Scholarlink

<p>In this study, we present a high-precision hybrid framework for classifying Parkinsonian gait. The proposed architecture fuses the local feature-extraction capabilities of a Temporal Convolutional Network (TCN) with the global sequence memory of a Recurrent Neural Network (RNN). To move beyond the limitations of deterministic models and better manage sensor noise, our design employs Variational Bayesian Expectation Maximization (VB–EM) for probabilistic state optimization and integrates Multifractal Detrended Fluctuation Analysis (MFDFA) to isolate critical non-linear biomarkers. This framework was rigorously evaluated on a comprehensive clinical repository (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(N=45\)</EquationSource> </InlineEquation>), including 28 Parkinson’s Disease patients and 17 healthy controls. Under a strict subject-independent 5-fold cross-validation protocol, the TCN–RNN framework demonstrated superior performance, reaching a peak accuracy of 92.29%, an F1-score of 92.30%, and an ROC–AUC of 0.98. Paired statistical testing (<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(p &lt; 0.01\)</EquationSource> </InlineEquation>) verifies that this approach consistently exceeds the capabilities of standard baselines such as CNN–LSTM, BiLSTM, and XGBoost. Furthermore, qualitative analysis confirms the model’s sensitivity to subtle transitional freezing phases, while a recorded inference latency of just 5.2&#xa0;ms highlights its readiness for real-time clinical deployment. Our results suggest that combining multi-scale dilated convolutions with recurrent state-switching logic offers a significant advancement in the development of objective, high-fidelity diagnostic tools for neurodegenerative monitoring.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A neural hopping state-space model for multimodal motor variability in parkinson’s disease: variational inference and deep temporal integration

  • 期刊论文

  • 发表日期：

    2026年5月12日

查看全文
  • Gauri Chandra,
  • Tapan K. Gandhi,
  • Bhim Singh

摘要

In this study, we present a high-precision hybrid framework for classifying Parkinsonian gait. The proposed architecture fuses the local feature-extraction capabilities of a Temporal Convolutional Network (TCN) with the global sequence memory of a Recurrent Neural Network (RNN). To move beyond the limitations of deterministic models and better manage sensor noise, our design employs Variational Bayesian Expectation Maximization (VB–EM) for probabilistic state optimization and integrates Multifractal Detrended Fluctuation Analysis (MFDFA) to isolate critical non-linear biomarkers. This framework was rigorously evaluated on a comprehensive clinical repository ( \(N=45\) ), including 28 Parkinson’s Disease patients and 17 healthy controls. Under a strict subject-independent 5-fold cross-validation protocol, the TCN–RNN framework demonstrated superior performance, reaching a peak accuracy of 92.29%, an F1-score of 92.30%, and an ROC–AUC of 0.98. Paired statistical testing ( \(p < 0.01\) ) verifies that this approach consistently exceeds the capabilities of standard baselines such as CNN–LSTM, BiLSTM, and XGBoost. Furthermore, qualitative analysis confirms the model’s sensitivity to subtle transitional freezing phases, while a recorded inference latency of just 5.2 ms highlights its readiness for real-time clinical deployment. Our results suggest that combining multi-scale dilated convolutions with recurrent state-switching logic offers a significant advancement in the development of objective, high-fidelity diagnostic tools for neurodegenerative monitoring.