Modeling the Psychosocial Determinants of Problematic Social Media Use Risk Among High School Students Using Artificial Neural Networks

Abstract Background Problematic social media use during adolescence is shaped by interacting psychological and behavioral factors. However, evidence regarding the relative predictive contributions of these factors remains limited. This study aimed to examine academic expectation stress, loneliness, mobile device ownership, and physical activity level as predictors of Bergen Social Media Addiction Scale (BSMAS)-based risk status among high school students. To this end, an artificial neural network (ANN) model was employed, complemented by binary logistic regression analysis. Methods Data from 899 high school students were analyzed. Participants who scored 19 or higher on the Bergen Social Media Addiction Scale (BSMAS) were classified as being at risk for problematic social media use. A multilayer perceptron artificial neural network (ANN) model, comprising four standardized input variables, a single hidden layer with three neurons, and two output units, was trained using an internal training–test split of the dataset. Model performance was evaluated using accuracy, sensitivity, specificity, precision, F1 score, balanced accuracy, and the area under the receiver operating characteristic curve (AUC). Binary logistic regression analysis was additionally used as a complementary model to enhance interpretability. Results The ANN model achieved an accuracy of 93.06% in the training subset and 93.54% in the test subset. In the test subset, sensitivity was 92.80%, specificity was 96.55%, precision was 99.10%, the F1 score was 95.84%, balanced accuracy was 94.67%, and the AUC was .971. Academic expectation stress had the highest normalized importance (100.0%), followed by loneliness (58.7%), mobile device ownership (51.9%), and physical activity level (35.9%). The logistic regression model was statistically significant, χ²(4) = 672.06, p < .001, with a Nagelkerke R² of .808 and an apparent AUC of .971 (95% CI [.956, .986]). Variable ranking based on the Wald statistic reproduced the importance ranking obtained from the ANN model. Conclusion Academic expectation stress emerged as the most prominent predictor of risk for problematic social media use, while loneliness, mobile device ownership, and physical activity level provided complementary predictive information to the model. The convergence of the ANN and logistic regression findings supports a multidimensional understanding of risk status among adolescents. However, external validation studies and longitudinal research designs are needed before the model can be used for individual-level screening or decision-making.