Using Large Language Models for Data Cleaning: An Evaluation of ChatGPT-4o's Performance.

Area of Science:

• Data Science
• Artificial Intelligence
• Health Informatics

Background:

• Manual data cleaning is essential for data quality but is time-consuming and prone to errors.
• Automated data cleaning approaches are needed to improve efficiency and accuracy.
• Large language models like ChatGPT offer potential for automating data cleaning tasks.

Purpose of the Study:

• To evaluate the performance of ChatGPT-4o in automating data cleaning.
• To assess the accuracy and consistency of ChatGPT-4o across different data variables.

Main Methods:

• Utilized ChatGPT-4o for automated data cleaning.
• Evaluated cleaning performance on gender, hemoglobin, route, and urine glucose variables.
• Conducted three trials to assess consistency and identify variations.

Main Results:

• ChatGPT-4o achieved high mean accuracies: 94.3% (gender), 92.5% (hemoglobin), 92.8% (route).
• Lower accuracy (70.0%) was observed for the urine glucose variable.
• Consistent accuracy was noted for gender, hemoglobin, and route across trials.
• Significant variation in accuracy was found for urine glucose across trials.

Conclusions:

• ChatGPT-4o demonstrates significant potential for automated data cleaning, particularly for structured variables.
• The performance for complex or variable data (e.g., urine glucose) requires further investigation and refinement.
• Future research should focus on understanding and mitigating the limitations of AI in data cleaning.