Medication Combination Prediction via Attention Neural Networks with Prior Medical Knowledge

With the adoption of electronic health records (EHR), deep learning technologies have the potential to employ the EHR data to assist experts in better understanding the complex mechanisms underlying the health and disease. Existing studies have made progress on the research of medication combination prediction from the medical data, but few of them take into account the prior medical knowledge. This paper proposes a PKANet model that integrates the prior medical knowledge into the deep learning architecture to predict the medication combination. The prior medical knowledge is calculated from the mapping relation between diagnoses and medications hidden in the EHR data. It can provide the heuristic medications to help the PKANet model learn optimal parameters. In order to predict the possible medication combination, the PKANet model utilizes attention neural networks to obtain the relationship between different elements in the medical sequence data. The experiment results have demonstrated that the proposed PKANet model outperforms the state-of-the-art baselines on evaluation metrics.