Parkinson’s disease classification with CWNN: Using wavelet transformations and IMU data fusion for improved accuracy

2.1Feature-based approaches

Feature-based methods are commonly used for PD classification with IMU data. These methods involve selecting specific features of the IMU sensor data that are relevant to PD classification and using machine learning algorithms to classify the data based on these features [7, 8]. The features selected can have a large variation depending on the study such as:

Feature based methods have been using several machine learning algorithms for PD classification including Support Vector Machine (SVM) which is the most common algorithm for classification tasks, it performs by finding the hyperplane that best separates the two classes of data. Trabassi et al. [8] show that SVM based classification method outperformed Decision Trees (DT) in predicting subjects with Parkinson’s from healthy subjects. DT is a simple yet effective algorithm for classification, it works by recursively partitioning the data into smaller subsets until the desired classification or regression is achieved. DT has been used in gait analysis studies for classifying gait abnormalities in people with Parkinson’s disease [8]. Random Forest (RF) is an set of learning algorithm that combines multiple DTs to improve the classification accuracy and make predictions [8], by randomly selecting features and splitting points when constructing the DTs. K-Nearest Neighbors (KNN) is a non-parametric algorithm that classifies data based on the majority vote of its k nearest neighbors. It has been explored for PD classification using IMU data [7]. The main problems related to using machine learning algorithms for PD classification is having lower classification accuracy compared to Deep Learning (DL) approaches and they may not be able to capture complex patterns in the IMU data.

2.2Deep learning approaches

Identify applicable funding agency here. If none, delete this text box.

Different deep learning algorithms used in literature for PD classification and are more suitable to IMU data:

Several research studies have investigated the use of CNNs for PD classification using IMU data. Yeon-Wook et al. [13] demonstrated a CNN-bidirectional-GRU model that showed better classification performance for clinical balance assessment using IMU data. In a study by Celik et al. [14], CNN architectures with a small number of deep layers achieved high accuracy for inertial sensor-based activity recognition in neurological populations. In a paper by Carvajal-Castaño et al. [15], raw gait signals captured using IMU sensors were used to assess the ability of different deep learning architectures to classify PD patients vs. healthy controls. The study found that CNN and GRU based model achieved the highest accuracy of 92.6% for temporal and spectral information extracted from IMU data. In the work suggested by Bikias et al. [16], a deep learning model was used to detect freezing of gait (FoG) episodes in PD patients using IMU data. The study found that the model achieved high accuracy and sensitivity in detecting FoG episodes. Alissa et al. [17] suggest to use used CNNs for PD diagnosis using figure-copying tasks. The study aimed to contribute to the PD diagnosis process by using CNNs to classify PD patients based on their ability to copy figures. A systematic review by Sigcha et al. [18] found that different deep learning models, including hybrid models, showed higher performance for the diagnosis and monitoring of PD using wearable sensors.

Convolutional wavelet neural network (CWNN) is a type of neural network that combines combine the spatial and frequency-domain processing capabilities of CNNs wavelet transformation (WT). This type of deep learning is usually used for pattern recognition tasks and image classification, which it shows higher performance and improved accuracy [19, 20]. To our knowledge, there is no study that shows the performance of CWNN for PD classification with IMU data. Therefore, this work shows preliminary performance results for PD classification based on CWNN.

Figure 1.

Shape of the Mexican hat, Morlet and Gaussian mother wavelets.

3.1Background

Wavelet Transform (WT) is a transformation technique used for analyzing signals and data in both the time and frequency domains. It decomposes a signal into a set of wavelet functions, which are scaled and shifted versions of a mother wavelet function according to defined scale factor [21]. The general function of the WT can be described as follows. Given a signal x⁢(t), the wavelet transform can be defined as:

where W⁢(a,b) represents the wavelet coefficients, a and b are the scaling and translation parameters, and ψ⁢(a,b,t) represents the wavelet function. The mother wavelet function ψ⁢(t) can be described as follow:

WT has several properties including: time-frequency localization where wavelets offer good localization properties by capturing both temporal and frequency information of a signal. This allows for a detailed analysis of signal characteristics at different scales. Multi-Resolution Analysis properties where WT can analyze the different levels of detail in the signal. This is achieved by dilating and translating the wavelet function to cover different time-frequency regions. Also, Shift Invariance property which is useful for analyzing signals with non-stationary properties. The above properties are suitable for the nature of signals collected from the IMU sensors. The scale factor defines signal scaling over time and it is inversely proportional to frequency, where small scale factor captures higher frequencies while large scale factor shows the lower frequencies in the signal [19]. Most common mother wavelet functions in continuous wavelet transformation (CWT) are Mexican hat, Morlet and Gaussian wavelet functions.

