ChatGPT improves usability, effectiveness, scalability, interpretability and accessibility, in early diagnosis of metabolic dysfunction-associated fatty liver disease.

BACKGROUND

Metabolic dysfunction-associated fatty liver disease (MAFLD) is one of the most common chronic liver diseases. We aimed to conduct a pilot study to evaluate the potential utility of ChatGPT in early diagnosis of MAFLD.

METHODS

We retrospectively analyzed data from participants undergoing annual health examinations at the First Affiliated Hospital, Zhejiang University School of Medicine in 2022. Questionnaires, laboratory tests, physical examinations, and liver ultrasonography were conducted. We built MAFLD-GPT by leveraging the GPT-3.5-turbo model provided by OpenAI. Zero-shot and few-shot learning approaches were used to identify patterns within the MAFLD data samples to construct an MAFLD model. We compared MAFLD-GPT with Machine Learning Technique (MAFLD-ML).

RESULTS

In total, 7,571 participants were included in the study, with 123 cases containing non-numeric data. MAFLD-GPT generates predictions using prompt engineering with zero- or few-shot training samples, whereas machine learning methods require massive training data. MAFLD-GPT demonstrated superior performance in recall and F1-score compared to traditional ML models, highlighting its potential for MAFLD diagnosis. The initial in-context examples (3-shot) brought the largest performance gain (F1 score = 0.724) in MAFLD-GPT, which outperformed SVM (the best performance achieved by traditional ML algorithms) by 6.1%. In addition, the evaluation results of the MAFLD-GPT model on 123 feature-missing samples demonstrated that MAFLD-GPT achieved promising scalability on non-numeric features. Importantly, MAFLD-GPT has the potential to be accessible to the public and to present its results in a clear and understandable manner.

CONCLUSIONS

MAFLD-GPT surpasses the MAFLD-ML model in terms of usability, effectiveness, scalability, interpretability and accessibility, demonstrating its tremendous potential for early diagnosis of MAFLD.