Smartphone Based Digital Phenotyping in Predicting Risk factor of Self-harm: application of Machine Learning

The present findings are based on data collected and analyzed through T2. This study examined predictors of nonsuicidal self-injury (NSSI) among adults aged 19–29 years with a history of NSSI. Negative affect, NSSI urges and behaviors, sleep, physical activity, and physiological indicators were repeatedly assessed using smartphone-based ecological momentary assessment and wearable devices. The final analysis included 151 participants, comprising 88 participants at Wave 1 and 63 participants at Wave 2. Partial least squares regression with cross-wave validation showed that the prediction models explained 55.9%–67.2% of the variance in NSSI urges. Across both waves, negative emotional states, including depression, sadness, anxiety, fear, emptiness, self-directed anger, and rejection, were identified as the primary predictors, whereas wearable-derived sleep, heart rate, and activity indicators showed relatively smaller contributions. The model predicting subsequent NSSI occurrence from Wave 1 to Wave 2 demonstrated significant predictive performance, with an AUC of 0.764 and a sensitivity of 0.792. Participants who engaged in NSSI during the observation period reported consistently higher levels of negative affect than those who did not. At Wave 2, the NSSI occurrence group also tended to show shorter REM sleep duration and slower walking speed.