Authors: Miguel Temboury¹; Gerard Encina-Llamas²; Torsten Dau¹
¹Technical University of Denmark (DTU); ²Universitat de Vic – Central University of Catalonia
Background: Auditory evoked potentials (AEP) are widely used clinically but offer limited insight into the specific cochlear mechanisms underlying hearing loss. Because outer hair cells (OHC), inner hair cells (IHC), and auditory nerve fibers (ANF) are functionally coupled, it is difficult to disentangle their individual contributions from recorded responses. Computational forward models can link cochlear pathologies to predicted AEPs. Inverting this mapping using machine learning (ML) could enable more precise and individualized cochlear diagnostics.
Method: A physiologically detailed auditory nerve model was used to simulate AEPs across diverse cochlear damage profiles derived from standard audiograms and human histopathological data. Each profile was encoded as an 18-dimensional vector representing OHC, IHC, and ANF integrity across six cochlear segments. A total of 55,000 profiles, with responses to acoustic stimuli at multiple intensities, were used to train and test a convolutional neural network. The network predicted continuous cell-status values (0-1) for each cochlear cell population and segment.
Results: The network accurately decoded OHC, IHC, and ANF damage from simulated AEPs in ears with previously unseen audiograms. It could also distinguish different ratios of IHC-to-OHC damage, as well as distinct ANF profiles, even across ears with identical audiograms. Ongoing work identifies a minimal stimulus subset that preserves decoding performance while reducing test time.
Conclusion: We present a physics-informed ML framework for mapping AEPs onto cochlear pathology. Physiologically validated simulations provide labeled training data unavailable from clinical recordings, enabling ML models to identify pathology-specific response signatures. These results demonstrate the feasibility of distinguishing OHC, IHC, and ANF dysfunction non-invasively and represent a step toward precision audiology and pathology-targeted rehabilitation.
