Applications of machine learning algorithms to detect digital addiction: a meta-analysis

Digital addiction (DA) has emerged as a significant global concern, yet traditional diagnostic methods relying on self-report questionnaires face subjective bias and threshold inconsistencies. Recent advances in machine learning (ML) offer promising alternatives for automated DA detection. This study conducted a systematic meta-analysis of 64 eligible studies (75 independent datasets; N = 165,624), employing both single-group proportion and bivariate diagnostic test accuracy (DTA) models. The pooled classification accuracy was 0.87 (95% CI [0.85, 0.90]), and the DTA framework yielded a robust AUC of 0.92, with balanced sensitivity and specificity (both 0.86). Subgroup analyses showed high accuracy across subtypes, particularly for internet (0.90) and social media addiction (0.86). Accuracy was comparable between survey-based and physiological data, though physiological markers demonstrated superior specificity (0.90). These findings underscore the potential of ML-driven tools as scalable screening instruments while emphasizing the need for representative sampling and standardized diagnostic criteria to advance digital mental health practice.