Healthcare Event Prediction via Iterative in-Context Learning of LLM

Large Language Models (LLMs) demonstrate the ability to perform a wide range of natural language tasks being crucial in healthcare event prediction applications, particularly from wearable device data. However, the current fine-tuning method has great limitations. Inadequate generalization performance leads to difficulty in migrating tasks, and the solution of using GPT for direct reasoning has privacy problems. This paper presents an Iterative In-Context Learning framework applied to 10 tasks across four wearable device datasets (PMData, LifeSnaps, GLOBEM, and AW_FB). The framework iteratively filters use cases and validates them using GPT-4, performing inference on the Llama3-8B model. It achieves SOTA performance on 6 of the tasks under the same scale model, with an average performance increase of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1 2. 5 4 \%}$</tex>. Ablation studies demonstrate the framework's efficacy. We also compared the fine-tuning model and the utilization of our method on training time and resources, and the results show that our method is more efficient.