Prediction of post-Schroth Cobb angle changes in adolescent idiopathic scoliosis patients based on neural networks and surface electromyography

Abstract

Introduction: To develop a temporal-convolutional-LSTM (TCN-LSTM) hybrid model integrating surface electromyography (sEMG) signals for forecasting post-Schroth Cobb angle progression in adolescent idiopathic scoliosis (AIS) patients, thereby offering accurate feedback for personalized treatment.

Methodology: A total of 143 AIS patients were included. A systematic Schroth exercise training program was designed. sEMG data from specific muscles and Cobb angle measurements were collected. A neural network model integrating Temporal Convolutional Network (TCN), Long Short-Term Memory (LSTM) layers, and feature vectors was constructed. Four prediction models were compared: TCN-LSTM hybrid model, TCN, LSTM, and Support Vector Regression (SVR).

Results: The TCN-LSTM hybrid model demonstrated superior performance, with Cobb angle-Thoracic (Cobb Angle-T) prediction accuracy reaching R² = 0.63 (baseline) and 0.69 (Week 24), achieving overall R² = 0.74. For Cobb angle-Lumbar (Cobb Angle-L), accuracy was R² = 0.61 (baseline) and 0.65 (Week 24), with overall R² = 0.73. The SVR model showed lowest performance (R² < 0.12).

Conclusion: The TCN-LSTM hybrid model can precisely predict Cobb angle changes in AIS patients during Schroth exercises, especially in long-term predictions. It provides real-time feedback for clinical treatment and contributes to optimizing treatment plans, presenting a novel prediction approach and reference basis for evaluating the effectiveness of Schroth correction exercises in AIS patients.

Keywords: Cobb angle; Schroth exercises; TCN-LSTM hybrid model; adolescent idiopathic scoliosis (AIS); neural networks; surface electromyography (SEMG).

FIGURE 1

Schroth exercise intervention program. (A): Quadruped position exercise; (B): Squatting on the bar exercise; (C): Unilateral kneeling position exercise; (D): Lateral flexion sitting position exercise.

FIGURE 2

Placement of the EMG.

FIGURE 3

Architecture of neural networks.

FIGURE 4

Selection of predictive variables.

FIGURE 5

Mean value of R-squared for cross-validation of different models.

FIGURE 6

Performance of TCN+LSTM for predicting Cobb Angle-T.

FIGURE 7

Performance of TCN+LSTM for predicting Cobb Angle-L.

FIGURE 8

Scatter plots of cross - validation for predicting Cobb Angles using different models.

FIGURE 9

Training time for different models to predict Cobb angle.