PVminerLLM: Extracting Structured Patient Voices from Patient-Generated Text Using Large Language Models

Core Views

PVminerLLM is an innovative framework that uses large language models to automatically extract structured patient voice signals from unstructured patient-generated text. It addresses the limitations of traditional questionnaires and provides a new path for patient feedback analysis in the medical field. This project is open-source, offering pre-trained models of various scales to support multi-scenario applications.

Project Basic Information

• Original author/maintainer: SarielMa
• Source platform: GitHub
• Release time: June 11, 2026
• Original link: https://github.com/SarielMa/PVminerLLM

Research Background

Traditional patient feedback relies on structured questionnaires, which struggle to capture real, personalized expressions. The popularity of internet-based healthcare has led to an explosive growth of unstructured text, but extracting structured patient voices remains a challenge in medical NLP. Thus, PVminerLLM was developed.

Core Dimensions of Patient Voices

1. Patient Concerns: Health issues, treatment doubts, prognosis anxiety, etc., to facilitate doctor-patient communication.
2. Treatment Experience: Drug side effects, medical process, healthcare provider attitude, etc., to guide service improvement.
3. Contextual Signals: Emotional state, health literacy, social support, etc., to help fully understand patient expressions.

Three-Stage Pipeline Architecture

1. Supervised Fine-Tuning (SFT): Using LoRA/QLoRA techniques from the PEFT library, keeping original model parameters unchanged while training low-rank matrices. Advantages: High parameter efficiency, avoids overfitting, easy deployment; supports multi-GPU distributed training.
2. Model Merging: Merge the LoRA adapter back into the base model to generate a dedicated extraction model.
3. FinBen Evaluation Framework: Precisely measure accuracy, evaluate performance of different signals, and provide fine-grained error analysis.

Code and Usage

• Training scripts: sft_peft_ddp.py (distributed training), merge_lora.py (adapter merging), etc.
• Environment setup: conda env create -f environment.yml to activate the finben_vllm3 environment.
• Training command: torchrun --nproc_per_node=2 sft_peft_ddp.py (specify model, dataset path, and other parameters).

Pre-trained Models (Released on Hugging Face)

• voice_70b_llama3.3_instruct (high-precision offline tasks)
• voice_8b_llama3.1_instruct
• voice_3b_llama3.2_instruct
• voice_qwen2.5_1.5b_instruct (real-time applications)

Application Scenarios

1. Online Patient Community Analysis: Extract patient concerns and experiences from forums/social media.
2. Electronic Health Record (EHR) Information Extraction: Structured processing of chief complaints and medical history in EHRs to support clinical decision-making.
3. Satisfaction Survey Enhancement: Analyze open-ended feedback to identify issues not covered by preset options.
4. Adverse Drug Reaction Monitoring: Identify spontaneously reported side effects from patients.

Technical Contributions

1. Domain-specific fine-tuning strategy: Designed data construction, prompts, and evaluation metrics for patient voice extraction.
2. Multi-model scale coverage: 1.5B to 70B parameters, adapting to different computing resource requirements.
3. Open-source and reproducible: Complete code and pre-trained models are open-source, supporting follow-up research.

Limitations

1. Data privacy: Strict desensitization and privacy protection measures are required.
2. Cross-language adaptability: Currently focused on English; needs adaptation to multi-language and cultural contexts.
3. Clinical validation: Extracted information requires verification by clinical experts for accuracy and relevance.

Conclusion

PVminerLLM combines the capabilities of large language models with medical needs, providing a feasible solution for extracting structured information from massive patient texts and promoting the implementation of patient-centered medical concepts.

Future Directions

1. Strengthen data privacy and ethical protection.
2. Improve cross-language and cross-cultural adaptability.
3. Conduct clinical validation to ensure the clinical value of extracted information.

This open-source project provides new ideas and tools for the digital transformation of healthcare, and will play an important role in improving medical quality and optimizing resource allocation.