Performance comparison across models.

Background
Digital media usage has become an integral part of daily life, but prolonged or emotionally driven engagement—especially during late-night hours—may lead to concerns about behavioral and mental health. Existing predictive systems fail to account for the nuanced interplay between users’ internal psychological states and their surrounding ecological contexts.
Objective
This study aims to develop a psychologically and ecologically informed behavior prediction model to identify high-risk patterns of digital media usage and support early-stage intervention strategies.
Methods
We propose a Dual-Channel Cross-Attention Network (DCCAN) architecture composed of three layers: signal identification (for psychological and ecological encoding), interaction modeling (via cross-modal attention), and behavior prediction. The model was trained and tested on a dataset of 9,782 users and 51,264 behavior sequences, annotated with labels for immersive usage, late-night activity, and susceptibility to health misinformation.
Results
The DCCAN model achieved superior performance across all three tasks, especially in immersive usage prediction (F1-score: 0.891, AUC: 0.913), outperforming LSTM, GRU, and XGBoost baselines. Ablation studies confirmed the critical role of both psychological and ecological signals, as well as the effectiveness of the cross-attention mechanism.
Conclusions
Incorporating psychological and ecological modalities through attention-based fusion yields interpretable and accurate predictions for digital risk behaviors. This framework shows promise for scalable, real-time behavioral health monitoring and adaptive content moderation on media platforms.