Surrogate models based on machine learning methods for parameter estimation of left ventricular myocardium.

Area of Science:

• Biomechanics
• Computational Biology
• Machine Learning

Background:

• Estimating material properties from clinical data is a challenge in biomechanics.
• Fast methods are needed for analyzing left ventricular (LV) myocardium.
• Machine learning (ML) offers potential for efficient parameter estimation.

Purpose of the Study:

• To develop and evaluate ML-based surrogate models for rapid estimation of LV myocardial material properties.
• To compare the performance of K-nearest neighbour (KNN), XGBoost, and multi-layer perceptron (MLP) models.
• To assess the accuracy and uncertainty of parameter inference using these ML models.

Main Methods:

• A finite-element simulator of LV diastolic filling was used to generate training data.
• Training data were projected into a low-dimensional parametric space.
• Three ML models (KNN, XGBoost, MLP) were trained to emulate pressure-volume and pressure-strain relationships.

Main Results:

• All three ML models accurately learned the pressure-volume and pressure-strain relationships.
• Parameter inference was achieved within minutes using the surrogate models.
• XGBoost and MLP models demonstrated significantly lower parameter uncertainties compared to the KNN model.
• XGBoost showed superior performance in predicting LV diastolic dynamics and estimating passive parameters.

Conclusions:

• ML-based surrogate models provide a fast and effective method for estimating myocardial material properties.
• XGBoost is a promising tool for accurate prediction of LV diastolic function and parameter estimation.
• Further research into XGBoost for cardiac function emulation in multi-physics frameworks is warranted.