Freezing of gait (FoG) is a common symptom of Parkinson's disease (PD) that causes intermittent absence of forward progression of patient's feet while walking. Accordingly, FoG momentary episodes are always accompanied with falls. This chapter presents a novel multi-feature fusion model for early detection of FoG episodes in patients with PD. In this chapter, two feature engineering schemes are investigated, namely time-domain hand-crafted feature engineering and convolutional neural network (CNN)-based spectrogram feature learning. Data of tri-axial accelerometer sensors for patients with PD is utilized to characterize the performance of the proposed model through several experiments with various machine learning (ML) algorithms. Obtained experimental results showed that the multi-feature fusion approach has outperformed typical single feature sets. Conclusively, the significance of this chapter is to highlight the impact of using feature fusion of multi-feature sets through investigating the performance of a FoG episodes early detection model.
TopBackground
Parkinson's Disease (PD) is a degenerative disorder, which affects patient's movements. It is marked by decreased dopamine levels in the brain and considered as the second most common symptom after Alzheimer's Disease (AD). A lack of dopamine, which results in abnormal nerve functioning, causes a loss in the ability to control body movements. The PD has influenced about 1% or 2% of elderly people worldwide (Nilashi, 2016). This study aimed to detect Freezing of Gait (FoG) attacks in patients with PD, using different features and classifiers, for increasing the detection performance and decreasing the social costs that face the patients with PD symptoms. The PD patients usually spend almost two more days in hospitals, 43 more days in care institutions, and fill more than 20 medical therapies than the non-PD subjects do. The total cost for PD patients is more than a double of the non-PD subjects. On the other hand, the productivity loss recorded for PD patients reaches 49.4% (Dua et al., 2006). Furthermore, as shown in Figure 1, among several consequences related to elderly falling, the loss of independence risk represents a significant social consequence. That is, the PD patient will constantly be dependent on one of the family members or a medical center caregiver (El-Bendary et al., 2013). According to the World Health Organization (WHO), the percentage of Global Disability Adjusted Life Years (DALYs) by 2030 will increase by 0.13% for patients with PD, coming after the percentage of Alzheimer’s Disease that has been predicted to increase by 1.2%. Also, the deaths for PD patients will reach 23% by 2030 as the total deaths globally for neurological disorders will reach 12.22% (Dua et al., 2006).
Figure 1. The main consequences related to elderly falling
Parkinson's Disease contains two types of symptoms that affect the quality of daily life; namely, motor and non-motor symptoms. Motor symptoms or cardinal symptoms contain resting tremor, rigidity, bradykinesia (movement slowness), postural instability (balance problems), and Freezing of Gait (FoG). Whereas non-motor symptoms contain cognitive impairment, sleep behavior disorder, olfactory loss, constipation, speech & swallowing problems, unexplained pains, drooling, low blood pressure when standing, and rapid eye movement. Freezing of Gait occurs in most patients with PD in early stages, as patients suddenly feel an inability to step forward while walking. During the FoG episodes that are characterized by a short period of inability to initiate a gait, PD patients intermittently feel that their feet are stuck to the floor as being held by magnets when trying to walk. Accordingly, FoG momentary episodes are always accompanied by patients falling, the case that affects activities in daily life, and quality of life by having a significant negative impact on PD patients with FoG symptoms (Sveinbjornsdottir, 2016). The FoG could be observed in some patients who experience brief trembling in their feet followed by short small steps. Other PD patients may experience total immobility in body movements and are unable to move at all for a few seconds.
For proposing state-of-the-art solutions to soundly handle the problem of detecting/predicting Parkinson's Disease, various studies were proposed using Machine Learning (ML) techniques. In (Kumar, 2016), the author aimed to predict PD using Random Forest classification algorithm with 20 different features using a dataset of a range of biomedical voice measurements from 31 subjects, 23 with PD. Also, the authors in (Sujatha & Rajagopalan, 2017) aimed to classify the patients as healthy and PD subjects using fundamental frequency, amplitude, hitter, noise-to-harmonic ratio, harmonic-to-noise ratio, Detrended Fluctuation Analysis (DFA), and spread biomedical voice features. Machine Learning algorithms such as ZeroR, OneR, Bayes Net, Radial Basis Function (RBF), Hidden Markov Model, Naïve Bayes, Logistic Regression, Multilayer Perceptron, AdaBoost, Decision Tree, J48, and Random Forest were used for classification of PD patients.
The authors in (Karan et al., 2020) proposed a model based on empirical mode decomposition for PD detection using two speech datasets. Different features were extracted; namely, acoustic features, mel-frequency cepstral coefficients (MFCC), statistical features, spectral entropy, energy, entropy, and intrinsic mode function-cepstral coefficient, with Support Vector Machine (SVM) and Random Forest ML classifiers.