Page 20 - AIH-1-4

Artificial Intelligence in Health AI scribe in clinical documentation

transcripts of clinical encounters will likely produce 3.7. Involvement of clinicians in co-design and
significantly improved accurate results. Some of these implementation
models are already available, for example, Alphabet’s Clinicians’ input in the design and implementation of
MedPalm. Some companies are even offering end-to-end any new system in healthcare is crucial for its success.
9
pipelines starting with speech-to-text transcription to note As the primary users of AI scribe applications, clinicians
generation. 10 will have a deep understanding of their operational needs
3.4. Retrieval-augmented generation (RAG) and workflow requirements. This will not only make such
applications more effective but also improve adoption by
LLMs store data in their parameters. However, their ability clinicians. Similarly, ongoing education and training of
to retrieve and present precise information remains limited, clinicians along with the acquisition of their feedback, will
leading to subpar performance in knowledge-intensive ensure seamless integration as well as improvement of the
tasks compared to more task-specific architectures. This application.
can be overcome by providing the model access to “non-
parametric” data, known as RAG. The combination of 4. Conclusion and prospects
parametric information with explicit non-parametric
information can lead to much more accurate output. Clinical documentation is a crucial component of
11
This, when applied to AI scribes, can potentially improve modern healthcare, but it also contributes significantly
the quality of the generated note significantly. to the burnout of clinicians. AI-based technologies, like
AI scribes provide a potential solution to alleviate this
3.5. Small LLMs burden. Even though current technology is not without
challenges, the prospects are promising. Anticipating
Another potential is the use of “tiny LLMs” or “small widespread demand, EHR vendors will likely incorporate
LLMs.” These are LLMs with a smaller number of AI models in the core of their software, enabling not
parameters. The idea is that LLMs contain large amounts just AI scribes but also improving clinical decision
of generic data that may add little value to a specific support systems, automatic summarization of medical
task. Therefore, the models are trained and fine-tuned history, and research. Furthermore, AI also enables
on smaller amounts of more specific high-quality data the advancements in patient-facing EHR systems that
to improve their performance while keeping their size
and thus computational expense low. The performance allow for documentation and provision of personalized
of these smaller LLMs for text summarization has information.
been shown to be poor compared to larger LLMs. Acknowledgments
12
However, there is potential for improvement through
various methodologies. For example, knowledge can be None.
transferred from a larger LLM to a smaller LLM to achieve
better performance through improving “reasoning” by the Funding
smaller model. This methodology showed that the smaller None.
LLM can even outperform some of the larger LLMs for
certain tasks. In the context of AI scribes, this can be Conflict of interest
13
very beneficial. Not only can the output be improved, the The author is the founder of a company that specializes
financial burden involved with implementing in-house in AI scribe services, which is relevant to the topic of
LLMs can be reduced significantly. this article. This has not influenced the content of the

3.6. In-house AI solutions manuscript. No reference to the author’s company is made,
but it is declared for full transparency.
To ensure true data security, complete control over
data and customization, in-house AI solutions could Author contributions
be implemented. Training and implementing LLMs is a This is a single-authored article.
computationally heavy task, which necessitates a significant
financial investment in the initial phase. However, this Ethics approval and consent to participate
initial investment will pay off in the long term and may
even prove more profitable by reducing physician burnout, Not applicable.
improving efficiency, and ensuring data security. It will also Consent for publication
offer seamless integration with in-house systems, reducing
technical difficulties. Not applicable.

Volume 1 Issue 4 (2024) 14 doi: 10.36922/aih.3103