Knowledge-Informed AI group IDSIA USI-SUPSI

Project

GUIDE

Machine learning and language models for personalised diabetes management

  • Healthcare
Activity and glucose data from an app are analysed by machine learning models and a large language model that produces a personalised recommendation.
Machine learning models analyse glucose and activity data, and a large language model turns the results into personalised, actionable advice.

Challenge

Managing type 1 diabetes requires constant attention to glucose levels, especially during physical activity. Language models open new possibilities for personalised recommendations based on activity data, but the advice they generate must be accurate, appropriate to the context and relevant for the individual.

Approach

We studied the feasibility of generating personalised and actionable advice from the physical activity data of people with type 1 diabetes. Machine learning models extract the relevant patterns, and a set of prompts guides open-weight language models to produce recommendations, which were evaluated for medical soundness and personalisation.

Results

The study showed that the task is feasible, that results depend strongly on prompt design and on the choice of language model, and that safety and generalisability require further work.

Publications