Explainable Deep Learning for Readmission Prediction with Tree-GloVe Embedding

Preventable hospital readmissions have been identified as one of the primary targets for improving the efficiency of the current healthcare system. Over the past decade, several data-driven solutions for predicting readmissions have been presented. While maintaining high predictive accuracy is the obvious main goal for such solutions, ensuring explainability of the model and its predictions, is equally important for adoption in the healthcare domain. Unfortunately, most solutions have struggled to strike an optimal balance between accuracy and explainability. Linear models only provide moderately accurate results while complex machine learning models are non-explainable black boxes, which precludes them from being used effectively within the decision support systems in the hospitals. We propose a solution that integrates domain knowledge, in the form of a hierarchical taxonomy defined for disease codes, into the learning framework to advance state-of-the-art in readmission prediction. We first propose a novel tree-structured embedding method to map disease codes into an explainable domain-guided representation. Next, we propose an attention-driven recurrent deep learning architecture. Results on two healthcare claims data sets show that the proposed model outperforms state-of-the-art methods proposed for this task, both in terms of accuracy and explainability.