Nibeditaa SpeaksResearchers develop a voting-based optical character recognition system to automate audiometric data extraction, potentially reducing manual errors and accelerating hearing loss research.
With more than 1.5 billion people worldwide experiencing some degree of hearing loss, researchers are increasingly looking for ways to improve the collection and analysis of hearing-related medical data. A new automated method using optical character recognition (OCR) could help researchers process hearing test records faster, reducing the time, cost and errors associated with manual data entry.
The method, detailed in a study titled Development of an Automated Method for Audiometric Data Extraction and Analysis Using Voting Optical Character Recognition, combines the open-source Tesseract OCR engine with an improved voting-based algorithm to extract numerical information from images of audiometric reports stored in electronic medical records (EMRs).
The researchers reported an overall image-level accuracy of 97.60%. The system also achieved high accuracy across key categories of audiometric data, including pure tone audiometry (PTA), speech reception threshold (SRT) and word recognition score (WRS).
Hearing assessments generate important information that helps clinicians evaluate patients’ hearing abilities and develop appropriate treatment plans. However, these results are frequently stored as images or tables within electronic medical records, making them difficult to use directly in large-scale research. Researchers often have to manually transfer the information into structured datasets, a process that can be expensive, time-consuming and vulnerable to transcription errors.
How the voting OCR method works
OCR technology converts text contained in images into machine-readable information. Although existing systems can recognise printed characters with high accuracy, extracting medical data presents particular challenges when symbols, percentages, units and numerical values appear in structured tables.
To address these limitations, the researchers developed a voting optical character recognition (VOCR) algorithm. The method generates multiple OCR candidates for each table cell and uses a voting mechanism to select the most reliable value.
The system also removes certain symbols, including “dB” and “%”, that can interfere with character recognition. By concentrating on numerical tables rather than relying exclusively on graphical audiograms, the approach addresses a form of medical data that has received comparatively less attention in automated hearing-test analysis.
The researchers reported an accuracy of 99.95% for PTA data, 98.95% for SRT, 98.05% for WRS percentages and 99.80% for WRS stimulus values.
An additional feature is its use of agreement between OCR outputs to assess reliability. When multiple readings converge on the same value, the result has stronger consensus. When they disagree, the system can flag the entry for targeted human review rather than requiring researchers to check every extracted value manually.
Why automated hearing data extraction matters
Previous automated approaches have focused largely on digitising graphical audiograms, where hearing thresholds are represented by symbols and curves plotted against frequency and intensity. While these methods can recover numerical thresholds from graphs, they do not address all the tabulated results found in clinical and occupational-health records.
The new approach is designed to complement existing methods by extracting numerical data directly from report tables. Its developers suggest that the technique could support multi-centre studies involving records from different medical institutions.
The system could also be incorporated into web-based data collection services, enabling more standardised handling of audiometric information and reducing dependence on manual transcription.
However, the reported accuracy does not eliminate the need for clinical oversight. Low-consensus results and incorrectly recognised values may still require verification before the extracted information is used for research or clinical analysis.
A step towards more accessible hearing research
The World Health Organization has highlighted the substantial health and economic consequences of untreated hearing loss, including effects on communication, education, quality of life and economic independence.
Automating the extraction of hearing-test information could help researchers assemble larger datasets and investigate hearing-related conditions more efficiently. Better access to structured data may also support future studies into hearing impairment and its management.
The researchers’ approach does not replace hearing assessments or clinical judgement. Instead, it tackles a less visible barrier to progress: the difficulty of turning existing medical records into usable research data.
If validated across diverse report formats and clinical settings, voting-based OCR could help make audiometric information easier to analyse at scale, allowing researchers to spend less time transcribing records and more time investigating the causes, consequences and treatment of hearing loss.
Add comment