PathoSummarize AI: An LLM Fine-Tuning Framework for Intelligent Medical Record Summarization

PathoSummarize AI is an open-source framework for clinical medical scenarios, maintained by doowenskysintilus and released on the GitHub platform in June 2026 (link: https://github.com/doowenskysintilus/PathoSummarize_AI). This framework uses LoRA/QLoRA techniques to fine-tune open-source LLMs such as Mistral and Llama, generating structured course summaries from longitudinal medical record data. It also provides complete experiment management (Hydra) and deployment solutions (FastAPI, Streamlit, Docker), addressing challenges in medical data intelligence such as professional term understanding, temporal logic, factual accuracy, and data privacy.

In modern healthcare systems, patient medical records are scattered, forming a large amount of longitudinal clinical data. Manual summary整理 is time-consuming and prone to missing key information. General LLMs applied in the medical field face four core challenges: 1. Professional term understanding (medical texts contain a large number of professional terms and abbreviations); 2. Temporal logic (the course of the disease requires the model to understand the timeline); 3. Factual accuracy (summaries must strictly adhere to original records and avoid hallucinations); 4. Data privacy (requires local processing capabilities). PathoSummarize AI provides a complete solution for these issues.

The project aims to build a reproducible framework to fine-tune instruction-following LLMs to generate concise, accurate, and temporally clear course summaries. The tech stack includes: base models (Mistral, Llama), fine-tuning techniques (LoRA/QLoRA), configuration management (Hydra), deployment (FastAPI, Streamlit), and evaluation metrics (ROUGE, BERTScore). LoRA reduces parameter updates by adding low-rank matrices, lowering memory usage and training speed; QLoRA further quantizes base model weights (4-bit precision), enabling consumer-grade GPUs to fine-tune models with 70B parameters.

The architecture adopts a strategy of separating environment variables from configurations: .env files store confidential information (HF_TOKEN, WANDB_API_KEY, etc.); Hydra configuration management handles experiment parameters (model, training, data, experiment configurations), supporting strict reproducibility and parameterized scanning. The data pipeline converts raw medical data into instruction fine-tuning format ({"input":"...","output":"..."}), supporting multiple formats (JSONL, JSON, CSV), text cleaning, quality checks, and dataset splitting.

Model evaluation combines ROUGE (n-gram overlap) and BERTScore (semantic similarity) to comprehensively assess summary quality. The optional RAG enhancement layer supports FAISS/ChromaDB vector databases. The process is: historical medical record encoding → storage → retrieval of relevant records when generating summaries → input as context to the model, suitable for handling complex medical records of chronic disease patients.

Deployment includes: 1. FastAPI service: encapsulates RESTful API, POST /summarize receives medical record text and returns summaries; 2. Streamlit interactive interface: medical staff can paste medical records, view summaries in real-time, and compare output differences; 3. Docker containerization: provides Dockerfile and docker-compose.yml for one-click deployment to ensure environment consistency.

The framework has wide applications in the medical field: 1. Outpatient pre-diagnosis assistance: doctors browse summaries in advance to improve consultation efficiency; 2. Medical record quality control: automatically check completeness and consistency; 3. Research data organization: extract structured information to accelerate retrospective studies; 4. Referral handover: generate concise and comprehensive summaries to ensure accurate information transmission.

Technical highlights include modular design (clear responsibilities for each module), configuration-driven (easy to reproduce and compare), progressive optimization (LoRA → QLoRA → RAG expansion), and production readiness (emphasis on deployment and user experience). Summary: PathoSummarize AI provides a complete technical solution for intelligent medical text processing, serving as a reference example for the reliable application of LLMs in medical scenarios, suitable for researchers and developers to refer to.