Validity of an Automatic Approach for Acceleration Signal Segmentation during Weight Lifting

The increasing adoption of wearable sensors in biomechanics has enhanced the acquisition and analysis of physiological signals, enabling continuous real-time monitoring in naturalistic settings. In this context, automatic segmentation methods have emerged as efficient alternatives to manual segmentation, helping to overcome limitations such as time consumption and subjective variability. This study assessed the validity of an automatic segmentation approach for acceleration signals recorded during load-lifting tasks, comparing it with manual segmentation. Four healthy female participants (aged 20 to 30) each performed 20 load-lifting tasks, conducted under safe conditions in accordance with the guidelines of the National Institute for Occupational Safety and Health. An inertial sensor positioned on the lumbar region recorded acceleration data along the vertical axis. The automatic segmentation approach consisted of a subject-specific threshold algorithm, which used a 4th-order Butterworth filter and Savitzky-Golay smoothing to detect lifting events. For comparison, three time-domain features - rectified signal area, root mean square, and peak-to-peak amplitude - were extracted from both manual and automatic segmentations. Statistical analyses revealed strong agreement between manual and automatic segmentations for rectified signal area (p=0.535; bias =0.731, CI: -2.689 to 1.226) and peak-to-peak amplitude (p=0.371; bias =0.003, CI: -0.002 to 0.008), with no observable trends or systematic errors. However, root mean square values exhibited a systematic constant error (p < 0.001; bias CI: 0.003 to 0.007), with the automatic method consistently overestimating compared to manual segmentation. These findings underscore the potential of the proposed automatic segmentation method as a fast, accurate, and effective tool for signal analysis, although further refinements are needed to improve its generalizability.