Anatomy-Based Assessment of Spinal Posture Using IMU Sensors and Machine Learning
Özet
Highlights What are the main findings? IMU-based posture angles were used to derive proxy deviation labels relative to reference ranges from the literature, and machine learning models predicted these labels using demographic and anthropometric variables. Daily posture habits, particularly prolonged desk and smartphone usage, significantly influence deviations in cervical lordosis, thoracic kyphosis, and lumbar lordosis. What is the implication of the main finding? The IMU-based monitoring of spinal posture may motivate future studies on preventive health strategies, but current findings remain exploratory and underpowered. Integrating sensor-based posture monitoring into daily life and occupational settings could be explored in future research, while emphasizing that it is not yet a validated diagnostic tool.Highlights What are the main findings? IMU-based posture angles were used to derive proxy deviation labels relative to reference ranges from the literature, and machine learning models predicted these labels using demographic and anthropometric variables. Daily posture habits, particularly prolonged desk and smartphone usage, significantly influence deviations in cervical lordosis, thoracic kyphosis, and lumbar lordosis. What is the implication of the main finding? The IMU-based monitoring of spinal posture may motivate future studies on preventive health strategies, but current findings remain exploratory and underpowered. Integrating sensor-based posture monitoring into daily life and occupational settings could be explored in future research, while emphasizing that it is not yet a validated diagnostic tool.Abstract Background: This study used inertial measurement unit (IMU)-based posture angle estimates to define proxy risk labels and investigated whether these labels can be predicted from demographic, anthropometric, and lifestyle variables through machine learning analysis. Methods: Thirty healthy individuals aged 18-25 years were included. Demographic and anthropometric data and information on daily living activities were collected. The IMU sensors were placed at vertebral levels C1, C7, T5, T12, and L5. Participants were instructed to stand in an upright posture, followed by a relaxed daily posture. Anatomic postural changes between these positions were analyzed. Cervical lordosis, thoracic kyphosis, lumbar lordosis, and scoliosis risks were predicted using machine learning algorithms, including Random Forest (RF) and Artificial Neural Networks (ANN). Results: Incorrect postures during desk work and phone use were associated with an increased likelihood of posture-related deviations, such as cervical lordosis, thoracic kyphosis, and lumbar lordosis. Conversely, daily physical activity reduced these deviations. Using LOSO and stratified cross-validation with imbalance handling, balanced accuracies ranged between 0.55 and 0.82 across targets, with majority-class baselines between 0.53 and 0.87. For cervical lordosis risk, RF achieved a 0.82 balanced accuracy (95% CI: 0.74-0.97), while other categories showed a moderate but consistent performance. AUPRC values exceeded baseline levels across all models. Conclusions: IMU-based posture angle estimates can be used to identify posture-related risk categories. In this study, ML models have demonstrated predictive relationships with demographic, anthropometric, and lifestyle variables. These findings provide exploratory evidence based on IMU-derived proxy labels in a small cohort of healthy young adults. They represent exploratory indicators of postural deviation rather than clinical outcomes and may motivate future studies on preventive strategies. Importantly, the results remain underpowered relative to the a priori power targets and should be interpreted qualitatively.
