Self-supervised detection of disease reporting events in outbreak reports

State-of-the-art supervised approaches for automatically detecting disease reporting events are typically constructed using manual training examples. Such systems suffer from high initial, and sustainability costs. This paper addresses the aforementioned problem, laying the groundwork for a new approach to disease reporting classification for Epidemic Intelligence. Instead of building a classifier strictly from manually labeled data, we exploit outbreak reports, to build a self-supervised classifier, one that labels its own training examples. We measure the performance of our self-supervised classifier and find that it achieves an accuracy of 88%, comparable with existing, state-of-the-art systems. The implications for this work is that by using a self-supervised learner, Epidemic Intelligence systems can build and deploy reliable classifiers; bringing them one step closer to detecting infectious disease threats from on-line informal sources, more quickly.