1. Large foreign medical models
1. Google Medical Large Model (Med-PaLM)
Researchers from Google and DeepMind published a study in the journal Nature. According to the results of the study, a group of clinicians scored 92.6% of the answers to the Google and DeepMind team’s large medical model Med-PaLM, which is consistent with real-life humans. The level of clinicians (92.9%) was comparable.
2、BioMedLM（PubMedGPT）
Stanford's Center for Research on Fundamental Models (CRFM) and MosaicML jointly developed the BioMedLM (PubMedGPT) model, a large language model trained to interpret biomedical language. CRFM used the MosaicML platform to train a 2.7B parameter GPT based on PubMed's biomedical data, achieving state-of-the-art results on the medical question and answer text of the United States Medical Licensing Examination (USMLE). BioMedLM was trained on the PubMed Abstracts and PubMed Central parts of the Pile dataset. The dataset contains ~50B tokens covering 16 million abstracts and 5 million full-text articles in the biomedical literature curated by the National Institutes of Health.
3、GatorTron
GatorTron is an electronic health record (EHR) big data model developed at the University of Florida, using an LLM developed from scratch (not based on other pre-trained models) using 8.9 billion parameters and >90 billion words of text from electronic health records. Improve 5 clinical natural language processing tasks, including medical question answering and medical relationship extraction.
While much smaller than Med-PaLM's model, this is the first medical-based model developed by an academic medical institution rather than a large tech company like Google, OpenAI, or Meta.
This data source pulls a total of 290 million clinical notes from 2.47 million patients from the UF Health Integrated Data Repository (IDR), the UF Health system's enterprise data warehouse. The notes were created from 2011-2021 from more than 126 clinical departments and approximately 50 million contacts, covering healthcare settings including but not limited to inpatient, outpatient and emergency department visits. After preprocessing and de-identification, the corpus includes more than 82 billion medical words.
4、CLINICAL QA BIOGPT (JSL)
John Snow Labs has long been a leader in natural language processing (NLP) tools and algorithms for medical use cases. In addition to data labeling and extraction, they also have tools for de-identifying clinical notes and medical data. JSL recently announced an LLM based on BioGPT, an older, smaller large-scale language model trained on medical information (BIOGPT (JSL)) , through fine-tuning based on JSL data and NLP tools. The model may perform better, and may even outperform ChatGPT, in areas such as patient de-identification, entity resolution (such as extracting operation codes and medical terms), and accuracy of clinical summaries.
https://nlp.johnsnowlabs.com/2023/04/12/biogpt_chat_jsl_en.html
5、ChatDoctor
ChatDoctor: A large medical model fine-tuned on the large language model LLaMA using medical domain knowledge.
More than 700 diseases and their corresponding symptoms + required medical examinations + recommended drugs were collected to generate a data set of 5k doctor-patient conversations. In addition, a dataset of 200k real doctor-patient conversations was obtained from an online Q&A medical consultation website.
Use 205k doctor-patient dialogue data set to fine-tune LLM, and the generated model can be understood in Significant improvement in ability to respond to patient needs, provide sound advice and provide assistance in various medical related areas.
In addition, in order to improve the credibility of the model, the project also designed a knowledge brain based on Wikipedia and medical field databases, which can access authoritative information in real time and answer patients' questions based on this reliable information. The medical field is crucial.
Experiments show that the fine-tuned model of doctor-patient dialogue exceeds ChatGPT in terms of precision, recall and F1.
https://www.yunxiangli.top/ChatDoctor/
2. Chinese Medical Large Model
1、DoctorGLM
Chinese consultation model based on ChatGLM-6B
Based on the Chinese consultation model of ChatGLM-6B, fine-tuning is carried out through the Chinese medical conversation data set, including fine-tuning and deployment of lora, p-tuningv2, etc.
Github address: https://github.com/xionghonglin/DoctorGLM
2、BenTsao
The LLaMA-7B model that has been fine-tuned/instruct-tuned with Chinese medical instructions is open sourced. A Chinese medical instruction data set was constructed through the medical knowledge graph and GPT3.5 API, and on this basis, LLaMA instructions were fine-tuned to improve LLaMA's question-answering effect in the medical field.
Address: https://github.com/SCIR-HI/Huatuo-Llama-Med-Chinese
3. BianQue
A large medical dialogue model that has been fine-tuned through instructions and multiple rounds of inquiry dialogues. It is based on ClueAI/ChatYuan-large-v2 as the base and fine-tuned using a mixed data set of Chinese medical question and answer instructions and multiple rounds of inquiry dialogues.
Address: https://github.com/scutcyr/BianQue
4、HuatuoGPT
A GPT-like model that has been fine-tuned/instruct-tuned (Instruct-tuning) of Chinese medical instructions is open sourced.
