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

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Decision Support Systems

Background:

• Drug-disease contraindications (DDCC) in patients with comorbidities present a significant clinical challenge due to their long-tail distribution and individual variability.
• Existing electronic health record-based medication recommendation models struggle with DDCC, increasing patient safety risks.

Purpose of the Study:

• To develop KATMed, a knowledge-augmented transformer model designed for contraindication-aware medication recommendation in patients with comorbidities.
• To improve the accuracy and safety of medication recommendations by effectively handling sparse and diverse comorbidity data.

Main Methods:

• Utilized Transformer-based encoding of patient records with two self-supervised tasks to capture temporal and semantic dependencies.
• Developed a hybrid knowledge-augmented framework integrating bidirectional medication-related clinical associations.
• Incorporated positive disease-procedure associations via a dynamic semantic relevance matrix and negative DDCC rules as differentiable logical constraints.

Main Results:

• KATMed demonstrated significant performance improvements on the MIMIC-III and MIMIC-IV datasets.
• Achieved a 5.2% increase in recommendation accuracy.
• Reduced safety violations by 2.04% compared to existing methods.

Conclusions:

• KATMed effectively addresses the challenge of drug-disease contraindications in comorbidities for medication recommendation.
• The knowledge-augmented approach enhances feature learning on sparse comorbidity data, leading to safer and more accurate clinical decisions.