Introduction: Neurodegenerative diseases (NDDs) present substantial challenges due to their impact on movement, emphasizing the critical role of biomedical engineering research in clinical diagnosis. Measuring the biomechanical properties of gait during walking can provide valuable insights into the movement pattern of NDDs and has great promise for developing non-invasive automated NDD classification techniques. 
Methods: Based on the GaitNDD database, two experimental trials were conducted on healthy controls (HCs) and three NDDs: Parkinson disease (PD), amyotrophic lateral sclerosis (ALS), and Huntington disease (HD), showcasing a comprehensive analysis of 1-dimensional and 2-dimensional force gait features. In the first trial, two time-frequency feature sequences were extracted from right, left, and combined feet during a walking task, feeding a bidirectional long short-term memory (BiLSTM) network. The second trial involves constructing spectrogram images of the gait signal as input for 3 popular pre-trained convolutional neural networks (CNNs): AlexNet, GoogLeNet, and VGG16. 
Results: VGG16 emerges as the standout performer, achieving a remarkable accuracy of 99.91%, sensitivity of 99.93%, and specificity of 99.97% for automatic 4-class NDD detection using high-level features from the right foot gait signal. BiLSTM performance significantly improved when fed with VGG16-extracted high-level features, surpassing hand-crafted features.
Conclusion: The study underscores the superiority of CNNs, particularly VGG16, in extracting high-level features from spectrogram-derived vertical ground reaction force (vGRF) signals for robust NDD classification. The hybrid VGG16-BiLSTM approach demonstrates enhanced performance, affirming the synergistic benefits of combining deep learning techniques. Overall, the CNN high-level features derived from vGRF signal spectrograms provide valuable insights into NDD groups, offering a promising avenue for understanding diverse mechanisms underlying gait-related conditions.