Multi-feature machine learning classification of sonotubometry for eustachian tube dysfunction assessment

Eustachian tube dysfunction (ETD) is a critical factor for otitis media with effusion (OME) in children, yet there is currently no gold standard for its direct clinical measurement. Sonotubometry serves as a physiological diagnosis method for ETD, assessing sound transmission from the nasopharynx to the ear canal during swallowing. However, its usability is limited by challenges in interpreting the recordings. In this work, a machine learning (ML) model is applied to analyze the audio features from sonotubometry for the detection and classification of ETD. Various audio feature extraction techniques were employed, with the Mel-frequency cepstral coefficients (MFCC) features yielding the best results. Specifically, when combined with the convolutional neural network (CNN) model, MFCC achieved a sensitivity of 0.975 (95 % CI: 0.906, 1.000), which significantly outperformed the traditional threshold-based method 0.645 (95 % CI: 0.293, 0.997). Through feature heatmaps generated via masking, it was found that classification of normal ET opening primarily relies on the acoustic response of 6 to 8 kHz. This work demonstrates the potential of ML-based sonotubometry to provide an objective, non-invasive, and efficient diagnostic tool for ETD.