<p>Neurodegenerative diseases (NDs) such as Parkinson’s disease (PD), Huntington’s disease (HD), and Amyotrophic Lateral Sclerosis (ALS) induce complex alterations in gait dynamics, making reliable gait-based discrimination challenging, particularly in small and highly imbalanced datasets. We propose a hybrid generative–discriminative framework combining a class-specific Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) and a CNN-LSTM architecture for spatiotemporal gait classification. The class-conditional augmentation strategy mitigates data imbalance by generating distribution-consistent synthetic gait signals,while the CNN-LSTM captures both local gait patterns and long-range temporal dependencies.Experiments conducted using a 5-fold Group Cross-Validation protocol on the PhysioNet Gait in Neurodegenerative Diseases dataset (Healthy Controls, PD, HD, and ALS) demonstrate consistent performance improvements. GAN-based augmentation increased accuracy from <InlineEquation ID="IEq1"> <EquationSource Format="TEX">$94.62 \pm 1.34\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>94.62</mn> <mo>±</mo> <mn>1.34</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> to <InlineEquation ID="IEq2"> <EquationSource Format="TEX">$97.48 \pm 0.96\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>97.48</mn> <mo>±</mo> <mn>0.96</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> for HC vs. PD, from <InlineEquation ID="IEq3"> <EquationSource Format="TEX">$95.91 \pm 1.22\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>95.91</mn> <mo>±</mo> <mn>1.22</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> to <InlineEquation ID="IEq4"> <EquationSource Format="TEX">$98.36 \pm 0.72\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>98.36</mn> <mo>±</mo> <mn>0.72</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> for HC vs. ALS, and from <InlineEquation ID="IEq5"> <EquationSource Format="TEX">$91.24 \pm 1.52\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>91.24</mn> <mo>±</mo> <mn>1.52</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> to <InlineEquation ID="IEq6"> <EquationSource Format="TEX">$95.18 \pm 1.04\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>95.18</mn> <mo>±</mo> <mn>1.04</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> in the four-class classification task. Signal-level and classification-based analyses confirm that the generated samples are statistically consistent with real gait data and contribute to improved model robustness. These results highlight the potential of generative deep learning as a data-efficient support tool for gait-based analysis of neurodegenerative diseases, particularly in small and imbalanced clinical datasets.</p>

A generative–discriminative deep learning framework for multi-class gait analysis in neurodegenerative disorders

期刊论文
发表日期：

2026年5月21日

查看全文

Sana Trigui,
Hala Bezine,
Basant Agarwal

摘要

Neurodegenerative diseases (NDs) such as Parkinson’s disease (PD), Huntington’s disease (HD), and Amyotrophic Lateral Sclerosis (ALS) induce complex alterations in gait dynamics, making reliable gait-based discrimination challenging, particularly in small and highly imbalanced datasets. We propose a hybrid generative–discriminative framework combining a class-specific Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) and a CNN-LSTM architecture for spatiotemporal gait classification. The class-conditional augmentation strategy mitigates data imbalance by generating distribution-consistent synthetic gait signals,while the CNN-LSTM captures both local gait patterns and long-range temporal dependencies.Experiments conducted using a 5-fold Group Cross-Validation protocol on the PhysioNet Gait in Neurodegenerative Diseases dataset (Healthy Controls, PD, HD, and ALS) demonstrate consistent performance improvements. GAN-based augmentation increased accuracy from $94.62 \pm 1.34\%$ $94.62 \pm 1.34 %$ to $97.48 \pm 0.96\%$ $97.48 \pm 0.96 %$ for HC vs. PD, from $95.91 \pm 1.22\%$ $95.91 \pm 1.22 %$ to $98.36 \pm 0.72\%$ $98.36 \pm 0.72 %$ for HC vs. ALS, and from $91.24 \pm 1.52\%$ $91.24 \pm 1.52 %$ to $95.18 \pm 1.04\%$ $95.18 \pm 1.04 %$ in the four-class classification task. Signal-level and classification-based analyses confirm that the generated samples are statistically consistent with real gait data and contribute to improved model robustness. These results highlight the potential of generative deep learning as a data-efficient support tool for gait-based analysis of neurodegenerative diseases, particularly in small and imbalanced clinical datasets.