SHAPE: A Sample-Adaptive Hierarchical Prediction Network for Medication Recommendation.

Area of Science:

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

Background:

• Medication recommendation for patients with complex multimorbidity is challenging.
• Existing methods often overlook intra-visit event relationships and variable patient longitudinal data structures.

Purpose of the Study:

• To propose a novel Sample-adaptive Hierarchical medicAtion Prediction nEtwork (SHAPE).
• To address limitations in encoding intra-visit medical events and learning variable-length patient sequences for accurate medication recommendation.

Main Methods:

• Developed a compact intra-visit set encoder for visit-level representation.
• Designed an inter-visit longitudinal encoder for efficient patient-level longitudinal representation.
• Implemented a soft curriculum learning method to handle variable visit lengths.

Main Results:

• The SHAPE model demonstrated superior performance compared to state-of-the-art baselines.
• Experimental results on a benchmark dataset validate the model's effectiveness.

Conclusions:

• SHAPE offers an effective approach for medication recommendation in complex multimorbidity cases.
• The model's adaptive hierarchical structure and learning strategies improve prediction accuracy and handle data variability.