The research introduces a novel broad language model aimed at understanding and reasoning health-related questions and personal data. To systematically evaluate the model, three reference datasets were selected to test domain knowledge, alignment with patient-reported outcomes, and the ability to produce quality human recommendations.
Mobile and wearable devices can provide continuous, granular, and longitudinal data on an individual’s physiological state and behaviors. Examples include step counts, raw sensor measurements like heart rate variability, sleep duration, and more. This data can be used by individuals to monitor personal health and motivate healthy behavior. AI generative models can further provide personalized information and recommendations to help individuals achieve their health goals.
The study demonstrates building upon Gemini models’ advanced capabilities in multimodality and long-context reasoning, leading to cutting-edge performance in a diverse range of medical tasks. By leveraging next-generation Gemini model capabilities, the research highlights two complementary approaches to providing accurate personal health insights with LLMs.
The first article showcases the Personal Health Large Language Model (PH-LLM), an enhanced version of Gemini designed to generate insights and recommendations for improving personal health behaviors related to sleep and fitness patterns. The model’s performance was evaluated across tasks of personalized health recommendations, expert domain knowledge, and predicted self-reported sleep quality assessments.
The evaluation revealed that PH-LLM’s performance was not statistically different from expert performance in fitness tasks, although expert-written recommendations scored higher for sleep. Adjusting PH-LLM significantly improved its ability to use relevant domain knowledge and personalize information for generating insights and predicting potential causal factors accurately.
Furthermore, PH-LLM demonstrated superior performance compared to human experts in physical fitness and sleep knowledge domain assessments, surpassing benchmarks for receiving continuing education credits for professional licenses in those domains.
The study highlights the importance of fine-tuning PH-LLM to contextualize physiological data for personal health applications. Additionally, a framework for a personalized health knowledge agent based on Gemini Ultra 1.0 was presented to extend model capabilities for analyzing complex temporal wearable sensor data and offering personalized insights and recommendations for health queries. The agent’s iterative reasoning capabilities significantly improved performance across various evaluation metrics compared to baseline models in providing health information.
The agent’s focus on sleep and physical fitness data suggests potential to expand its framework to analyze a wider range of health information, including medical records, nutritional data, and user-provided journal entries. As LLMs advance, agents hold the promise of delivering even more sophisticated insights and effective guidance for personal health management.
In conclusion, the research aims to investigate features and capabilities to support individuals in leading longer and healthier lives. Sleep and physical exercise play crucial roles in population health and are predictors of premature mortality globally. The capabilities demonstrated in the research represent a significant step towards AI models that provide personalized insights and recommendations, empowering individuals to draw accurate and actionable conclusions from their health data. Further testing is needed to understand which capabilities are most beneficial for users.
The collaboration for this research involved Google Research, Google Health, Google DeepMind, and associated teams.
Article Source
http://research.google/blog/advancing-personal-health-and-wellness-insights-with-ai/
