Qingnang Smart Diagnosis - MedFlow All - Process Medical Framework

| English | 中文 |

0. Latest News 🎉🎉

• [2025-03-28] Qingnang Smart Diagnosis - MedFlow All - Process Medical Framework published!

1. Introduction

Qingnang Smart Diagnosis is an end-to-end AI healthcare framework with field-proven application capabilities, designed to provide efficient and intelligent solutions for the medical industry. By integrating advanced large language model (LLM) technologies, this framework has realized core functions including intelligent triage, patient registration, medical record creation, diagnosis, ordering lab tests, prescriptions, and treatment plan development. We are dedicated to enhancing healthcare efficiency, improving patient experiences, and driving intelligent transformation in the medical industry. With a focus on ease of use and adaptability, this framework supports seamless integration via APIs with leading LLMs such as DeepSeek, Qwen, Llama, and Baichuan. Whether deployed in cloud-based or on-premises environments, it enables rapid construction of intelligent, healthcare-centric service systems.

Core Features

• End-to-End Interface: Supports HTTP service interfaces for pre-consultation, in-consultation, and post-consultation processes.
• Flexible Data Integration: Enables customization of user-specific databases (e.g., drugs, disease names) and enhances implementation effectiveness through LLM + retrieval.
• High-Compatibility Deployment: Allows direct access to various models via OpenAI services.
• Multi-Scenario Adaptability: Covers diverse application scenarios such as hospitals, telemedicine, and pharmaceutical distribution.
• Multi-Role Support: Offers comprehensive functionality for medical staff, patients, and administrators, streamlining workflows for all parties.

Technical Advantages

• Agent Capabilities: Features meticulously optimized prompt instructions and reasoning workflow orchestration.
• Key Challenge Solutions: Provides all solutions combining large model intent recognition, intent fallback, SQL distribution interfaces, and multi-channel retrieval.
• Developer-Friendly Design: Focused on practical results, offering lightweight technical examples (e.g., intent, retrieval, database) with application tips for easy secondary development and upgrades.
• Ready-to-Use: Comes pre-equipped with fine-tuned and enhanced training for medical domain-specific models.
• Demo Interface: Visual business trial interface (user-friendly for non-technical personnel).

2. Model Downloads

All large language models can be accessed via standard OpenAI services. Our compatibility testing has confirmed support for: Deepseek v3, Deepseek R1, Qwen2.5, Llama 3.1, Baichuan4, QwQ, etc. For models with strong reasoning capabilities like Deepseek R1 and QwQ, format parsing components may require adjustments.

Our fine-tuned and enhanced medical large language models, along with voice interaction-based Chinese ASR, Cantonese ASR, and TTS models, are available via the following links:

Model	Foundational Model	Format	Download Link
Qingnang-72B	Qwen2.5-72B-Instruct	HuggingFace	ModelScope
Qingnang-9B	GLM4-9B-Instruct	HuggingFace	ModelScope
Qingnang-ASR	Whisper V3	HuggingFace	ModelScope
Qingnang-TTS	GPT-Sovits2	HuggingFace	ModelScope

3. Main Features

Main Features 🏆

This framework demonstrates exceptional performance in two core functional modules— Dialogue Interaction and Clinical Inference —delivering benefits to hospitals, physicians, and patients:

• Intelligent Dialogue System: Enables precise and natural AI-driven patient-doctor interactions, supporting tasks such as patient registration, pre-consultation triage, and follow-up care through conversations with patients. Supports medical record generation, quality check, and revision via dialogues with physicians.
• Clinical Reasoning Engine: Provides core reasoning capabilities including diagnosis, medication recommendations, and ordering lab tests. Critical workflows achieve over 95% accuracy.

Functional Module List

Function	Process Stage	Type	Users
Patient Registration	Pre-consultation	Dialogue Interaction	Patients, Hospitals
Intelligent Appointment	Pre-consultation	Dialogue Interaction	Patients, Hospitals
Symptom Pre-Consultation	Pre-consultation	Dialogue Interaction	Patients, Physicians
Department Recommendation	Pre-consultation	Dialogue Interaction	Patients, Hospitals
Medical Record QC	In-consultation	Dialogue Interaction & Clinical Inference	Physicians, Hospitals
Medical Record Generation	In-consultation	Dialogue Interaction	Physicians
Specialist Record Gen.	In-consultation	Dialogue Interaction & Clinical Inference	Physicians
Disease Diagnosis	In-consultation	Clinical Inference	Physicians
Medication Recommendations	In-consultation	Clinical Inference	Physicians
Medical Test Ordering	In-consultation	Clinical Inference	Physicians
Prescription Generation	In-consultation	Dialogue Interaction & Clinical Inference	Physicians
Treatment Plan Development	In-consultation	Dialogue Interaction & Clinical Inference	Physicians
Follow-up Management	Post-consultation	Dialogue Interaction	Patients, Hospitals

