A deep learning approach for predicting sensory-motor integration in postural stability.

To maintain postural stability, reliable sensory input and appropriate motor responses are essential. In this study, a novel deep learning approach was proposed to evaluate postural stability by estimating sensory input capabilities (equilibrium score) and corresponding motor responses (movement strategy) from standing center of pressure (COP) signals. The model inputs were COP signals from quiet standing, with short-time Fourier transform (STFT) and wavelet transform (WT) employed as feature extraction methods. The models included GoogleNet and residual attention network (RAN) architectures, with feature distribution smoothing (FDS) applied to mitigate performance degradation caused by class imbalance in the target data. The starting point and bin number for FDS were incorporated as design parameters, with the mean squared error serving as the cost function during the hyperparameter optimization process. The STFT + RAN + FDS model achieved average errors of 0.89 for the equilibrium score and 0.72 for the movement strategy, demonstrating improvements of approximately 48% and 38% compared to the WT + RAN and STFT + RAN models, respectively. The proposed approach effectively predicts postural stability by leveraging COP signal features and advanced model architectures. It holds potential as a cost-effective and accessible solution for addressing postural stability issues, particularly for the growing population of patients with dizziness disorders.