Prediction of aided speech recognition using Random Forest Regression

Hearing aid (HA) users show largely unexplained variability in aided speech recognition. Until now, there is no clear recommendation for evaluating HA outcomes, particularly regarding the achievable speech recognition with HA. This study aimed to identify the most influential factors affecting the word recognition score with HA at 65 dB sound pressure level (SPL), referred to hereinafter as WRS65(HA). Retrospective data from clinical routine measurements of 635 HA users were analysed, including 18 demographic, audiological, and HA-related features. A Random Forest Regression (RFR) was applied to predict WRS65(HA); and an iterative feature-selection process was used to determine the feature combination with the lowest mean absolute error (MAE). Audiological features such as maximum word recognition score (WRSmax), pure-tone average (PTA), and unaided word recognition score at 65 dB SPL (WRS65) showed the highest individual predictive accuracy. Demographic features such as age and sex performed considerably worse. The lowest statistically significant MAE (9.8 percentage points, pp) was achieved with a three-feature combination: WRSmax, WRS65 and the fit-to-target accuracy in real-ear measurements for medium frequencies at 65 dB SPL input level. The inclusion of additional features appears to yield limited benefit and may increase the risk of overfitting. The simple prediction model by Hoppe et al. (2014) based on the PTA achieved an MAE of 14.4 pp and was outperformed by 4.6 pp when using the best three-feature combination. These findings highlight that PTA alone is insufficient for accurately predicting WRS65(HA). Combining speech audiometric data in combination with HA-specific parameters provides substantially better results.