Development of a Machine Learning Model for the Estimation of Hip and Lumbar Angles in Ballet Dancers.

Abstract

Accurate field-based assessment of dance kinematics is important to understand the etiology, and thus prevention and management, of hip and back pain. The study objective was to develop a machine learning model to estimate thigh elevation and lumbar sagittal plane angles during ballet leg lifting tasks, using wearable sensor data. Female dancers (n=30) performed ballet-specific leg lifting tasks to the front, side, and behind the body. Dancers wore six wearable sensors (100 Hz). Data were simultaneously collected using an 18-camera motion analysis system (250 Hz). Due to synchronization and hardware malfunction issues, only 23 dancers had usable data. Using leave-one-out cross-validation, machine learning models were compared with the optic motion capture system using root mean square error (RMSE) in degrees and correlation coefficients (r) over the complete movement profile of each leg lift and mean absolute error (MAE) and Bland Altman plots for peak angle accuracy. The average RMSE for model estimation was 6.8° for thigh elevation angle and 5.6° for lumbar spine sagittal plane angle, with respective MAE of 6.3°and 5.7°. There was a strong correlation between the machine learning model and optic motion capture for peak angle values (thigh r=0.86, lumbar r=0.96). The models developed demonstrated an acceptable degree of accuracy for the estimation of thigh elevation angle and lumbar spine sagittal plane angle during dance-specific leg lifting tasks. This provides potential for a near-real-time, field-based measurement system.