Knowledge-Informed AI group IDSIA USI-SUPSI

Main project

PRAESIIDIUM

Physics-informed machine learning for the prediction and prevention of prediabetes

Diagram showing differential equations and patient data combined into a patient digital twin, analysed with physics-informed neural networks and structural causal models.
Differential equation models and patient data are combined into a patient digital twin. Physics-informed neural networks estimate unobserved dynamics, and structural causal models support counterfactual analysis.

Challenge

Prediabetes affects hundreds of millions of adults worldwide and can be reversed through timely changes in lifestyle. Mathematical models based on differential equations describe metabolic and inflammatory processes in detail, but their computational cost makes them impractical for real-time, patient-specific risk prediction.

Approach

PRAESIIDIUM combines physics-informed machine learning with multi-scale, multi-organ mathematical models to build a digital twin of each patient, using clinical records and data from wearable sensors. Our group develops the machine learning components:

  • physics-informed neural networks that estimate patient-specific risk and unobserved quantities, such as insulin dynamics, from glucose data;
  • computationally efficient surrogates of complex models of prediabetes progression;
  • causal models learned from electronic medical records and refined with expert knowledge, used to simulate the effect of lifestyle interventions;
  • counterfactual analysis that couples differential equations with structural causal models, to suggest physical activity plans expected to reduce the risk of type 2 diabetes for an individual patient.

Outcome

The methods are designed to be integrated in a platform for healthcare professionals and patients, linked to common wearable sensors that monitor physical activity.

Causal graph linking lifestyle, BMI, cholesterol, triglycerides, fasting glucose and blood pressure to type 2 diabetes, in the factual and in the counterfactual world.
Causal model of type 2 diabetes risk factors learned from electronic medical records and expert knowledge, used to simulate a lifestyle intervention (factual and counterfactual world).
Differential equations, a causal graph and simulated glucose trajectories under different physical activity plans.
Physics-informed counterfactual reasoning: simulated glucose trajectories for different physical activity plans. Green, no intervention needed; orange, reversion obtained with a personalised plan; red, progression that physical activity alone cannot reverse.

Publications

Software

  • Physics-informed neural networks for hidden insulin dynamics

    Physics-informed neural networks that estimate insulin dynamics, which are not measured directly, from glucose data.

    GitLabPaper

  • Reduced model of type 2 diabetes progression

    Simulation code for a two-scale model of the long-term effects of physical activity on type 2 diabetes and for its reduced (homogenised) version, which allows fast long-term predictions.

    GitLabPaper

  • Surrogate model of prediabetes progression

    Deep learning surrogate of a mechanistic model of prediabetes progression, for fast patient-specific simulations. Data and trained model are not included.

    GitLabPaper

  • Counterfactual inference with differential equation models

    Counterfactual inference for ordinary differential equation models, applied to a model of glucose regulation to assess personalised physical activity plans for the prevention of type 2 diabetes.

    GitLabPaper

  • Causal discovery for type 2 diabetes risk

    Bootstrap-based causal discovery of a global causal model of type 2 diabetes risk factors from primary care data.

    GitLabPaper