Wavelet Decomposition and Feedforward Neural Network for Classification of Acute Ischemic Stroke based on Electroencephalography

Stroke is one of the leading causes of death in Indonesia. From 2013 to 2018, the prevalence of stroke increased from 7% to 10.9%. There are two types of strokes, namely Hemorrhagic and Acutte Ischemic Stroke (AIS) with the majority of it being AIS. Early detection and diagnosis are essential in stroke as it is a life-threatening disease, and the stroke treatment is based on its type. Currently, the gold imaging standards in stroke diagnosis are Computed Tomography (CT) scan and Magnetic Resonance Imaging (MRI). However, the mentioned services for stroke diagnosis are primarily available in hospitals classified as “class A” (general hospitals with extensive facilities and medical services).  Compared to CT scans and MRI, electroencephalography (EEG) is a cost-friendly, non-invasive device studied for various brain-related diseases. This study aimed to determine the optimal epoch length to classify four stroke classes (healthy, minor, moderate, and severe) during  resting condition for a machine learning-based AIS computer-aided diagnostics system. 32-channel EEG, CT scan, and NIHSS Scores were the obtained data. The features were delta-theta to alpha-beta ratio (DTABR), delta to alpha ratio (DAR), relative power ratio (RPR), and asymmetry, which were extracted using wavelet decomposition technique. The epoch length was varied by 1s, 2s, 10s, 30s, 60s, and 120s. The severity of stroke were classified using a feedforward neural network. The best performance was obtained at the 60-second epoch length with 89% accuracy using 15 hidden layers. This EEG-based diagnostic system would be expected to be implemented in “class C” hospitals, where only essential medical services and facilities are available, usually in rural areas.

Acute Ischemic Stroke (AIS); Electroencephalography (EEG); Epoch length; Feedforward Neural Network (FNN); Wavelet decomposition

EEG is a non-invasive device that has been used in several research, such as for depression detection (Apsari & Wijaya, 2020), seizure detection (Srivastava et al., 2020), and stroke detection (van Putten, 2007).  An EEG device is more affordable to procure than a CT scan or an MRI, and consequently less costly to operate. Ischemic stroke brain may present abnormalities in the EEG signal (Finnigan & van Putten, 2013). Typically, AIS EEG signals exhibit high delta activity (1-4 Hz). The delta power has been consistently identified as the marker for ischemic stroke. The alpha power, Brain Symmetry Index (BSI), delta-alpha ratio (DAR) and delta-theta alpha-beta ratio (DTABR) were also found to be good markers for stroke (Rahma et al., 2017; Finnigan et al., 2007; Müller et al., 2002). The delta band power and alpha to delta band ratio can distinguish patients who have sizeable ischemic stroke from other stroke patients (Shreve et al., 2019).

EEG comes with complex brain signals, and thus machine learning (ML) approach has been implemented as means to make sense of them. Extreme ML has been used to classify stroke features (Chan et al., 2019; Rahma et al., 2017), while XGBoost and principal component analysis were applied to improve classification accuracy in a small number of selected EEG channels (Fitriah et al., 2017). Convolutional Neural Networks, including Multi-Layered Perceptron, as well as Decision Tree and Artificial Neural Networks, have been used for EEG-based classification (Nurfirdausi et al., 2022; Dewi et al., 2020; Qureshi et al., 2018; Omar et al., 2014). One of the issues with classifying EEG signals using the ML approach is looking for the most optimal feature settings to obtain the best performance. Thus, this study explored variations of epoch lengths to investigate which length would yield the best results.

