Abstract
Some individuals complain of listening-in-noise difficulty despite having a normal audiogram. In this study, machine learning is applied to examine the extent to which hearing thresholds can predict speech-in-noise recognition among normal-hearing individuals. The specific goals were to (1) compare the performance of one standard (GAM, generalized additive model) and four machine learning models (ANN, artificial neural network; DNN, deep neural network; RF, random forest; XGBoost; eXtreme gradient boosting), and (2) examine the relative contribution of individual audiometric frequencies and demographic variables in predicting speech-in-noise recognition. Archival data included thresholds (0.25-16 kHz) and speech recognition thresholds (SRTs) from listeners with clinically normal audiograms (n = 764 participants or 1528 ears; age, 4-38 years old). Among the machine learning models, XGBoost performed significantly better than other methods (mean absolute error; MAE = 1.62 dB). ANN and RF yielded similar performances (MAE = 1.68 and 1.67 dB, respectively), whereas, surprisingly, DNN showed relatively poorer performance (MAE = 1.94 dB). The MAE for GAM was 1.61 dB. SHapley Additive exPlanations revealed that age, thresholds at 16 kHz, 12.5 kHz, etc., on the order of importance, contributed to SRT. These results suggest the importance of hearing in the extended high frequencies for predicting speech-in-noise recognition in listeners with normal audiograms.