Skjødt Mathias, Brønd Jan Christian, Tully Mark A, Tsai Li-Tang, Koster Annemarie, Visser Marjolein, Caserotti Paolo
Department of Sports Science and Clinical Biomechanics, University of Southern Denmark, Odense, Denmark.
Department of Sports Science and Clinical Biomechanics, Center for Active and Healthy Ageing (CAHA), University of Southern Denmark, Odense, Denmark.
Scand J Med Sci Sports. 2025 Jan;35(1):e70009. doi: 10.1111/sms.70009.
Physical activity (PA) reduces the risk of negative mental and physical health outcomes in older adults. Traditionally, PA intensity is classified using METs, with 1 MET equal to 3.5 mL O·min·kg. However, this may underestimate moderate and vigorous intensity due to age-related changes in resting metabolic rate (RMR) and VOmax. VOreserve accounts for these changes. While receiver operating characteristics (ROC) analysis is commonly used to develop PA, intensity cut-points, machine learning (ML) offers a potential alternative. This study aimed to develop ROC cut-points and ML models to classify PA intensity in older adults. Sixty-seven older adults performed activities of daily living (ADL) and two six-minute walking tests (6-MWT) while wearing six accelerometers on their hips, wrists, thigh, and lower back. Oxygen uptake was measured. ROC and ML models were developed for ENMO and Actigraph counts (AGVMC) using VOreserve as the criterion in two-third of the sample and validated in the remaining third. ROC-developed cut-points showed good-excellent AUC (0.84-0.93) for the hips, lower back, and thigh, but wrist cut-points failed to distinguish between moderate and vigorous intensity. The accuracy of ML models was high and consistent across all six anatomical sites (0.83-0.89). Validation of the ML models showed better results compared to ROC cut-points, with the thigh showing the highest accuracy. This study provides ML models that optimize the classification of PA intensity in very old adults for six anatomical placements hips (left/right), wrist (dominant/non-dominant), thigh and lower back increasing comparability between studies using different wear-position. Clinical Trial Registration: clinicaltrials.gov identifier: NCT04821713.
身体活动（PA）可降低老年人出现负面身心健康结果的风险。传统上，PA强度是使用代谢当量（METs）来分类的，1个MET等于3.5毫升氧气·分钟·千克。然而，由于静息代谢率（RMR）和最大摄氧量（VOmax）随年龄变化，这可能会低估中等强度和剧烈强度的活动。VO储备考虑了这些变化。虽然常用受试者工作特征（ROC）分析来制定PA强度切点，但机器学习（ML）提供了一种潜在的替代方法。本研究旨在开发ROC切点和ML模型，以对老年人的PA强度进行分类。67名老年人在髋部、手腕、大腿和下背部佩戴六个加速度计的情况下进行日常生活活动（ADL）和两项六分钟步行测试（6-MWT）。测量了摄氧量。使用VO储备作为标准，在三分之二的样本中为等效噪声加速度（ENMO）和活动记录仪计数（AGVMC）开发了ROC和ML模型，并在其余三分之一的样本中进行了验证。ROC开发的切点在髋部、下背部和大腿处显示出良好至优秀的曲线下面积（AUC，0.84 - 0.93），但手腕切点未能区分中等强度和剧烈强度。ML模型在所有六个解剖部位的准确性都很高且一致（0.83 - 0.89）。与ROC切点相比，ML模型的验证结果更好，大腿部位的准确性最高。本研究提供了ML模型，该模型针对六个解剖位置（髋部（左/右）、手腕（优势手/非优势手）、大腿和下背部）优化了对非常老年人PA强度的分类，提高了使用不同佩戴位置的研究之间的可比性。临床试验注册：clinicaltrials.gov标识符：NCT04821713。
