Quantifying postprandial glucose responses using a hybrid modeling approach: Combining mechanistic and data-driven
Balázs Erdős1,2, Bart van Sloun1,2, Gijs H Goossens1,3
1TiFN, Wageningen, Netherlands.
Plos One
|July 27, 2023
Summary
Mechanistic models accurately predict glucose and insulin responses during oral glucose tolerance tests. Combining these with data-driven models did not improve predictions, showing the strength of mechanistic approaches for understanding glucose regulation.
Area of Science:
- Physiology
- Computational Biology
- Endocrinology
Background:
- Inter-individual variability in glucose regulation is complex.
- Mechanistic and data-driven models offer different insights into glucose homeostasis.
- A combined modeling approach has not been fully explored.
Purpose of the Study:
- To quantify glucose and insulin responses to oral glucose tolerance tests using a combined modeling approach.
- To compare the predictive performance of mechanistic, data-driven, and combined models.
- To assess the utility of bottom-up mechanistic models in dynamic glucose regulation.
Main Methods:
- A sequential combination of mechanistic and data-driven modeling was proposed.
- Cross-sectional data from 2968 individuals in the Maastricht Study were used.
- Predictive performance was measured by R2 and mean squared error of prediction.
Main Results:
- Personalized mechanistic models alone achieved the best predictive performance.
- The addition of a data-driven model did not enhance predictive accuracy.
- Mechanistic models consistently outperformed data-driven and combined approaches.
Conclusions:
- Bottom-up mechanistic models are strong and suitable for describing dynamic glucose and insulin responses.
- Personalized mechanistic models are highly effective for predicting glucose homeostasis.
- Further exploration of combined modeling may not be necessary for this specific application.
