Deep learning based screening and regular assessment of adolescent idiopathic scoliosis using wearable IMU sensors.

OBJECTIVE

The study's objective is to propose a novel non-invasive method for rapid screening and regular assessment of adolescent idiopathic scoliosis (AIS) through development of a wearable system integrated with multiple inertial measurement units (IMUs) and deep learning models. The system is designed to automatically distinguish between healthy individuals and AIS patients, and subsequently predict the Cobb angle based on continuous temporal kinematic angular sequences acquired during gait.

METHODS

Gait kinematic data were acquired from 124 participants (104 patients with average Cobb angle of 21.62 ± 7.93° and 20 healthy subjects) using a 9-IMU wearable device. Extracted angular features were analyzed to quantify bilateral asymmetry, compare differences across severity subgroups, and evaluate their linear correlation with Cobb angles. A two-stage deep learning framework was implemented with a convolutional neural network (CNN) classification model developed for the rapid screening of scoliosis, and a CNN-Transformer model was designed and compared with other five model architectures to predict Cobb angles from the acquired temporal angular sequences.

RESULTS

Scapular kinematics emerged as the most prominent marker of asymmetry, and knee joint kinematics served as the strongest indicator of severity. Meanwhile, angular features of knee, hip and ankle joints demonstrated weak negative linear correlations with Cobb angle. In addition, the scoliosis screening model achieved high predictive performance, with an accuracy of 96.59% and a precision-recall AUC of 0.94 for scoliosis detection. For Cobb angle prediction, the CNN-Transformer model regularized with Gaussian noise during training proved most effective, yielding a mean absolute error of 2.14 ± 0.28° and R 2 value of 0.85 ± 0.03, outperforming other architectural alternatives evaluated in this study.

CONCLUSION

Kinematic analysis of angular data validated the efficacy of the wearable system and effectively captured gait characteristics specific to AIS. The deep learning models accurately distinguished scoliosis patients from healthy cases and predicted Cobb angles using temporal kinematic angular sequences, providing a safe, non-invasive, operator-friendly approach suitable for rapid screening and regular assessment.