A FEM-ANN framework to estimate the on-diagonal elements of the impedance matrix in a Cochlear Implant

Abstract

Accurate estimation of the impedance matrix is essential for optimizing cochlear implant (CI) performance, yet the on-diagonal terms, that represent the contact impedances of electrodes, remain poorly characterized in existing models. In this work, we first analyze these on-diagonal terms and highlight their impact on electric field distribution. We then revisit the classic linear extrapolation approach and introduce two novel extrapolation methods to enhance prediction accuracy. To capture patient-specific variability, real impedance measurements are incorporated into a resistive–conductive finite-element method (FEM) model, whose matrices serve as the basis for a supervised neural network. The network is trained and validated on a diverse dataset of FEM-derived impedance matrices, enabling robust generalization across electrode configurations. Benchmarking against state-of-the-art techniques shows that our hybrid FEM-ANN framework reduces prediction error for diagonal terms. Moreover, when used in multipolar stimulation strategies, the ANN-based impedance matrices yield comparable focalization while requiring lower electrical power. Our results demonstrate that combining physical modeling with data-driven methods produces more reliable and efficient impedance estimates, paving the way for improved CI fitting and patient outcomes.