Cell and wearable units can present steady, granular, and longitudinal knowledge on a person’s physiological state and behaviors. Examples embody step counts, uncooked sensor measurements akin to coronary heart charge variability, sleep period, and extra. People can use these knowledge for private well being monitoring in addition to to encourage wholesome habits. This represents an thrilling space wherein generative AI fashions can be utilized to offer further customized insights and proposals to a person to assist them attain their well being objectives. To take action, nonetheless, fashions should be capable of motive about private well being knowledge comprising complicated time collection and sporadic info (like exercise logs), contextualize these knowledge utilizing related private well being area information, and produce customized interpretations and proposals grounded in a person’s well being context.
Contemplate a standard well being question, “How can I get higher sleep?” Although a seemingly easy query, arriving at a response that’s personalized to the person entails performing a collection of complicated analytical steps, akin to: checking knowledge availability, calculating common sleep period, figuring out sleep sample anomalies over a time frame, contextualizing these findings throughout the particular person’s broader well being, integrating information of inhabitants norms of sleep, and providing tailor-made sleep enchancment suggestions. Just lately, we confirmed how constructing on Gemini fashions’ superior capabilities in multimodality and long-context reasoning might allow state-of-the-art efficiency on a various set of medical duties. Nonetheless, such duties not often make use of complicated knowledge sourced from cell and wearable units related for private well being monitoring.
Constructing on the next-generation capabilities of Gemini fashions, we current analysis that highlights two complementary approaches to offering correct private well being and wellness info with LLMs. The primary paper, “In the direction of a Private Well being Massive Language Mannequin”, demonstrates that LLMs fine-tuned on skilled evaluation and self-reported outcomes are in a position to efficiently contextualize physiological knowledge for private well being duties. The second paper, “Remodeling Wearable Knowledge into Private Well being Insights Utilizing Massive Language Mannequin Brokers”, emphasizes the worth of code technology and agent-based workflows to precisely analyze behavioral well being knowledge by pure language queries. We imagine that bringing these concepts collectively, to allow interactive computation and grounded reasoning over private well being knowledge, can be vital parts for creating really customized well being assistants. With these two papers, we curate new benchmark datasets throughout a variety of non-public well being duties, which assist consider the effectiveness of those fashions.