Simulation-driven deep learning for the diagnosis of middle ear pathologies using wideband acoustic immittance

Wideband acoustic immittance (WAI) provides comprehensive frequency-dependent information for diagnosing middle ear pathologies. However, the scarcity of clinical data and complex response patterns significantly hinder automated diagnosis, particularly in data-limited scenarios. To address this issue, this study proposes a simulation-driven computer-aided diagnosis framework for WAI based on finite element (FE) modeling. Latin hypercube sampling was employed to systematically perturb key physiological parameters of the human ear FE models, generating a standardized virtual WAI dataset comprising 12 000 samples across the 0.2-6 kHz frequency range. The dataset includes four middle ear conditions: normal ear, ossicular chain discontinuity, ossicular chain fixation, and otitis media with effusion. Based on this dataset, a lightweight convolutional neural network tailored for multi-channel WAI inputs, termed WAIHybrid, was developed. It was benchmarked against traditional feature-based machine learning models and five representative deep learning architectures on simulated data and subsequently evaluated on an external clinical dataset comprising 206 ear-level WAI records. WAIHybrid achieved a macro-F1 of 96.30% and a balanced accuracy of 96.29% on an independent simulated test set. On the external clinical dataset, the corresponding values were 87.76% and 88.07%, respectively. Response-level comparisons, learned-representation analyses, and Integrated Gradients maps identified partial class-related correspondence between the simulated and clinical data, residual simulation-to-clinical discrepancy, and class-dependent channel-frequency attribution patterns. These findings support a simulation-driven proof of concept for automated WAI analysis in data-limited middle ear assessment. Further evaluation in larger, more balanced, and clinically heterogeneous cohorts is needed.