The Electroencephalogram (EEG), which is used in the diagnosis of brain related disease, which reflects the physiological processes of the human brain. Any human expert must take their time visually examining a large amount of EEG data, and there is a chance that their assessments will differ or be inconsistent across doctors. So, using deep learning techniques, a process of automated identification and diagnosis is started. The Neuro Fuzzy Framework (NFF) classifier is discovered to be the superior classifier when the results are studied and contrasted based on the different factors. This work might make the categorization process stronger in practise, advancing medical diagnosis. The diagnosed information can be further utilized by the supply chain management system for providing the medicines and other needed materials to the user on appropriate time. The enrichment of supply chain management system is analysed in this article.
Research Article
Neuro Fuzzy Framework for EEG Signal Classification to improve Supply Chain efficiency
https://doi.org/10.21203/rs.3.rs-2608353/v1
This work is licensed under a CC BY 4.0 License
Version 1
posted
You are reading this latest preprint version
EEG signals
neural network
disease
brain activity
blockchain
supply chain
and fuzzy rules
The Electroencephalogram (EEG) is a non-invasive method used as a clinical diagnostic and research tool for the analysis of the brain. In spite of recent advancements in technology, the classification of EEG signal is significant and the skilled trainers are required for the study of brain. Although, many works have been done based on EEG signal, states the use of automatic classification, but the task remains still tricky for the few reasons. The conventional EEG signals are analyzed only by the technical expertise [1]. The main focuses on the EEG signals are to be made effective and efficient classifiers are used for the classification of disease patient and healthy control [2]. The challenging issue in the classification problem is the feature extraction. The most important dispute is to extract the important features of the large set of EEG data. In addition, the major concern is about the classifiers for the best classification.
The Electroencephalogram (EEG) is a non-invasive tool, which records the electrical activity of the human brain from the scalp. The waveform recorded and reflects the cortical activity from the surface of the brain [3, 4].With the help of modern technology and software, it is simple to manage the recording of EEG signals. With the naked eye, it is impossible to detect any abnormalities in these signals. Therefore, the most crucial issue is to appropriately analyse these signals and retrieve the internal hidden properties. An irregular EEG flow in a human brain might result in a neurological condition [5]. It is important to carefully examine this anomaly in order to identify the disease's pattern so that similar disorders in human brains can be predicted. EEG recording typically results in the gathering of a massive quantity of numerical data that represents the electrical activity at various times [6].
In essence, the amplitudes of all forms of signals produced during medical diagnosis are analysed in the temporal domain [7]. EEG signals are typically a collection of amplitudes with relation to time, just as ECG signals. The graph generated utilising the data allows one to see a patient's pathological state. By displaying the graph, any substantial shape variation may be appropriately displayed and observed [8]. The observed signals are investigated by deploying computational techniques without the assistance of medical experts. Further, these information are utilized in the medical related supply chain management [9, 10].
The point of this exploration work is to make a relative investigation of existing classifiers, namely Artificial Neural Network (Back Propagation Neural Network), Support Vector Machine (SVM), and proposed Neuro-Fuzzy Framework (NFF) which set apart the Healthy Control and Disease patients, based on the features extracted from the EEG signal and the performance measures of the neuroimaging techniques are compared with these classifiers [11, 12].Internet of Medical Things (IoMT) and e-health advancements can play a vital, as well as growing role in enabling quick responses to epidemics and public health emergencies [13, 14].
Intelligent healthcare systems are able to collect enormous volumes of data because to the extensive and simple deployment of IoMT devices, which calls for rapid processing and transmitting. The creation of intelligent, secure health systems with effective medical data interchange abilities across many organisations that is global significance in light of the current epidemic. Automating the majority of healthcare functions to deliver effective healthcare services can make a significant contribution to the progression of intelligent health systems.This vision can become a reality thanks to cutting-edge technologies like edge computing, blockchain, and artificial intelligence (AI) [15].
The existing healthcare system might be replaced with an intelligent healthcare system that allows for effective medical data collection, processing, and exchange owing to emerging technology. An intelligent health system may perform a wide range of functions, including event categorisation, detection, real-time remote monitoring, and expediting clinical trials for innovative drugs. Blockchain also provides transaction traceability and audibility across several organisations; for example, this might be very helpful for tracing the supply chain of certain drugs or vaccinations during clinical research. On the other hand, being decentralised opens up the possibility of applying edge computing, which makes it possible to quickly and portably identify unfavourable occurrences at the edge and to deliver an instant reaction from healthcare entities even when hospitals are packed [16].
Antony, M. J., et al. (2022) [17] examines the generally utilized non-invasive method for the investigation of motor imagery EEG data. For feature extraction technique, the author uses the Common Spatial Pattern (CSP) and welch transformation method for spectrum estimation and these features are handled through the classifiers. In this work, the author utilizes SVM classifiers and Linear Discriminant Analysis (LDA). The performance analysis of Support vector machine classifier with the spectral features from welch method results high accuracy when compared to the linear discriminant analysis with both spectral features.
