Deep Holistic Representation Learning from EHR

In recent years there has been a surge of interest in applying deep neural networks to electronic health records (EHRs) for predictive clinical tasks. EHR data cannot be mined like traditional image or text data because it has unique characteristics including temporality, irregularity, heterogeneity (both structured and unstructured) and incompleteness. We begin by identifying weaknesses in the way deep learning is currently being applied to health data. Then, leveraging these insights, we propose an end-to-end strategy for extracting complimentary deep feature representations from EHRs. This strategy is based on a "bringing model to data" machine learning approach instead of "transforming data to model". It uses multiple neural networks, that have each been optimised for the characteristics of their input data, to extract features. Then, the output of these neural networks is combined. We show that prediction accuracy improves as the output of each neural network is contributed. This work demonstrates the value of extracting relevant insights from different aspects of a patients record, which is analogous to how a clinician makes decisions.