Research map / Large language models / Retrieval and knowledge
Enhancing Small Medical Learners with Privacy-preserving Contextual Prompting
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.
Extracts medical keywords to protect patient privacy, prompts GPT-3.5 to generate clinician-style contexts, and fine-tunes small biomedical language models with these contexts for medical multiple-choice QA.
Problem
Small domain-specific language models underperform LLMs on medical tasks, while sending full patient data to LLMs violates privacy restrictions in healthcare settings.
Approach
Feed only extracted medical keywords and candidate answers to an LLM, using clinician-written few-shot demonstrations, to generate privacy-restricted medical contexts (overall and per-answer) that are appended as extra input for fine-tuning small models.
Key results
- FTC with BioMedLM reaches 55.90% on MedQA test and 63.17% on HEADQA test, the reported SOTA among privacy-restricted methods, versus 50.3% for SFT on MedQA test.
- FTC improves over SFT by up to 22.57% absolute accuracy in few-shot settings on HEADQA.
- On general-domain CSQA and OBQA with T5-base, FTC beats SFT by 3.24% and 11.67% respectively.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add LLM-generated context to SLM training | BioLinkBERT-Base · MedQA test | accuracy: 42.21_{0.91} -> 50.17_{0.42} | Helped |
| Add LLM-generated context to SLM training | BioLinkBERT-Base · HEADQA test | accuracy: 41.00_{0.34} -> 60.21_{0.47} | Helped |
| Add LLM-generated context to SLM training | BioLinkBERT-Base · MedMCQA dev | accuracy: 32.15_{2.23} -> 49.20_{0.45} | Helped |
| Use only overall context instead of overall and candidate-specific contexts (ablation) | FTC with BioLinkBERT-Base · MedQA, HEADQA, and MedMCQA | accuracy | Hurt |
| Use only candidate-specific context instead of overall and candidate-specific contexts (ablation) | FTC with BioLinkBERT-Base · MedQA, HEADQA, and MedMCQA | accuracy | Hurt |
| Remove relationship information from candidate-specific contexts (ablation) | FTC with BioLinkBERT-Base · MedQA, HEADQA, and MedMCQA | accuracy | Hurt |
| Use context only when the LLM preliminary decision is correct (ablation) | FTC with BioLinkBERT-Base · MedQA dev | accuracy: 48.56_{0.62} -> 50.73_{0.35} | Hurt |
| Use context only when the LLM preliminary decision is correct (ablation) | FTC with BioLinkBERT-Base · HEADQA dev | accuracy: 58.17_{0.78} -> 61.35_{0.16} | Hurt |
| Represent raw data with keywords rather than random spans or random words (ablation) | Privacy-restricted LLM context generation and FTC · MedQA | accuracy | Helped |
| Reduce keyword usage ratio (ablation) | LLM prompting and FTC · MedQA | accuracy | Hurt |
| Use full medical context rather than preliminary decisions alone (ablation) | BioLinkBERT-Base SLM · MedQA | accuracy: 47.21_{0.31} -> 50.17}_{0.42} | Helped |
| Add LLM-generated context to SLM training | T5-base · CSQA dev and OBQA test | accuracy: +3.24% and +11.67% | Helped |
Surprising findings
- Contexts from LLM predictions that are incorrect still help the SLM; FTC outperforms FTC with rejection (keeping only contexts with correct LLM decisions).
- Full-data SFT does not surpass the LLM on HEADQA and MedMCQA, yet FTC with only 100 training examples beats both.
Limitations
- Privacy is only mitigated, not eliminated, since keywords are still sent to the LLM.
- LLM-generated medical knowledge may be inaccurate or biased and outdated due to training cutoff.
- Evaluated only on multiple-choice QA with a single LLM (gpt-3.5-turbo) and relatively small SLMs.
Experimental setup
Datasets: MedQA HEADQA MedMCQA MMLU professional medicine CommonsenseQA OpenbookQA
Benchmarks: MedQA HEADQA MedMCQA MMLU professional medicine CommonsenseQA OpenbookQA
Metrics: accuracy
Baselines: QA-GNN GreaseLM DRAGON MurKe MOEBQA HDRN VOD LLM prompting (gpt-3.5-turbo) SFT
Builds on: in-context learning BioLinkBERT Fusion-in-Decoder
Closest papers by meaning
- CARES: Comprehensive Evaluation of Safety and Adversarial Robustness in Medical LLMs (NeurIPS 2025 · similarity 0.86)
- MedAlign: A Clinician-Generated Dataset for Instruction Following with Electronic Medical Records (AAAI 2024 · similarity 0.85)
- DearLLM: Enhancing Personalized Healthcare via Large Language Models-Deduced Feature Correlations (AAAI 2025 · similarity 0.85)
- Small Models are LLM Knowledge Triggers for Medical Tabular Prediction (ICLR 2025 · similarity 0.85)
- CareBot: A Pioneering Full-Process Open-Source Medical Language Model (AAAI 2025 · similarity 0.85)
- Enhancing Healthcare Recommendations: A Privacy-Protective and Interpretable Cross-Domain Framework (AAAI 2025 · similarity 0.85)
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.