Voting OCR recovered complete hearing-test tables from 97.6% of report images

A hearing test may already be stored in an electronic medical record, yet still be unusable for large-scale research. The result is a picture of a table, not values that a computer can search, compare or analyse. A Korean research team has shown that recovering those numbers may not require training another language model. On a set of 2,000 digital audiometric-report images, an existing optical-character-recognition system correctly extracted every required cell in just 24.95 percent of reports. An added preprocessing and voting procedure raised that strict image-level accuracy to 97.60 percent.[1]

The work, published in Scientific Reports on 26 October, studied files from 710 patients who attended one institution between 2013 and 2018. The reports were originally generated electronically as PDFs and converted to raster images for storage. They were not photographs of paper records or an unstructured collection from dozens of hospitals. All 2,000 images were reserved for validation, and the recognition engine itself was a pretrained open-source Tesseract model used without fine-tuning. That limited but clearly defined setting explains both the large result and its boundaries.[1]

What makes a hearing table difficult to read

Audiometric reports store several different types of clinical measurement. Pure-tone audiometry describes hearing thresholds at selected frequencies. Speech reception threshold is another measured quantity, and word-recognition fields can contain percentages and stimulus levels. A robust extraction system must handle the layout, distinguish nearby columns and preserve the meaning of units such as decibels and percent signs. One wrong cell in a multi-field report can spoil an entire extracted patient record.[1]

The researchers used a demanding definition of success. Every cell across four report regions had to match the human-verified reference for the image to count as correct. The pure-tone region alone contained 48 cells. Under an all-or-nothing standard, a conventional recogniser can get most digits right yet fail a large proportion of full records. The headline difference between 24.95 and 97.60 percent is therefore not a comparison of individual characters recognised; it is the proportion of reports whose required tables were entirely correct.[1]

The voting approach generated multiple OCR candidates for a cell after image preprocessing and compared them rather than trusting the first output. It included handling for unit symbols and character-length expectations, with consensus used to select values. The logic is familiar outside medicine: repeated imperfect readings can be more reliable when their mistakes are not identical. The achievement is engineering around an existing model's failure points in a constrained document format.[1]

A strict test on a consistent format

The ground truth was not taken on faith from the OCR output. Otolaryngology residents transcribed the audiometric values, those entries were independently checked, and discrepancies were resolved by consensus. Each report then went through both recognition methods, allowing a paired statistical comparison. The authors report a statistically significant improvement using McNemar's test, with p below 0.001. They also say the proposed method never made a previously correct basic-OCR image worse in their 2,000-image evaluation.[1]

The gains were substantial across individual groups: 99.95 percent image-level accuracy for the pure-tone region, 98.95 percent for speech reception threshold, 98.05 percent for word-recognition percentage and 99.80 percent for stimulus level. Those are within-region all-cells-correct figures, not the chance that any arbitrary audiogram in the world can be read automatically. The authors specifically note reliance on a fixed table format and monospaced-font widths; performance on other hospitals' layouts, proportional fonts and scanned paper has not yet been established.[1]

The hospital’s closed network prevented the use of external commercial APIs. A local pretrained recogniser therefore served a practical constraint as well as a technical one: the extraction could run where the records already lived.[1]

The gain came from understanding the document, not retraining the recogniser. Fixed table positions, symbol removal and multiple readings turned mostly correct digits into far more completely correct records. Other archives would need validation on their own layouts, but this one demonstrates how much useful engineering can sit around an existing model.[1]

Report an error or suggest a correction