To build a functional generalized estimating equation (GEE) model to detect glaucomatous visual field progression and compare the performance of the proposed method with that of commonly employed algorithms.

Totally 716 eyes of 716 patients with primary open angle glaucoma (POAG) with at least 5 reliable 24-2 test results and 2y of follow-up were selected. The functional GEE model was used to detect perimetric progression in the training dataset (501 eyes). In the testing dataset (215 eyes), progression was evaluated the functional GEE model, mean deviation (MD) and visual field index (VFI) rates of change, Advanced Glaucoma Intervention Study (AGIS) and Collaborative Initial Glaucoma Treatment Study (CIGTS) scores, and pointwise linear regression (PLR).

The proposed method showed the highest proportion of eyes detected as progression (54.4%), followed by the VFI rate (34.4%), PLR (23.3%), and MD rate (21.4%). The CIGTS and AGIS scores had a lower proportion of eyes detected as progression (7.9% and 5.1%, respectively). The time to detection of progression was significantly shorter for the proposed method than that of other algorithms (adjusted P≤0.019). The VFI rate displayed moderate pairwise agreement with the proposed method (k=0.47).

The functional GEE model shows the highest proportion of eyes detected as perimetric progression and the shortest time to detect perimetric progression in patients with POAG.

To evaluate long-term visual field (VF) prediction using K-means clustering in patients with primary open angle glaucoma (POAG).

Patients who underwent 24-2 VF tests ≥10 were included in this study. Using 52 total deviation values (TDVs) from the first 10 VF tests of the training dataset, VF points were clustered into several regions using the hierarchical ordered partitioning and collapsing hybrid (HOPACH) and K-means clustering. Based on the clustering results, a linear regression analysis was applied to each clustered region of the testing dataset to predict the TDVs of the 10^th VF test. Three to nine VF tests were used to predict the 10^th VF test, and the prediction errors (root mean square error, RMSE) of each clustering method and pointwise linear regression (PLR) were compared.

The training group consisted of 228 patients (mean age, 54.20±14.38y; 123 males and 105 females), and the testing group included 81 patients (mean age, 54.88±15.22y; 43 males and 38 females). All subjects were diagnosed with POAG. Fifty-two VF points were clustered into 11 and nine regions using HOPACH and K-means clustering, respectively. K-means clustering had a lower prediction error than PLR when n=1:3 and 1:4 (both P≤0.003). The prediction errors of K-means clustering were lower than those of HOPACH in all sections (n=1:4 to 1:9; all P≤0.011), except for n=1:3 (P=0.680). PLR outperformed K-means clustering only when n=1:8 and 1:9 (both P≤0.020).

K-means clustering can predict long-term VF test results more accurately in patients with POAG with limited VF data.