Enhancing Medication Recommendation with Hierarchical Network and Patient Visit Histories.

Abstract

Prescribing medications is an essential part of patient care and requires precision and personalization in selection. Our study introduces a hierarchical medication recommendation system that aims to improve the prescribing process. We use FastText to embed medical contexts and employ a hierarchical attention-based model to manage the hierarchical structure of medication codes. The system takes input data from the current visit and the three previous visits to make recommendations. We trained and evaluated our model on 99,417 anonymized primary care outpatient visits. Our model achieved a mean average precision (mean AP) of 0.8724, 0.7419, 0.6805, and 0.6184 at the first, second, third, and fourth levels of the ATC system, respectively. We demonstrate that incorporating patient visit histories can improve predictions. Our results provide a solution to improve medication prescribing and suggest possible extensions for more comprehensive recommendations.