Research map / Label-efficient and robust learning / Recognition and long-tail
VITA: ‘Carefully Chosen and Weighted Less’ Is Better in Medication Recommendation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes VITA, a medication recommendation framework that selects only relevant past visits via Gumbel-softmax-based relevant-visit selection and weights them with target-aware attention, outperforming six baselines on MIMIC-III and MIMIC-IV.
Problem
Medication recommendation needs an accurate estimate of each past visit's relevance to the patient's current visit when building her current health status; existing longitudinal methods either use visit order or ignore relevance.
Approach
Learn per-patient which past visits are relevant (Gumbel-softmax selection, excluding irrelevant ones entirely) and weight the selected visits with a target-aware attention whose query is the current visit and whose softmax includes the current visit itself, so weights can sum well below 1 when all past visits are weakly relevant.
Key results
- On MIMIC-IV, VITA reaches Jaccard 0.5218 vs 0.4943 for the best competitor COGNet (up to 5.56% improvement); on MIMIC-III it reaches 0.5282 vs 0.5087 (3.83%).
- VITA outperforms all six competitors on all measures (Jaccard, PRAUC, F1) on both datasets.
- Adding the two core ideas to GAMENet, MRSC, SafeDrug, and COGNet improves their Jaccard accuracy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use VITA instead of COGNet | Medication recommendation · MIMIC-III | Jaccard: 0.5087 -> 0.5282 | Helped |
| use VITA instead of COGNet | Medication recommendation · MIMIC-IV | Jaccard: 0.4943 -> 0.5218 | Helped |
| add relevant-visit selection (ablation) | VITA · MIMIC-III | Jaccard: 0.5163 -> 0.5282 | Helped |
| add relevant-visit selection (ablation) | VITA · MIMIC-IV | Jaccard: 0.5197 -> 0.5218 | Helped |
| select a variable number of relevant visits instead of only the most similar past visit (ablation) | VITA · MIMIC-III | Jaccard: 0.5140 -> 0.5282 | Helped |
| select a variable number of relevant visits instead of only the most similar past visit (ablation) | VITA · MIMIC-IV | Jaccard: 0.4803 -> 0.5218 | Helped |
| use relevant-visit selection instead of sharpening attention weights (ablation) | VITA · MIMIC-III | Jaccard: 0.5188 -> 0.5282 | Helped |
| use relevant-visit selection instead of sharpening attention weights (ablation) | VITA · MIMIC-IV | Jaccard: 0.5199 -> 0.5218 | Helped |
| add target-aware attention instead of mean pooling (ablation) | VITA · MIMIC-III | Jaccard: 0.5119 -> 0.5282 | Helped |
| add target-aware attention instead of mean pooling (ablation) | VITA · MIMIC-IV | Jaccard: 0.5130 -> 0.5218 | Helped |
| use target-aware attention instead of a GRU (ablation) | VITA · MIMIC-III | Jaccard: 0.5152 -> 0.5282 | Helped |
| use target-aware attention instead of a GRU (ablation) | VITA · MIMIC-IV | Jaccard: 0.5167 -> 0.5218 | Helped |
Surprising findings
- Using only the most similar past visit (VITA-RS Top-1) performs worse than the variant without the selection module, showing a fixed number of selected visits hurts.
- Using past visits irrelevant to the current visit has a negative effect on accuracy, while using all past visits is still better than none.
Limitations
- Relevance is assessed mainly against Jaccard similarity in motivation; the learned selection is not directly validated against clinical relevance.
- Experiments limited to two MIMIC datasets and evaluated on the medication recommendation task only.
- Main text reports only accuracy measures; DDI rate results and some analyses are deferred to the Appendix.
Experimental setup
Datasets: MIMIC-III MIMIC-IV
Benchmarks: MIMIC-III MIMIC-IV
Metrics: Jaccard PRAUC F1 DDI rate
Baselines: Nearest LEAP GAMENet MRSC SafeDrug COGNet
Builds on: GAMENet COGNet SafeDrug Gumbel-softmax
Closest papers by meaning
- HI-DR: Exploiting Health Status-Aware Attention and an EHR Graph+ for Effective Medication Recommendation (AAAI 2025 · similarity 0.88)
- Fine-grained List-wise Alignment for Generative Medication Recommendation (NeurIPS 2025 · similarity 0.79)
- SAFER: A Calibrated Risk-Aware Multimodal Recommendation Model for Dynamic Treatment Regimes (ICML 2025 · similarity 0.79)
- Enhancing Healthcare Recommendations: A Privacy-Protective and Interpretable Cross-Domain Framework (AAAI 2025 · similarity 0.78)
- GraphCare: Enhancing Healthcare Predictions with Personalized Knowledge Graphs (ICLR 2024 · similarity 0.77)
- Teaching LLMs to Recommend and Defer in Underrepresented Epilepsy Care (NeurIPS 2026 · similarity 0.76)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.