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

Area of Science:

• Digital mental health
• Computational psychiatry
• Behavioral addiction research

Background:

• Digital addiction (DA) is a growing global concern.
• Traditional self-report methods for DA lack objectivity and consistency.
• Machine learning (ML) presents a novel approach for automated DA detection.

Purpose of the Study:

• To systematically evaluate the diagnostic accuracy of ML models for digital addiction.
• To assess the performance of ML across different subtypes of digital addiction.
• To compare the effectiveness of survey-based versus physiological data in ML-driven DA detection.

Main Methods:

• A systematic meta-analysis of 64 studies (75 datasets; N=165,624) was performed.
• Single-group proportion and bivariate diagnostic test accuracy (DTA) models were employed.
• Subgroup analyses were conducted for different addiction subtypes and data types.

Main Results:

• The pooled classification accuracy for ML-based DA detection was 0.87 (95% CI [0.85, 0.90]).
• The DTA framework achieved an AUC of 0.92, with balanced sensitivity and specificity (0.86).
• High accuracy was observed for internet (0.90) and social media addiction (0.86); physiological data showed superior specificity (0.90).

Conclusions:

• ML-driven tools demonstrate significant potential as scalable screening instruments for digital addiction.
• Further research requires representative sampling and standardized diagnostic criteria for advancing digital mental health.
• Automated detection can overcome limitations of subjective self-report measures in diagnosing DA.