The current study is an extension of previous studies (Chan et al., 2019; Fitriah et al., 2017; Rahma et al., 2017) using the AIS dataset. The EEG data were obtained for healthy control and stroke patients as a 30-minute recording. NIHSS scores were referenced to separate patients into classes based on stroke severity, i.e., healthy, mild, moderate, and severe. EEG data were read and analyzed in MATLAB R2020A. The 30-minute recording was segmented into epochs of a certain length (120, 60, 30, 10, 2, and 1 second epochs) and features—such as relative power ratio (RPR), delta-theta alpha-beta ratio (DTABR), delta-alpha ratio (DAR), and asymmetry—were computed for each epoch. In the future, this program is expected to be implemented into the EEG hardware so that the most accurate prediction of stroke type can be obtained automatically in the shortest time possible. The features were classified into four classes based on stroke severity, and classification performance was obtained from that. Feedforward neural network was chosen as a classification algorithm because of its simplicity and advantage in handling nonlinear data.

2.1. Data Acquisition

The EEG devices used in this study were the Biologic Netlink System and the Xltek EEG 32U Natus, both had 32 channels and 512 Hz sampling frequency. The electrode placement was based on the international 10–20 system and was saved in European Data Format (.edf). Each AIS patient’s EEG recording was accompanied by their NIHSS score, CT scan, and EEG interpretation from the supervising physician. NIHSS score defines the level of stroke severity: 0 indicates no stroke, 1–4 indicates minor stroke, 5–15 indicates moderate stroke and above  15 indicates severe stroke (Rahma et al., 2017).

2.2. Signal Processing and Features

After data collection, pre-processing using Independent Component Analysis (ICA) was done automatically in the device. The signals were then decomposed into their respective frequency bands using wavelet transformation, and their features were calculated before inputting into the ML algorithm. The features calculated in this study were relative power ratio (RPR), delta-alpha ratio (DAR), delta-theta alpha-beta ratio (DTABR), and asymmetry.

2.2.1.  Segmenting into Multiple Epochs

The standard procedures for EEG recording were based on the conditions shown in Table 2. The total recording time was 30 minutes, which were divided as follows:
Table 2 EEG data timepoints and their recording conditions

In this study, only eyes closed conditions were used. The reason for this was to investigate AIS markers during resting condition. The total eyes closed recording was 18 minutes. This recording was then divided into multiple epochs of various lengths, which are 120, 60, 30, 10, 2, and 1 second epochs. Each epoch was transformed to a frequency domain using wavelet transformation.

2.2.2.  Wavelet Transformation

Wavelet transformation is a time-to-frequency domain transformation that utilizes wavelets called the “mother” and “daughter” wavelets. Wavelet is concentrated in both time and frequency, whereas standard Fourier transforms is localized only in frequency (Agarwal et al., 2017). It is known to localize signal components better than other methods (Sanei & Chambers, 2007). The wavelet is represented as:

where ?_? is the angular frequency, ? is the wavenumber, and ?_{??, ?}(?) is the wavelet function.

Wavelet transformation can be classified into two types: Continuous Wavelet Transformation (CWT) and Discrete Wavelet Transformation (DWT). For this study, DWT was used to decompose the EEG signals into their respective frequency bands: delta, theta, beta, alpha, and gamma. DWT discreetly samples the signal and acts like a filter bank that takes the input signal and outputs the coefficients of the signal. Decomposing EEG signals into their respective frequency components can be done using a multi-level DWT decomposition (Tumari et al., 2013).

To decompose a signal of EEG, the window function and the level of decomposition must be appropriate for the signal. The frequency sampling of the acquisition was 512 Hz, which means that the recorded signal is 0-512 Hz. A 7-level decomposition was required to decompose these signals. The window function used in this study was Daubechies 4 (db4) due to its small mean square error (MSE) when used for EEG signals (Tumari et al., 2013).

The decomposed signals consisted of approximation (A) and detail (D) signals. The first three decompositions (D1, D2, and D3) were not used because they were considered as noise (Tumari et al., 2013). EEG frequency bands were obtained from the decomposition of AD3, which were D4 (gamma band, 32 – 64 Hz), D5 (beta band, 16 – 32 Hz), D6 (alpha band, 8 – 1 6 Hz), D7 (theta band, 4 – 8 Hz), and AD7 (delta band, 0 – 4 Hz).

