Automated Labeling Function Generation using Distance Functions for Physiological Alarm Suppression

Deep neural networks (DNNs) have the ability to transform inference with medical data. However, large DNN models need sufficient labeled data to be effective at generalization. This involves considerable manual efforts to generate quality labeled data in copious amounts for data hungry learning tasks. Data programming addresses this issue by using weak labeling functions obtained from experts to label unlabeled data. Such weak labeling functions now become a key component in this pipeline. A step further in this direction is to ask the question: is it possible to automatically generate these labeling functions? A possible path we explore and found effective was through the use of distance functions. In medical cyber-physical systems, it turns out that it is often easier to capture distance between two samples. Such distance functions can in turn be used to generate multiple labeling functions, which can aid data programming techniques further, and produce better accuracy for the learnt models. In this work, we demonstrate this on five physiological monitor alarm datasets for alarm suppression and present results to show its effectiveness. In the future we would like to extend this work to higher dimensional datasets such as images.