Scholarlink

<p>Depression represents a leading global health burden affecting over 280 million people worldwide, yet traditional diagnostic approaches rely heavily on subjective assessments susceptible to bias and recall limitations. Wearable technology and actigraphy-based monitoring offer objective alternatives for continuous depression assessment, though current methodologies predominantly employ daily aggregation strategies that potentially discard critical temporal information about disruptions in activity patterns characteristic of mood disorders. This study presents a systematic evaluation of temporal segmentation strategies for actigraphy-based depression classification, comparing full-day, 2-segment, and 3-segment divisions using machine learning approaches. This study utilized the <i>Depresjon</i> dataset containing motor activity recordings from 55 participants including 23 patients with depression and 32 healthy controls. Activity intensity and variability features were extracted from the segments and five distinct classification models across multiple temporal configurations were implemented and cross-validated. Results demonstrate that temporal segmentation enhances classification performance compared to traditional full-day aggregation approaches, with the 3-segment division achieving optimal discrimination using standard stratified cross-validation (F1-score: 0.830). Subject-wise group cross-validation, which provides a more clinically realistic estimate for generalization to unseen study participants, yielded an F1-score of 0.744. These findings establish temporal segmentation as a fundamental component to take into account for optimizing digital biomarkers in depression classification, providing baselines for the development of wearable device and machine-learning-based monitoring systems for potential clinical applications.</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

An evaluation of motor activity time-segmentation approaches for depression classification using machine learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月6日

查看全文
  • Cristóbal Estrada-Salinas,
  • Enrique Garcia-Ceja

摘要

Depression represents a leading global health burden affecting over 280 million people worldwide, yet traditional diagnostic approaches rely heavily on subjective assessments susceptible to bias and recall limitations. Wearable technology and actigraphy-based monitoring offer objective alternatives for continuous depression assessment, though current methodologies predominantly employ daily aggregation strategies that potentially discard critical temporal information about disruptions in activity patterns characteristic of mood disorders. This study presents a systematic evaluation of temporal segmentation strategies for actigraphy-based depression classification, comparing full-day, 2-segment, and 3-segment divisions using machine learning approaches. This study utilized the Depresjon dataset containing motor activity recordings from 55 participants including 23 patients with depression and 32 healthy controls. Activity intensity and variability features were extracted from the segments and five distinct classification models across multiple temporal configurations were implemented and cross-validated. Results demonstrate that temporal segmentation enhances classification performance compared to traditional full-day aggregation approaches, with the 3-segment division achieving optimal discrimination using standard stratified cross-validation (F1-score: 0.830). Subject-wise group cross-validation, which provides a more clinically realistic estimate for generalization to unseen study participants, yielded an F1-score of 0.744. These findings establish temporal segmentation as a fundamental component to take into account for optimizing digital biomarkers in depression classification, providing baselines for the development of wearable device and machine-learning-based monitoring systems for potential clinical applications.