In human–machine interaction, continuous assessment of attentional states enables the development of responsive and adaptive interfaces. However, current approaches often rely on subjective or intrusive measures, limiting their applicability in real-world systems. To address this, we focus on mindful attention, a cognitive state characterized by present-moment awareness and nonreactive focus, which reflects stable, internally regulated attention during interactive tasks. While most studies assess mindfulness through self-report questionnaires, behavioural metrics can provide objective, complementary insights. This study introduces a framework that integrates Trail Making Test (TMT) performance, dynamic mouse movement features, and self-reported mindful attention (MAAS) scores. Mouse trajectory features such as velocity, acceleration, and curvature were extracted and analyzed using multiple regression models, combined with different feature selection strategies. Results showed that wrapper-based and Boruta-style methods substantially outperformed filter-based approaches, with progressive feature selection yielding the highest accuracy. The Gradient Boosting Regressor with progressive selection achieved a strong predictive performance (R^2score = 0.85 on test data), demonstrating that mouse dynamics can serve as reliable behavioural indicators of mindfulness. These findings highlight the potential of integrating behavioural features and machine learning for multidimensional attention assessment.