A paper accepted in AAAI 2024 (full paper)

Title: VITA: ‘Carefully Chosen and Weighted Less’ Is Better in Medication Recommendation
Author: Taeri Kim*, Jiho Heo*, Hongil Kim*, Kijung Shin, and Sang-Wook Kim
Abstract
In this paper, we consider the medication recommendation problem, where we aim to recommend effective medications for a patient at the current visit by utilizing the current and past visit information (i.e., diagnoses and procedures) of the patient. While there exist a number of recommender systems designed for this problem, we point out the following two limitations of them: (L1) they may reflect the patient’s past visit information that is not related to the current visit information in the patient representation based on which medications are recommended, (L2) they do not appropriately capture how relevant each past visit information is to the current visit information. To address the above limitations, we propose a novel medication recommendation framework, named VITA, based on the following two core ideas: relevant-visit selection and target aware attention. Through extensive experiments using a real-world dataset, we demonstrate (1) the superiority of VITA over state-of-the-art competitors and (2) the effectiveness of its two core ideas.