Computer Vision-based Extraction of Structured Data From Scanned Audiograms in the Electronic Health Record.

OBJECTIVE

To develop and evaluate a computer vision (CV) method for extracting structured hearing threshold data from scanned audiogram test sheets stored in the electronic health record (EHR).

STUDY

DESIGN

Algorithm development and validation study using a contour-based CV pipeline.

SETTING

Tertiary academic health system.

PATIENTS: A total of 907 hand-filled audiogram test sheets (January 1, 2014, to December 31, 2022) were selected via stratified random sampling to ensure balanced representation of normal hearing, bilateral sensorineural, asymmetric sensorineural, conductive, and mixed hearing loss configurations. Thresholds for all symbols on the pure-tone plots were manually extracted to serve as ground truth.

INTERVENTIONS: The CV pipeline accepted scanned audiograms in PDF format and returned estimated frequency (Hz) and threshold (dB HL) values for each symbol.

MAIN OUTCOME MEASURES: Accuracy and mean absolute error in frequency and threshold between CV-estimated and human-labeled values on a test set of 30 audiograms comprising 618 thresholds.

RESULTS

The CV pipeline consisted of 5 steps: image cropping, symbol detection via grayscale preprocessing and contour analysis, axis label detection using optical character recognition, coordinate calibration, and symbol digitization. Across test set audiograms, the mean absolute error was 136 Hz for frequency and 1.3 dB HL for threshold. Exact threshold accuracy exceeded 85% for unmasked air conduction symbols.

CONCLUSIONS

CV can be used to accurately extract pure-tone threshold data from scanned audiogram test sheets without reliance on deep learning or manual preprocessing. The method offers a scalable solution for transforming legacy audiograms into structured data sets suitable for population-level hearing research, clinical decision support, and epidemiologic surveillance.