DADA-MED: Data-Augmented Dual Attention Model for Enhanced Medication Recommendations

Krati Saxena, Tomohiro Shibata

Abstract: Medication recommendations are vital for healthcare AI, enhancing patient care and alleviating healthcare professional’s workloads. Accurate drug suggestions support physicians in making informed choices. Existing research focuses on limited features, overlooking the rich data in Electronic Health Records (EHRs). Also, physicians typically prescribe based on similar cases, a consideration insufficiently addressed in current predictive models, potentially leading to treatment inconsistencies. This research introduces three advancements to the medication prediction system. Firstly, we enhance the framework by incorporating additional data sources, such as lab events and patient demographics, for a comprehensive view of health and treatment. Secondly, we propose a patient-similarity-based retrieval framework to derive candidate medications from similar cases as supplementary features. Thirdly, we develop an intra-visit differential attention mechanism for a refined understanding of patient responses during a visit. We also implement inter-visit attention to track patient health progression over time, allowing our model to adjust to changing conditions and improve medication predictions. We validate our approach using the established MIMIC-III EHR dataset. Our results indicate significant improvements over leading state-of-the-art models across various evaluation metrics, highlighting the efficacy of our enhancements in medication prediction.

External IDs:doi:10.1007/978-3-031-96235-6_7