Address: https://github.com/FreedomIntelligence/HuatuoGPT
5、Med-ChatGLM
ChatGLM model fine-tuning based on Chinese medical knowledge, the fine-tuning data is the same as BenTsao.
Address: https://github.com/SCIR-HI/Med-ChatGLM
6. QiZhenGPT
This project uses the Chinese medical instruction data set constructed by Qizhen Medical Knowledge Base, and based on this, fine-tunes the instructions on the LLaMA-7B model, which greatly improves the effect of the model in Chinese medical scenarios. It first releases evaluation data for drug knowledge questions and answers. After the collection, the follow-up plan is to optimize the Q&A effect on diseases, surgeries, tests, etc., and expand applications such as doctor-patient Q&A and automatic generation of medical records.
Address: https://github.com/CMKRG/QiZhenGPT
7、ChatMed
This project launches the ChatMed series of Chinese medical large-scale language models. The model backbone is LlaMA-7b and uses LoRA fine-tuning. Specifically, it includes ChatMed-Consult: 500,000+ online consultations + ChatGPT responses based on the Chinese medical online consultation data set ChatMed_Consult_Dataset as the training set; ChatMed-TCM: Based on the traditional Chinese medicine instruction data set ChatMed_TCM_Dataset, based on the open source traditional Chinese medicine knowledge graph, using the entity-centric self-instruct method, calling ChatGPT to obtain 2.6w+ information surrounding traditional Chinese medicine Obtained by training on instruction data.
Address: https://github.com/michael-wzhu/ChatMed
8. XrayGLM, the first Chinese multi-modal medical model that can read chest X-rays
In order to promote the research and development of large multi-modal medical models in the Chinese field, this project released the XrayGLM data set and model, which has shown extraordinary potential in medical imaging diagnosis and multi-round interactive dialogue.
Address: https://github.com/WangRongsheng/XrayGLM
3. Large-scale medical models in the domestic industry
1. Baidu Spiritual Medicine Large Model
On September 19, 2023, Baidu officially released the first "industrial-level" medical model in China-the Spiritual Medicine Model. The large model of Lingyi focuses on three major directions: intelligent health housekeeper, intelligent doctor assistant, and intelligent enterprise services, providing AI native applications for patients, hospitals, enterprises, etc.
The large model of the spiritual doctor can combine free text to generate structured medical records in seconds, and accurately analyze the doctor-patient dialogue to generate content such as chief complaints and current history. In addition, the large model of spiritual medicine is also the only large model in the industry that supports the simultaneous analysis of multiple Chinese and English documents, and realizes intelligent question and answer based on the content of document analysis. In terms of auxiliary diagnosis and treatment, the large model of spiritual medicine can understand the patient's condition through multiple rounds of dialogue, assist doctors in real-time diagnosis of diseases, recommend treatment plans, improve the efficiency and experience of the entire medical treatment process, and become a 24-hour "health steward" for patients, providing intelligent Customer service. In addition, the large model of spiritual medicine can also provide multiple empowerments for pharmaceutical companies, including professional training, medical information support, etc.
2. Jing Tokyo Medical Qianxun
JD Health has released the "Jingyi Qianxun" large-scale medical model, which can quickly complete the migration and learning of various scenarios in the medical and health field, and achieve comprehensive AI-based deployment of products and solutions.
3. Tencent Hunyuan Medical Large Model
The data used for pre-training of Tencent's Hunyuan large model is as high as 2 trillion tokens, which is an order of magnitude higher than many models. The training data covers 2.85 million medical entities, 12.5 million medical relationships, a medical knowledge graph covering 98% of medical knowledge, and Chinese and English medical literature. This knowledge not only extracts knowledge from a large number of papers, encyclopedias, and medication instructions, but also incorporates targeted medical articles written by various medical experts in Tencent Medical Dictionary. All knowledge sources are verified, thus providing an authoritative basis for the results output by the large model.
On the one hand, it comes from patient scenarios, such as online consultation, medical Q&A, guidance, and pre-consultation; on the other hand, it comes from doctor scenarios, such as medical examination questions, medical record generation, discharge summary, examination recommendations, diagnosis results, and medication recommendations.
4. MedGPT
More than 2 billion medical text data were used in the pre-training phase, 8 million pieces of high-quality structured clinical diagnosis and treatment data were used in the fine-tuning training phase, and more than 100 doctors were invested in manual feedback supervision Fine-tuned training.
5. Shangtang “Great Doctor” model
Based on massive medical knowledge and clinical data, a large Chinese medical language model "Big Doctor" has been created, which can provide multi-scenario and multi-round conversation capabilities such as guidance, health consultation, and assisted decision-making. In addition, SenseTime has also launched a variety of vertical basic model groups such as medical imaging large models and bioinformatics large models, covering different medical data modalities such as CT, MRI, ultrasound, endoscopy, pathology, medical text, and bioinformatics data. .
