Accurate Prediction of Humphrey 10-2 Visual Fields in Glaucoma from a Single, Rapid IMOvifa 24plus(1-2)

<h2>Abstract</h2><h3>Purpose</h3> Predicting the Humphrey Field Analyzer (HFA) 10-2 visual field (VF) using machine learning (ML) based on IMOvifa 24plus(1-2) VF data. <h3>Design</h3> Retrospective cross-sectional study. <h3>Participants</h3> 70 actual IMOvifa 24plus(1-2) tests from 25 patients (Jikei University School of Medicine) and 3,472 synthesized IMOvifa 24plus(1-2) tests from 884 patients who underwent HFA 24-2 and HFA 10-2 VF measurements at four affiliated hospitals. <h3>Methods</h3> Synthesized 24plus(1-2) data were created by merging 54 points from HFA 24-2 and 24 points from HFA 10-2 tests. An XGBoost model, trained on the synthetic dataset, predicted thresholds at the 68 HFA 10-2 test locations. Model performance was assessed on the actual dataset using Leave-One-Out Cross-Validation (LOOCV). Stratified patient-level bootstrap analyses accounting for multiple tests per patient were used to compare results across three glaucoma severity groups (mild, moderate, advanced) based on HFA 10-2 Mean Deviation (MD). Four models utilizing different input subsets were evaluated: Model 1 (all 78 IMOvifa 24plus(1-2) points), Model 2 (54 points of 24-2), Model 3 (central 40 points of 24plus(1-2)), and Model 4 (central 16 points of 24-2). Test durations were also compared. <h3>Main Outcome Measures</h3> Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Coefficient of Determination (R²) between predicted and measured HFA 10-2 sensitivities. <h3>Results</h3> Model 1 (all 24plus(1-2) points) achieved the highest overall accuracy (MAE 3.59 dB, RMSE 5.71 dB, R² 0.76), significantly outperforming Model 2 (24-2 points; MAE 4.15 dB, RMSE 6.31 dB, R² 0.70) (p < 0.05 for all metrics). Similarly, Model 3 significantly outperformed Model 4 (p < 0.05). Stratified analysis indicated that adding central test points yielded significant accuracy improvements, consistently across all metrics in the moderate group, whereas results varied by metric in the mild and advanced groups. The IMOvifa 24plus(1-2) test duration (mean 155 seconds) was significantly shorter than the HFA 10-2 SITA Standard test (mean 376 seconds) (p < 0.001). <h3>Conclusion</h3> Incorporating the additional central test points from the IMOvifa 24plus(1-2) significantly enhances the accuracy of ML based HFA 10-2 VF prediction. This approach offers an efficient strategy for obtaining detailed central VF information from a single, rapid test, potentially improving glaucoma management.