Data-driven Uncertainty Quantification in Computational Human Head Models.

Area of Science:

• Biomechanics
• Computational modeling
• Uncertainty quantification

Background:

• Computational head models are vital for predicting traumatic brain injury (TBI).
• Significant uncertainties in model parameters (geometry, material properties, boundary conditions) challenge prediction reliability.
• High computational costs of biofidelic head models limit traditional uncertainty quantification (UQ) methods.

Purpose of the Study:

• To develop and validate a novel two-stage, data-driven manifold learning framework for UQ of computational head models.
• To address the computational expense of UQ in high-dimensional biofidelic simulations.
• To quantify uncertainty in simulated brain strain fields due to variability in material properties.

Main Methods:

• A two-stage framework combining manifold learning with surrogate modeling.
• Stage 1: Data-driven generation of input random vector realizations using kernel-density estimation and diffusion maps.
• Stage 2: Training surrogate models with nonlinear dimensionality reduction (Grassmannian diffusion maps, Gaussian process regression, geometric harmonics) for efficient input-output mapping.

Main Results:

• The proposed framework significantly reduces computational cost while maintaining high accuracy in approximating computational head model responses.
• Monte Carlo simulations using surrogate models effectively propagate uncertainty.
• UQ analysis revealed significant spatial variations in model uncertainty and differences among brain injury predictor variables.

Conclusions:

• The data-driven manifold learning framework offers an efficient and accurate approach for UQ in complex computational head models.
• This method enhances the reliability of TBI prediction by characterizing uncertainty in simulated brain responses.
• The findings highlight the importance of considering spatial variations in uncertainty for accurate TBI risk assessment.