2.2.3. Relative Power Ratio

Relative power ratio is the ratio between a certain frequency band's power and all bands' total power. The RPR of a certain frequency band is computed using the equation: 

where f₁ and f₂ are the low and high boundaries of the frequency band, f_L­ and f_H are the low and high boundaries of all bands. P(f₁, f₂) refers to the band power of a particular frequency band, while P(f_L, f_H) refers to the total power of all bands.

2.2.4. Delta-Alpha Ratio (DAR)

Delta-alpha ratio (DAR) is the ratio between the delta band's power and the alpha band's power. DAR is calculated using the equation:

DAR in AIS patients’ EEG signals was found to be higher than healthy controls, with higher variability as well (Finnigan et al., 2016).

2.2.5. Delta-Theta Alpha-Beta Ratio (DTABR)

2.2.6. Asymmetry

Asymmetry is the measure of activity between the left and right brain, calculated using the band power of each EEG frequency band. A high asymmetry indicates that the brain in that frequency band is more dominant in the left or right hemispheres. When asymmetry is positive, it indicates that the right hemisphere is more dominant. When it is negative, then the left hemisphere is more dominant (Allen et al., 2004). The calculation of asymmetry is as follows:

where N is the number of electrodes on each hemisphere and P_x is the band power of frequency band x for each electrode.

This study calculated nine asymmetry features from eight left-right electrode pairs and the total left-right hemispheric asymmetry. These features can be further explained as follows: prefrontal (FP2 – FP1); frontal 1 (F4 – F3); frontal 2 (F8 – F7); temporal 1 (T2 – T1); temporal 2 (T4 – T3); central (C4 – C3); parietal (P4 – P3); occipital (O2 – O1).

2.3. Machine Learning Classification

2.3.1. Feedforward Neural Network

Feedforward neural network or multilayer perceptron (MLPs) is the first artificial neural network that does not have a loop for their connections. In this network, information only flows forward from the input nodes to the hidden layers (if any) and the output node, hence the name “feedforward”. It is also divided into two groups depending on the number of the layers, which are single-layer and multi-layer (SAZLI, 2006). Feedforward neural network is considered simple compared to a recurrent neural network (RNN), in which is constructed as a loop. This chain's length is called the network's depth.

2.3.2. K-Fold Cross Validation

Cross-validation is a sampling method to estimate the performance of a predictive model in testing with the advantage of giving insight into performance from an independent dataset. It separates the dataset into portions and utilizes different parts of that data as either testing or training data in each of its iterations. The estimation of cross-validation accuracy is the number of correct classifications divided by the total data in the dataset. In k-fold cross-validation, the dataset will be randomly divided into equal sizes of subsets or folds.

The results of k-fold cross-validation are averaged from the results of k number of subsets. In k-fold cross-validation, all subsets will take turns as training and validation data, but each subset will only be used once for validation. The standard parameter -k on k-fold cross-validation is 10, however, it remains an undetermined parameter (Seni & Elder, 2010). 

3.1. Features Analysis

Figure 1 Comparison of DAR and DTABR among stroke severity level: normal, mild, moderate, and severe

In addition to DAR and DTABR, RPR was calculated using Decomposition Wavelet Transform (DWT). Figure 2 shows RPR in different frequency bands of EEG: alpha RPR, beta RPR, delta RPR, and theta RPR based on the level of stroke severity. The slight dominance of delta RPR is shown in mild stroke patients compared to normal subjects. In moderate and severe stroke patients, delta dominated the brain signals. Severe stroke patients had an abnormally dominant delta band, especially in the prefrontal areas (FP1, FP2, and FPZ), which proves that stroke patients have unusually high power of slow waves in their brains. Figure 2 shows that the alpha RPR of moderate stroke is slightly lower compared to the alpha RPR of mild stroke.

