Active Learning for Classification of Medical Signals

Hospitals are increasingly capturing and storing medical signals produced by patients. Machine learning classifiers could be used on these signals to improve patient care, but while unlabeled data is often plentiful, obtaining labels for such a large quantity of data is prohibitively expensive. Active learning attempts to achieve good classification performance with only a minimal number of labelings by choosing which data instances to label. In this work, we survey the state-of-the-art in applying active learning techniques to medical signals. In particular, we examine five active learning query strategies: (1) expected error reduction sampling, (2) variance reduction sampling, (3) density-weighted sampling, (4) maximal uncertainty sampling, and (5) query-by-committee. Though they are applied to various medical signals (including electrocardiogram, electroencephalogram, and medical images), each provides insight into the viability of active learning in the domain. Taken together they provide a comprehensive overview of current work in active learning. Finally, we compare the performance of these techniques and discuss how active learning could improve real-time machine learning in the medical domain.