wireless low-cost digital stethoscope. Our proposed pipeline explores and provides dimensionally reduced interpertable visualizations for functional understanding of the effect of various outlier filtering methods on deep learning model hypothesis space and fusion strategies for multiple views of heart sound data namely raw time-series signal and Mel Frequency Cepstrum Coefficients achieving 98.19% state-of-the-art testing accuracy.