Enhancing Diagnosis Prediction in Healthcare With Knowledge-Based Recurrent Neural Networks

The objective of diagnosis prediction involves foreseeing the potential diseases/conditions according to analyzing patients’ historical Electronic Health Records (EHRs). The primary challenge in this task is to develop a predictive model that is both sturdy and accurate, while also being interpretable. The most advanced models usually take recurrent neural networks (RNNs) as backbones and then utilize other techniques, such as attention mechanisms, to address this challenge. However, the effectiveness of these models heavily relies on having ample EHR data. Consequently, when the data is insufficient, the performance of these models declines significantly. Recently, graph-based attention models have been proposed to mitigate the issues caused by insufficient data, although they do not fully capitalize on the knowledge present in medical ontologies. To address these problems,
k
nowledge-b
a
sed
r
ecurrent
n
eural networks (named KARNS) are introduced, which is an end-to-end, robust, and accurate deep learning-based architecture designed to predict patients’ future health information. KARNS explicitly leverages the high-level representations of medical codes within the medical ontologies to enhance the accuracy of predictions. Experimental outcomes demonstrate that the proposed KARNS outperforms existing approaches on three real-world medical datasets. It ensures robustness even with limited training data and learns disease representations that are interpretable.