KATMed: a knowledge-augmented transformer model for contraindication-aware medication recommendation in comorbidities

Abstract

Drug-disease contraindications in comorbidities (DDCC) pose a significant challenge and priority in clinical treatment. These contraindications exhibit a prototypical long-tail distribution, characterized by low-frequency, highly diverse, and substantial individual variability. Such distinct properties impose significant limitations on electronic health record-based medication recommendation modeling, ultimately elevating safety risks in clinical practice. To address this challenge, this study proposes KATMed, a knowledge-augmented transformer model for contraindication-aware medication recommendation in comorbidities. The model employs Transformer-based encoding of patient records and leverages two self-supervised tasks to capture rich temporal and semantic dependencies. Based on this foundation, a hybrid knowledge-augmented framework is developed to integrate bidirectional medication-related clinical associations. Positive disease-procedure associations are modeled by using a dynamic semantic relevance matrix to expand the input information, thereby enhancing the model’s feature learning capability on sparse yet diverse comorbidity records. Negative DDCC rules are incorporated as differentiable logical constraints in the loss function to suppress unsafe medications. Experiments on the MIMIC-III and MIMIC-IV datasets show that KATMed significantly improves performance, achieving a 5.2% increase in accuracy and a 2.04% reduction in safety violations.