Multimodal Machine Learning for Glaucoma Detection in a Sub-Saharan African Clinical Population

PurposeTo evaluate the performance of machine learning models for automated glaucoma detection using multimodal clinical, structural, and functional data from a West African clinical cohort. MethodsIn this retrospective observational study, we analyzed clinical records from two major eye care centers in Ghana. A total of 605 eyes from 417 patients who underwent comprehensive glaucoma evaluation were included. Extracted features included demographic data, intraocular pressure, optical coherence tomography (OCT) structural parameters, and Humphrey visual field indices. We assessed the diagnostic performance of individual parameters using receiver operating characteristic (ROC) analysis. Supervised machine learning classifiers, including support vector machine (SVM), random forest (RF), gradient boosting machine (GBM), and a multi-layer perceptron (MLP), were trained using a forward feature selection approach and evaluated using five-fold cross-validation. We assessed model performance by computing performance metrics like sensitivity, specificity, and area under the ROC curve (AUC). ResultsAmong the 605 eyes analyzed, 361 (59.7%) were glaucomatous, and 244 (40.3%) were healthy. Individual structural and functional parameters demonstrated moderate discriminative ability, while age showed no significant diagnostic value (AUC = 0.49, p = 0.841). Among machine learning models, the MLP achieved the highest diagnostic performance (AUC = 0.90 [95% CI: 0.86-0.92], sensitivity = 0.88, specificity = 0.86), outperforming SVM (AUC = 0.82), RF (AUC = 0.78), and GBM (AUC = 0.77). Multimodal integration of clinical, structural, and functional features improved discrimination compared with individual parameters. ConclusionsMultimodal machine learning models can effectively automate glaucoma detection using routinely collected clinical data. In this West African cohort, an MLP-based approach demonstrated superior diagnostic performance compared with traditional machine learning models and individual clinical metrics. These findings highlight the potential of clinically grounded artificial intelligence tools to support glaucoma diagnosis and triage in resource-constrained eye care settings.