Synthetic Skeleton Data Generation using Large Language Model for Nurse Activity Recognition

In this paper, we improve nurse activity recognition by employing a Large Language Model (LLM) to generate synthetic pose estimation data. Keypoint data extracted from recorded videos of a single nurse performing Endotracheal suctioning (ES) activities using You Only Look Once v7 (YOLOv7) is used as a database. We explore the issue of data imbalances that hinder the effectiveness of activity recognition algorithms. To counter this, we utilize LLMs to artificially augment the dataset by generating varied synthetic samples through prompting strategies with different content and context. A Random Forest (RF) classifier is trained on annotations of medical activities and corresponding keypoints. Additionally, we generate synthetic datasets in equal volumes using Random Sampling and Generative Adversarial Networks (GAN) to benchmark against our LLM-based approach. To evaluate, we compared the performance between baseline data and different augmentation approaches. The similarity between original and synthetic data is measured using the Kolmogorov-Smirnov (K-S) test. The proposed approach using LLM with prompts containing the explanation of the task with the description of the datasets to generate synthetic data has improved the overall ES classification performance. Our study illustrates the critical role of context and content in prompts for optimizing LLMs for synthetic data generation.