An Organized Approach to Using Large Language Models for Medical Information

TL;DR

This paper introduces a framework for using large language models in healthcare by defining prompting terms and demonstrating their use in patient-specific scenarios.

Contribution

The paper introduces a novel framework for structuring prompts in medical LLM applications using defined terms like 'variable' and 'clause'.

Findings

Precise combinations of variables and clauses can generate personalized medical outputs.

The framework allows for patient-specific information to be efficiently implemented.

LLMs can generate educational material that may improve healthcare outcomes.

Abstract

ChatGPT and other large language models (LLM) have increased in popularity. Despite the rapid rise in the implementation of such technologies, frameworks for implementing appropriate prompting techniques in medical applications are limited. In this paper we establish the nomenclature of “variable” and “clause” in the prompting of a LLM, while providing example interviews that outline the utility of such an approach in medical applications.

In this study assessing the LLM ChatGPT-4, we define terms used in prompting procedures including “input prompt,” “variable,” “demographic variable and clause,” “independent variable and clause,” “dependent variable and clause,” “generative clause,” and “output.” This methodology was implemented with three sample patient cases from both a patient and physician perspective.

As demonstrated in our three cases, precise combinations of variables and clauses that consider the patient’s age, gender, weight, height, and education level can yield unique outputs. The software can do so quickly and in a personalized, patient-specific manner. Our findings demonstrate that LLMs can be used to generate comprehensive sets of educational material to augment current limitations, with the potential of improving healthcare outcomes as the use of LLM is further explored.

The framework we describe represents a unique attempt to standardize a methodology for medical inputs into a large language model. Doing so expands the potential for outlining patient-specific information that can be implemented in a query by either a patient or a physician. Most notably, future projects should consider the specialty- and presentation-specific input changes that may yield the best outputs for the desired goals.

Figures