MACHINE LEARNING-BASED MULTICLASS CLASSIFICATION OF HEARING-LOSS SEVERITY USING AUDIOMETRIC AND CLINICAL FEATURES

Abstract

Hearing loss is a heterogeneous condition that varies in severity, laterality, frequency configuration, and clinical cause. Pure-tone audiometry remains central to hearing assessment, but interpretation can require specialist knowledge and may be difficult to scale when large numbers of records must be screened. This study develops and evaluates an artificial-intelligence-oriented classification framework that combines audiometric thresholds with selected clinical variables to classify hearing-loss severity. A synthetic dataset of 600 records was constructed for methodological demonstration, with five outcome classes: normal hearing, mild, moderate, moderately severe, and severe/profound hearing loss. Audiometric thresholds at 500, 1000, 2000, 4000, and 8000 Hz for both ears were combined with age, sex, noise exposure, ototoxic medication exposure, diabetes, family history, air-bone gap, speech-discrimination score, and an interaural laterality index. Four supervised models—multinomial logistic regression, support vector machine, random forest, and a three-hidden-layer neural network—were trained using an 80:20 stratified split. Accuracy, precision, recall, F1-score, and confusion matrices were used for evaluation. The illustrative analysis showed that nonlinear models captured the interaction between audiometric and clinical variables more effectively than a simple linear classifier. Random forest provided strong interpretability through feature importance, while the neural network represented the deep-learning direction of the framework. The results should not be interpreted as clinical performance because the dataset is synthetic. Instead, the study demonstrates a complete research design that can be transferred to real audiology records after ethical approval, expert annotation, data cleaning, and external validation. The paper also discusses class imbalance, explainability, data privacy, bias, clinical integration, and the need for prospective validation.