Artificial Intelligence Scribe and Large Language Model Technology in Healthcare Documentation: Advantages, Limitations, and Recommendations

Abstract

Artificial intelligence (AI) scribe applications in the healthcare community are in the early adoption phase and offer unprecedented efficiency for medical documentation. They typically use an application programming interface with a large language model (LLM), for example, generative pretrained transformer 4. They use automatic speech recognition on the physician-patient interaction, generating a full medical note for the encounter, together with a draft follow-up e-mail for the patient and, often, recommendations, all within seconds or minutes. This provides physicians with increased cognitive freedom during medical encounters due to less time needed interfacing with electronic medical records. However, careful proofreading of the AI-generated language by the physician signing the note is essential. Insidious and potentially significant errors of omission, fabrication, or substitution may occur. The neural network algorithms of LLMs have unpredictable sensitivity to user input and inherent variability in their output. LLMs are unconstrained by established medical knowledge or rules. As they gain increasing levels of access to large corpora of medical records, the explosion of discovered knowledge comes with large potential risks, including to patient privacy, and potential bias in algorithms. Medical AI developers should use robust regulatory oversights, adhere to ethical guidelines, correct bias in algorithms, and improve detection and correction of deviations from the intended output.

Fig. 1.

Diagram of how AI scribe works.

Fig. 2.

Advantages of AI scribe.

Fig. 3.

Limitations of AI scribe.

Fig. 4.

Recommendations for improvement of AI scribe.