Efficient Iterative Data Generation Using Evaluation of Prioritized Input Parameters in ANNs for SAR Prediction in Human Head Models at 13.56 MHz

In this work, an efficient iterative dataset generation strategy is proposed, considering prediction accuracy and high impact input parameters as figures of merit to define a sufficient number of samples for reliable machine learning (ML) results. Parameter prioritization in combination with artificial neural networks (ANNs) is designed for examination of multiparameterized simulation setups in bioelectromagnetic (BEM), aiming to avoid redundant parameters and excessive sample sizes and to provide an alternative to expensive measurements, full-wave simulations, and the limitations of adaptive sampling methods in high-dimensional BEM problems. Specifically, the variation of mass-averaged specific absorption ratio (SAR) in each individual tissue in human head models is studied, considering up to <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\pm$</tex-math></inline-formula>95% uncertainty in a uniform distribution of electrical properties of tissues. Up to 3500 full-wave simulations for seven different scenarios are performed. Utilizing parameter prioritization in ANNs enables high accuracy results with fewer input parameters, allowing improved physical interpretability. By applying this method, the required number of numerical simulations (samples) for an optimal dataset is approximately 5 to 10 times the total number of input parameters. The results of this innovative method demonstrate that the reduced dataset successfully encapsulates the core aspects of the SAR problem under investigation, resulting in ML prediction accuracy surpassing 95% while reducing time and memory consumption by approximately 60%.