Hearing loss is a prevalent and growing global health concern affecting approximately 466 million people. By 2050, this number is projected to exceed 700 million, impacting nearly one in ten individuals worldwide (World Health Organization, 2024). These trends highlight the urgent need for effective public health strategies, early intervention, and enhanced global accessibility to hearing healthcare.

Accurate diagnosis and management of hearing loss are essential and rely on the systematic utilization of hearing data. In recent decades, healthcare services have undergone a significant shift from paper-based services to electronic medical records (EMRs) (Ge et al., 2022). The adoption of EMRs has substantially increased the volume of hearing data, transforming healthcare delivery and patient outcomes (Saunders et al., 2021). These developments enable more precise and personalized diagnostic and therapeutic approaches through the comprehensive integration, management, and analysis of large-scale hearing data.

However, in addition to audiograms, approximately 80 % of data in EMRs comprise unstructured information (e.g., medical images and videos), necessitating data extraction tools to derive meaningful information (Martin-Sanchez and Verspoor, 2014). Medical researchers often struggle to fully utilize the vast amounts of healthcare data embedded within fragmented clinical databases for both clinical and research purposes (Negro-Calduch et al., 2021). Despite their clinical significance, hearing data, particularly audiograms, are predominantly stored as image files, which limits their numerical analysis and integration into EMRs (Yang et al., 2024). This limitation hinders the effective use of hearing data to construct CDMs and develop AI-based diagnostic tools. Therefore, addressing the challenges posed by unstructured data is critical for advancing auditory diagnostics.

Deep-learning-based image analysis offers solutions for extracting and structuring meaningful information from unstructured data (Hoover et al., 2023). Recent studies have demonstrated potential for various aspects of auditory data analysis. For instance, several previous studies (Charih and Green, 2022; Crowson et al., 2020; Kassjański et al., 2024; Yang et al., 2024) have utilized deep learning models trained on audiograms to (1) classify types of hearing tests and (2) quantify handwritten audiograms. Other studies focused on predicting speech audiometry results from pure-tone audiometry (PTA) data (Kim et al., 2021; Shin et al., 2024b). Some studies have focused on the automated classification of audiograms to accurately determine the degree, type, and configuration of hearing loss, thereby assisting in clinical diagnosis (Dou et al., 2024; Shin et al., 2024a). In addition, advanced machine learning strategies have been integrated into PTA protocols, such as the implementation of automated masking techniques that reduce the test duration and improve the consistency of test results (Wallaert et al., 2024). Comprehensive reviews of digital hearing assessment methods further indicate that although these approaches can perform at levels comparable to or better than conventional techniques, persistent challenges remain in terms of data quality, model standardization, and seamless integration with existing clinical workflows (Wasmann et al., 2022). Notably, most of these studies addressed discrete tasks, such as prediction, classification, or test automation, without integrating the extracted data into broader clinical infrastructures such as EMRs and CDMs.

Our research addresses this gap by developing a comprehensive deep-learning framework that automates data extraction and ensures robust data security through pseudonymization and de-identification. This comprehensive framework facilitates the construction of large-scale hearing datasets and CDM, ensuring the reliability of digitized data and its applicability for longitudinal clinical analyses. By emphasizing accurate numerical transcription, this approach ensures the reliability and consistency of hearing data, which are essential for the effective analysis and interpretation of research outcomes, supported by pseudonymization and de-identification measures. It offers significant potential to save time and resources and provides high efficiency, particularly in clinical settings. This model will ultimately contribute to the development of improved treatment options for individuals with hearing loss by enhancing the accuracy and efficiency of hearing data collection in clinical and research settings.