Detection of Freezing of Gait in Parkinson's Disease Based on Deep Neural Networks and Machine Learning Models

This paper introduces an innovative methodology for the detection of freezing of gait (FOG) in individuals with Parkinson's disease (PD) by leveraging wireless inertial sensors. Utilizing the Daphnet Parkinson's database, which provides three-axis six-degree-of-freedom acceleration data, we present an advanced approach to address the intricate challenges associated with FOG detection. The methodology encompasses the extraction of acceleration time series data for both normal and freezing gaits through sliding window segmentation with varying widths of 1s, 2 $s$, 3s, and 4 $s$, each incorporating a fixed overlap ratio of 0.5. This deliberate variation in window widths aims to scrutinize the influence of temporal granularity on the efficacy of FOG detection. In the data preprocessing phase, we employ the Synthetic Minority Over-sampling Technique (SMOTE) to augment the FOG dataset, thereby bolstering the robustness of our model. The core of our methodology lies in the application of a deep neural network (DNN) for feature extraction. In this pivotal phase, we introduce the Convolutional Block Attention Module (CBAM) to enhance the discriminative capabilities of the network. CBAM incorporates channel and spatial attention mechanisms, enabling the network to selectively focus on salient features within the acceleration time series data. For classification of normal and FOG episodes, nine machine learning classifiers are employed, achieving a remarkable accuracy rate of 100% using a 10-fold cross-validation method. This comprehensive approach not only showcases the effectiveness of the CBAM-enhanced DNN but also underscores the meticulous consideration of temporal granularity in optimizing FOG detection efficacy.