Professional Report Generation Using Lexeme Theories and Openai's Generative Pretrained Transformer, GPT-4: A Comparison

Abstract

The public release of OpenIA's GPT-4 has caused an explosion of interest in the ability of large language models to generate text documents in response to a simple text prompt. These documents can appear to be genuine professional reports, such as medical case notes. Expert-written templates guided by lexeme theories (TGLT) is a system under development which creates professional notes exploiting a set of theories and a lexicon which converts a clinician's ideas into text. We explored the differences between the two systems to determine if they can be used in clinical practice. Every element in a document created by TGLT is triggered by the user, whereas GPT-4 created documents may include invented text. LGLT can generate complex clinical notes that are more complete, more orderly, and less error-prone than conventionally written notes. The lexicon constructed for TGLT can be updated or corrected rapidly by end-users, whereas GPT-4 uses a huge library that may take many months to update. TGLT notes are concise, complete, and organized in a defined order, whereas GPT-4 may be incomplete and poorly ordered. TGLT can alert the user to recent best practice advisories. TGLT issues computer codes for every text element in the document and does not include confidential identity information, enabling the facile aggregation of the content of notes for down-stream analysis. GPT-4 does not issue computer codes, and generates text that may include patient identifiers. TGLT costs vastly less than GPT-4. We conclude that TGLT has none of the manifold disadvantages that GPT-4 has for creating professional reports.