목적: Tinnitus is a common auditory disorder frequently linked to hearing impairment. Identifying how tinnitus characteristics correlate with hearing thresholds across different frequencies is essential for improving diagnosis and treatment. This study explores the ability of tinnitus-related factors to predict hearing loss at specific frequencies using linear and non-linear machine learning techniques. 방법:Data from 744 tinnitus patients were analyzed, including demographic details (age, sex), tinnitus attributes (nature, type, duration), diagnosis, and audiometric results (PTA values). Linear regression evaluated the association between predictors and hearing thresholds at higher frequencies (PTA-2000, 3000, 4000, 6000, 8000 Hz). Additionally, Random Forest and Neural Networks were applied to detect non-linear relationships. Hyperparameter optimization was performed to enhance model accuracy. 결과:Linear regression yielded strong predictive accuracy at higher frequencies, with R² values of 0.70 for PTA-2000, 0.84 for PTA-3000, 0.76 for PTA-4000, 0.62 for PTA-6000, and 0.45 for PTA-8000, suggesting a largely linear association between tinnitus characteristics and hearing loss at these frequencies. Non-linear models, including Random Forest and Neural Networks, showed inferior predictive performance, with R² values below 0.05 across all frequencies. Adjustments to model parameters and feature selection did not improve performance, indicating minimal non-linear effects. While tinnitus type did not significantly influence average hearing thresholds, it showed a meaningful correlation with boundary frequency values, where hearing levels dropped below 30 dB. Moreover, tinnitus nature (e.g., unilateral vs. bilateral) was significantly related to PTA values, reinforcing the role of hearing thresholds in tinnitus. 결론:Linear models demonstrated higher efficacy in predicting hearing loss at higher frequencies compared to non-linear approaches. The findings suggest that the relationship between tinnitus characteristics and hearing thresholds in the analyzed cohort is largely linear, which seems to have a predictive value in diagnosing sensorineural tinnitus.