Crowdsourcing the creation of an audio dataset for human and machine medical diagnosis training

We describe a method for creating and maintaining an open, annotated, community-moderated dataset of audio recordings of heart and lung sounds, with which to train machine listening systems to perform medical diagnosis. This is achieved by partnering with education programs for nursing and medical professionals who will receive training in diagnosis using digital stethoscopes. We developed a low-cost digital stethoscope using a peer-reviewed, open-source, 3-D printed design. With sufficiently numerous examples supplied and tagged by nursing and medical students, it is possible to employ machine learning classifiers such as our convolutional neural network code—originally developed for music information retrieval—to identify diagnosis classes. While primary intent of this dataset-creation and moderation system is for medical audio, the underlying functionality of community moderation could be applied to other waveform content, including a variety of musical datasets.