Attention-based Explanations for Hospital Readmission Prediction

Hospital readmission has become a significant challenge for hospitals and governments due to the substantial costs incurred. Deep learning technologies have been widely applied to building readmission prediction models. But the complexity of the models often makes it difficult to understand the outputs of the models.We investigate the use of attention mechanisms in deep learning prediction models as a means of generating explanations to the predictions made by the prediction models.We utilize the attention mechanism in a deep-learning-based prediction model to generate explanations to the readmission predictions.We evaluate the explainability of the attention mechanisms in comparison with contemporary explainable AI methods. We provide a solid comparative analysis to challenge the prevailing view against using attention mechanisms as explanations.Our result shows that, the explanations generated by using attention mechanism are more effective than those produced by several conventional XAI methods.At the same time, the observed variability and disagreement among different XAI methods in selecting and ranking important features highlight the inherent difficulty of relying on any single XAI method as a universally faithful explanation.Our results demonstrate that attention weights can provide explanations for the model’s decisions.