6. Yunzhisheng Mountain and Sea Model
Based on the Shanhai model, Yunzhisheng will enhance its capabilities in the Internet of Things, medical and other industries, and provide customers with smarter and more flexible solutions. In the medical scenario, three major medical product applications have been released: surgical medical record writing assistant, outpatient medical record generation system, and commercial insurance intelligent claims system.
7. CareGPT
CareGPT is committed to giving full play to the value of health management in real medical service scenarios and realizing full-cycle intelligent health management capabilities of prevention, consultation, appointment and rehabilitation. The current parameter size is 7 billion, which can support multi-modal input and output in medical and health scenarios.
8. Neusoft Tianyi Medical
Doctors interact with Tianyi through natural language to quickly and accurately complete medical reports, medical records, and issue medical orders; for patients, Tianyi makes consultations more convenient and becomes the patient's private doctor all day long, providing comprehensive post-diagnosis healthy diet, Nutrition and exercise advice and other services. Tianyi's multi-modal data fusion capabilities will also provide hospital managers with conversational interactions and data insights, simplify data utilization, and make hospital management more refined.
9. Dingdang Health HealthGPT
Dingdang HealthGPT can be used as an AI health assistant to provide users with comprehensive answers to health-related questions and professional advice. Whether users are interested in medical procedures, disease treatment, drug use, interpretation of test results, or disease prevention, health care, diet and nutrition, beauty and fitness, home medical care, mental health and stress management, Dingdang HealthGPT can meet the needs of users. .
10. ChatDD
A new generation of conversational drug research and development assistant ChatDD and the world's first 100-billion-parameter multi-modal biomedical conversation large model ChatDD-FM 100B. ChatDD (Chat Drug Discovery & Design) is based on large model capabilities and can conduct multi-modal data analysis. Integrated understanding, natural interaction with experts and human-computer collaboration, connecting human expert knowledge and large model knowledge, with capabilities such as problem understanding, task dismantling, and tool invocation, may potentially redefine the drug research and development model.
11. Huawei Cloud Pangu drug molecule large model
Huawei Cloud Pangu large model has penetrated into more than 10 industries including finance, manufacturing, government affairs, electric power, coal mining, medical care, and railways, supporting the implementation of AI applications in more than 400 business scenarios. The Huawei Cloud Pangu drug molecule large model released in 2021 is a large model jointly trained by Huawei Cloud and the Shanghai Institute of Materia Medica, Chinese Academy of Sciences. It can realize artificial intelligence-assisted drug design for the entire process of small molecule drugs. Experimental verification results show that the druggability prediction accuracy of Pangu's large drug molecule model is 20% higher than that of traditional methods, thereby improving research and development efficiency, shortening the development cycle of lead drugs from several years to one month, and reducing research and development costs by 70%.
12. Zhiyun Health: ClouD GPT
Relying on basic platforms such as big data platform, machine learning platform, model development platform, and model training platform, Zhiyun Health has developed the medical industry model ClouD GPT, which has been implemented in medical application scenarios of Zhiyun AI-assisted diagnosis and AI drug and device research and development.
13. Weining Health: WiNEX Copilot
Weining Health has launched the research and development and training of WiNGPT, a large language model in the medical vertical field, in January 2023. As of April, June and September, the number of model training parameters has reached or will reach 6 billion, 15.6 billion, 650 Billion, is currently exploring more medical application scenarios, and plans to officially release the new product WiNEX Copilot powered by GPT technology in October.
14. Entrepreneurship Huikang BSoftGPT
BSoftGP will aggregate and utilize the general GPT model through API calls combined with local deployment. At the same time, through local deployment of the embedding vector database and the company's own domain knowledge base, it will gradually realize product power through language model training and fine-tuning in the medical vertical field, and provide the company with Internal and external application scenarios, such as outputting AI intelligent services in medical services and personal health scenarios.
In terms of clinical medical services, BSoftGPT can automatically generate clinical decision-making suggestions and treatment plans based on medical record information and clinical data provided by doctors, thereby assisting doctors in making clinical decisions and improving the intelligence level of the existing clinical decision support system CDSS; In terms of patient services, BSoftGPT can realize intelligent guidance and management throughout the entire process of patient pre- and post-diagnosis through natural language interaction with patients.
15. iFlytek: Spark Cognition
The upgraded iFlytek post-diagnosis and rehabilitation management platform based on the Spark cognitive large model extends professional post-diagnosis management and rehabilitation guidance outside the hospital. Based on automatic analysis of patients' health portraits, the platform can intelligently generate personalized rehabilitation plans for patients and urge patients to implement them as planned. At present, the pilot of iFlytek's post-diagnosis rehabilitation management platform has achieved remarkable results: the management efficiency of doctors in cooperative hospitals has been improved by more than 10 times, the follow-up rate and consultation response rate during patient recovery have reached 100%, and the satisfaction rate of discharged patients has reached more than 98%. .
