Predicting 10-2 Swedish Interactive Thresholding Algorithm Standard from 24-2C SITA Faster: Development and Validation of Deep24-2C.

Abstract

Purpose: To develop and validate Deep24-2C, a machine learning (ML) model that reconstructs comprehensive Humphrey Field Analyzer (HFA) 10-2 visual field (VF) thresholds from HFA 24-2C.

Design: A retrospective study.

Participants: A total of 3653 HFA 24-2C tests, including 181 actual tests from 136 eyes of 130 patients with glaucoma or suspected glaucoma, and 3472 synthesized tests at the Jikei University School of Medicine were included. Thirty-five actual HFA 24-2C tests from 24 eyes of 21 patients at Tajimi Iwase Eye Clinic were used for pilot external validation.

Methods: The HFA 24-2C testing grid incorporates 10 additional central test points into the conventional 24-2 pattern. Deep24-2C was trained to predict all 68 threshold values of the HFA 10-2 test using subsets of 24-2C points (22-26 within the 10° region) along with age. Three ML architectures-random forest (RF), XGBoost, and multilayer perceptron (MLP)-were developed, and their ensemble combination was evaluated. Model training and internal validation were conducted using the Jikei dataset, while external validation used the Tajimi dataset. Data augmentation was applied using synthesized 24-2C tests, which combined paired HFA 24-2 and 10-2 tests. Model performance was evaluated using mean absolute error (MAE).

Main outcome measures: Mean absolute error for 10-2 point-wise thresholds and mean deviation.

Results: We first trained and validated Deep24-2C using the Jikei dataset. Incorporating synthesized data improved predictive accuracy from MAE 2.52 ± 0.07 dB to 2.42 ± 0.07 dB. Furthermore, models utilizing additional central test points outperformed those based on conventional 24-2 points across all ML architectures (RF, XGBoost, and MLP); the best-performing XGBoost model achieved MAE 2.30 ± 0.01 dB for predicting 10-2 thresholds. Subsequently, we performed external validation on the Tajimi dataset using an ensemble model that combined all models utilizing 10 additional central test points; the ensemble model yielded MAE 1.81 (95% confidence interval [1.69-1.93]) dB.

Conclusions: Deep24-2C enables reconstruction of 10-2 VF thresholds from a 24-2C Swedish Interactive Thresholding Algorithm Faster test. Incorporating additional central points enhances predictive performance compared with conventional 24-2 grids. Deep24-2C may serve as a practical and efficient tool for evaluating central VF defects, reducing testing burden, and supporting individualized glaucoma management.

Financial disclosure(s): Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.