Predicting 10-2 Swedish Interactive Thresholding Algorithm Standard from 24-2C SITA Faster: Development and...

Area of Science:

• Ophthalmology
• Medical Technology
• Artificial Intelligence

Background:

• Glaucoma diagnosis relies on visual field (VF) testing.
• Humphrey Field Analyzer (HFA) 10-2 and 24-2C are standard VF tests.
• Reconstructing 10-2 thresholds from 24-2C data can improve central VF assessment.

Purpose of the Study:

• Develop and validate Deep24-2C, a machine learning model.
• Reconstruct comprehensive HFA 10-2 VF thresholds from HFA 24-2C data.
• Evaluate the model's performance in glaucoma patients.

Main Methods:

• Retrospective study using 3,653 HFA 24-2C tests (Jikei University) and 35 HFA 24-2C tests (Tajimi Iwase Eye Clinic).
• Trained Deep24-2C using subsets of HFA 24-2C points and age to predict HFA 10-2 thresholds.
• Utilized Random Forest, XGBoost, and MLP architectures, evaluating their ensemble combination.
• Employed data augmentation with synthesized tests and evaluated performance using Mean Absolute Error (MAE).

Main Results:

• Incorporating synthesized data improved predictive accuracy (MAE from 2.52 to 2.42 dB).
• Models using additional central test points outperformed conventional 24-2 points.
• The best XGBoost model achieved MAE 2.30 ± 0.01 dB; the ensemble model yielded MAE 1.81 (95% CI, [1.69-1.93]) dB on external validation.

Conclusions:

• Deep24-2C successfully reconstructs HFA 10-2 VF thresholds from HFA 24-2C SITA Faster tests.
• Additional central points enhance predictive performance over conventional grids.
• Deep24-2C offers a practical tool for central VF defect evaluation, reducing testing burden and supporting glaucoma management.