Attention Network with PMI Value for Medication Recommendation Algorithm

Medication recommendation plays an essential role to help doctors make diagnose in real clinic events more efficient with more and more EHR data in our daily life which contains rich information. Given the patients' medical data such as diagnosis, procedure and medication, we propose a novel model APVAN to utilize the medicine background knowledge and the patients' historical records in this work. Point-wise Mutual Information value is introduced to construct a medicine graph to illustrate the co-relationship of each medicine. Attention mechanism is adopted to design a new query method to combine different kinds of medical data. Raw data will be embedded first and then use the RNN module to temporally encoding; next, the medicine graph and historical module are combined with a novel query module to make a prediction. A little trick multi-label margin loss is used to solve the label imbalance between the positive and negative samples. Through the experiments, our method achieves the best performance 0.6879, 0.6072, 0.4483 on metrics including, PR-AUC, F1 score and Jaccard.