Automatically extracted valuesLightGBM
94.72%test F1
Journal of Otology · 2025 Co-author
From marks on paper
to meaningful data.
Can a model read an audiogram?
We combined object detection with machine learning to turn audiogram images into numerical features and classify hearing-loss severity.
Follow the extraction ↓Reading between the gridlines.
01 / Explore the methodImage → symbols → valuesSynthetic example
AUDIOMETRYIllustration / 001
First, find the frameGive the marks
Give the marks
a coordinate system.
The first detector locates graphs and tables. A second detector reads the symbols inside the graph.
The outline shows the graph region in this illustration.
Two symbol classes are shown here. The study detects eight, including air- and bone-conduction symbols.
| Hz | Right ○ | Left × |
|---|
An original schematic with invented values. No patient data or live model prediction.
Then, classify
The extracted features feed a classifier with six severity categories, from normal to profound.
Explore the implementation ↗02 / Published evidenceTable 5 in the paper ↗
How much survives
the move from
image to numbers?
A retrospective study of 1,959 audiograms from Vajira Hospital. Detection used 200 images; classification used 1,407 training and 352 test images.
Manually extracted valuesRandom forest
96.43%test F1
Best model in each group. Different classifiers, so the gap does not isolate extraction error.
Compare all five reported classifiers
| Classifier | Automatic | Manual |
|---|---|---|
| Gradient boosting | 93.88 | 96.29 |
| LightGBM | 94.72 | 96.29 |
| Random forest | 93.17 | 96.43 |
| Decision tree | 93.35 | 92.75 |
| Extra trees | 90.33 | 95.00 |
What remains difficult
Overlapping symbols, handwriting, and unfamiliar chart formats can disrupt extraction. These single-hospital results do not establish performance elsewhere.
03 / Read the research
Developing a diagnostic support system for audiogram interpretation using deep learning-based object detection
Achakulvisut et al. · Journal of Otology 20 (2025), 26–32
Source note
Model assignments follow Table 5 and the repository; the abstract reverses them. Dataset counts follow the results and Figure 3 because split percentages in the prose are inconsistent.