# Simulation-based inference for subject-specific tuning of middle ear finite-element models towards personalized objective diagnosis

## TL;DR

This paper introduces a new method using simulation-based inference to personalize middle ear models for better diagnosis and treatment.

## Contribution

The study applies neural posterior estimation to tune finite-element models with subject-specific data for personalized diagnosis.

## Key findings

- The method generates probability distributions of parameters instead of single values, capturing experimental variability.
- The approach enables individualized differential inference by comparing inferred parameters with normal ranges.
- SBI supports personalized diagnosis and treatment by uncovering parameter interactions and compensating for limited clinical data.

## Abstract

Computational models, particularly finite-element (FE) models, are essential for interpreting experimental data and predicting system behavior, especially when direct measurements are limited. Tuning these models is particularly challenging when a large number of parameters are involved. Traditional methods, such as sensitivity analyses, are time-consuming and often provide only a single set of parameter values, focusing on reproducing averaged trends rather than capturing experimental variability. New approaches are needed to make computational models more adaptable to patient-specific clinical applications. We applied simulation-based inference (SBI) using neural posterior estimation (NPE) to tune an FE model of the human middle ear against subject-specific data. The training dataset consisted of 10,000 FE simulations of stapes velocity, ear-canal input impedance, and absorbance, paired with seven FE parameter values sampled within plausible ranges. By using simulated data, we generated a diverse training dataset, enabling efficient learning by the neural network (NN). The NN learned the association between parameters and simulation outcomes, providing a probability distribution of parameter values, which could be used to produce subject-specific computational inferences. By accounting for noise and test–retest variability, the method provided a probability distribution of parameters, rather than a single set, fitting three experimental datasets simultaneously. Importantly, examining the inferred parameter distributions alongside prior knowledge of normal ranges enables individualized differential inference used for diagnosis. SBI offers an objective alternative to sensitivity analyses, uncovering parameter interactions, supporting personalized diagnosis and treatment, and compensating for limited clinical training data. This method is applicable to any computational model, enhancing its potential for improved patient outcomes.

The online version contains supplementary material available at 10.1038/s41598-025-22164-2.