Kanchapogu Naga Raju, Nandan Mohanty Sachi
School of Computer Science & Engineering (SCOPE), VIT-AP University, Amaravati, Andhra Pradesh, India.
Dialogues Clin Neurosci. 2025 Dec;27(1):16-35. doi: 10.1080/19585969.2025.2524337. Epub 2025 Jun 30.
Depression, including Bipolar and Unipolar types, is a widespread mental health issue. Conventional diagnostic methods rely on subjective assessments, leading to possible underreporting and bias. Machine learning (ML) and deep learning (DL) offer automated approaches to detect depression using behavioral and demographic data.
This study proposes a hybrid AI framework combining structured demographic features with synthetic actigraph time-series data. Demographic data is modeled using an XGBoost ensemble, while temporal data is analyzed through a deep convolutional neural network (CNN). The training pipeline includes stratified k-fold cross-validation, hyperparameter tuning, and statistical testing. Model explainability is enhanced using SHAP (XGBoost) and Grad-CAM (CNN).
The hybrid model demonstrated strong classification performance across metrics like accuracy, sensitivity, and specificity. Integrating temporal and static features improved prediction of Bipolar and Unipolar Depression. Interpretability tools revealed key features and time patterns influencing predictions.
This work introduces a robust and interpretable framework for depression classification using synthetic multimodal data. While not clinically validated, the model serves as a methodological foundation for future research with real-world datasets.
抑郁症，包括双相情感障碍和单相抑郁症，是一个普遍存在的心理健康问题。传统的诊断方法依赖主观评估，可能导致报告不足和偏差。机器学习（ML）和深度学习（DL）提供了利用行为和人口统计学数据检测抑郁症的自动化方法。
本研究提出了一种将结构化人口统计学特征与合成活动记录仪时间序列数据相结合的混合人工智能框架。人口统计学数据使用XGBoost集成模型进行建模，而时间数据则通过深度卷积神经网络（CNN）进行分析。训练流程包括分层k折交叉验证、超参数调整和统计测试。使用SHAP（XGBoost）和Grad-CAM（CNN）增强模型的可解释性。
混合模型在准确率、灵敏度和特异性等指标上表现出强大的分类性能。整合时间和静态特征提高了对双相情感障碍和单相抑郁症的预测能力。可解释性工具揭示了影响预测的关键特征和时间模式。
这项工作引入了一个使用合成多模态数据进行抑郁症分类的强大且可解释的框架。虽然该模型尚未经过临床验证，但它为未来使用真实世界数据集的研究奠定了方法基础。
