Enhancing data quality in medical concept normalization through large language models.

OBJECTIVE

Medical concept normalization (MCN) aims to map informal medical terms to formal medical concepts, a critical task in building machine learning systems for medical applications. However, most existing studies on MCN primarily focus on models and algorithms, often overlooking the vital role of data quality. This research evaluates MCN performance across varying data quality scenarios and investigates how to leverage these evaluation results to enhance data quality, ultimately improving MCN performance through the use of large language models (LLMs). The effectiveness of the proposed approach is demonstrated through a case study.

METHODS

We begin by conducting a data quality evaluation of a dataset used for MCN. Based on these findings, we employ ChatGPT-based zero-shot prompting for data augmentation. The quality of the generated data is then assessed across the dimensions of correctness and comprehensiveness. A series of experiments is performed to analyze the impact of data quality on MCN model performance. These results guide us in implementing LLM-based few-shot prompting to further enhance data quality and improve model performance.

RESULTS

Duplication of data items within a dataset can lead to inaccurate evaluation results. Data augmentation techniques such as zero-shot and few-shot learning with ChatGPT can introduce duplicated data items, particularly those in the mean region of a dataset's distribution. As such, data augmentation strategies must be carefully designed, incorporating context information and training data to avoid these issues. Additionally, we found that including augmented data in the testing set is necessary to fairly evaluate the effectiveness of data augmentation strategies.

CONCLUSION

While LLMs can generate high-quality data for MCN, the success of data augmentation depends heavily on the strategy employed. Our study found that few-shot learning, with prompts that incorporate appropriate context and a small, representative set of original data, is an effective approach. The methods developed in this research, including the data quality evaluation framework, LLM-based data augmentation strategies, and procedures for data quality enhancement, provide valuable insights for data augmentation and evaluation in similar deep learning applications.

AVAILABILITY

https://github.com/RichardLRC/mcn-data-quality-llm/tree/main/evaluation.