Challenges with Audio Classification using Image Based Approaches for Health Measurement Applications

Image classification has had huge success in recent years, mainly due to the vast array of databases available. The lack of audio databases presents a problem when it comes to creating a deep neural network classifier aimed at measurement and monitoring of health-related sounds. Such sounds (i.e. cough) can be indicative of worsening health conditions, specifically as it relates to remote monitoring of older adults. The application of pre-existing deep neural network image classifiers to audio classification has been presented as a potential solution. This paper describes some of the issues associated with utilizing audio spectrograms to retrain the AlexNet image classifier for the purpose of remote patient monitoring. The spatial invariance assumption of the classifier is further investigated by creating two different classification tasks based on spectrograms computed from notes on a classical piano at four different noise levels; (1) octave classification and (2) note classification. As expected, the AlexNet classifier with clean data performs better when classifying octaves (98%), when compared to the note classification (83 %). When evaluating on audio with noise, the note classifier performance decreases more than the octave classification performance.