Scholarlink

<p>This study aimed to develop and evaluate a hybrid machine learning (ML) model for screening depression using visual features from pre-drawn structured mandala colorings to complement the limitations of conventional self-report depression assessments. The Beck Depression Inventory-II (BDI-II) scores of 1044 Korean adults (M age = 44.7, SD = 19.7; 59.8% female) were used to label mandala coloring images for depression severity. EfficientNet-B1 convolutional neural networks (CNNs) were used to extract four visual features: Black Color Ratio, Color Entropy, Completeness, and Core Completeness. These features were then categorized using conventional ML algorithms (Support Vector Machine, Logistic Regression, Random Forest). Accuracy, precision, recall, F1-score, and Pearson correlations between features and BDI-II scores were used to evaluate the performance. The Random Forest achieved the best performance with an accuracy of 81.5% and an F1-score of 75.9%. While Color Entropy displayed a weak negative correlation (<i>r</i> = − .065, <i>p</i> = .035), the Black Color Ratio showed a significant positive correlation with depression severity (<i>r</i> = .115, <i>p</i> &lt; .001). There were negative trends toward significance for Completeness and Core Completeness. These results suggest that visual characteristics of mandala colorings may provide potential cues of depressive symptoms. The proposed hybrid CNN-ML approach must be considered a preliminary, supplementary, and auxiliary screening tool rather than a diagnostic instrument. Future studies incorporating clinical diagnostic measures and more diverse populations are needed to establish its broader clinical utility.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Machine learning-based depression screening model from pre-drawn structured mandala colorings

  • 期刊论文

  • 发表日期：

    2026年8月3日

查看全文
  • Se-Ryun Park,
  • JongHan Kim,
  • Ujin Jeon,
  • Chaehee Park,
  • Sang-Shin Lee,
  • Jinyoung Han,
  • Yu-Jung Cha

摘要

This study aimed to develop and evaluate a hybrid machine learning (ML) model for screening depression using visual features from pre-drawn structured mandala colorings to complement the limitations of conventional self-report depression assessments. The Beck Depression Inventory-II (BDI-II) scores of 1044 Korean adults (M age = 44.7, SD = 19.7; 59.8% female) were used to label mandala coloring images for depression severity. EfficientNet-B1 convolutional neural networks (CNNs) were used to extract four visual features: Black Color Ratio, Color Entropy, Completeness, and Core Completeness. These features were then categorized using conventional ML algorithms (Support Vector Machine, Logistic Regression, Random Forest). Accuracy, precision, recall, F1-score, and Pearson correlations between features and BDI-II scores were used to evaluate the performance. The Random Forest achieved the best performance with an accuracy of 81.5% and an F1-score of 75.9%. While Color Entropy displayed a weak negative correlation (r = − .065, p = .035), the Black Color Ratio showed a significant positive correlation with depression severity (r = .115, p < .001). There were negative trends toward significance for Completeness and Core Completeness. These results suggest that visual characteristics of mandala colorings may provide potential cues of depressive symptoms. The proposed hybrid CNN-ML approach must be considered a preliminary, supplementary, and auxiliary screening tool rather than a diagnostic instrument. Future studies incorporating clinical diagnostic measures and more diverse populations are needed to establish its broader clinical utility.