Multi-speed transformer network for neurodegenerative disease assessment and activity recognition

Area of Science:

• Computer Vision
• Machine Learning
• Neurology

Background:

• Neurodegenerative diseases are common age-related conditions requiring early diagnosis for better patient outcomes.
• Gait analysis is a key diagnostic tool for neurologists to assess disease severity.
• Timely detection of neurodegenerative diseases is crucial to prevent compromised quality of life.

Purpose of the Study:

• To develop a computer vision technique for early-stage neurodegenerative disease detection using gait analysis.
• To present the "Beside Gait" dataset featuring time-series joint coordinate data from patients and controls.
• To introduce and evaluate the novel Multi-Speed Transformer technique against existing deep and shallow learning models.

Main Methods:

• Pose estimation was used to extract human skeleton joint coordinates from video data.
• Time-series joint data were analyzed using novel deep neural network and Shallow Learning architectures.
• The Multi-Speed Transformer was benchmarked against Temporal Convolutional Networks, Transformers, and various classifiers (Random Forests, SVM, etc.).

Main Results:

• The Multi-Speed Transformer achieved 96.9% accuracy in binary classification and 71.6% in multi-class classification for neurodegenerative disease detection.
• It outperformed all tested models, including state-of-the-art techniques, on activity recognition datasets (SHREC, JHMDB).
• The model demonstrated effectiveness in learning short and long-term patterns for pathological gait analysis.

Conclusions:

• The Multi-Speed Transformer is a powerful tool for neurodegenerative disease assessment via computer vision.
• "Beside Gait" dataset provides a foundation for future research in automatic gait-based disease recognition.
• This technology holds potential for integration into smartphone applications for accessible, early diagnosis.