Scanning faces: a deep learning approach to studying eye movements in prosopagnosia.

Background: Healthy individuals show fixation biases when scanning faces, likely toward the regions that are most informative for identifying faces. Some reports suggest that subjects with prosopagnosia, an impairment in face recognition, have anomalous face scanning. Objective: Our goal was to determine whether an approach using artificial intelligence could identify key scanning markers of prosopagnosia. Methods: We used an image-classification technique based on deep learning to study the fixations of subjects with and without prosopagnosia during a face recognition task. We identified the number of fixations that maximizes classification performance and developed two methods of displaying scanpaths as images, each used to train a convolutional neural network. Results: Optimal classification of acquired prosopagnosic from control trials required four fixations, with an AUC of 80%. The model showed a greater tendency to fixate the lower face and the right eye in acquired prosopagnosia. Optimal classification of developmental prosopagnosic from control trials required 16 fixations, with an AUC of 69%. Fixations on developmental prosopagnosic trials were shifted more toward peripheral regions. When the classifier trained to discriminate acquired prosopagnosia from controls was asked to analyze the developmental prosopagnosic trials, the latter were classified as being more like control scanpaths. Conclusion: Only a few fixations during face scanning are required to differentiate controls from acquired prosopagnosia, with the latter showing anomalous biases. Developmental prosopagnosic scanpaths resemble degraded control scanpaths rather than anomalous biases. This study shows the potential of deep learning to identify abnormal behavioral markers in a disorder of complex visual processing.