A novel system for the automatic reconstruction of visual field based on eye tracking and machine learning

Eye movement perimetry (EMP) is a paradigm developed to assess the visual field without the necessity of suppressing the natural eye movements during the test. Unlike the standard automated perimetry (SAP) where the patient's responses are recorded using a button, EMP uses the natural eye movements reflex as responses during the evaluation. The reliability of EMP depends on correctly determining whether a stimulus is seen or not which, in turn, depends on an adequate analysis of the eye movement data. However, many studies in EMP have focused on characterizing eye movements and only a few authors have documented their methods to determine whether a peripheral stimulus was seen during the test. Furthermore, many of them use static thresholds to perform the classification, but it is not clear how these threshold values were obtained. Based on the foregoing, we develop a threshold test based on FASTPAC C24-2 and EMP for the visual field assessment. Our method uses two machine learning techniques: (1) cascaded K-Means and Bayesian classifiers (KBC) and (2) an Artificial Neural Network (ANN) to classify whether a stimulus was seen or not. Our method was validated with twenty healthy participants (13 women and 7 men) aged 19-43 years (mu = 26 +/- 5 years), where the participants performed both an EMP test and an SAP emulation test. Results were compared with gaze trajectories annotations performed by an expert, obtaining accuracy values between 96.8% and 98.9% for KBC and ANN, and values between 90.5% and 92% for SAP emulation.