3.2Proposed method

Two main steps constitute the proposed method. The first step is related to data collection and preprocessing. IMU sensors are attached to the lower limbs for both left and right side. The data collected is in the form of 3D accelerometer signals and 3D gyroscope signals along with x, y and z axis. The signals are segmented and synchronized. A sliding window with predefined size and 50% overlap is used to apply CWT on each signal to get 2D image representation with three channels for both accelerometer data and gyroscope data Fig. 2. The size of the sliding windows is chosen to be equal to the predefined scaling factor to get squared shape of the input images.

Figure 2.

Placement of IMU sensors and wavelet extraction from preprocessed IMU data.

The second step is training the CNN model. The architecture contains four sets of convolutional layers, two sets of fully connected layers and max-pooling layers between each of the above layers Fig. 3.

4.1Dataset

Lower limb kinematics data was collected at Vilnius University Hospital Santaros Klinikos Centre of Neurology, under the supervision of medical professionals [22]. The research involved two groups of subjects: (PD) group represents individuals with Parkinson’s disease, and (CO) group consisting of healthy control individuals. PD groups are characterized with UPDRS III as clinical assessment. The gait difficulty was defined as normal gait or walks slowly with possibility of shuffle with short steps, but no festination or propulsion and no major difficulties were found.

The PD group’s inclusion criteria comprised being an adult with a disease severity between 2 and 3 on the Hoehn and Yahr scale and having the ability to walk without assistance. Exclusion criteria included the presence of other movement-impairing diseases. For the CO group, participants were included if they had no injuries or illnesses that could affect their movement. All procedures adhered to the ethical guidelines set by the local (Vilnius University Hospital Santaros Klinikos).

Figure 3.

CNN with LeNet-5 architecture with max pooling and ReLU activation.

To measure motor tasks, six wireless IMUs (Shimmer Research, Dublin, Ireland) was attached to the subjects’ limbs using mounting straps Fig. 2. The IMU’s kinematic data was sampled at a frequency of 51.2 Hz and transmitted via Bluetooth to a laptop for processing. The movement task involved walking five meters in a straight line, with clearly marked start and end points. Participants were instructed to begin walking upon a verbal command. Each participant performed the task three times at their chosen comfortable speed.

4.2Training and testing results

The scale factor chosen for this experiment is 64. In general, a smaller size of scales enables more focus of abrupt changes on the signal. These fast changes are usually the most important characteristics. In addition, a large range of scales provides more information about slow changes, which can produce better classification accuracy. However, it needs a deeper and complex CNN. After an exhaustive training of the model with different scaling factors, a scale of 64 helps to reach a good prediction accuracy.

Accordingly, sliding windows in the preprocessing steps are with size 64 and overlap of 50%. The proposed CWNN is trained for input data of 3D accelerometer signals, 3D gyroscope signals and data fusion which is a combination of accelerometer and gyroscope signals leading to have input image with six channels. The model was evaluated with different mother wavelet function such as Marlet, Mexican hat and Gaussian wavelets.

Figure 4.

Accuracy of the training and validation of CWNN with fusion data using Morlet function.

Figure 5.

Loss of the training and validation of CWNN with fusion data using Morlet function.

The comparison results are demonstrated in terms of several metric values such as: accuracy (acc), precision (Prec), recall, F1-score and loss functions Table 2. The above values in the comparison table shows that:

In general, CWNN achieves higher accuracy performance regardless of the many critical factors such as: type of signals (acceleration, angular rate or fusion), the position of the sensors in different part of lower limbs (thigh, shank or foot) and the type of mother wavelet function.

In particular, CWNN with fusion signals using Morlet function, outperform in term of predefined comparison metrics the other type of signals along with the remaining tested mother wavelet function.