Scholarlink

Background <p>Falls are a leading cause of disability and mortality in older adults. The 30-second sit-to-stand test (30STS) traditionally relies on repetition counts, limiting its ability to capture subtle postural control deficits. We applied computer vision and machine learning to extend this test into a multidimensional assessment framework for functional status discrimination.</p> Methods <p>Thirty-two community-dwelling older adults (18 normal, 14 weak) performed two 30STS tests at baseline (June–July 2022) and follow-up (January–February 2023). Smartphone videos were processed using Keypoint R-CNN to extract 17 joint coordinates. Center of mass (CoM), kinematic parameters, and multiscale sample entropy (MSE) were derived. Paired t-tests assessed within-group changes; logistic regression, random forests, and support vector machines (SVM) evaluated classification using repeated stratified five-fold cross-validation.</p> Results <p>CoM extremal and shape metrics (maximum, minimum, skewness, kurtosis) were more sensitive than mean values, revealing transient instability and compensatory strategies in weak participants (<i>p</i> &lt; 0.05). Kinematic analysis showed consistent control in normal participants, whereas weak participants adopted conservative amplitude-reducing strategies (e.g., acc-y CV, <i>p</i> = 0.022). MSE values suggested increased horizontal temporal complexity in the normal group, whereas reduced complexity was observed in the weak group, which may reflect altered movement coordination patterns. Random forests achieved the highest cross-validated performance within this cohort (AUC = 0.993); logistic regression was consistently strong (AUC ≈ 0.93, CA ≈ 0.90); SVM improved to high accuracy (AUC = 0.961, CA = 0.886).</p> Conclusions <p>Smartphone-based pose estimation expanded the 30STS from a repetition count to a multidimensional assessment, providing additional quantitative information beyond repetition count. This approach shows promise for longitudinal monitoring, community screening, and remote care, with machine learning models demonstrating preliminary feasibility within this cohort.</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

Quantitative assessment of the 30-second sit-to-stand test using computer vision technology for physical therapists

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月24日

查看全文
  • Chin-Ying Hsu,
  • Ying-Po Hsu,
  • Tien-Lung Sun,
  • Chia-Hsuan Lee

摘要

Background

Falls are a leading cause of disability and mortality in older adults. The 30-second sit-to-stand test (30STS) traditionally relies on repetition counts, limiting its ability to capture subtle postural control deficits. We applied computer vision and machine learning to extend this test into a multidimensional assessment framework for functional status discrimination.

Methods

Thirty-two community-dwelling older adults (18 normal, 14 weak) performed two 30STS tests at baseline (June–July 2022) and follow-up (January–February 2023). Smartphone videos were processed using Keypoint R-CNN to extract 17 joint coordinates. Center of mass (CoM), kinematic parameters, and multiscale sample entropy (MSE) were derived. Paired t-tests assessed within-group changes; logistic regression, random forests, and support vector machines (SVM) evaluated classification using repeated stratified five-fold cross-validation.

Results

CoM extremal and shape metrics (maximum, minimum, skewness, kurtosis) were more sensitive than mean values, revealing transient instability and compensatory strategies in weak participants (p < 0.05). Kinematic analysis showed consistent control in normal participants, whereas weak participants adopted conservative amplitude-reducing strategies (e.g., acc-y CV, p = 0.022). MSE values suggested increased horizontal temporal complexity in the normal group, whereas reduced complexity was observed in the weak group, which may reflect altered movement coordination patterns. Random forests achieved the highest cross-validated performance within this cohort (AUC = 0.993); logistic regression was consistently strong (AUC ≈ 0.93, CA ≈ 0.90); SVM improved to high accuracy (AUC = 0.961, CA = 0.886).

Conclusions

Smartphone-based pose estimation expanded the 30STS from a repetition count to a multidimensional assessment, providing additional quantitative information beyond repetition count. This approach shows promise for longitudinal monitoring, community screening, and remote care, with machine learning models demonstrating preliminary feasibility within this cohort.