Figure 2 Relative Power Ratio (RPR) comparison among all stroke severity level

Figure 3 Asymmetry scores between healthy and stroke patients

This study calculated features that could identify stroke from healthy controls from resting EEG signals, which were delta theta to alpha beta ratio (DTABR), delta to alpha ratio (DAR), relative power ratio (RPR), and asymmetry. Stroke patients had higher delta RPR value as compared to healthy controls. Differences between control and stroke patients were identified. Stroke patients had higher delta RPR value as compared to healthy controls. In general, as stroke severity increases, so does the dominance of the delta band. Severe stroke showed a very dominant delta compared to the other bands, especially in the prefrontal region. Feedforward neural network and AIS EEG features were utilized to identify stroke and predict its severity from resting EEG data. The EEG signals were segmented into different epoch lengths, and the neural network's hidden layers varied. For this study, the optimum network configuration was 60-second epochs with 15 hidden layers. This simple configuration could classify stroke into four different classes with the best accuracy of 89%, specificity of 100%, and sensitivity of 88%. Further studies could implement feature-selection methods such as Principal Component Analysis (PCA) to reduce dimensionality and improve classifier performance. The result of this study shows a promising future for a more robust AIS computer-aided diagnostic system that uses EEG as an alternative neuroimaging device for stroke diagnosis.

This study was supported by the Department of Education and Culture of the Republic of Indonesia through DRPM Universitas Indonesia by PDUPT 2020 with contract number NKB2822/UN2.RST/HKP.05.00/2020.

Agarwal, S., Singh, O. P., Nagaria, D., 2017. Analysis and comparison of wavelet transforms for denoising MRI image. Biomedical and Pharmacology Journal, Volume 10(2), pp. 831–836

Allen, J.J.B., Coan, J.A., Nazarian, M., 2004. Issues and Assumptions on the Road From Raw Signals to Metrics of Frontal EEG Asymmetry in Emotion. Biological Psychology, Volume 67(1–2), pp. 183–218

Apsari, R.A., Wijaya, S.K., 2020. Electroencephalographic Spectral Analysis to Help Detect Depressive Disorder. In: Proceedings of the 37^th International Conference on Biomedical Engineering

Chan, J.Y., Wijaya, S.K., Rahma, O.N., 2019. ELM (Extreme Learning Machine) Method for Detecting Acute Ischemic Stroke Using Conventional and Specific Asymmetry BSI (Brain Symmetry Index) Features based on EEG Signals ELM ( Extreme Learning Machine ) Method for Detecting Acute Ischemic Stroke Us. AIP Conference Proceedings, Volume 2092(1), p. 020023

Dewi, F.Y., Faza, A., Prajitno, P., Kusuma Wijaya, S., 2020. Stroke severity classification based on EEG signals using 1D convolutional neural network. Journal of Physics: Conference Series, Volume 1528, p. 012006

Finnigan, S.P., Walsh, M., Rose, S.E., Chalk, J.B., 2007. Quantitative EEG Indices of Sub-Acute Ischaemic Stroke Correlate with Clinical Outcomes. Clinical Neurophysiology, Volume 118(11), pp. 2525–2532

Finnigan, S., van Putten, M.J.A.M., 2013. EEG in Ischaemic Stroke: Quantitative EEG Can Uniquely Inform (Sub-)Acute Prognoses and Clinical Management. Clinical Neurophysiology, Volume 124(1), pp. 10–19

Finnigan, S., Wong, A., Read, S., 2016. Defining Abnormal Slow EEG Activity in Acute Ischaemic Stroke: Delta/Alpha Ratio as an Optimal QEEG Index. Clinical Neurophysiology, Volume 127(2), pp. 1452–1459

Fitriah, N., Wijaya, S.K., Fanany, M.I., Badri, C.,Rezal, M., 2017. EEG Channels Reduction Using PCA to Increase Xgboost’s Accuracy for Stroke Detection. AIP Conference Proceedings, Volume 1862(1), p. 030128

Health Research and Development Department, 2013. Basic Research of Health 2013. Ministry of Health of Indonesia.

