<p>Freezing of Gait (FoG) is one of the most disabling motor symptoms of Parkinson’s Disease (PD), characterized by brief episodes of motor arrest that significantly elevate the risk of falls and loss of mobility. Early and accurate detection of FoG events is critical for timely intervention and personalized care. This study presents a novel deep learning framework leveraging a Bidirectional Gated Recurrent Unit (BiGRU) architecture integrated with an attention mechanism to detect FoG episodes using inertial sensor data. The model was trained and evaluated on the Daphnet FoG dataset, which contains multi-channel accelerometer signals captured from multiple body locations. To enhance robustness, the dataset was balanced and evaluated using a fixed subject-wise 70/30 split (14 subjects for training, 3 subjects for testing). Comparative experiments against conventional deep learning models, including CNN, GRU, LSTM, and BiLSTM, revealed that the proposed BiGRU-attention model achieved superior performance across multiple evaluation metrics, with a test accuracy of 95.66%, Cohen’s Kappa of 0.79, and ROC-AUC of 0.98. These improvements are attributed to the model’s capacity to capture bidirectional temporal dependencies and assign dynamic attention weights to salient gait patterns. The findings highlight the effectiveness of the BiGRU-based model for real-time and clinically relevant FoG detection, offering promising applications in wearable-based monitoring systems and decision support tools for PD management.</p>