Adversarial Learning for Improved Patient Representations

Abstract

In recent years, there has been an explosion in the amount of patient Electronic Health Records (EHR) made publicly available. This presents an opportunity to create predictive models that leverage the large amount of data to help guide healthcare worker’s decision-making capacity. However, Patient EHR data is often high-dimensional, sparse, temporal and multimodal. This presents an issue for predictive modeling with Machine Learning models, as many ML model archetypes are not built to handle these types of data. The temporality of EHR data also presents a complicating factor, as it contains multiple time series at different resolutions, which most ML models are not built to handle. This brings us to the topic of patient representation, which is the process of converting this raw EHR data into a dense, mathematical representation. Previous work in this field, however, has not leveraged the full potential of the data, since they opt to only deal with a single modality of data, or do not leverage the temporality of the data. In this paper, we attempt to create a network that creates a multimodal representation of EHR data by modeling it as a multiple sparse time series fusion task. We leverage transformers for sparse time series data, using a custom time-based positional encoding. We then fuse the data into a low-dimensional vector, which serves as our representation. We train the model on 2 separate tasks – Mortality prediction using a classification head, and we attempt to leverage a form of adversarial learning to improve the quality of the representation. We show that the patient representation extracted is meaningful and useful for downstream classification tasks. To our knowledge, this is the first attempt to leverage both adversarial learning and multimodality to create a patient representation. Source code can be found at: https://github.com/BharathSShankar/Patient-Rep-UROPS.