Deep-Learning Based Clinical Decision Simulator for Electronic Health Data

Abstract

In this work we propose a novel clinical simulation technology that can revolutionize the way medicine is practiced and researched. We propose a deep learning predictive EHR solution that centers on a recursive autoregressive model paired with a conditional prediction mechanism that can simulate how a patient will respond to any treatment plan by simulating the full downstream sequence of events, while taking into account individual medication administrations along the way. In addition, we describe the supportive data processing pipelines, and a user interface, as part of a broader system architecture. Finally, we conduct an experiment of the core algorithm on the MIMIC-III dataset, and successfully predict 1959 diagnosis codes and 145 medication codes with an AUC of 0.87 and 0.93 respectively, at the next visitation given all previous visitations. Furthermore, we demonstrate the model is able to recursively predict (using previous predictions as inputs) all medical codes at the next visitation with an average degradation of 0.87% per each successive prediction.