User-Adaptive Meta-Learning for Cold-Start Medication Recommendation with Uncertainty Filtering

TL;DR

MetaDrug addresses the patient-level cold-start problem in EHR-based medication recommendations by introducing a two-level, uncertainty-aware meta-learning framework that combines self-adaptation on a target patient’s history with peer-adaptation using similar visits from peers, augmented by a Transformer-based uncertainty filtering module. The model employs a dual-transformer recommender architecture (Patient-Transformer and Visit-Transformer), and updates the task-specific head through meta-learning while maintaining a shared initialization for fast adaptation; uncertainty filtering further refines support visits used during adaptation. Across MIMIC-III and AKI, MetaDrug achieves higher PRAUC, F1, and Jaccard for cold-start patients than strong baselines, with acceptable DDI trade-offs, demonstrating the value of uncertainty-guided personalization in sparse, sequential EHR data. These findings highlight the practical impact of combining self- and peer-adaptation with uncertainty quantification to deliver more reliable, patient-tailored medication recommendations in real-world clinical settings.

Abstract

Large-scale Electronic Health Record (EHR) databases have become indispensable in supporting clinical decision-making through data-driven treatment recommendations. However, existing medication recommender methods often struggle with a user (i.e., patient) cold-start problem, where recommendations for new patients are usually unreliable due to the lack of sufficient prescription history for patient profiling. While prior studies have utilized medical knowledge graphs to connect medication concepts through pharmacological or chemical relationships, these methods primarily focus on mitigating the item cold-start issue and fall short in providing personalized recommendations that adapt to individual patient characteristics. Meta-learning has shown promise in handling new users with sparse interactions in recommender systems. However, its application to EHRs remains underexplored due to the unique sequential structure of EHR data. To tackle these challenges, we propose MetaDrug, a multi-level, uncertainty-aware meta-learning framework designed to address the patient cold-start problem in medication recommendation. MetaDrug proposes a novel two-level meta-adaptation mechanism, including self-adaptation, which adapts the model to new patients using their own medical events as support sets to capture temporal dependencies; and peer-adaptation, which adapts the model using similar visits from peer patients to enrich new patient representations. Meanwhile, to further improve meta-adaptation outcomes, we introduce an uncertainty quantification module that ranks the support visits and filters out the unrelated information for adaptation consistency. We evaluate our approach on the MIMIC-III and Acute Kidney Injury (AKI) datasets. Experimental results on both datasets demonstrate that MetaDrug consistently outperforms state-of-the-art medication recommendation methods on cold-start patients.

Figures (6)

• Figure 1: In MetaDrug, a two-level adaptation approach is employed to address the cold-start problem: (1) self-adaptation: the model adapts to a target patient using their own previous visits, and (2) peer-adaptation: the model further adapts by leveraging visits from peer patients with similar medical histories to the target patient.
• Figure 2: In the MetaDrug framework, patient embeddings are first generated using a Patient Transformer, while visit embeddings are obtained through a Visit Transformer. In the self-adaptation phase, the prediction layer is then adapted using the patient's historical visits, by a local update of the layer. During peer-adaptation, similar visits from other patients are identified based on Jaccard similarity, and the prediction layer undergoes a second round of local adaptation using these visits. The resulting adapted model is used to make predictions on the query set. Finally, the trained model is employed for Uncertainty Filtering, where a relevance score is computed for each visit. These scores are used to filter visits during meta-testing, thereby improving generalization performance.
• Figure 3: Self-adaptation and peer-adaptation for the target patient based on their previous visits and similar peers.
• Figure 4: The uncertainty filtering model, trained on meta-training predictions, estimates uncertainty scores for each visit. During meta-testing, visits with high uncertainty are filtered out.
• Figure 5: Performance comparison of MetaDrug and other baselines for cold-start patients across different percentiles, ranked by the number of medical codes, evaluated using PRAUC, F1, Jaccard, and DDI metrics for both MIMIC-III and AKI datasets.