Development of an automated method for audiometric data extraction and analysis using voting optical character recognition

Hearing loss affects over 1.5 billion people worldwide, significantly impacting language development, quality of life, and economic independence. Research on audiometric data, such as pure tone audiometry (PTA), the speech reception threshold (SRT), and the word recognition score (WRS), is essential for understanding and preventing hearing loss. However, this data is frequently stored as image files within electronic medical records (EMRs), thus requiring manual extraction for research use. Manual extraction is time-consuming, costly, and prone to errors, limiting data utility and impeding large-scale data acquisition. In this paper, we propose an automated extraction method for audiometric report images using Tesseract optical character recognition (OCR) and an improved voting OCR (VOCR) algorithm. VOCR mitigates OCR variability by removing symbols (e.g., “dB,” “%”) and using tabulated data, generating multiple OCR candidates per cell and utilizing a voting mechanism to select reliable values. Our results indicate that VOCR achieves an image-level accuracy of 97.60% and high accuracy across all audiometric data categories, including PTA (99.95%), SRT (98.95%), WRS % (98.05%), and WRS stimulus (99.80%). The voting-based accuracy metric allows for the assessment of reliability, enabling a targeted review of areas with low consensus. This method has the potential to be employed in web-based data collection services, supporting real-time, standardized data management and reducing the need for manual extraction, making it a valuable tool for otolaryngological research.