Research Objectives Social inferencing (SI) is often impaired after a traumatic brain injury (TBI). Mechanisms underlying these deficits are unknown. This study explored the accuracy of machine learning classification using eye tracking data to predict SI performance in participants with and without TBI. Design Retrospective data analysis. Setting N/A. Participants Forty-five participants with moderate to severe TBI and fifty-five healthy controls (HC). Only participants with acceptable eye-tracking data were included. Interventions N/A. Main Outcome Measures Eye-tracking data (e.g., fixation number, saccadic direction and amplitude) was collected (Tobii 300x) while participants completed The Awareness of Social Inferencing Test (TASIT), which includes 3 parts: Emotion Evaluation Test (EET), Social-inference-Minimal (SI-M), and Social inference-enriched (SI-E). EET requires participants to infer actors’ emotions from nonverbal cues (e.g., facial and vocal expressions). SI-M and SI-E require participants to infer characters’ thoughts, intentions, and feelings in situations with limited or enriched context, respectively. Results Eye-tracking data was used as part of machine learning classification to predict correct versus incorrect behavioral responses to each TASIT scenario. Eye tracking data significantly correlated with TASIT performance. Using salient time-vectors of the TASIT videos, a voting mechanism on support vector machine and random forest algorithms, we achieved a .71 sensitivity and .75 specificity at predicting correct/incorrect TASIT responses in participants with and without TBI. Conclusions Machine learning and eye-tracking data appear to be useful for predicting SI performance. Findings suggest that post-TBI SI errors are at least partly related to what they are visually attending to during SI tasks. This line of research may help clinicians detect and correct visual scanning problems in patients with TBI, which may improve SI. Author(s) Disclosures Nothing to disclose. Social inferencing (SI) is often impaired after a traumatic brain injury (TBI). Mechanisms underlying these deficits are unknown. This study explored the accuracy of machine learning classification using eye tracking data to predict SI performance in participants with and without TBI. Retrospective data analysis. N/A. Forty-five participants with moderate to severe TBI and fifty-five healthy controls (HC). Only participants with acceptable eye-tracking data were included. N/A. Eye-tracking data (e.g., fixation number, saccadic direction and amplitude) was collected (Tobii 300x) while participants completed The Awareness of Social Inferencing Test (TASIT), which includes 3 parts: Emotion Evaluation Test (EET), Social-inference-Minimal (SI-M), and Social inference-enriched (SI-E). EET requires participants to infer actors’ emotions from nonverbal cues (e.g., facial and vocal expressions). SI-M and SI-E require participants to infer characters’ thoughts, intentions, and feelings in situations with limited or enriched context, respectively. Eye-tracking data was used as part of machine learning classification to predict correct versus incorrect behavioral responses to each TASIT scenario. Eye tracking data significantly correlated with TASIT performance. Using salient time-vectors of the TASIT videos, a voting mechanism on support vector machine and random forest algorithms, we achieved a .71 sensitivity and .75 specificity at predicting correct/incorrect TASIT responses in participants with and without TBI. Machine learning and eye-tracking data appear to be useful for predicting SI performance. Findings suggest that post-TBI SI errors are at least partly related to what they are visually attending to during SI tasks. This line of research may help clinicians detect and correct visual scanning problems in patients with TBI, which may improve SI.