A reduced model for the long-term effects of physical activity on type 2 diabetes

Type 2 diabetes progresses slowly and may be reversed through lifestyle changes, but quantifying the long-term impact of regular physical activity remains challenging due to sparse longitudinal data. Mechanistic models offer a powerful tool by simulating metabolic processes over extended timescales. However, multi-scale formulations that capture both the short-term effects of exercise sessions and the slow evolution of disease tend to be computationally demanding, limiting their practical use in personalized decision support. To address this limitation, we derived a reduced version of a two-scale model that captures the short- and long-term effects of physical activity on blood glucose regulation. By analytically averaging the short-term effects induced by exercise, we developed a homogenized formulation that transmits the average contribution of physical activity to the slower glucose-insulin dynamics. This reduction preserves the key model dynamics while decreasing computational complexity by almost a factor 2000. We prove that the approximation error remains bounded and confirm the model’s accuracy through a parameter-based simulation study. The resulting model provides a mathematically grounded reduction that retains key physiological mechanisms while enabling fast long-term simulations. This substantial computational gain makes it suitable for integration into medical decision support systems, where it can be used to design and evaluate personalized physical activity plans aimed at reducing the risk of type 2 diabetes.

💡 Research Summary

**
The paper addresses a critical gap in diabetes research: quantifying the long‑term impact of regular physical activity on type‑2 diabetes progression when longitudinal data are scarce. To this end, the authors start from a previously published mechanistic model that couples a fast, minute‑scale subsystem describing the acute physiological response to an exercise bout with a slow, year‑scale subsystem governing glucose‑insulin dynamics, β‑cell function, insulin sensitivity, and inflammatory signaling. The full model comprises twelve ordinary differential equations (ODEs): five short‑term state variables (oxygen consumption V̇O₂, hepatic glucose production G_pr, peripheral glucose uptake G_up, exercise‑induced insulin removal I_e, and interleukin‑6 IL‑6) driven by a periodic Heaviside control u(t) that represents repeated exercise sessions of duration δ and period ν; and seven long‑term variables (long‑term activity effect V_L, insulin sensitivity SI, two auxiliary variables Σ and Γ, β‑cell mass B, plasma insulin I, and plasma glucose G). The short‑term subsystem feeds forward into the long‑term subsystem, while there is no feedback from the slow variables to the fast ones.

The authors first prove existence and uniqueness of solutions for the coupled system using the Picard‑Lindelöf theorem, handling the discontinuities of u(t) by restarting the ODE integration at each switching point. They then apply a rigorous averaging (homogenization) technique to replace the rapidly oscillating short‑term dynamics with their mean over one exercise period. By solving the short‑term ODEs analytically for the on‑phase (u=1) and off‑phase (u=0), they derive closed‑form expressions for each fast variable and compute the period‑averaged vector μ = (μ_VO₂, μ_Gpr, μ_Gup, μ_Ie, μ_IL6). The averaging yields a constant vector that depends only on the model parameters (θ, α₂, α₄, α₆, κ_IL6, λ_t, δ, ν) and can be pre‑computed.

The reduced model is then constructed by fixing the fast subsystem at its average value: the differential equations for the five fast variables are replaced by dy₁/dt = 0 with initial condition y₁(0)=μ, while the long‑term equations retain their original functional forms but receive μ as a constant input. This yields a 7‑dimensional ODE system that no longer depends on the high‑frequency control u(t). The authors prove that the approximation error ‖y(t)−ŷ(t)‖ remains bounded, exploiting the fact that the decay rates α₂, α₄, α₆, κ_IL6 are much smaller than the primary activation rate θ, which guarantees that the fast variables stay within