IMU-based Pelvic Rotation Detection: Insights From Machine Learning and RoM Analysis

Background and Aim: The integration of wearables and machine learning (ML) has advanced human activity recognition (HAR), allowing robust detection of the common activities of daily life such as sitting, standing, and walking. While excelling in HAR is essential for daily health monitoring, current solutions often overlook subtle but clinically significant movements like pelvic rotation (PR), which plays an important role in posture control and is essential for low back pain (LBP) prevention. This study explores using wearable IMUs and ML to classify and analyze PR in different directions and ranges of motion (RoM). Methods: Motion data were collected from 39 sedentary participants (21 females, 18 males), including 5h23m31s of valid recordings sampled at 60Hz using Xsens, an IMU-based Motion Capture System. Participants performed sitting, standing, walking, 5-time sit-to-stand-to-sit (5TSTS), and anterior and posterior PR during sitting and standing. Data were proceeded into 159510 windows (0.5s length, 0.1s stride). Six classical ML classifiers were applied to classify PR across tasks. To complement the ML analysis, PR RoM was extracted from the “Ergonomic Joint Angles ZXY-Vertical_Pelvis Flexion/Extension” feature. For each repetition, the difference between anterior and posterior PR was computed by peak detection, with maximum values extracted as the RoM per posture and subject. Results: XGBoost, the best-performing classifier, achieved a class-specific F1 score of 0.965 for PR in binary classification (PR vs. the other activities), demonstrating the effectiveness of the approach. When extended to 3 classes (anterior PR, posterior PR, and other activities), the class-specific F1 scores of 0.823 for anterior PR and 0.793 for posterior PR. This reduction highlights the added complexity of differentiating between PR directions. We also observed a significant directional difference. During sitting, the maximum absolute angle of anterior PR from the neutral (vertical) position was 18.0° ± 8.3° vs. 13.5° ± 4.5° for posterior PR (p= 0.004); during standing, the anterior was 17.5° ± 5.8° compared to the posterior at 10.3° ± 3.8° (p < 0.001). The larger anterior angles resulted in a greater number of windows classified as being in anterior postures, regardless of the actual direction of PR, making identifying anterior PR relatively easier than posterior PR. Additionally, RoM (max anterior to max posterior) between sitting (30.0° ± 9.7°) and standing (24.1° ± 6.8°) revealed considerable variability across participants. Conclusions: This study demonstrates the feasibility of IMU-based ML models to recognize PR. The observed directional differences from the vertical orientation provide insights into the biomechanical and anatomical factors influencing PR. These findings pave the way for future work aiming to integrate precise PR detection into prolonged posture monitoring, with promising applications in LBP prevention and personalized daily healthcare solutions.