[Paper Review] Differentiate ChatGPT-generated and Human-written Medical Texts

Wenxiong Liao, Zhengliang Liu+9 more|arXiv (Cornell University)|Apr 23, 2023

Artificial Intelligence in Healthcare and Education24 citations

TL;DR

The paper builds datasets of human- and ChatGPT-generated medical texts, analyzes linguistic differences, and shows a BERT-based detector achieving F1 > 95% in distinguishing ChatGPT-generated medical text from human-written text.

ABSTRACT

Background: Large language models such as ChatGPT are capable of generating grammatically perfect and human-like text content, and a large number of ChatGPT-generated texts have appeared on the Internet. However, medical texts such as clinical notes and diagnoses require rigorous validation, and erroneous medical content generated by ChatGPT could potentially lead to disinformation that poses significant harm to healthcare and the general public. Objective: This research is among the first studies on responsible and ethical AIGC (Artificial Intelligence Generated Content) in medicine. We focus on analyzing the differences between medical texts written by human experts and generated by ChatGPT, and designing machine learning workflows to effectively detect and differentiate medical texts generated by ChatGPT. Methods: We first construct a suite of datasets containing medical texts written by human experts and generated by ChatGPT. In the next step, we analyze the linguistic features of these two types of content and uncover differences in vocabulary, part-of-speech, dependency, sentiment, perplexity, etc. Finally, we design and implement machine learning methods to detect medical text generated by ChatGPT. Results: Medical texts written by humans are more concrete, more diverse, and typically contain more useful information, while medical texts generated by ChatGPT pay more attention to fluency and logic, and usually express general terminologies rather than effective information specific to the context of the problem. A BERT-based model can effectively detect medical texts generated by ChatGPT, and the F1 exceeds 95%.

Motivation & Objective

• Motivate trustworthy use of AI in medicine by understanding differences between human and ChatGPT-generated medical texts.
• Construct dedicated datasets of medical abstracts and radiology reports consisting of human- and ChatGPT-generated content.
• Characterize linguistic features (vocabulary, POS, dependencies, sentiment, perplexity) that distinguish the two text types.
• Develop and evaluate machine learning detectors to reliably identify ChatGPT-generated medical texts.

Proposed method

• Create two datasets: medical abstracts from Kaggle and radiology reports; generate ChatGPT counterparts via text continuation with demonstrations (in-context learning).
• Perform linguistic analyses: vocabulary/stem counts, POS tagging (NLTK), dependency parsing (Stanford CoreNLP), sentiment (Cardiff NLP model), and perplexity (BioGPT).
• Quantify differences in fluency, specificity, and statistical properties between human and ChatGPT texts.
• Develop detectors using: perplexity threshold (Perplexity-CLS), CART (TF-IDF features), XGBoost (TF-IDF features), and fine-tuned BERT (bert-base-cased).
• Evaluate detectors on train/validation/test splits (7:2:1) and report precision, recall, and F1 scores.

Experimental results

Research questions

• RQ1 What linguistic differences exist between medical texts written by humans and those generated by ChatGPT?
• RQ2 Can machine learning models accurately detect ChatGPT-generated medical texts, and which models perform best?
• RQ3 Are the two datasets (medical abstracts and radiology reports) consistently distinguishable using the proposed features and models?

Key findings

• Humans produce more diverse and concrete medical content; ChatGPT produces text that is fluent, logical, and more general.
• ChatGPT-generated medical texts tend to be more neutral/positive in sentiment and have lower perplexity, reflecting pattern replication from training data.
• POS and dependency analyses show ChatGPT uses more determiners, conjunctions, and coordination, with shorter dependency distances.
• A BERT-based detector achieves the best performance, with F1 > 0.95 across both datasets; Perplexity-based methods are weaker.
• XGBoost and CART provide competitive performance and offer interpretable feature insights; BERT provides the strongest overall discrimination.