Koh, J. E. W., et al. (2022) [18] have carried out a non-invasive technique for the detection of brain disorder using SVM classifier. In this, the author applied wavelet transform as the feature extraction technique and for the classification of EEG signals, the machine learning algorithm are used. The proposed work uses 8 channel EEG signal system for the detection of Alzheimer Disease using SVM classifier and performs 90.9% accuracy. The main advantage for the detection of Alzheimer Disease is based on the frequency bands, the author prefer wavelet transform for the analysis both in time and frequency domain.
Li Zhiwei & Shen Minfen (2007) [19] have addressed the classification of mental tasks EEG signal and it is one of the major issues in Brain Computer Interface (BCI). The approach was based on wavelet packet entropy and SVM. The author discussed the major drawbacks of ANN. It requires large set of training data and limits the performance. The SVM classifier uses three kernel functions namely polynomial kernel, linear kernel and radial basis function (RBF) kernel for the classification of mental task. The author concludes that the linear kernel SVM classifier produces better performance rate when compared to RBF kernel.
Ahsan Bin Tufail et al. (2012) [20] proposed to treat the individuals suffered from Alzheimer Disease (AD) based on the structural Magnetic Resonance Imaging (sMRI) technique. The images taken from the MRI technique discriminates the AD patients, mild cognitive impairment and normal control with the clinical dementia rating of 1. The sample of tissue classification images is taken from the dataset is further proceeded with principal component analysis for the dimensionality reduction of phase images. Subsequently, the images are processed by the independent component analysis, to extract the important features for the classification. The widely used classification algorithms are proximal SVM, K-nearest neighbor and multilayer ANN, it results in segregating the phase images of the different stages of AD and improves the accuracy.
3.1. Pre-processing
The aim of the pre-processing signal is to improve the broad-spectrum quality of the EEG signal, which results in the précised measurement. The EEG signal determines the cerebral function of the brain using the non-invasive technique. The raw input of the EEG signal before pre-processing is shown in the Fig. 1. The major intent of the pre-processing system is to remove the noises from the acquired EEG signal. There are different methods to extract the EEG parameters from the noisy EEG signal.
Low Pass Filter: It removes the artifacts with high frequency and it is the inverse operation of the high pass filter. The recorded EEG signals contaminates with the muscular movements, eye blink movements etc.,
High Pass Filter: It involves in filtering the low frequency artifacts occurred in the electro galvanic signals and electro oculogram signals. During recording, the drifts in electrode impedance, breathing and sweating are capable of measuring slow variations in the voltage.
Band Pass Filter: It allows the pass band frequencies in the range of 0.1 Hz to 50 Hz. These signals are filtered using the band pass filter and digitized with the sampling rate of 1KHz.
3.2. Feature Selection and Dimensionality Reduction
The idea of instantaneous linear aliasing blind signal separation serves as the foundation for independent component analysis (ICA). The nearly independent source signals' outputs are statistically distinct from one another. The non-Gaussian maximisation criteria, which employs a fixed-point recursive algorithm, is the foundation of the FastICA algorithm. determining the split signals' best mutually independent values and the cost function's optimum value. Kurtosis is used to identify a fixed-point algorithm, and a fixed-point algorithm based on negative entropy is created.
The non-Gaussian signal is measured using negative entropy, which has a better convergence impact than kurtosis. By batch processing to continually update the weight WI, the Newton iterative method determines the ideal separation matrix OPSmax, ensuring that the negative entropy is as high as possible.If the data is centralised then the information collected is determined as A, the mean value is zero and Z is replaced with A. Estimation of e, selection of the required component count, and designation of iteration count as t. Initial phase is chosen at random by Wt and arranged according to
$${W}_{t}=E\left\{{Z}_{g}\left(\frac{W}{Z}\right)\right\}-E\left\{{g}^{{\prime }}\left(\frac{W}{Z}\right)\right\}W$$
The values are acquired while incorporating the Newton Iterative Approach. P-I
$${W}_{t}={W}_{t}-L\left({W}_{t}^{T}{W}_{j}\right){W}_{j}$$
Data regularisation is completed, and the initial step is repeated if convergence is achieved. If p > m, step 2 is returned and the value of p is allocated as p = p + 1.
By modifying the successive biassed approximation of the autocorrelation data function, the autoregressive parameters in this Yule Walker method estimate the coefficients. The next step of this strategy is to minimise the forward prediction error, as shown below:
where exx can be expressed as,
Since the frequency resolution of asystematically determined AR spectrum is unlimited and independent of the length of the studied data, spectral analysis based on the AR technique is particularly useful when small data segments are evaluated.
3.3. Classification – Neuro Fuzzy System
Numerous researchers in this field are becoming interested in the neuro-fuzzy or fuzzy-neuro method because it combines the advantages of neural networks with fuzzy logic. The primary study topics for this AI method include defining the shape and scale of neural networks, as well as systematically determining fuzzy rules and membership functions. The neuro-fuzzy hybrid system combines the advantages of fuzzy logic, which deals with knowledge that can be explained and understood openly, with neural networks, which deal with knowledge that may be learnt intuitively. Many applications, particularly in process control, have used these structures.
