Title: No One Left Behind: Improving Metric Learning to Predict Health Risks from a Single Lab Visit
Abstract: The main goal of our work is to predict subjects’ future health risk using a single lab visit—a challenging but crucial task which has not been done before at scale. We found that current representations/transformations for Electronic Health Records (EHR) are inadequate for this task. To improve EHR representations, we propose a novel deep metric learning framework that improves representation learning across various tasks and datasets, including EHR. We use the improved representations to set the state-of-the-art performance in many tasks, including the prediction of healthy patients’ risk of developing conditions in the future using a single time point.
Technical TL;DR: We propose a novel triplet objective that improves representation learning on a variety of applications without requiring additional sample mining or overhead costs.
Link to Preprint: https://arxiv.org/abs/2210.09506
