Non-intrusive Reduced Order Modeling of Patient-specific Cochlear Implantations.

Abstract

Objective: Cochlear implants successfully treat severe to profound hearing loss patients. Patient-specific numerical simulations can yield important insights that could guide surgical planning and the interpretation of post-operative measurements. However, these simulations have a high computational effort.

Methods: A non-intrusive reduced-order model has been used to replace the patient-specific model generation and simulation of different electrical stimulation sources, reducing the computational time and enabling fast response simulations. The reduced-order model combines proper orthogonal decomposition with radial basis function interpolation. The dataset used to build the reduced order model consists of 528 different solutions, also referred to as snapshots, from 24 cochlear models, with each cochlea subjected to 22 simulations with varying electrical stimuli. Each simulation is characterized by five parameters, three specifying the cochlea geometry and two specifying the electrode array position and the active electrode.

Results: A leave-one-out strategy was used to verify the accuracy of the reduced-order model. The presented approach reduces the time for the patient-specific model generation and simulation from nearly 1.5 hours to less than a second while providing a high accuracy of the solutions with a relative error of 2.5% compared to the finite element solution.

Conclusion: The presented non-intrusive reduced order model can predict the 3D intracochlear voltage distribution for new patients and implant positions.

Significance: This work demonstrates the feasibility of fast patient-specific simulations. These numerical investigations could support the fitting of cochlear implants, the design of individualized sound coding strategies and surgery-dependent decision-making.