Risk Prediction of Freezing of Gaits Episodes in Parkinson's Patients Using Gated Recurrent Unit

Freezing of gait (FOG) presents a temporary impact on the movement of Parkinson's disease patients. This is associated with sudden falls leading to injuries. Preventive measures can be useful in tackling such unforeseen situations. The early prediction and detection of FOG episodes are essential for improving the quality of life of the patients. This study presents two innovative hybrid deep learning models named Fuzzy Activation Function Bidirectional Gated Recurrent Unit (FAF-Bi-GRU) and Fuzzy Activation Function Self-Attention Bidirectional Gated Recurrent Unit (FAF-Self-Attention-Bi-GRU). The hybrid FAF-Bi-GRU has incorporated the fuzzy logic principles with the Bidirectional Gated Recurrent unit (Bi-GRU). Similarly, the proposed FAF-Self-Attention-Bi-GRU Network integrates both attention modules with fuzzy logic. This study also evaluates the performance of Bi-GRU with proposed hybrid models for the identification of better-performing recurrent neural networks. In the accelerometer dataset, the hybrid FAF-Self-Attention-Bi-GRU outperforms with an accuracy of 98.142%, precision of 98.409%, recall of 97.891%, f1 score of 98.149%, specificity of 98.399%, and G-mean of 98.144%. While in the gyroscope dataset, the hybrid FAF- Bi-LSTM outperforms with an accuracy of 99.195%, precision of 99.275%, recall of 99.119%, f1 score of 99.196%, specificity of 99.276%, and G-mean of 99.197%. This study presents the potential of the proposed hybrid models for the prediction of FOG using wearable sensor data.