Research map / Large language models / Retrieval and knowledge
DearLLM: Enhancing Personalized Healthcare via Large Language Models-Deduced Feature Correlations
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 DearLLM, which uses conditional perplexity of a medical LLM (HuatuoGPT-II) deduced feature correlations as a patient-specific graph prior, learned by a GCN with frequency-aware (CF-IPF) graph pooling, to improve EHR mortality prediction.
Problem
Mortality prediction from sparse EHR data is data-hungry; existing external-knowledge methods use fixed, qualitative feature correlation priors and dilute patient-specific knowledge when pooling.
Approach
Use an LLM's conditional perplexity under a personalized visit-history prompt to quantify correlation strength between diagnosis codes, build a directed patient-specific prior graph encoded by GCN, and pool knowledge with a TF-IDF-style feature-frequency-aware weighting.
Key results
- On MIMIC-III with GRU backbone, DearLLM reaches AUPRC 0.3079 and AUROC 0.6096 versus 0.2087 and 0.4998 for vanilla GRU.
- On MIMIC-IV with GRU backbone, DearLLM reaches AUPRC 0.1735 and AUROC 0.7659, versus GraphCare at 0.1401 AUPRC.
- Ablation shows randomly permuting edge weights drops MIMIC-III AUPRC from 0.3079 to 0.1909, indicating LLM-deduced correlations matter.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DearLLM with GRU instead of the strongest listed GRU knowledge baseline | GRU mortality prediction · MIMIC-III | AUPRC: .2571 -> .3079 | Helped |
| Use DearLLM with GRU instead of the strongest listed GRU knowledge baseline | GRU mortality prediction · MIMIC-IV | AUPRC: .1478 -> .1735 | Helped |
| Add personalized visit history context to the LLM prompt (ablation) | DearLLM · MIMIC-III | AUPRC: 0.2893 -> 0.3079 | Helped |
| Add personalized visit history context to the LLM prompt (ablation) | DearLLM · MIMIC-IV | AUPRC: 0.1633 -> 0.1735 | Helped |
| Use Feature-Frequency-Aware Graph Pooling instead of mean pooling (ablation) | DearLLM · MIMIC-III | AUPRC: 0.2822 -> 0.3079 | Helped |
| Use Feature-Frequency-Aware Graph Pooling instead of mean pooling (ablation) | DearLLM · MIMIC-IV | AUPRC: 0.1619 -> 0.1735 | Helped |
| Use LLM-deduced quantitative correlation weights rather than randomly permuted weights (ablation) | DearLLM · MIMIC-III | AUPRC: 0.1909 -> 0.3079 | Helped |
| Use LLM-deduced quantitative correlation weights rather than randomly permuted weights (ablation) | DearLLM · MIMIC-IV | AUPRC: 0.0593 -> 0.1735 | Helped |
Limitations
- Relies on hallucination-prone domain LLM knowledge, which the authors acknowledge may occur.
- Evaluated only on mortality prediction with two MIMIC datasets and ICD diagnosis codes.
- Compute hardware not reported.
Experimental setup
Datasets: MIMIC-III MIMIC-IV
Benchmarks: MIMIC-III MIMIC-IV
Metrics: AUPRC AUROC F1-score
Baselines: Vanilla GRU Vanilla StageNet Vanilla HiTANet Vanilla Transformer MedRetriever RAM-EHR GraphCare
Builds on: GraphCare RAM-EHR MedRetriever GCN TF-IDF
Closest papers by meaning
- GraphCare: Enhancing Healthcare Predictions with Personalized Knowledge Graphs (ICLR 2024 · similarity 0.86)
- BoxLM: Unifying Structures and Semantics of Medical Concepts for Diagnosis Prediction in Healthcare (ICML 2025 · similarity 0.85)
- Enhancing Small Medical Learners with Privacy-preserving Contextual Prompting (ICLR 2024 · similarity 0.85)
- VILA-M3: Enhancing Vision-Language Models with Medical Expert Knowledge (CVPR 2025 · similarity 0.84)
- Toward Better EHR Reasoning in LLMs: Reinforcement Learning with Expert Attention Guidance (AAAI 2026 · similarity 0.84)
- Reasoning-Enhanced Healthcare Predictions with Knowledge Graph Community Retrieval (ICLR 2025 · 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.