Dyskinesia Estimation of Imbalanced Data Using a Deep-Learning Model.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Neurology

Background:

• Parkinson's Disease (PD) data collection often yields imbalanced and incomplete time-series datasets.
• Imbalanced data leads to overfitting and poor generalizability in deep learning models for PD severity estimation.
• Traditional Convolutional Neural Networks struggle with the geometric distribution of PD complication severity scores.

Purpose of the Study:

• To investigate a novel Conditional Generative Adversarial Network (cGAN) framework.
• To enhance the extrapolation and generalizability of regression models for imbalanced PD time-series data.
• To accurately estimate Dyskinesia severity in patients with Parkinson's Disease.

Main Methods:

• Developed a Conditional Generative Adversarial Network (cGAN) framework.
• Utilized a real-world Parkinson's Disease dataset for training and validation.
• Applied the cGAN to estimate Dyskinesia severity in PD patients.

Main Results:

• The developed cGAN demonstrated significantly improved generalizability to unseen data.
• Achieved a 34% improvement in generalizability compared to traditional Convolutional Neural Networks.
• Successfully addressed challenges posed by imbalanced and incomplete PD time-series data.

Conclusions:

• Conditional Generative Adversarial Networks (cGANs) offer a robust solution for imbalanced time-series data in healthcare.
• This approach improves the reliability of deep learning models for estimating Parkinson's Disease severity.
• The framework is applicable to similar datasets lacking balanced and uniformly distributed samples.