Valence-specific oculomotor abnormalities induced by dynamic video stimuli as a robust biomarker for schizophrenia

Deficits in visual attention and emotional processing are core to schizophrenia (SZ). Although oculomotor abnormalities during static viewing are well documented, it remains unclear whether they persist or differ by valence under dynamic emotional stimuli that better mimic real-world social situations. Identifying emotion-related differences in oculomotor responses may help characterize candidate behavioral markers in SZ.

We enrolled 104 SZ patients and 104 demographically matched healthy controls (HCs). All participants viewed stimuli within the Dynamic Emotional Video Paradigm (DEVP), from which 14 oculomotor metrics were extracted. Using two feature selection strategies and nested 5-fold cross-validation (CV), we compared all-emotional and emotion-specific machine learning models for SZ classification.

SZ patients showed oculomotor abnormalities across all-emotional conditions, including altered fixation duration, changed saccade patterns, and decreased pupil responsiveness. These findings extend existing static-task evidence to more ecologically valid dynamic conditions. Oculomotor impairments varied by valence, with relatively more evident changes under sad stimuli. Random forest (RF) selection identified pupil dynamics and fixation duration as key features, whereas correlation-variance filtering retained a wider range of oculomotor indices. The sad-specific model achieved accuracy of 0.79 and AUC of 0.86, with a marginal advantage over global models.

Our findings highlight that oculomotor impairments in SZ exhibit clear valence specificity under dynamic emotional stimuli. Oculomotor metrics measured under the sad emotion condition show favorable biomarker potential for supporting auxiliary early screening of SZ.