Recently Large Language Models (LLMs), with the most famous one (chat)GPT, made leaps in AI's capabilities to understand medical language and context; from passing the US medical licensing exam to summarizing clinical notes (https://sites.research.google/med-palm/).

While they are powerful general-purpose models for language understanding, LLMs come with several issues; a lack of transparency, high costs, environmental impact, errors due to hallucinations and increased vulnerability to hacking and data privacy breaches.  

These concerns prevent more wide-spread adoption and block many use cases.
‍

Quality and accuracy in clinical contexts

LynxCare provides fit-for-purpose, smaller specialized models that are tailored to the specificities of the healthcare sector and contexts of use, which is vital to creating high-quality Real-World Evidence from text.  

Clinical text includes specific use of language, acronyms and medical jargon that varies across different healthcare providers and departments, while LLMs are pre-trained on vast amounts of open-source materials on the web. An NLP system that performs well for one use case might not provide quality results for another due to the change in language and terminology.

At LynxCare, by precise fine-tuning and thus reducing the size of large models, and training the NLP models in the specific domains of healthcare, we are able to support new therapeutic areas using only modest amounts of labeled, annotated data, while providing optimal quality and accuracy, matching that of much larger models.
‍

Transparency

Large Language Models lack transparency, LynxCare’s rigorous approach, encompassing clinical and medical validations along with a comprehensive clinical validation dashboard, guarantees data quality transparency. Our validation dashboard serves as a powerful tool, allowing clinicians to gain valuable insights into the association between clinical concepts and their corresponding data points (CUIs), while also offering visibility into the utilized data sources. This ensures a robust foundation of trust and clarity in our model's performance and the clinical integrity of the insights it provides.
‍

Infrastructure costs

Using state-of-the-art LLMs requires specialized hardware and calling an external cloud-API requires sharing data with cloud providers. Due to the confidentiality of patient data, many RWE projects need to be carried out entirely on premise i.e., using the existing clinical IT infrastructure where computational resources are limited.

By training specialized models, we reduce the requirements on infrastructure, hence the cost, and deliver models that can be deployed privately.
‍

Data safety

LynxCare is in line with (inter)national data privacy regulations across the globe, as well as recently renewed its ISO 27001 and NEN 7510 certifications. These certifications underscore our ongoing dedication to preserving the confidentiality, integrity, and availability of sensitive healthcare data. The ISO 27001 and NEN 7510 standards serve as crucial benchmarks in our industry.  

In conclusion, it’s essential for healthcare providers to carefully consider the strengths and limitations of large and general-purpose AI models.  

While LLMs have pushed the boundaries of language understanding by machines, they are not as effective as specialized models when it comes to extraction and generation of qualitative Real-World Evidence.