ELHS Institute

We are committed to democratizing generative AI (GenAI) based on large language models (LLMs) in healthcare to reduce health disparities.

Enabling you to conduct cutting-edge LLM clinical research, generate new GenAI knowledge, and publish papers in top journals. More information on ELHS platform fine-tuned LLMs page.

New clinical evidence and knowledge are needed for every disease in the following areas, which you can explore using fine-tuned open-source LLMs controlled by you:

• Improving accuracy in disease diagnosis and prediction.
• Enhancing disease risk prediction for routine screening and early detection.
• Enabling personalized healthcare through optimized treatment selection.
• Advancing disease prognosis and management to improve patient outcomes.
• Empowering primary care in community and rural areas to enhance screening, early diagnosis, personalized treatment, prognosis, and more.

The Urgent Need for Clinical Evaluation Research of LLMs

We envision that doctors will benefit from GenAI in delivering better clinical care, including preventive screening, early detection, diagnosis, personalized treatments, and disease management. By integrating GenAI tools into every step of routine clinical workflows, patient outcomes can be significantly improved.

The first step in this healthcare revolution is the integration of LLM-enabled prediction for all diseases encountered in clinical settings. However, there is currently a lack of validation data for disease prediction by LLMs in real-world clinical settings. Leading journals such as JAMA, NEJM, and Nature are urgently calling for more clinical evaluation studies of LLMs. For the responsible use of GenAI, clinical validation data is essential before GenAI can be applied in clinical settings involving patients.

The Bottleneck Facing Most Doctors

Doctors are highly interested in applying GenAI in clinical care, yet they face significant barriers:

1. Patient Privacy and Data Security: Publicly available high-performance LLMs like ChatGPT cannot be used in clinical workflows due to data privacy and security concerns.
2. Accuracy Issues: While smaller open-source LLMs like Llama3.1-8B are affordable, their diagnostic accuracy does not meet the high standards required for clinical use.
3. Technical Barriers: To achieve high accuracy, most clinical teams need to fine-tune small open-source LLMs. However, the technical challenges involved in fine-tuning are often too complex to tackle independently.

Our LLM Fine-Tuning Services Using Synthetic Patient Data Remove the Bottleneck

To remove the bottleneck for clinical teams, we have developed a new technology that uses synthetic patient data to fine-tune Llama3.1-8B, achieving over 90% accuracy in predicting a wide range of diseases. By providing preclinically validated high-accuracy fine-tuned Llama3.1-8B models (i.e. theoretical LLMs) for free and the necesary LLM fine-tuning services, we essentially reduce both technical and cost barriers, enabling clinical teams to start LLM clinical research immediately.

Overcoming this bottleneck allows clinical teams to evaluate the benefits of GenAI in clinical care, generate new evidence, and publish high-quality papers. Our step-by-step support includes:

1. Defining the Prediction Task: For the disease interested by a clinical team, we help define the prediction task and goal. We will create an initial fine-tuned Llama3.1-8B model using synthetic patient data and provide preclinical validation data for clinical review.
2. Deploying the Model: We deploy the fine-tuned theoretical LLM along with a Gradio-powered chatbot on an AWS cloud server for easy testing by the clinical team. The team retains control over the data and LLM and can terminate the server as needed.
3. Validating with Retrospective Data: The clinical team validates the fine-tuned LLM using their retrospective patient data. We guide the team in preparing datasets from electronic health records for training and validation.
4. Fine-Tuning with Real Data: We fine-tune the Llama3.1-8B model using real patient data and update the chatbot with the newly fine-tuned LLM.
5. Evaluating Impact: The clinical team validates the updated fine-tuned LLM with retrospective data. If successful, they evaluate its impact on prospective real-world data during routine clinical delivery. This comparison of outcomes before and after using LLM predictions can generate valuable evidence for clinical use of GenAI.
6. If needed, we can help analyze study results, summarize clinical evidence, and prepare manuscripts for submission to top journals.
   
   

Preclinically Validated Fine-Tuned LLMs Open for Research Collaborations

To make it easy for clinical teams worldwide to begin LLM clinical research, we have been fine-tuning Llama3.1-8B for selected diseases and making these fine-tuned models available for research collaborations. As shown in our results (table below), some diseases are difficult to predict with baseline models but can be accurately predicted after fine-tuning. These fine-tuned models enable productive clinical research, clearly demonstrating how fine-tuning improves prediction accuracy and helps clinical teams integrate GenAI into specific steps of clinical delivery.

Contact for Collaboration

For collaboration on GenAI studies and applications or any technical questions, visit our ELHS GenAI Copilot Platform.