Background/Objectives: Assessment of Internet Gaming Disorder (IGD) relies largely on retrospective self-report and clinical interviews, which may be affected by recall bias, social desirability bias, and limited sensitivity to within-person behavioral change. This study evaluated an AI-enabled, privacy-preserving digital phenotyping framework for personalized IGD risk stratification under controlled simulation assumptions. Methods: A reproducible synthetic dataset of 1,000 virtual user profiles was generated; 20% were assigned to an elevated-risk class, and 5% balanced stochastic label noise was introduced to approximate imperfect ground truth. Four aggregated telemetry features were modeled: average session duration, sessions per week, Late-Night Index, and application-switching rate. Random Forest, Logistic Regression, and Gradient Boosting classifiers were evaluated against playtime-only baselines using a stratified 80:20 train–test split. Results: In the primary Random Forest model, accuracy was 0.880, balanced accuracy was 0.850, sensitivity was 0.800, specificity was 0.900, area under the receiver operating characteristic curve (ROC-AUC) was 0.909, area under the precision–recall curve (PR-AUC) was 0.779, and the Brier score was 0.089. All-feature models substantially outperformed playtime-only baselines. Feature-importance analyses recovered the known signal hierarchy encoded in the synthetic data-generating process, with application-switching rate and Late-Night Index showing the largest Gini-based and permutation-importance values. Performance degraded progressively as label noise increased from 0% to 20%. Conclusions: The framework demonstrates the methodological feasibility of transforming aggregated, privacy-preserving behavioral telemetry into interpretable simulated risk signals for IGD. The findings are hypothesis-generating and do not establish clinical validity or diagnostic performance. Longitudinal validation in clinically characterized cohorts using validated psychometric instruments and person-level calibration is required before practical deployment.
Background/Objectives: Assessment of Internet Gaming Disorder (IGD) relies on retrospective self-reports and clinical interviews, which may be affected by recall and social desirability biases and may be insensitive to behavioral change. This study evaluated an artificial intelligence (AI)-enabled, privacy-preserving digital phenotyping framework for personalized IGD risk stratification under controlled simulation assumptions. Methods: A synthetic dataset of 1000 virtual user profiles was generated with a 20% elevated-risk prevalence and 5% balanced stochastic label noise. Four aggregated telemetry features were modeled: average session duration, sessions per week, Late-Night Index, and application-switching rate. Random Forest, Logistic Regression, and Gradient Boosting classifiers were evaluated using a stratified 80:20 hold-out split, five-fold cross-validation, playtime-only baselines, label-noise sensitivity analysis, and 200 synthetic realizations. Results: The primary Random Forest model achieved a balanced accuracy of 0.850, a sensitivity of 0.800, a specificity of 0.900, an area under the receiver operating characteristic curve (ROC-AUC) of 0.909, an average precision (AP) of 0.779, and a Brier score of 0.089. As an internal consistency check under the pre-specified synthetic signal structure, all-feature models showed higher performance than playtime-only baselines, and feature importance analyses recovered the encoded signal hierarchy. Performance declined with increasing label noise. Across 200 realizations, mean ROC-AUC values for the three all-feature models ranged from 0.888 to 0.904, with overlapping empirical 95% intervals. Conclusions: The framework demonstrates the methodological feasibility of transforming aggregated telemetry into interpretable risk signals while avoiding content-level monitoring. These findings are hypothesis-generating and do not establish clinical validity or diagnostic performance. Longitudinal validation in clinically characterized cohorts is required before deployment.