A hybrid CNN-machine learning model for anxiety screening with mandala coloring patterns: Feature extraction and classification performance

Traditional anxiety assessments typically rely on self-report questionnaires and often emphasize somatic symptoms, potentially overlooking nuanced behavioral indicators. This study presents a hybrid convolutional neural network (CNN) and machine learning (ML) model to classify visual features of coloring accuracy, color transitions, color entropy, and symmetry from 1044 digitized mandala images. Under stratified 5-fold cross-validation, the Random Forest (RF) classifier achieved an overall accuracy of 84.56%. However, performance under class-imbalanced conditions was limited, with a sensitivity of 0.216 for the high-anxiety group. These findings suggest that quantifiable visual features may function as exploratory visual markers to support anxiety severity classification rather than as a standalone diagnostic tool. This approach provides preliminary evidence for integrating computational modeling with art-based methods as an exploratory assessment support framework.