Parkinson's Disease Diagnosis and Severity Assessment from Gait Signals via Bayesian-Optimized Deep Learning

Area of Science:

• Biomedical Engineering
• Machine Learning
• Neurology

Background:

• Early Parkinson's Disease (PD) diagnosis is crucial for timely interventions and improved quality of life.
• Gait analysis offers a non-invasive method to detect subtle motor impairments indicative of PD.
• Vertical Ground Reaction Force (VGRF) signals contain valuable information for PD assessment.

Purpose of the Study:

• To evaluate and compare Bayesian-optimized Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) models for PD detection and staging.
• To apply these deep learning models directly to VGRF signals.
• To assess the models' performance across different VGRF signal window lengths.

Main Methods:

• VGRF recordings were segmented into 5, 10, 15, 20, and 25-second windows.
• Segments were normalized and used as input for CNN and LSTM networks.
• Bayesian optimization with five-fold cross-validation was employed for hyperparameter tuning.

Main Results:

• The LSTM model achieved 99.42% accuracy (AUC=1.000) for PD vs. control detection at 10s and 98.24% accuracy (AUC=0.999) for Hoehn-Yahr staging at 5s.
• The CNN model reached 98.46% accuracy (AUC=0.998) for binary classification and 96.62% accuracy (AUC=0.998) for multi-class staging.
• Both models demonstrated high efficacy, with LSTM showing a slight advantage in temporal pattern recognition.

Conclusions:

• Bayesian-optimized CNN and LSTM models effectively detect and stage Parkinson's disease using VGRF data.
• LSTM models excel at capturing temporal gait dynamics, while CNNs offer comparable performance with lower computational cost.
• End-to-end deep learning on gait data presents a promising avenue for non-invasive PD assessment.