Machine learning identification of tinnitus-related features in auditory peripheral spontaneous activity in a guinea pig noise-induced tinnitus model.

OBJECTIVES

Tinnitus affects millions globally, yet its clinical assessment relies on subjective reports, limiting diagnostic accuracy and treatment development. This study aimed to identify objective, tinnitus-related features within ensemble spontaneous activity (ESA) recorded from the cochlear surface in a guinea pig model and to evaluate their reversibility using extracochlear electrical stimulation (EES) and machine learning.

DESIGN

ESA was recorded from four groups: normal controls, noise-exposed animals with tinnitus, noise-exposed animals without tinnitus, and tinnitus animals after EES. Spectral features-central frequency, bandwidth, skewness, and kurtosis-were extracted using Fast-Fourier Transform and sliding window analysis. Behavioral tinnitus was assessed using the gap-prepulse inhibition of the acoustic startle reflex (GPIAS). Five machine learning models were trained to classify tinnitus status based on EES-reversible spectral features, with SHAP analysis used to identify key predictors.

RESULTS

Tinnitus-related spectral alterations were observed in frequency bands associated with the exposure noise and its harmonic/subharmonic ranges. These changes were reversed by EES, coinciding with behavioral improvement in GPIAS. The support vector machine achieved the highest classification performance (AUC = 0.962). SHAP analysis identified central frequency (1400-1600 Hz) and skewness (8000-9000 Hz; 16,000-17,000 Hz) as the most informative features.

CONCLUSIONS

These findings suggest that specific ESA spectral features serve as objective and reversible biomarkers of tinnitus in a guinea pig model, offering potential for translation to clinical diagnostics and therapeutic monitoring.