CogProg: Utilizing Large Language Models to Forecast In-the-moment Health Assessment.

Area of Science:

• Artificial Intelligence
• Health Informatics
• Computational Psychology

Background:

• Forecasting future health status aids in understanding health patterns and proactive support for cognitive and physical challenges.
• Generative large language models (LLMs) are emerging as capable tools for various forecasting tasks, including those involving unstructured data and explainable reasoning.

Purpose of the Study:

• To investigate the efficacy of large language models (LLMs) in accurately forecasting future self-reported health states.
• To evaluate the performance of LLMs against traditional numerical methods in health status prediction.

Main Methods:

• Utilized in-the-moment assessments of mental sharpness, fatigue, and stress from multiple studies.
• Constructed prompt/response pairs using daily responses (N=106) and text descriptions of activities (N=32) to predict subsequent self-reported health.
• Fine-tuned several LLMs and applied chain-of-thought prompting to assess forecasting accuracy and explainability.

Main Results:

• LLMs achieved the lowest overall mean absolute error (MAE) of 0.851.
• With additional text context, multimodal LLMs demonstrated the lowest MAE for mental sharpness (0.862), fatigue (1.000), and stress (0.414).
• Multimodal LLMs outperformed numeric baselines in RMSE for stress prediction (0.947), while traditional algorithms were superior for mental sharpness and fatigue.

Conclusions:

• LLMs, particularly when augmented with text-based contextual information, can be effective for enhancing health forecasting accuracy.
• This study provides valuable insights into the potential applications of LLMs for predictive health monitoring and personalized interventions.