Back

LLMs for analyzing open text in global health surveys: why children are not accessing vaccine services in the DRC

Burstein, R. L.; Mufata, E.; Proctor, J. L.

2024-11-15 public and global health
10.1101/2024.11.14.24317253 medRxiv
Show abstract

This study evaluates the use of large language models (LLMs) to analyze free-text responses from large-scale global health surveys, using data from the Enquete de Couverture Vaccinale (ECV) household coverage surveys from 2020, 2021, 2022, and 2023 as a case study. We tested several LLM approaches varying from zero-shot and few-shot prompting, fine-tuning, and a natural language processing approach using semantic embeddings to analyze responses on reasons caregivers did not vaccinate their children. Performance ranged from 61.5% to 96% based on testing against a curated benchmarking dataset drawn from the ECV surveys, with accuracy improving when LLM models were fine-tuned or provided examples for few-shot learning. We show that even with as few as 20-100 examples, LLMs can achieve high accuracy in categorizing free-text responses. This approach offers significant opportunities for reanalyzing existing datasets and designing surveys with more open-ended questions, providing a scalable, cost-effective solution for global health organizations. Despite challenges with closed-source models and computational costs, the study underscores LLMs potential to enhance data analysis and inform global health policy.

Matching journals

The top 6 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.