AI-GUIDED ENDPOINT SELECTION FOR NEUROPROTECTION TRIALS IN GLAUCOMA

Standard Automated Perimetry (SAP) is the mainstay for monitoring glaucoma progression and has been accepted by the U.S. Food and Drug Administration (FDA) as a trial endpoint, but only under stringent criteria of ≥7 dB loss in five pre-specified test locations. Identifying such locations a priori has remained a major barrier for neuroprotection trials. We developed an attention-based graph neural network (GNN) to predict the visual field points most likely to deteriorate (High-5) using baseline SAP data. The model was trained in the Bascom Palmer Ophthalmic Registry (BPOR; 6,996 eyes, 5,405 patients, 40,914 tests) and externally validated in the Duke Glaucoma Registry (DGR; 5,211 eyes, 3,933 patients, 31,225 tests) and the University of Washington Humphrey Visual Field dataset (UWHVF; 2,030 eyes, 1,195 patients, 10,310 tests). In internal validation, the mean slope at High-5 points among progressors was −2.16±0.80 dB/year, compared to −0.55±0.44 dB/year for Low-5 and −1.02±0.40 dB/year for mean deviation (MD). Similar results were observed in DGR (–2.05 vs −0.45 vs −0.93 dB/year) and UWHVF (−2.32 vs −0.66 vs −1.14 dB/year). High-5 showed superior discrimination of progressors from non-progressors with areas under the ROC curve of 0.883, 0.898, and 0.937 across the three cohorts, consistently outperforming MD (0.871–0.911) and Low-5 (0.668–0.731). Nearly all progressing eyes exhibited a repeatable ≥ 7 dB loss in average High-5 sensitivity during follow-up, compared to fewer than 30% when using MD. In sample size projections, High-5 increased the absolute effect size and lowered the σ ² /Δ ² ratio, translating into an estimated 42% reduction in required trial size compared to MD. In conclusion, this GNN-based framework enables data-driven identification of high-risk SAP locations, aligning with regulatory definitions of progression while substantially improving trial efficiency and sensitivity to detect meaningful visual field change.