Interpretable Recurrent Neural Networks for Heart Failure Re-hospitalisation Prediction

Abstract

Interpretability and predictive performance are important aspects of a machine learning model. Typically, there is a trade-off between interpretability and predictive performance. This trade-off results in a choice between accurate but opaque models such as multilayer perceptron (MLP) and less accurate but more transparent models such as logistic regression (LR). In the healthcare domain, model interpretability is especially important because the real-life goals (e.g. patient well-being) are hard to model (and thus optimize) formally. Traditional methods such as LR \& MLP use aggregate features and are therefore not able to effectively model temporal dimension that is inherent in Electronic Health Records (EHR) data. Recently, Recurrent Neural Network (RNN) approaches have been successful in modelling healthcare data because they are able to effectively take the temporal dimension into account. However, the RNN model is notoriously hard to interpret. We have looked at three recently proposed RNN-based models for medical event prediction that claim to be interpretable (Dipole, GRNN-HA and RETAIN). The interpretability of these models is tied to the implementation of a neural attention mechanism. Having considered how well the models are able to relate the input to the output in understandable terms, we devised an ordering of the interpretability of these models. Then, we compared performance in predicting 30 re-hospitalisation on an EHR dataset with 37,287 medical histories using admission and diagnosis data. The interpretability/performance trade-off within the three `interpretable' models was partly observed. Although the performance of the RNN-based models was quite similar, the difference in interpretability is more substantial. Therefore we believe that the interpretable RNN-based models are the better overall option to use for predicting events in the healthcare domain.