Digital Health

Section 3 - How Can Generative AI, Specifically LLMs Aid in Documentation

Large Language Models (LLMs) are adept at analyzing and manipulating written language. While LLMs are classified within the domain of generative artificial intelligence (AI), the term “generative” can be misleading. Beyond generating text, LLMs can also summarize, classify, and extract discrete information. This is especially transformative for clinical documentation, which has historically required expensive and time-consuming human labor to perform these specialized tasks. Creation of clinical documents can be streamlined by distilling ambient transcripts into the familiar SOAP format. Sets of documents can be summarized—either to expedite chart review or to draft hospital discharge summaries. Notes can be classified in terms of complexity, supporting efforts to ensure accurate coding and billing practices. Discrete data can be extracted from unstructured text to enhance clinical decision support tools and predictive models. The capabilities of LLMs listed above are just a sample of how LLMs can aid in patient care and the clinician workload, with the result of improved patient care. This section of the 6-part series further explores the application of LLMs to clinical documentation. 

Utilizing LLMs to streamline the process of summarizing medical dialogues can potentially enhance understanding and retention of information exchanged during appointments, providing both physicians and patients with concise points from their discussions. Similarly, LLMs can be employed to respond efficiently to patient queries, offering quick and accurate answers to questions regarding diagnoses, treatments, or medications, thereby saving time for healthcare providers while keeping patients well-informed.  

In the documentation of patient histories and physical assessments, LLMs can aid physicians in creating detailed records based on interviews and past medical data, facilitating the structuring of comprehensive patient profiles. Complementing this, the technology can recover missing patient data from unstructured texts, helping to create a fuller picture of a patient's health and potentially aiding in more accurate diagnoses.  

Through the effective analysis of patient data, including predictive analytics, LLMs can facilitate the early identification of potential health issues before they escalate. By highlighting individuals who are at higher risk for certain conditions, healthcare providers can leverage this tool for early interventions and to plan more effective treatments. One practical application of this is in the simplification of radiology reports, where LLMs can generate plain language summaries, rendering them more accessible to both physicians and patients and aiding in a more informed and proactive healthcare approach.  

The planning and monitoring stage of treatment can also benefit from LLM integration. The technology can aid in generating treatment plans and recommending medications efficiently, automating the process to save time and convey strategies clearly to patients. Furthermore, it can be utilized in predicting disease progression, aiding healthcare providers in monitoring patients’ health over time and facilitating early interventions when necessary.  

LLMs can support clinical decision-making by providing data-driven insights into the potential risks and benefits of different treatment paths, helping to inform balanced clinical decisions. On the administrative end, these models can automate the coding of medical procedures and services, reducing the risk of denied claims and financial penalties through the identification and correction of coding errors, thereby improving the efficiency of billing processes. 

The value of generative AI, specifically LLMs, will also aid in documentation through interoperability, pointedly a patient’s longitudinal/profile data being available, which will be used it to supply truly “high value relative information” sorted from the data at the point of care. The concept of having LLMs in electronic healthcare information technology do the work in searching all available proper healthcare data and presenting that in a high-value relative informational form to the clinician at the point of care presents, potentially, an invaluable tool in caring for patients. This simply translates as the electronic HIT working for the caregiver and the patient, not vice versa. 

In achieving these improvements, advances, and innovations with LLMs in documentation, and as applied to other aspects of care, will be vital for a “trust” in the technology be developed and occur in its use for the delivery of care. Medicine and healthcare are based on trust. Patients trust caregivers, and caregivers trust technology and AI, specifically LLMs, in helping deliver that care. A key point to keep in mind is that clinical validation in this process is necessary for the data source(s) and the synthesis of the final outputs.  

The vast potential power to improve healthcare exists with LLMs and all of the AI technologies, if and when it is used to achieve the highest quality of care possible in the most appropriate ways. As Winston Churchill stated, “With Great Power Comes Great Responsibility.” Stay tuned for section 4 where we will be covering the issues and concerns with LLMs. 

Section 4 of the series will cover the vast promises and vast considerations that need to be managed when using LLMs and other generative AI models in healthcare. You can access previously published sections in this series via the links below.  

Other Sections

The views and opinions expressed in this content or by commenters are those of the author and do not necessarily reflect the official policy or position of HIMSS or its affiliates.