Validating a Medical AI Fine-Tuning Platform for Major Diseases
Trial statusRecruiting
Trial phaseNot listed
Trial typeObservational
Biological sexAll
Age18+
SponsorBeijing Friendship Hospital
About this trial
The goal of this observational study is to leverage the abundant patient resources and standardized medical records from Beijing Friendship Hospital, Xuanwu Hospital, and Beijing Anzhen Hospital, combined with the existing data and knowledge platform of guidelines, consensus, medical literature, and dialogue data from Beijing Haitian Ruisheng Science Technology Co.,Ltd, with Beijing Zhilan Medical Technology Co., Ltd. conducting the fine-tuning, optimization, and validation of the medical large language model. The model is fine-tuned according to the consultation and diagnostic needs of different departments to improve the quality and efficiency of hospital medical services, enhance intelligence, and elevate the level of medical care. Through deployment to hospitals at all levels, it aims to achieve standardized services and support graded diagnosis and treatment. The overall research includes medical big data construction, medical knowledge graph construction, medical large model training and fine-tuning, and large model application platform development and deployment.
Eligibility criteria
This trial does not accept healthy volunteersQualifiers
Retrospective Historical Medical Records: patients with stomach, cardiovascular and cerebrovascular diseases from September 2014 to August 2024 were enrolled.
Prospective Historical Medical Records: Patients with stomach, cardiovascular and cerebrovascular diseases from September 2024 and August 2027 are enrolled.Outpatient and emergency records are used for large model training, and inpatient records are used for both training and internal validation. Inpatient records are allocated to the training set and internal validation set at a 3:1 ratio. Block randomization is used to reduce bias. Each sample is assigned a raw random number uniformly distributed between 0 and 1. Under the block design, each block contains 4 samples. Within each block, samples are ranked by the raw random number and assigned a random code from 1 to 4. The randomization schedule is prepared by a statistician on a computer system before the start of the study and printed on opaque, sealed envelopes.
Age ≥ 18 years; (2) Clinically diagnosed with one of the following: chronic gastritis, gastric cancer, gastro esophageal reflux, coronary artery disease, or stroke; (3) Patient or legally authorized representative able to understand the study and provide informed consent; (4) Clinically stable and able to complete the study procedures
Disqualifiers
Participant requests to withdraw from the study during the research process
Investigator determines that the study should be terminated based on consideration of the participant's best interests
Trial population
The study population consists of patients diagnosed with chronic gastritis, gastric cancer, coronary artery disease, or stroke, enrolled from three hospitals in Beijing: Beijing Friendship Hospital, Beijing Anzhen Hospital, and Beijing Xuanwu Hospital. The study includes approximately 80,000 retrospective historical medical records and approximately 16,000 prospectively enrolled patients. All participants are aged 18 years or older.
Trial design
Design model
Cohort
Time perspective
Other
Treatments tested in this trial
Smart Medical Big Data Dataset Construction
Other interventionIntegrating multimodal medical data (electronic medical records, medical images and reports, laboratory results, genetic information) with medical guidelines, expert consensus, clinical databases, medical literature, encyclopedias, patents, and doctor-patient dialogue data. Datasets are constructed in phases: pre-training general knowledge datasets (guidelines, textbooks, literature, historical records) and fine-tuning disease-specific datasets (doctor-patient dialogues, disease-specific records, treatment plans, follow-up records). Techniques include data cleaning, standardization, transformation, annotation, and augmentation. The platform adopts a human-machine collaboration strategy to reduce data annotation costs, combining medical experts' professional knowledge with AI capabilities.
Large-Scale Medical Knowledge Graph Construction
Other interventionExtracting multimodal knowledge graphs from guidelines, consensus, and literature using the QLora training framework. The process involves medical knowledge modeling (defining entities, relations, attributes), entity recognition (disease names, drug names), relation extraction (disease-symptom, drug-treatment relationships), attribute extraction (incidence, dosage), and knowledge fusion and completion to build a tens-of-millions-level knowledge graph. The construction incorporates multimodal data integration (text, images, audio) through image recognition and speech recognition technologies.
Multi-Disease Medical Large Model Fine-Tuning Platform Construction and Fine-Tuning
Other interventionDeveloping fine-tuning techniques for different disease areas using the ChatGLM large model, with integration of expert feedback. Pre-training uses large-scale medical datasets with language modeling tasks. Fine-tuning incorporates physician annotation, data augmentation (multi-task learning, adversarial training), and prompt-tuning to generate disease-specific responses. Deployment applies quantization to reduce model size and improve inference speed. A quality evaluation system is developed as a standardized method for measuring model performance, incorporating customized methods tailored to specific diseases and drawing on existing frameworks such as Ragas and CMB.
Medical History Collection System Development
Other interventionBuilding an intelligent medical history collection system using voice interaction, natural language processing, and large model technologies. Speech recognition and NLP techniques collect patient information through multi-round dialogue with strategies including questioning, paraphrasing, feeling reflection, and self-disclosure. Physicians supervise and validate the model through Reinforcement Learning from Human Feedback to improve collection efficiency. After collection, the model extracts key information (chief complaint, history of present illness, past medical history, personal history, allergy history) and uses Retrieval-Augmented Generation (RAG) combined with guidelines for preliminary triage.
