Enhancing FoG Detection with Wavelet Transforms Using Random Forest and Neural Networks

Freezing of Gait (FoG) is a prevalent and debilitating symptom associated with Parkinson's disease, marked by sudden and brief periods of immobility. Accurate detection of these episodes is essential for timely intervention and treatment. For this study, machine learning method and the proposed NN were used for identifying FoG using time series data from the UCI Freezing of Gait dataset. The approach employs Discrete Wavelet Transform (DWT) for feature extraction, along with variance-based techniques for selecting important features. Specifically, the Daubechies wavelet (db2) at level 2 decomposition was utilized to derive frequency-domain features from the time series data, and statistical features such as standard deviation, variance, skewness, and kurtosis were computed. Random Forest (RF) and proposed Neural Network (NN) models were trained using both raw and DWT-derived features. The results showed that DWT features significantly enhanced the model's performance, achieving perfect precision (1.00), recall (1.00), F1-score (1.00), and accuracy (1.00) with the RF model and proposed NN. Furthermore, k-fold cross-validation revealed that DWT features offered more consistent and reliable performance across different folds compared to raw features, minimizing variability and over-fitting observed in models trained on raw data. These findings highlight the significance of feature extraction through DWT and variance-based feature selection for FoG detection, indicating the potential for developing more accurate and robust machine learning models for diagnosing gait-related disorders.