MR-KPA

MR-KPA recommends medications from electronic medical records by combining knowledge-enhanced pre-training and a deep adversarial network to improve feature representation and robustness for medication recommendation tasks.

Key Features:

• Knowledge-Enhanced Pre-training Visit Model: Employs a visit-level pre-training model that leverages ontology embedding to fuse external domain knowledge with internal EMR features for improved representation.
• Deep Adversarial Network: Uses a deep adversarial network that introduces adversarial examples during fine-tuning to mitigate overfitting and enhance generalization when pre-training and recommendation tasks differ.
• Complementary Optimizations: Integrates ontology embedding, the pre-training visit model, and adversarial training to synergistically enhance medication recommendation accuracy on small-scale longitudinal EMR data.

Scientific Applications:

• Medication recommendation on limited longitudinal EMRs: Applies to medication recommendation from EMRs that lack large-scale temporal correlations.
• Feature representation learning for EMR data: Enhances extraction and fusion of EMR features using ontology embedding and pre-training.
• Robust fine-tuning against task mismatch: Improves model robustness during fine-tuning when pre-training tasks differ from the target recommendation task.
• Empirical validation: Demonstrated improved performance in experiments using EMRs from medical and health institutions in Hainan Province, China.

Methodology:

Implements a knowledge-enhanced visit-level pre-training model with ontology embedding to fuse external and internal EMR features; mines features via pre-training; applies a deep adversarial network that injects adversarial examples during fine-tuning to mitigate overfitting and improve generalization.