Hierarchical Coarse-to-Fine cGAN for Subtype-Specific Freezing of Gait Signal Generation

Area of Science:

• Biomedical Engineering
• Machine Learning in Healthcare
• Neurology

Background:

• Freezing of gait (FOG) is a severe Parkinson's disease symptom with diverse subtypes (shuffling, trembling, akinesia).
• Deep learning (DL) models for FOG detection struggle with data scarcity and class imbalances across subtypes, limiting robustness and generalization.
• Existing data augmentation methods are insufficient for addressing the complexities of FOG subtype data.

Purpose of the Study:

• To develop a subtype-aware data augmentation technique for improving deep learning-based FOG detection.
• To generate realistic and diverse FOG-like ankle acceleration data conditioned on specific FOG subtypes.
• To enhance the consistency and accuracy of DL models across different FOG subtypes and patient profiles.

Main Methods:

• Introduction of Hierarchical Coarse-to-Fine conditional Generative Adversary Network (Hi-CF cGAN), a two-stage model for generating subtype-conditioned FOG data.
• Validation of generated data realism and diversity using visualization, UMAPs, and Maximum Mean Discrepancy.
• Training Convolutional Neural Networks (CNNs) for FOG detection using Hi-CF cGAN-generated data (general and personalized augmentation) and benchmarking against classical methods and baseline.

Main Results:

• Hi-CF cGAN generated data significantly improved FOG detection rates compared to baseline and classical augmentations.
• General augmentation boosted shuffling FOG detection from 66.8% to 81.6% and akinesia FOG from 58.7% to 77.9%.
• Personalized augmentation further enhanced accuracy for specific patient subtypes, demonstrating tailored optimization potential.

Conclusions:

• The proposed Hi-CF cGAN technique effectively addresses data imbalances and scarcity in FOG detection.
• Subtype-aware augmentation using Hi-CF cGAN significantly improves DL model performance across all FOG subtypes, particularly minor ones.
• This approach offers a promising strategy for developing more robust and personalized FOG detection systems for Parkinson's disease patients.