Topological Knowledge Enhanced Personalized Ranking Model for Sequential Medication Recommendation

Yanda Wang, Lin Yue, Ying Li

Abstract: Sequential medication recommendation is a crucial healthcare application designed to administer medications in appropriate sequences. While clinical records offer extensive information on patient conditions, they often lack precise prescribing times needed for constructing target sequences. Previous studies have achieved advances in item ranking for recommendation, but fail to address personalized patient preference for medication diversity or translate global medication associations into ranking scores without target sequences. To address these limitations, we propose the TOpological Knowledge Enhanced PErsonalized Ranking (TOKEPER) model for sequential recommendation. TOKEPER first constructs a key-value Memory Neural Network based on historical clinical records of patients, and employs multi-hop reading to infer patient preferences for medication diversity, generating personalized probabilistic potentials for medications. Then TOKEPER calculates the topological potentials of medications based on concurrent prescription across patient cohort, translating global associations of medications into ranking criteria without relying on target sequences. By integrating these heterogeneous potentials, TOKEPER constructs supervised sequences for sequential medication recommendation. Experiments on real-world data demonstrate that TOKEPER yields satisfactory outcomes and facilitate effective sequential medication recommendation.