Teaching AI to Predict How Your Body Responds to Food

Identical meals can trigger very different metabolic responses, both between individuals and within the same person over time. To capture this variability, we have developed a proprietary algorithm that estimates blood glucose levels in real time from smartwatch data, without invasive CGM. Combined with our molecular food database, this allows us to link dietary intake to physiological responses at high resolution. Building on these technologies, we aim to develop prediction models that simulate individual responses to meals and support recommendations for the right meal, for the right person, at the right moment.

A Fellow who develops AI models to predict how a given meal affects an individual’s blood glucose over the following hours. The Fellow will work with existing and newly collected datasets and use time-series foundation models such as TimeGPT to build, compare and validate these prediction models.

• How accurately can the postprandial glucose curve be predicted from molecular food data, wearable signals and prior glucose dynamics?
• How well can models such as TimeGPT be adapted to this task and generalize across individuals and new meals?
• Which inputs matter most: meal composition, physiological state or glucose history?

The Fellow reviews the literature, prepares the datasets, defines input features and builds, compares and validates different modelling approaches. The work is carried out independently within a clearly defined project, supported by regular supervision, research group meetings and exchange with experts from nutrition, medicine and data science. By the end, the Fellow delivers an overview of relevant time-series foundation models and a validation showing which approach best fits the task.

The Fellow has access to our proprietary non-invasive glucose algorithm, our molecular food database, continuous glucose and wearable datasets, and expertise in nutritional medicine, digital health and computational modelling.