Gait phase recognition of children with cerebral palsy via deep learning based on IMU data from a soft ankle exoskeleton

Accurate gait-phase identification in children with Cerebral Palsy (CP) constitutes a pivotal prerequisite for evidence-based rehabilitation. Addressing the precise detection of gait disturbances under natural ambulation, we propose a deep-learning framework that integrates a stacked denoising autoencoder (SDA) with a long short-term memory network (SDA–LSTM) to classify four canonical gait phases. A community-oriented dataset was constructed by synchronizing ankle-mounted inertial measurement units (IMU) with plantar-pressure insoles; natural gait sequences of six children with mild CP were acquired in open environments. The SDA layer robustly extracts discriminative representations from non-stationary, high-noise signals, whereas the LSTM module models inter-phase temporal dependencies, thereby enhancing generalization cross-user. In noise-free conditions the SDA–LSTM framework attained 97.83% accuracy, significantly exceeding SVM (94.68%), random forest (96.05%), and standalone LSTM (95.86%). Under additive Gaussian noise with SNR ranging from 5 to 30 dB, the model preserved stable performance; at 10 dB SNR (Signal-to-Noise Ratio), accuracy remained 90.96%, corroborating its exceptional robustness. These findings demonstrate that SDA–LSTM effectively handles the complex, heterogeneous gait patterns of children with CP and is readily deployable for clinical assessment and exoskeletal assistance systems, indicating substantial translational potential.