3 - (SYM 61) Employing Computer Vision to Objectively Assess the Severity of Self-injury to Augment Suicide Risk Prediction

Background: A history of prior self-injury, including both nonsuicidal self-injury (NSSI) and suicidal self-injury (e.g., suicide attempts [SA]), has been consistently found to be the strongest predictor of future suicidal behavior, with evidence suggesting that the more severe such behaviors are, the greater the risk for future self-injury. Importantly, however, our current means of assessing severity of prior self-injury is almost entirely reliant on self-report, despite the fact that self-injury frequently leaves visible physical markings. Although applications of machine learning in medical image analysis are growing exponentially, none have attempted to augment suicide risk detection through automated analysis of self-directed tissue damage. The objective of this novel proof-of-concept study is to utilize computer vision techniques to automate the assessment of hypothesized self-injury visual severity indicators (scar frequency, method, recency) that may indicate risk for prospective SAs.

Methods: Adolescents (N = 2,545) ages 16 to 18 years old were recruited via social media advertising if they had currently visible self-injury scarring. Participants uploaded images of their arms and annotated their images, outlining scars on their arms and providing information about scar origin, frequency, method, and recency. Clinical research staff annotated a subset of participant images to serve as scar presence ground truth. Participants with a past-month self-injury history (n = 1,497) were followed longitudinally for three months to assess prospective self-injury.

Results: Findings demonstrate our computer vision scar detector model employing FasterRCNN evidenced a mean average precision of .81 and performed well in detecting scar frequency (ground truth scar detection rate = 91%). Results indicate our scar origin detection models evidenced an 87% accuracy in classifying scar origin as either SA or NSSI. Further, our models suggest a 96% accuracy in classifying self-injury method and a 76% accuracy in classifying self-injury scar recency.

Conclusions: Findings offer initial evidence that a computer vision scar detector model can validly detect scars, scar origin, and self-injury severity indices. Clinical applications of these findings will be discussed, including the promise of this technology to serve as a clinical decision-support tool to help augment suicide risk detection.