Prediction of acoustic tinnitus suppression using resting-state EEG via explainable AI approach.

Tinnitus is defined as the perception of sound without an external source. Its perceptual suppression or on/off states remain poorly understood. This study investigates neural traits linked to brief acoustic tinnitus suppression (BATS) using naive resting-state EEG (closed eyes) from 102 individuals. A set of EEG features (band power, entropy, aperiodic slope and offset of the EEG spectrum, and connectivity) and standard classifiers were applied achieving consistent high accuracy across data splits: 98% for sensor and 86% for source models. The Random Forest model outperformed other classifiers by excelling in robustness and reduction of overfitting. It identified several key EEG features, most prominently alpha and gamma frequency band power. Gamma power was stronger in the left auditory network, while alpha power dominated the right hemisphere. Aperiodic features were normalized in individuals with BATS. Additionally, hyperconnected auditory-limbic networks in BATS suggest sensory gating may aid suppression. These findings demonstrate robust classification of BATS status, revealing distinct neural traits between tinnitus subpopulations. Our work emphasizes the role of neural mechanisms in predicting and managing tinnitus suppression. Moreover, it advances the understanding of effective feature selection, model choice, and validation strategies for analyzing clinical neurophysiological data in general.