Automated perimetry is a key functional test for evaluating visual field impairment. In conventional clinical practice, the visual field pattern and testing strategy are selected before acquisition according to diagnostic suspicion and specialist criteria. However, when an initial examination is performed by non-specialist personnel and a specialist is not immediately available, the decision to repeat, expand, or terminate the study may be deferred until later review, generating bottlenecks and possible rescheduling. This study proposes a point-of-care artificial intelligence system that uses an initial rapid 24-2 examination as standardized input to recommend the subsequent campimetric protocol without replacing the specialist’s definitive diagnosis.
A database of 28,943 24-2 visual fields was analyzed. After preprocessing and hierarchical labeling, 67 features were selected, including age, evaluated eye, mean sensitivity, 52 point-wise sensitivities, and 12 derived spatial variables related to suspect points, clusters, defect intensity, eccentricity, centroid location, and central or peripheral involvement. A hierarchical AI architecture was designed with two tasks: Model A recommended the subsequent field, including no expansion, repeat 24-2, expand to 10-2, or expand to P50-2; Model B selected either the ZETA or ZETA_FAST strategy. An interpretable Decision Tree was compared with two deep learning approaches, a Multilayer Perceptron and a Convolutional Neural Network.
For field selection, the Decision Tree achieved 0.9868 accuracy and 0.9449 macro-F1, while the MLP and CNN reached 0.97 accuracy and 0.92 macro-F1. For strategy selection, the Decision Tree obtained 0.9933 accuracy and 0.9923 macro-F1, followed by the MLP with 0.93 accuracy and 0.92 macro-F1, and the CNN with 0.90 accuracy and 0.88 macro-F1. These results suggest preliminary feasibility for AI-assisted campimetric triage at the point of care. External validation and specialist review remain necessary before clinical implementation.