16. Zidong Taichu, Institute of Automation, Chinese Academy of Sciences
“Zidong Taichu” is positioned as a cross-modal general artificial intelligence platform and will be officially released in 2021. On June 16 this year, Zidong Taichu released version 2.0. At present, the "Zidong Taichu" large model has shown broad industrial application prospects in neurosurgery navigation, short video content summary, legal consultation, and medical multi-modal identification. A series of leading and exemplary applications have been launched in the fields of diagnosis and traffic image reading.
In the medical field, based on the Zidong Taichu large model open service platform, intelligent data annotation, efficient model training, and flexible model deployment are realized, automatic identification and inventory of orthopedic instruments/consumables are realized, and intelligent and refined management is realized, with efficiency higher than that of traditional The method has been improved by 6 times, and the accuracy rate is as high as over 97%.
17. Shenzhen Big Data Research Institute & Chinese University of Hong Kong (Shenzhen) Huatuo GPT
In June this year, the latest internal beta version of Huatuo GPT was released in Shenzhen. Huatuo GPT, jointly developed by Shenzhen Big Data Research Institute and the Chinese University of Hong Kong (Shenzhen), uses 100 million questions and answers (50G) and 10-20T medical texts, and is the largest medical question and answer data set. Mainly used in medical consultation and emotional companionship, including patient training, health consultation, medical triage, etc.
Huatuo GPT trains and open-sources a new large medical model by fusing the "distilled data" generated by ChatGPT and the data responded by real-world doctors. Automatic and manual evaluation results show that Huatuo GPT is better than the existing Chinese medical artificial intelligence model and GPT-3.5 in both single-round and multi-round consultation scenarios, fully proving its ability to handle complex consultation dialogues. In the next step, Huatuo GPT will support multi-modal input.
18. Beijing Zhipu Huazhang Technology Co., Ltd. & Beijing University of Chinese Medicine Oriental Hospital: Digital Traditional Chinese Medicine large model based on "GLM-130B"
On June 27, Beijing's first batch of 10 large-scale model application cases for the artificial intelligence industry were released, including a digital traditional Chinese medicine large-model demonstration application jointly developed by Beijing Zhipu Huazhang Technology Co., Ltd. and Beijing University of Chinese Medicine Oriental Hospital. This project uses the high-precision Chinese-English bilingual dense model "GLM-130B" based on Zhipu Huazhang to meet the needs of mining and collating the experience of famous doctors in the field of traditional Chinese medicine, build a digital traditional Chinese medicine service platform, and explore artificial intelligence clinical diagnosis and treatment of high-risk pulmonary nodules. Evaluation research and other solutions to achieve a new model of intelligent replication of clinical experience in traditional Chinese medicine. The project has initially developed a Q&A function in the medical vertical field to support intelligent knowledge Q&A on medical and health issues. It has also developed auxiliary diagnosis and treatment functions such as generating TCM prescriptions based on symptoms and providing medical explanations of the symptoms of the prescription.
19,Harbin Institute of Technology: "Compendium of Materia Medica" Chinese Medical Model (original name: Huatuo)
According to reports in May this year, a research team from Harbin Institute of Technology trained a large Chinese medical model and named it "Hua Tuo", which was later renamed "Materia Medica". The "Materia Medica" team mainly used the Chinese medical knowledge graph CMeKG and the Chinese medical literature on liver cancer in 2023. With the help of the OpenAI API, they constructed 8,000 question and answer data and 1,000 multi-round dialogue training data respectively. Then, based on the LLaMA-7B base model, supervised fine-tuning was performed to construct a large Chinese medical model of "Materia Medica".
20,Shanghai Artificial Intelligence Laboratory: OpenMEDLab Puyi
On June 29, Shanghai Artificial Intelligence Laboratory took the lead and cooperated with top domestic and foreign scientific research institutions, universities and hospitals to jointly release the world's first medical multi-modal basic model group "OpenMEDLab" and gradually open source it. "OpenMEDLab" integrates the world's top AI research and development capabilities, massive medical data and medical expert knowledge. The first batch of basic model groups released include more than 10 data models based on medical images, medical texts, biological information, protein engineering, etc. A basic model trained state-of-the-art. This model will promote cross-domain, cross-disease, and cross-modal scientific research breakthroughs based on basic medical models. It will also help solve long-tail problems in the medical field and promote the industrial implementation of large-scale medical models.
The heart of scientific research is dedicated to exploring the combination of large AI models and scientific research. Scientific Research Heart provides you with the latest AI information, the most practical AI tools, and the most in-depth AI analysis to help you explore the unlimited potential of AI in scientific research. Pay attention to the "heart of scientific research".