Every hearing test leaves behind a grid of numbers, yet in many hospitals those numbers are kept only as pictures, locked away from the researchers who could learn from them. Now a team has shown that an open-source reader, guided by a simple habit of consensus among its own guesses, can recover that grid from the images with striking reliability. It is a modest tool with a large purpose: hearing loss touches over a billion people, and understanding it begins with being able to read what was already written down.
Voting OCR Method Automates Extraction of Table-Formatted Audiometric Data from Clinical Report Images
Hearing results locked inside pictures, retyped by hand
Mark
Why does a paper about reading tables in images deserve attention?
Mimi
Because the data exists but is locked in pictures. Hearing loss affects over 1.5 billion people, and research on it stalls when someone has to retype every value by hand.
Mark
What's different here from earlier automation?
Mimi
Earlier work read the plotted audiogram. This handles the numeric table, which many settings export instead.
Luke
Before we get excited, the 97.60% figure is the authors' own, on reports from one hospital. Do we know it holds elsewhere?
Mimi
No. The paper says the method can adapt to other layouts by re-specifying the table region once, but the validation here is 2000 images from 710 patients at the focal hospital.
Mark
And they trained nothing?
Mimi
Right, the pretrained Tesseract model is used zero-shot. The gains come from preprocessing and voting.
Luke
Worth noting that 97.60% is per image, so a few percent of reports still contain at least one error. Is that acceptable for research?
Mimi
That's where the vote helps. Low-consensus cells get flagged for human review rather than trusted blindly.
Luke
Whether that review catches all the errors isn't shown in what we've read. The multi-center promise is also a stated goal, not a demonstrated result.
The Pulse
- Hearing records sit in hospital systems as flat images, forcing researchers into slow, costly hand transcription that invites error.
- Earlier automation worked only on standardized audiogram charts, leaving the common numerical tables untouched and the bottleneck in place.
- A high-security hospital network ruled out commercial cloud services, so the team relied on a free, pretrained Tesseract reader with no retraining.
- Multiple variant readings of each cell vote on the final value, reaching 97.60% image-level accuracy across 2000 reports from 710 patients.
- The method now points toward web-based collection and multi-center studies, though its evidence still comes from a single hospital.
Every hearing test leaves behind a grid of numbers, yet in many hospitals those numbers are kept only as pictures, locked away from the researchers who could learn from them. Now a team has shown that an open-source reader, guided by a simple habit of consensus among its own guesses, can recover that grid from the images with striking reliability. It is a modest tool with a large purpose: hearing loss touches over a billion people, and understanding it begins with being able to read what was already written down.
A hearing test produces a small grid of numbers: thresholds for each ear, air and bone conduction, and the scores for speech. In many hospitals that grid is saved in the electronic medical record only as a picture. Anyone who wants to study it must read each image and retype the values by hand, a process that is slow, expensive and prone to mistakes.
The stakes are considerable. More than 1.5 billion people have some degree of hearing loss, and many causes are preventable. Studying those causes depends on exactly this kind of data. Earlier automated methods focused on audiogram charts, which are standardized and therefore easier to process, but the numerical tables that many clinics and occupational health services produce were left out.
The team built its approach around Tesseract, a free and open-source text-recognition engine, and added a voting system they call VOCR. Because the hospital network was closed and secure, cloud services were not an option, and the pretrained model was used without any retraining. The 710 patients and 2000 report images served only to test the method. The pipeline finds the table using its standard layout, detects the grid lines, and determines each cell's meaning from its position. Speech-score tables are first cleaned of symbols such as "dB" and "%", then each cell is read several times and the readings vote.
The results are strong, with 97.60% accuracy overall and up to 99.95% for some categories. The vote also reveals uncertainty, so reviewers can check only the cells where readings disagree. The authors hope the tool will support shared, multi-center research, but its evidence so far comes from one hospital's reports, and that is the question still to be answered.
A hearing test leaves behind a small grid of numbers: thresholds for the left and right ear, air and bone conduction, spoken-word scores. In many hospitals that grid is stored in the electronic medical record not as data but as a picture of a report. A researcher who wants to study it has to read each image and type the values in by hand. A new paper in Nature's Scientific Reports describes a way to stop doing that.
The stakes are set out at the start. Over 1.5 billion people worldwide have experienced some degree of hearing loss, and at least 430 million need professional care. Left unaddressed, hearing impairment can damage language development, mental health, quality of life, educational achievement and economic independence. The World Health Organization puts the yearly global cost of untreated hearing loss above $980 billion, and many causes are preventable. Studying and preventing it depends on audiometric data: pure tone audiometry (PTA), the speech reception threshold (SRT) and the word recognition score (WRS).
The authors point to a gap in existing work. Earlier efforts to automate extraction have focused on the audiogram, the plotted chart in which thresholds appear as symbols and curves on frequency and intensity axes. Those methods combine convolutional neural networks, object detection and optical character recognition (OCR), and they generalize reasonably well because audiograms are so standardized. But many clinical and occupational-health settings also export results as a numerical table, and no existing automated method covers those reports. Researchers still transcribe them manually, which is slow, costly and error-prone.
The team built a program around Tesseract, the open-source OCR engine, and a voting algorithm they call VOCR. The hospital's network was closed and high-security, so commercial services from Google, Microsoft or AWS were off the table. Nothing was trained or fine-tuned: the publicly available pretrained Tesseract model was used as is. All 710 patients and 2000 report images served only for validation, so any gains come from preprocessing, candidate generation and voting.
The pipeline starts by locating the table from approximate coordinates, since the reports follow a standard layout. Canny edge detection, with thresholds of 50 and 200 and an aperture size of 3, picks out the table's horizontal and vertical lines. Pixel histograms then show peaks at those lines, which give the coordinates of each cell. A cell's meaning, whether ear side, air or bone conduction, or frequency, comes from its position in the grid rather than from reading a label. For a report with a different layout, the table region and its rows and columns are specified once through an interface. Thresholds are bounded, from -10 to 120 dB for air conduction and -10 to 80 dB for bone, and unmeasurable results are recorded as NR.
For the SRT and WRS tables, a symbol-removal step comes first, because marks such as "dB" and "%" are often misread. A horizontal histogram counts the characters in a cell, measures their widths and starting positions, and decides whether a symbol stands alone, in which case the cell counts as empty, or overlaps a digit. The remaining numeric content is then read multiple times from variant images of the same cell, and the candidates vote on the final value.
The reported results are strong. Image-level accuracy is 97.60%. By category, PTA reaches 99.95%, SRT 98.95%, WRS percentage 98.05% and WRS stimulus 99.80%. The vote itself doubles as a reliability measure, so data collectors can review only the cells where candidates disagreed.
The authors suggest the method could run inside web-based data collection services, supporting real-time, standardized data management and, they hope, multi-center studies across different report formats. The evidence so far comes from one hospital's reports, and that is the point to watch.
Notable Quotes
The voting-based accuracy metric allows reviewers to target areas with low consensus, reducing manual extraction.— Study authors, Scientific Reports
Want the full story?
Read the original at Nature ↗