Sloppiness Consistency in Biomechanical Models and Its Inspired Dual-Space Model Optimization

Abstract

Advanced medical solutions rely on dependable biomechanical modeling. An enduring challenge in the constitutive modeling of soft tissue is delicately balancing model complexity, goodness-of-fit, and parameter identifiability, all of which impact the reliability of material behavior predictions under mechanical loading. It is established that biomechanical constitutive models, whether physically motivated or neural network derived, are typically sloppy from the information theory perspective. By analyzing the sensitivity matrix associated with posterior distributions of the constitutive parameters, a consistent pattern revealing the regularity in parameter combinations across experimental protocols characterizing tissue mechanical behavior and prior beliefs with varying levels of informativeness is discovered. The discovered pattern inspires to construct a sloppiness-based parameter hyperspace and proposes a model reduction program that performs model optimization by exploring four sub-hyperspaces. The proposed program offers a guide for effectively simplifying models while tightly ensuring parameter identifiability and prediction accuracy. Clear improvements are showcased to the brain tissue constitutive models discovered by neural networks and a physically motivated constitutive model of the human patellar tendon.