Research map / Label-efficient and robust learning / Recognition and long-tail
BoxLM: Unifying Structures and Semantics of Medical Concepts for Diagnosis Prediction in Healthcare
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 BoxLM, which combines BioBERT semantic embeddings with ontology- and EHR-derived hierarchies in box embeddings and predicts next-visit CCS codes via Gumbel-smoothed intersection volume.
Problem
Diagnosis prediction from EHRs with language-model methods fails to capture hierarchical structures of medical concepts (ontology and visit-level), and performance degrades in few-shot settings.
Approach
Model diagnoses, CCS codes, visits and patients as box embeddings whose centers carry BioBERT semantics and whose offsets encode ontology hierarchy (via relation-aware GCN) and EHR hierarchy (visit offsets as max over concepts); patients evolve their center over time and memorize via max offset; predictions use log of Gumbel-smoothed intersection volume.
Key results
- On MIMIC-III with 5% training data, BoxLM reaches 43.88% visit-level P@10 vs 38.21% for the best baseline BoxCare.
- BoxLM outperforms BoxCare by 16.70% average at visit level and 13.96% at code level across both datasets.
- On MIMIC-IV, BoxLM improves P@10 by 19.67% over the second-best method.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ontology-based hierarchy modeling (ablation) | BoxLM · MIMIC-III | P@10: 34.42±0.25 -> 37.07±0.10 | Helped |
| Add EHR-based visit box embeddings (ablation) | BoxLM · MIMIC-III | P@10: 37.07±0.10 -> 41.40±0.24 | Helped |
| Use patient-level box embeddings and volume-based prediction (ablation) | BoxLM · MIMIC-III | P@10: 41.40±0.24 -> 43.88±0.23 | Helped |
| Add ontology-based hierarchy modeling (ablation) | BoxLM · MIMIC-III | Acc@10: 25.49±0.18 -> 26.83±0.13 | Helped |
| Add EHR-based visit box embeddings (ablation) | BoxLM · MIMIC-III | Acc@10: 26.83±0.13 -> 30.05±0.13 | Helped |
| Use patient-level box embeddings and volume-based prediction (ablation) | BoxLM · MIMIC-III | Acc@10: 30.05±0.13 -> 31.74±0.21 | Helped |
| Add ontology-based hierarchy modeling (ablation) | BoxLM · MIMIC-IV | P@10: 18.95±0.11 -> 32.14±0.03 | Helped |
| Add EHR-based visit box embeddings (ablation) | BoxLM · MIMIC-IV | P@10: 32.14±0.03 -> 40.28±0.05 | Helped |
| Use patient-level box embeddings and volume-based prediction (ablation) | BoxLM · MIMIC-IV | P@10: 40.28±0.05 -> 42.04±0.04 | Helped |
| Add ontology-based hierarchy modeling (ablation) | BoxLM · MIMIC-IV | Acc@10: 15.44±0.13 -> 23.29±0.03 | Helped |
| Add EHR-based visit box embeddings (ablation) | BoxLM · MIMIC-IV | Acc@10: 23.29±0.03 -> 28.73±0.04 | Helped |
| Use patient-level box embeddings and volume-based prediction (ablation) | BoxLM · MIMIC-IV | Acc@10: 28.73±0.04 -> 29.94±0.06 | Helped |
Limitations
- Evaluated only on two MIMIC datasets with CCS-level prediction
- Relies on a fixed BioBERT encoder rather than larger medical LLMs
- Multimodal and other medical tasks left for future work
Experimental setup
Datasets: MIMIC-III MIMIC-IV
Benchmarks: MIMIC-III MIMIC-IV
Metrics: Precision@k Accuracy@k Recall@k
Baselines: DoctorAI RETAIN StageNet TRANS KAME CGL HiTANet BoxCare BERT BioBERT VecoCare
Builds on: BoxCare GumbelBox BioBERT RETAIN
Closest papers by meaning
- Reasoning-Enhanced Healthcare Predictions with Knowledge Graph Community Retrieval (ICLR 2025 · similarity 0.86)
- Memorize and Rank: Elevating Large Language Models for Clinical Diagnosis Prediction (AAAI 2025 · similarity 0.86)
- DearLLM: Enhancing Personalized Healthcare via Large Language Models-Deduced Feature Correlations (AAAI 2025 · similarity 0.85)
- Multimodal Medical Code Tokenizer (ICML 2025 · similarity 0.84)
- Large Language Models Are Clinical Reasoners: Reasoning-Aware Diagnosis Framework with Prompt-Generated Rationales (AAAI 2024 · similarity 0.84)
- M3CoTBench: Benchmark Chain-of-Thought of MLLMs in Medical Image Understanding (ICLR 2026 · similarity 0.84)
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.