Grey Wolf Optimized Transformer Model for Enhanced Multi-class Prediction of Parkinson’s Freezing of Gait Events

Freezing of gait (FoG) is a debilitating symptom in Parkinson’s disease (PD), often leading to falls and reduced quality of life. Accurate, real-time detection and prediction of FoG events remain challenging due to patient variability, short-term onset, and noisy sensor data. This study proposes a grey wolf optimized transformer (Transformer-GWO) model for enhanced multi-class FoG event prediction using wearable sensor data from a publicly available dataset. The proposed model integrates the Transformer’s ability to capture long-range temporal dependencies with grey wolf optimizer (GWO) for hyperparameter tuning, improving classification accuracy and robustness. Comparative experiments against optimized convolutional neural network (CNN-GWO), long short-term memory (LSTM-GWO), and three recent state-of-the-art baselines demonstrate that Transformer-GWO achieves the highest accuracy (98.41%), precision (98.32%), recall (98.29%), F1-score (98.30%), and AUC (0.993), while maintaining competitive computational efficiency. Specifically, Transformer-GWO achieved a training time of 22.49 s, inference time of 2.60 s, and modest memory usage (~ 211 MB), outperforming other models in speed–accuracy balance. Detailed attention map analysis, feature importance rankings, and per-class confusion matrices illustrate the interpretability of the model and its relevance for clinical deployment. These findings suggest that Transformer-GWO can provide accurate, interpretable, and computationally efficient FoG event prediction, potentially aiding in patient monitoring and personalized PD management.