Scholarlink

<p>Analyzing tremor characteristics in patients with Parkinson’s Disease (PD) can assist in diagnosing and assessing disease progression. The current methods for collecting tremor data rely heavily on professional medical equipment and personnel, which are time-consuming and resource-intensive. The limited number of participants also leads to insufficient data diversity, resulting in suboptimal data quality and sample scarcity. We propose the Conditional TremorGANs model, a GAN-based method for generating synthetic tremor data from Parkinson’s disease patients, improving sample diversity and availability. The Conditional TremorGANs comprise a generator built upon time-based self-attention to model temporal dynamics and a discriminator equipped with multi-scale feature extraction to improve discrimination. The generator produces diverse and realistic samples through adversarial training with the discriminator. We evaluate the effectiveness of the generated data in terms of visualization, quantitative analysis, and usability assessment. Experimental results show that augmenting the training set with synthetic data increases accuracy from 74.4% to 80.1% (+ 5.7% points) for distinguishing between healthy controls (HC) and PD, and from 67.2% to 68.6% (+ 1.4% points) for differentiating PD from differential diagnoses (DD). These results confirm the effectiveness of our model in addressing data scarcity in PD tremor analysis, offering a generative solution for diagnostic support.</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 Generative Adversarial Network-Based Method for Synthesizing Tremor Data in Parkinson’s Disease

  • 期刊论文

  • 发表日期：

    2025年12月20日

查看全文
  • Guozhi Zhang,
  • Jiajia Shi,
  • Mingsen Hu,
  • Gaoyang Dai,
  • Zixuan Liu,
  • Lishi Zhang

摘要

Analyzing tremor characteristics in patients with Parkinson’s Disease (PD) can assist in diagnosing and assessing disease progression. The current methods for collecting tremor data rely heavily on professional medical equipment and personnel, which are time-consuming and resource-intensive. The limited number of participants also leads to insufficient data diversity, resulting in suboptimal data quality and sample scarcity. We propose the Conditional TremorGANs model, a GAN-based method for generating synthetic tremor data from Parkinson’s disease patients, improving sample diversity and availability. The Conditional TremorGANs comprise a generator built upon time-based self-attention to model temporal dynamics and a discriminator equipped with multi-scale feature extraction to improve discrimination. The generator produces diverse and realistic samples through adversarial training with the discriminator. We evaluate the effectiveness of the generated data in terms of visualization, quantitative analysis, and usability assessment. Experimental results show that augmenting the training set with synthetic data increases accuracy from 74.4% to 80.1% (+ 5.7% points) for distinguishing between healthy controls (HC) and PD, and from 67.2% to 68.6% (+ 1.4% points) for differentiating PD from differential diagnoses (DD). These results confirm the effectiveness of our model in addressing data scarcity in PD tremor analysis, offering a generative solution for diagnostic support.