The Predictive Role of Specific Audiometric Frequencies in Speech-in-Noise Perception: A Machine Learning Approach

Abstract

Purpose: Speech comprehension performance in noisy environments is a complex process that cannot be fully predicted by standard audiometric assessments. The aim of this study is to use machine learning (ML) algorithms to predict individuals’ difficulties in understanding speech in noise based on clinical data.Methods: The study utilized retrospective clinical data obtained from the Oldenburg Hearing Health Record (OHHR) dataset. Machine learning algorithms were trained to classify speech-in-noise performance based on frequencydependent hearing thresholds, cognitive status, and demographic factors, as assessed by the Digit Triple Test (DTT). Results: The trained models demonstrated high predictive accuracy in the classification task, with the highest performance achieved by logistic regression (LR) (AUC=0.952), gradient boosting (GB (AUC=0.951), and random forest (RF) (AUC=0.951). The analyses indicate that, despite the model being provided with cognitive data such as Vocabulary Size Test (VST), the algorithms derived their highest predictive power from pure-tone thresholds in the 1500–4000 Hz range.Conclusion: The findings confirm, based on the data, that the primary factor making it difficult to perceive speech in noise originates from the peripheral auditory system. With the high classification accuracy achieved (AUC> 0.95), standard hearing tests are transformed into a predictive clinical tool, providing a reliable digital decision-support mechanism for individuals who struggle with speech comprehension in noise.