Combating the Bucket Effect:Multi-Knowledge Alignment for Medication Recommendation

TL;DR

This work tackles the bucket effect in medication recommendation by introducing MKMed, a cross-modal encoder that aligns five modalities—image, text, structure, molecular properties, and knowledge graphs—into a unified medication representation using contrastive pre-training. The framework pre-trains a cross-modal encoder with five modalities and integrates these representations with longitudinal EHR data to predict safe medication combinations, directly addressing missing-modality challenges. Through extensive experiments on MIMIC-III and MIMIC-IV, MKMed outperforms state-of-the-art baselines on accuracy metrics (e.g., Jaccard, F1, PRAUC) while reducing DDI risk, demonstrating improved data utilization beyond the overlapping modality intersection. The approach offers a practical, scalable path toward more comprehensive and safer medication recommendations in real-world healthcare settings, with open-source implementation available on GitHub.

Abstract

Medication recommendation is crucial in healthcare, offering effective treatments based on patient's electronic health records (EHR). Previous studies show that integrating more medication-related knowledge improves medication representation accuracy. However, not all medications encompass multiple types of knowledge data simultaneously. For instance, some medications provide only textual descriptions without structured data. This imbalance in data availability limits the performance of existing models, a challenge we term the "bucket effect" in medication recommendation. Our data analysis uncovers the severity of the "bucket effect" in medication recommendation. To fill this gap, we introduce a cross-modal medication encoder capable of seamlessly aligning data from different modalities and propose a medication recommendation framework to integrate Multiple types of Knowledge, named MKMed. Specifically, we first pre-train a cross-modal encoder with contrastive learning on five knowledge modalities, aligning them into a unified space. Then, we combine the multi-knowledge medication representations with patient records for recommendations. Extensive experiments on the MIMIC-III and MIMIC-IV datasets demonstrate that MKMed mitigates the "bucket effect" in data, and significantly outperforms state-of-the-art baselines in recommendation accuracy and safety.

Figures (5)

• Figure 1: The sparse intersection of multiple data.
• Figure 2: MKMed: Left: Demonstrates the detailed process of knowledge embedding generation and alignment. Mid: The cross-modal encoder pre-training module, which integrates five distinct knowledge sources into the cross-modal encoder. Right: The prediction module, which first utilizes the pretrained encoder to generate medication representations, then combines them with the patient's current condition to produce the final medication recommendation.
• Figure 3: The low-dimensional data distribution represented by molecular embeddings. From left to right: No external knowledge, using one type of knowledge, and using all five types of knowledge.
• Figure 4: Differences in the performance changes of different alignment methods when the number of modaliteis changes.
• Figure 5: Experiments on parameter sensitivity on MIMIC-III dataset.