The major bottleneck in medical machine learning is curating massive hand-labeled datasets. Moreover, many medical applications involve analyzing various data modalities that include time-series, which creates a uniquely challenging labelling process. The project aims to alleviate this bottleneck by developing a weak supervision system that optimally deals with time-series data and takes advantage of multiple data modalities. The proposed system will facilitate programmatic large-scale labelling by aggregating cheaper, but less precise, labels using a novel stochastic filter. My research aims to make machine learning more feasible for real-world clinical settings, with a focused clinical use-case of automated seizure detection and localization.