From Simulations to Inference: Using Machine Learning to Tune Patient-Specific Finite-Element Models of the Middle...

Area of Science:

• Computational biology
• Biophysics
• Medical imaging and modeling

Background:

• Finite-element (FE) models are crucial for interpreting experimental data but are challenging to tune due to numerous parameters.
• Traditional sensitivity analyses are time-consuming, subjective, and often yield single parameter sets, failing to capture data variability.

Purpose of the Study:

• To apply simulation-based inference (SBI) with neural posterior estimation (NPE) for tuning a human middle ear FE model.
• To develop an objective and efficient method for parameter estimation in complex biological models.

Main Methods:

• Trained a neural network on 10,000 FE simulations of middle ear function (stapes velocity, EC input impedance, absorbance) with randomly sampled parameters.
• Utilized neural posterior estimation (NPE) to learn the relationship between model parameters and simulation outcomes.

Main Results:

• Successfully identified parameter sets that reproduced multiple experimental datasets simultaneously.
• Generated probability distributions of parameters, representing all valid combinations that fit the data, accounting for noise and variability.
• Demonstrated the neural network's robustness to noise and efficient learning with a large dataset.

Conclusions:

• SBI provides an objective, efficient alternative to traditional sensitivity analyses for tuning biological FE models.
• The method uncovers parameter interactions and provides probability distributions, crucial for understanding model behavior.
• This approach shows promise for objective differential diagnosis of conductive hearing loss by revealing middle ear mechanical properties.