Health Research and Development Department, 2018. Basic Research of Health 2018. In Ministry of Health of Indonesia. Available online at https://doi.org/http://repository.bkpk.kemkes.go.id/3514/1/Laporan%20Riskesdas%202018%20Nasional.pdf, Accessed on February 19, 2022

Kusuma, Y., Venketasubramanian, N., Kiemas, L.S., Misbach, J., 2009. Burden of Stroke in Indonesia. International Journal of Stroke, Volume 4(5), pp. 379–380

Müller, C., Achermann, P., Bischof, M., Nirkko, A.C., Roth, C., Bassetti, C.L., 2002. Visual and Spectral Analysis of Sleep EEG in Acute Hemispheric Stroke. European Neurology, Volume 48(3), pp. 164–171

Nurfirdausi, AF., Wijaya, S.K., Prajitno, P., Ibrahim, N., 2022. Classification of Acute Ischemic Stroke EEG Signal Using Entropy-Based Features, Wavelet Decomposition, and Machine Learning Algorithms. In: The 6^th Biomedical Engineering’s Recent Progress in Biomaterials, Drugs Development, and Medical Devices

Omar, W.R.W., Mohamad, Z., Taib, M.N., Jailani, R., 2014. ANN Classification of Ischemic Stroke Severity using EEG Sub Band Relative Power Ration. In: 2014 IEEE Conference on System, Process and Control, ICSPC 2014, December

Qureshi, A.A., Zhang, C., Zheng, R., Elmeligi, A., 2018. Ischemic Stroke Detection Using EEG Signals. In: Proceedings of the 28^th Annual International Conference on Computer Science and Software Engineering

Rahma, ON., Wijaya, S.K., Prawito, Badri, C., 2017. Electroencephalogram Analysis with Extreme Learning Machine as a Supporting Tool for Classifying Acute Ischemic Stroke Severity. In: International Seminar on Sensor, Instrumentation, Measurement and Metrology: Innovation for the Advancement and Competitiveness of the Nation

Rudd, A., 2016. National Clinical Guideline for Stroke. Royal College of Physicians, UK. Available online at: https://www.hse.ie/eng/about/who/cspd/ncps/stroke/ resources/2016-national-clinical-guideline-for-stroke-5th-edition.pdf

Sanei, S., Chambers, J., 2007. EEG Signal Processing. John Wiley & Sons

SAZLI, M. H., 2006. A brief review of feed-forward neural networks. Communications, Faculty Of Science, University of Ankara, January 2006, pp. 11–17. https://doi.org/10.1501/0003168

Seni, G. Elder, J.F., 2010. Ensemble Methods in Data Mining: Improving Accuracy Through Combining Predictions. In Synthesis Lectures on Data Mining and Knowledge Discovery. Springer Cham: Switzerland

Shreve, L., Kaur, A., Vo, C., Wu, J., Cassidy, J.M., Nguyen, A., Zhou, R.J., Thuong, B., Yang, D.Z., Medizade, A.I., Chakravarthy, B., Hoonpongsimanont, W., Barton, E., Yu, W., Cramer, S. C., 2019. HHS Public Access, Volume 28(8), pp. 2280–2286

Srivastava, G., Tripathi, A., Maurya, P.K., 2020. Fuzzy Entropy Based Seizure Detection Algorithms for EEG Data Analysis. In Smart Healthcare for Disease Diagnosis and Prevention. Academic Press

Tumari, S.Z.M., Sudirman, R.,Ahmad, A.H., 2013. Selection of a Suitable Wavelet for Cognitive Memory Using Electroencephalograph Signal. Engineering, Volume 05(05), pp. 15–19

van Putten, M.J.A.M., 2007. The Revised Brain Symmetry Index. Clinical Neurophysiology, Volume 118(11), pp. 2362–2367