MR-KPA: medication recommendation by combining knowledge-enhanced pre-training with a deep adversarial network

Area of Science:

• Artificial Intelligence in Healthcare
• Medical Informatics
• Computational Medicine

Background:

• Electronic Medical Records (EMR) are crucial for smart healthcare, but limited longitudinal data hinders medication recommendation.
• Developing robust computational methods for medication recommendation from EMRs is a significant research challenge.
• Existing methods struggle with the scarcity of time-correlated, large-scale longitudinal EMR datasets.

Purpose of the Study:

• To propose a novel EMR-based medication recommendation model, MR-KPA.
• To address the challenge of limited longitudinal EMR data for accurate medication recommendations.
• To enhance both feature representation and the fine-tuning process in recommendation models.

Main Methods:

• Developed a knowledge-enhanced pre-training visit model for external domain knowledge fusion and internal feature mining.
• Integrated a deep adversarial network to optimize the fine-tuning of the pre-training model.
• Utilized ontology embedding for knowledge enhancement and adversarial training to mitigate overfitting.

Main Results:

• The MR-KPA model demonstrated significant improvements in medication recommendation accuracy.
• Experiments were conducted on longitudinal EMR data from medical institutions in Hainan Province, China.
• MR-KPA outperformed existing representative methods, particularly on small-scale datasets.

Conclusions:

• The MR-KPA model's effectiveness stems from knowledge enhancement, a pre-training visit model, and adversarial training.
• The pre-training visit model provided the most substantial improvement in recommendation capability.
• The synergistic integration of these components resulted in the optimal recommendation performance.