4. Quick Start

4.1 Environment Configuration

We recommend the following Docker/Conda environment configurations:

• Ubuntu20.04、python3.12
• NVIDIA-SMI 535.154.05、Driver Version: 535.154.05、CUDA Version: 12.2

You can use your own environment that meets the requirements, or launch a container via the following commands:

docker pull nvcr.io/nvidia/cuda:12.2.0-devel-ubuntu20.04
sudo docker run -itd --name <container_name> -v /home/<username>:/home/workspace --gpus all --ipc=host -v /root/.ssh:/root/.ssh --cap-add=SYS_PTRACE --privileged=true --ulimit stack=68719476736 --network host --shm-size=20G nvcr.io/nvidia/cuda:12.2.0-devel-ubuntu20.04 /bin/bash

Code Download

git clone -b develop https://github.com/MedFlow2025/medflow.git

Dependency Installation

Please confirm your Python version and run:

pip install -r requirements.txt

4.2 Service Operation

OpenAI Service Preparation

You need to deploy your models in advance using tools such as ollama、fastchat、vllm. We will also release deployment documentation for reference in the ./docs directory.

After deployment, your OpenAI-compatible service will have a URL similar to:

http://<openai ip>:<openai port>/v1 

Data Preparation (Optional)

We offer a function to import a custom knowledge base, which includes information about diagnoses, medicines, examination forms, etc. For specific examples, please refer to the ./database directory. You can run the following script:

python3 create_database.py

We also provide a function to customize quality inspection rules. For specific examples, please refer to the quality/quality.json file. The changes will take effect after you modify the file and restart the service.

Service Startup

python3 inference.py --model <model_name> --model-url http://<openai ip>:<port>/v1 --fastbm25 --log --host <server ip> --port <server port> --max-round 30 --database ../data/processed/database

WebUI Interface Startup

python3 inference_gradio.py --host <server ip> --port <server port> --gradio-port <webui port> --model <model_name>

4.3 Function Experience

Request Testing

We provide documentation for curl request testing and scripts in the example directory. For detailed usage instructions, please refer to the curl request instruction document..

cd ./tests

# Dialogue Generation for Patient Registration
bash test-clientinfo.sh

# Dialogue Generation for Medical Guidance
bash test-hospitalguide.sh

# Dialogue Generation for Pre - consultation Report
bash test-basicmedicalrecord.sh

# Dialogue Generation for Appointment Booking
bash test-register.sh

# Diagnosis Generation
bash test-diagnosis.sh

# Generation of Examinations and Tests
bash test-examineassay.sh

# Generation of Treatment Plan
bash test-therapyscheme.sh

# Dialogue Generation for return Visit
bash test-returnvisit.sh

# Medical Record Generation
bash test-doctormedicalrecord.sh

Webui Experience

You can open the URL prompted after starting the WebUI interface to experience the functions of each module:

http://<webui ip>:<webui port>

5. Statement of Agreement

• The use of this framework's code and models must comply with the Apache 2.0 open-source license. Commercial use is permitted without requiring authorization. Please review and adhere to the following: Prohibited Uses: Do not use open-source models, code, or derivatives thereof for any purpose that may harm the country and society. Do not deploy these resources for services that have not undergone security assessments and filings.

• Although measures have been taken during training to ensure data compliance and accuracy, due to the large parameter size of the models and probabilistic randomness factors, we cannot guarantee the accuracy of outputs. The models are also susceptible to input instruction biases. This project assumes no liability for risks related to data security, public opinion risks, or consequences arising from model manipulation, misuse, dissemination, or improper utilization. You will assume full responsibility for all risks and consequences arising from the use, reproduction, distribution, and modification of the models and other components of this open-source project.

6. Contact Information

• This framework has been field-deployed in multiple top-tier hospitals across China. 🚀
• For any issues or suggestions during use, please submit them via the GitHub Issues page of this project—we will respond promptly!
• For medical or other domain data processing, routine training, or enhancement training, we recommend using the EPAI platform (Enhanced Platform for AI Innovation)!
• For private hospital deployments or HIS (Hospital Information System)-related customization needs, our partner with experienced medical vendors: https://www.skytek.com.cn. Contact us for referrals!
• We aspire to drive transformative changes in healthcare through large model inference deployments, delivering superior services to patients and medical staff. We sincerely invite your participation and support!