Fine-Tuned LLMs

Healthcare Fine-Tuned Open-Source LLMs

The Bottleneck Facing Most Clinical Teams

Most clinical teams are highly interested in applying GenAI in clinical care but face significant barriers:

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.
Accuracy Issues: Smaller open-source LLMs like Llama3.1-8B are computationally affordable, but their diagnostic accuracy does not meet the high standards required for clinical use.
Technical Barriers: To achieve high accuracy, clinical teams need to fine-tune the smaller open-source LLMs. However, the technical challenges involved in fine-tuning are often too complex to tackle independently.

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

To address these barriers, 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 and the necessary LLM fine-tuning services, we significantly reduce both technical and cost barriers. This enables clinical teams to start LLM clinical research immediately.

The Llama3.1-8B model can be trained and deployed on a single GPU server, making it an affordable option for most research projects. This ensures that every clinical team has the opportunity to conduct LLM studies and contribute to advancing predictive healthcare.

LLM Fine-Tuning Technology Services

If you require fine-tuned open-source LLMs for GenAI clinical research and applications, contact us through the platform or email us at support@elhsi.org. We will manage the technical aspects of your LLM project, including:

Helping develop GenAI clinical study plans.
Creating theoretical fine-tuned Llama3.1-8B first using synthetic patient data, removing the bottleneck, and then custom fine-tuned the LLM using your local patient data, tailored to your requirements.
Deploying the fine-tuned LLMs with a chatbot for easy integration and testing in real clinical settings.
Guiding your team in preparing datasets for training and validation from patient records.
Assisting in preparing findings and manuscripts for submission to top journals.

This collaboration allows you to focus on clinical evaluation while we handle the complexities of fine-tuning. By streamlining the process, you can conduct GenAI clinical studies more efficiently, publish papers sooner, and advance the responsible use of generative AI in healthcare. Our services are available via an annual subscription.

Preclinically Validated Fine-Tuned LLMs Open for Research Collaborations

To facilitate the initiation of LLM clinical research for clinical teams worldwide, we have fine-tuned Llama3.1-8B for selected diseases and made these models openly available on the Hugging Face ELHSI repository for research collaborations. As demonstrated by the accuracy comparison before and after fine-tuning (see table below), some diseases are challenging to predict with the baseline model but can be accurately predicted after fine-tuning. These fine-tuned models can support productive clinical research by clearly showing how fine-tuning enhances prediction accuracy and assists clinical teams in effectively integrating GenAI into specific steps of clinical delivery to support decision-making.

We encourage clinical teams to evaluate these preclinically validated LLMs in real-world clinical settings and contribute to publishing real-world evidence on the responsible use of generative AI to improve care delivery and patient outcomes. Feel free to contact us with any questions about using these LLMs in GenAI clinical research.

Preclinically Validated Fine-Tuned Llama3.1-8B Open Models

Accuracy comparison before and after fine-tuning using synthetic patient data for selected diseases (ongoing updates)

Disease	Task	Open-source Model	Accuracy Before Fine-tuning	Accuracy After Fine-tuning
Alzheimer Disease	Predict Alzheimer Disease	Llama3.1-8B	>90%	>90%
Amyotrophic Lateral Sclerosis	Predict Amyotrophic Lateral Sclerosis	Llama3.1-8B	>80%	>90%
Chronic Traumatic Encephalopathy	Predict Chronic Traumatic Encephalopathy	Llama3.1-8B	>80%	>90%
Corticobasal Syndrome	Predict Corticobasal Syndrome	Llama3.1-8B	>20%	>90%
Creutzfeldt-Jakob Disease	Predict Creutzfeldt-Jakob Disease	Llama3.1-8B	>50%	>90%
Fatal Familial Insomnia	Predict Fatal Familial Insomnia	Llama3.1-8B	>50%	>90%
Frontotemporal Dementia	Predict Frontotemporal Dementia	Llama3.1-8B	>90%	>90%
Ischemic Stroke	Predict Ischemic Stroke	Llama3.1-8B	>80%	>90%
Lewy Body Dementia	Predict Lewy Body Dementia	Llama3.1-8B	>30%	>90%
Mild Cognitive Impairment	Predict Mild Cognitive Impairment	Llama3.1-8B	>90%	>90%
Parkinson Disease	Predict Parkinson Disease	Llama3.1-8B	>90%	>90%
Breast Cancer	Predict Breast Cancer	Llama3.1-8B	>90%	>90%
Lung Cancer	Predict Lung Cancer	Llama3.1-8B	>80%	>90%
Nasopharyngeal Carcinoma	Predict Nasopharyngeal Carcinoma	Llama3.1-8B	>60%	>90%

Note: To eliminate the upfront development cost barrier for clinical teams, we have fine-tuned the smaller Llama3.1-8B (8 billion parameters) model for an expanding list of diseases and made them openly available on Hugging Face. The model's baseline prediction accuracy—prior to fine-tuning—ranges widely, from 20% to 100%. However, after fine-tuning with synthetic patient data, the fine-tuned Llama3.1-8B model consistently achieves over 90% accuracy in predicting target diseases across various patient cases. Accuracy is calculated based on the top-2 predicted disease scores using synthetic patient datasets.