A hybrid unsupervised approach improved the representation of central visual field loss

Background: Quantitative recognition of visual field loss is important for glaucoma patients and has implications for follow-up management. This study aimed to identify the characteristic patterns of 10 − 2 visual field (VF) test results and effectively express central VF loss using a machine learning approach. Methods We obtained 7,927 10 − 2 VF test data from 3,328 patients in five hospitals. We propose a hybrid approach that combines archetypal analysis (AA) and fuzzy c-means (FCM) to identify characteristic patterns and decompose VF without loss. To demonstrate the clinical usefulness of our approach, mean deviation (MD) change prediction was performed through supervised learning using decomposition coefficients change and a linear mixed model was performed to examine the relationship between the MD slope and baseline decomposition coefficients. Results We identified ten characteristic and representative archetypes (AT) for the central VF test results. FCM decomposition results outperformed the AA-only approach in MD change prediction based on mean squared error (MSE) and Pearson correlation coefficient (PCC) prediction evaluation metrics (all P ≤ 0.039 ). In the linear mixed model, the FCM is more suitable in the prediction of MD slope compared with the AA model for both Akaike (AIC) and Bayes information criteria (BIC) (AIC decrease:20.31, BIC decrease:13.33). The FCM baseline coefficients of AT 3, and AT 4 were significantly associated with a faster MD slope (both P ≤ 0.026 ). Conclusions In this study, we used a hybrid approach of unsupervised learning to identify hidden aspects of central VF loss via a characteristic archetype and lossless decomposition. We believe that our approach can help discover hidden clinical features of glaucoma.