The Promises and Challenges of Generative AI on Healthcare and the EMR

JASON LU–Recent advancements in computing and machine learning have brought about revolutionary changes in industries spanning from education to healthcare. The release of Large Language Models (LLMs) including OpenAI ChatGPT, Meta Llama, and Google Gemini has propelled AI into the public eye, showcasing impressive abilities to process text and perform tasks in natural language. The responses generated by these models can be modulated by a temperature parameter that adjusts the predictability of the response. To create such a model, developers and researchers train the system on large amounts of text data to predict and generate responses in natural language. As these technologies continue to improve, their transformative influence on healthcare is undeniable. In addition, the extent of AI’s impact on healthcare is still developing, making this an exciting time for supporters.

However, skepticism is shared across many industries where the rapid adoption of AI is applicable. Biases are often highlighted as a significant challenge to the widespread adoption of AI. Models are inherently limited by the data they are trained on. If those datasets are skewed or biased in any way, the outputs of the model will reflect those biases. On a different note, job displacement is a real concern. As the automation of tasks improves, many fear that AI will threaten jobs. This sentiment has been voiced within the medical field, with some physicians worried that certain diagnostic specialties, such as radiology and pathology, may be overtaken by AI. These physicians are concerned that AI could replace their jobs while providing subpar information. However, significant benefits have also been noted. As these models continue to evolve, there will be remarkable improvements in their capabilities. A study published in 2023 demonstrated that ChatGPT was able to pass the United States Medical Licensing Examination. While these advancements showcase the diagnostic potential of AI models, the most immediate and impactful application of AI in healthcare is likely within the electronic medical record (EMR) system.

The EMR, first implemented in the US in the early 1970s, has been a major technological advancement in healthcare. It has significantly improved accessibility to patient information, safety, and communication among providers. However, it also has limitations. Errors, often due to inconsistencies in data entry, make EMRs prone to inaccuracies. They often contain vast amounts of information that can be difficult to sift through, and their usability can be challenging. These limitations present applicable and simple opportunities for LLMs and generative AI to be implemented in healthcare. Such implementation has been highlighted as a potential solution to reduce the administrative burden, particularly in resource-limited practices. 

A recent study by Lake Country Medical Group showed a significant improvement in rural documentation burden in a chronic care management setting with the implementation of Freed AI for charting. Another study by the Department of Ophthalmology at Oregon Health & Science University further supports these benefits of AI in EMRs but notes that progress is limited in their field due to the lack of adequate training data in EMRs. Additionally, Epic Software, the most prominent EMR company, has been bolstering its development of generative AI in their products, which further suggests that the greatest immediate impacts of AI in healthcare will be in electronic health records. Beyond electronic charting, there are many other exciting applications of AI. For example, Johnson & Johnson is developing digital solutions that utilize AI to capture the most critical parts of a surgery, allowing for more efficient training of surgeons.

Generative AI models are not without their limitations. An interesting phenomenon has emerged in the development of LLMs, which researchers have termed “hallucinations.” Developers suggest that hallucinations occur due to these models being trained on probabilities rather than facts, the overgeneralization of responses, and incorrect training data. Another significant barrier to AI adoption is a financial one. Training AI models is extremely computationally expensive, requiring significant expenditure on graphical processing units (GPUs) and energy. Current-generation Blackwell GPUs from NVIDIA are priced between $30,000 and $70,000. Companies like Microsoft and Meta have invested a combined $9 billion in GPUs for their generative AI models in 2023.

Generative AI has remained a hot topic in recent years. The prospect of AI-assisted physicians and nurses is exciting for the future of healthcare. While the potential benefits are significant, there will undoubtedly be challenges to overcome. Currently, generative AI is expected to reduce the administrative burden of healthcare in the foreseeable future. However, the long-term impact of these advancements on healthcare in the coming decades remains open to speculation.

Copy Editor – Elizabeth Vaitl

Photography Source – https://nextgeninvent.com/blogs/use-cases-of-generative-ai-in-healthcare/