Large Language Model Usage in Healthcare | Databasin

Since the 1970s, data has been a cornerstone of decision-making in healthcare, transforming from simple statistical analyses to the backbone of sophisticated, personalized care models (1). The era saw the nascent stages of healthcare information systems, which, although primitive by today's standards, laid the groundwork for the advanced data-driven healthcare ecosystems we see today. The evolution of data science, particularly in the last few decades, has accelerated this transformation, spearheaded by academic medical centers aiming to become learning health systems. These systems continuously learn from each patient encounter to improve outcomes, an ambition that is now within reach for healthcare organizations at all levels, thanks to the lowered barriers to AI adoption.

However, the journey toward leveraging AI and data science in healthcare is nuanced. Achieving the best outcomes isn't solely about having the most advanced technology; it's equally about laying the right foundations and asking the right questions. It's crucial for health organizations to have mature cloud environments capable of uniting, governing, and infusing intelligence into data at scale, alongside sophisticated Electronic Medical Record systems. The true power of a health system's data ecosystem is unlocked only when it's applied to solving the right problems, guided by insightful questions from its people. This edition delves into a common leadership and technical debate across industries, particularly pertinent in healthcare: the choice between leveraging open-source solutions like Hugging Face, with its repository of pre-trained transformers, and opting for commercially vetted platforms such as GPT in Azure, part of Microsoft's Cognitive Services. Additionally, drawing from our hands-on experience, we'll share practical insights on identifying valuable AI use cases in healthcare. These tips are designed to empower technology and business leaders eager to shape the future of their organizations by leveraging the transformative power of AI.

Open Source vs. Commercial AI Solutions in Healthcare

The landscape of AI tools and platforms available for healthcare applications is vast and varied. On one end of the spectrum, we have open-source libraries like Hugging Face's transformers, which offer a treasure trove of pre-trained models ready for customization. These solutions provide flexibility and a low entry cost, making them appealing for organizations keen on experimenting with AI capabilities without significant upfront investment. On the other end, commercial solutions like GPT on Azure present a polished, enterprise-ready option, backed by comprehensive support and integration capabilities.

Choosing the Right Path

The decision between open-source and commercial solutions hinges on several factors:

Cost: Per Token Pricing vs. Free Open Source

• One of the most apparent distinctions between commercial and open-source AI solutions is the cost structure. Commercial solutions like GPT on Azure employ a per-token (character) pricing model. This approach means that costs scale with usage, providing a clear but not always predictable cost structure for organizations. Conversely, open-source models offered by Hugging Face and Meta's LLaMA are available freely. This access allows organizations to experiment and deploy these models without the upfront costs associated with commercial solutions. However, it's important to consider the hidden costs, such as the resources required for training, customizing, and maintaining these models. In our experience, this requires a fractional commitments across your IT or product team, from a DevOps engineer, front-end dev, a data scientist / data engineer with python skills.

Resources: The Ease of Implementation

• The ease with which data scientists can deploy and utilize AI models is crucial. Hugging Face models and Meta's LLaMA benefit from the simplicity and power of PyTorch, making them accessible to everyone with Python skills. For example, PyTorch's compatibility with the Fastai2 library simplifies the process of calling Hugging Face models and deploying additional training blocks of code. Similarly, Meta's LLaMA is supported by user-friendly Python wrappers, facilitating easy customization and training by data scientists.
• This accessibility assumes the presence of at least one skilled data scientist or data engineer within the organization who is proficient in Python. If your team is not there yet the complexities of deploying and customizing open-source AI solutions might present a significant hurdle, making commercial options the most viable choice.

Scalability and Integration: From Open Source Customization to Commercial Simplicity

• Scalability and integration are vital considerations for healthcare organizations looking to deploy AI solutions. Open-source tools offer a high degree of flexibility in this regard. The process of containerizing an open-source solution and wrapping it in a lightweight UI is well-documented, with resources readily available in repositories for both LLaMA and Hugging Face. This approach allows for significant customization to fit specific organizational needs but requires a non-trivial amount of technical skill and effort.
• Fine tuning was (and largely still is) the rallying cry for open-source LLMs, however, with the fine tuning capabilities recently released from Microsoft end users of GPT in Azure now have nearly the full suite of customization available to open-source end users without the need to write and deploy python code.
• Commercial solutions like GPT on Azure streamline the deployment process. Utilizing Azure's GPT is essentially leveraging an API that can be easily integrated across various interfaces, including custom UIs or those provided as part of the service. This ease of integration simplifies the process of scaling the solution across an organization, making it an attractive option for those seeking a straightforward, efficient deployment path.

In summary, the choice between open-source and commercial AI solutions in healthcare is nuanced, involving careful consideration of costs, resources, scalability, and integration capabilities. Open-source solutions offer a cost-effective, flexible option for organizations equipped with technical expertise and resources for customization. In contrast, commercial options provide a streamlined, user-friendly path to deployment, particularly suited to organizations looking for simplicity and scalability in their AI initiatives.

Identifying Valuable AI Use Cases in Healthcare

The implementation of AI in healthcare must be guided by value-driven considerations. Identifying high-impact use cases involves understanding patient needs, operational challenges, and research opportunities.

Framework for Impact

A systematic approach to identifying AI opportunities can be based on value and feasibility.

Original source: Gartner | Utilized for this use case by Dr. Thomas Maddox on February 1, 2024 in Innovation in Health: AI Use Cases

Implementation Blueprint

Deploying AI in healthcare involves ensuring readiness prior to project initialization. Four common areas across all use cases based on our experience are as follows:

• Assessing Data Readiness: Ensuring data is clean, unified, and accessible.
• Building Interdisciplinary Teams: Combining technical AI expertise with clinical insights.
• Prioritizing Ethical Considerations: Addressing privacy, bias, and fairness in AI applications.
• Adopting a Cycle of Continuous Improvement: Leveraging feedback loops to refine AI solutions.

Conclusion

Navigating the choice between open-source and commercial AI solutions in healthcare is a nuanced decision, influenced by an organization's specific needs, capabilities, and goals. Beyond this choice, the successful implementation of AI hinges on identifying valuable use cases that address real-world challenges in patient care and operational efficiency. As healthcare organizations strive to become learning health systems, the integration of AI and data science is the fuel for transformative patient outcomes and operational excellence. The path forward is clear: lay the right foundations, ask the right questions, and harness the power of AI to unlock the full potential of healthcare's data revolution.

Citation (1): Liyanage, Harshana, et al. "Artificial intelligence in primary health care: perceptions, issues, and challenges." Yearbook of medical informatics 28.01 (2019): 041-046.