Clinical Auxiliary Diagnosis Research
Other interventionBuilding an intelligent clinical auxiliary diagnosis system with multi-source data fusion integrating heterogeneous medical records, laboratory reports, and imaging text reports using OCR and natural language processing (NLP). Using chain-of-thought and reasoning capabilities of large language models combined with knowledge graph retrieval to generate evidence-based diagnostic recommendations. Physicians provide feedback through a data annotation interface to fine-tune the system, with customization settings to meet individual physician needs.
Medical Record Generation and Quality Control Research
Other interventionBuilding a medical record generation system enabling automatic entry of patient information and automatic medical record generation using large language models for entity recognition, text classification, and semantic understanding. The system generates records compliant with writing standards including chief complaint, history of present illness, past medical history, personal history, marital and reproductive history, family history, and auxiliary examinations. A quality control system automatically detects generated records and, when core fields are missing, alerts physicians through highlighting and submission blocking, prompting them to make necessary additions and modifications.
Doctor-Patient Consultation Audio Data Collection and Processing Research
Other interventionProviding a solution for consultation audio data collection and processing to support intelligent history collection, rapid triage, and clinical auxiliary diagnosis. The collected audio is converted to accurate textual content with preliminary analysis to identify key medical information (symptoms, medications, treatment processes). The study addresses speech recognition challenges including specialized terminology, dialects, accents, and elderly speech characteristics, using refined speech processing algorithms and deep learning approaches. Strict data protection measures (encryption, anonymization) are applied throughout collection, transmission, storage, and processing to ensure patient data security.
Cross-Center External Validation
Other interventionValidating the generalizability of the five intelligent diagnosis and treatment systems (gastric cancer, chronic gastritis, GERD, coronary artery disease, stroke) by transferring the fine-tuned systems from the training center to other centers for external cross-validation. Feasibility and accuracy are evaluated by comparing information collected from large model-patient interactions and physician-patient interactions across different center environments.
Treatment groups
96,000 Participants
are divided into 2 treatment groupsGroup A: Stomach Diseases8 interventions
Group B: Cardiovascular and Cerebrovascular Diseases8 interventions
Trial outcomes
Primary outcomes
1
Establishment of the Platform
The medical large language model fine-tuning technology platform is established based on the ChatGLM foundation model. The overall research includes medical big data construction, medical knowledge graph construction, medical large model training and fine-tuning, and large model application platform development and deployment.
Time frame
From enrollment to completion of the platform development, assessed up to 2 yearsSecondary outcomes
1
Establishment of the Medical Dialogue Audio Collection System
Development of a high-quality doctor-patient dialogue audio collection and recognition system for complex medical scenarios, providing the data foundation for medical history collection and medical record generation. The system uses a self-developed multi-microphone array and sound source localization technology for precise speaker positioning, with wiener filtering and echo cancellation to reduce environmental noise. Deep learning is used for speech recognition, and large models are applied for post-processing to correct recognition errors.
Time frame
From enrollment to completion of the platform development, assessed up to 2 years2
Establishment of the Medical Big Data Platform Database
This study has accumulated multimodal medical data including electronic medical records, medical images, laboratory results, and genetic information, as well as medical guidelines, expert consensus, clinical databases, medical literature, encyclopedias, patents, and doctor-patient dialogue data. A multi-source heterogeneous database is constructed covering images, audio, video, text, and tables from different stages inside and outside the hospital. A self-developed text recognition tool digitizes PDF guidelines and literature while preserving tables, images, and LaTeX-format formulas. A medical machine translation tool translates English literature into Chinese. Through knowledge extraction, fusion, and completion, a tens-of-millions-level medical knowledge graph is built. Data cleaning, standardization, transformation, annotation, and augmentation convert raw data into forms suitable for model training and diverse clinical applications
Time frame
From enrollment to completion of the platform development, assessed up to 2 years3
Establishment of the Large Model Fine-Tuning Technology
Based on the ChatGLM model, this study develops a full-cycle method and platform for domain-specific large model fine-tuning using technologies including Continue Pretrain, Multi-task Instruction Pretrain, Reinforcement Learning, In-context Learning, Mixture-of-Experts, Chain-of-Thought, Parameter-efficient fine-tuning, and prompt-tuning, establishing a comprehensive, objective, and effective medical model fine-tuning system in combination with real-world application scenarios. The technology can be applied to large model fine-tuning across different diseases, meeting the needs of pre-consultation, in-consultation, and post-consultation scenarios. Combined with task-specific annotation data, instruction fine-tuning enhances capabilities in knowledge question answering, summarization, knowledge extraction, medical record generation, and auxiliary diagnosis. The model is deployed and further optimized through reinforcement learning with human feedback from physicians and patients
Time frame
From enrollment to completion of the platform development, assessed up to 2 years4
Establishment of the Disease-Specific Knowledge Graph Construction Technology
Prompt Learning and Few-shot Learning enable few-shot knowledge extraction using human-in-the-loop iteration (pre-annotation, manual verification, continuous training). Self-developed knowledge linking and alignment algorithms integrate knowledge from guidelines, textbooks, consensus, and medical records. A knowledge graph extraction method combining local semantic and global structural information is proposed for digestive, cardiovascular, and cerebrovascular diseases. Entity recognition and representation are performed on the large model, and graph neural networks learn from existing knowledge graphs to mine entity relationships. Under an active learning framework, high-value data are manually annotated, enabling rapid model iteration with low cost to construct disease-specific knowledge graphs for auxiliary diagnosis and other scenarios
Time frame
From enrollment to completion of the knowledge graph construction, assessed up to 2 yearsSponsors and contacts
Click on the lead sponsor to view all of their trials.
Source ClinicalTrials.gov