Avaliação de justiça algorítmica na detecção de glaucoma a partir de métodos de aprendizado de máquina

Abstract

Glaucoma is a leading cause of irreversible blindness, and its early detection remains challenging. This study evaluated classical machine learning models applied to Optical Coherence Tomography (OCT) images of the peripapillary retinal nerve fiber layer (RNFLT) using the Harvard-GF public dataset. The pipeline included data standardization and dimensionality reduction via PCA (95\% explained variance). Among the tested models, SVM achieved the best overall performance (F1-macro = 0.73; accuracy = 0.73), followed by ensemble methods such as Random Forest and Gradient Boosting ($\approx$0.70). The class fairness analysis showed greater SVM stability across genders and higher intra-group performance by race, with significant drops in \textit{cross}-group scenarios involving Black patients. These findings suggest that classical models can provide viable, low-cost approaches for automated OCT-based glaucoma screening, provided fairness is systematically assessed.