Fuzzy logic and neural networks have several characteristics, such as distributed knowledge representation, model-free prediction, the capacity to handle input with uncertainty and imprecision, etc. While neural networks can handle noisy input, fuzzy logic can handle imperfect data. The capacity of a neural network to learn offers a useful method for modifying the knowledge of an expert, and it also automatically creates extra fuzzy rules and membership functions to satisfy specific requirements. This cuts down on both the cost and design time. The fuzzy logic technique, on the other hand, could improve a neural network's capacity to generalise by producing more accurate results when extrapolation is required outside the bounds of the training data.
Fuzzy modelling is the term used to describe the process of creating a fuzzy inference system. For the structure learning phase and the parameter learning phase, it makes use of domain knowledge. The method of using different learning approaches created in the literature on neural networks to fuzzy inference systems is referred to as neuro-fuzzy modelling. It denotes that NFIN is a fuzzy rule-based network that has the capacity to learn like a neural network. The network's primary characteristic is the absence of preassignment and the need for rule formulation. The rules are automatically created throughout the online operation. For the building job described below, two learning phases—the structural and parameter learning phases—are adopted online.
To adequately represent the available data and the number of membership functions for input and output variables, the structure learning phase entails determining the appropriate number of rules required, i.e., identifying how many rules are required and sufficient. The coefficients of each rule are adjusted throughout the parameter learning phase (like the shape and positions of membership functions). It is possible to get quick processing speed by using fewer adjustable parameters. Effective techniques are required for fine-tuning the membership functions to reduce output error measure or increase performance index.The architecture neuro fuzzy system is illustrated in Fig. 2.
In this case, one output f and two fuzzy if-then rules are taken into consideration together with two inputs x and y.
Rule 1: If x is A1 and y is B1, then f1 = p1x + q1y + r1,
Rule 2: If x is A2 and y is B2, then f2 = p2x + q2y + r2.
Following is a description of the roles played by each node in each tier.
Layer 1: Each node i in this layer has a node function and is a square node.
$${O}_{i}^{1}=\mu {A}_{i}\left(x\right)$$
where x is the input to node i, and Ai, is the linguistic label associated with this node function. In other words \({O}_{i}^{1}\), is the membership function of Ai.
Layer 2: Each node in this layer has the label "circle" is \({\Pi }\) and multiplies the incoming signals before sending the result out. For example,
$${w}_{i}=\mu {A}_{i}\left(x\right)+\mu {B}_{i}\left(Y\right), i=\text{1,2},\dots n$$
The strength of a rule's firing is represented by each node output.
Layer 3: This layer's nodes are all circle nodes with the label "N." The ith node determines the ratio of the firing strength of the ith rule to the total firing strength of all rules:
$$\stackrel{-}{{w}_{i}}=\frac{{w}_{i}}{{w}_{1}+{w}_{2}}, i=\text{1,2},\dots ..n$$
The outputs of this layer will be referred to as normalised firing strengths for simplicity.
Layer 4: Each node i in this layer has a node function and is a square node.
$${O}_{i}^{4}=\stackrel{-}{{w}_{i}}{f}_{i}=\stackrel{-}{{w}_{i}}({p}_{i}x,{q}_{i}y, {r}_{i})$$
where \(\stackrel{-}{{w}_{i}}\)is the output of layer 3, and (pi, qi, ri) is the parameter set. Parameters in this layer will be referred to as consequent parameters.
Layer 5: This layer has just one node, a circular node with the label " Σ " which computes the total output as the sum of all incoming signals.
$${O}_{i}^{5}=overall output={\sum }_{i}\stackrel{-}{{w}_{i}}{f}_{i}=\frac{{\sum }_{i}{w}_{i}{f}_{i}}{{\sum }_{i}{w}_{i}}$$
The response specification is determined by the intended percent overrun and settling time, and the Neuro-fuzzy system models the link between the controller gains and the goal output response in this way.
3.4. Supply Chain Management
Supply chain management is the practise of controlling the movement of materials and information along a supply chain to maximise customer satisfaction while minimising costs. Supply chain management enables a company to obtain a competitive edge in the market by streamlining its supply-side operations to optimise consumer value. In order to synchronise order creation, order taking, and order fulfilment, a solid guarantee that supply chain associates will cooperate closely is necessary for an effective supply chain. They so establish a vastly expanded enterprise that is independent of the manufacturer.
The pharmaceutical companies manufacture the drugs and send them to the company depot. These intermediaries store the products in their warehouse and support the companies in their inventory management. The main aim of establishing the model is to avoid the local taxes and push the products to other spaces where this facilitator will take care of its storage. Meantime there are a few small companies who take short-term leasing of the company productions as a loan licensee and produce the drugs on behalf of the main manufacture. And these companies supply the products to the end facilitators directly. At times some of these small companies do create counterfeits in the system by manufacturing more units than they are permitted to by the leasing company as per the signed agreement.
This section discusses the categorization of epilepsy using the proposed and current approaches. The Bern-Barcelona EEG database contains non-focal and focal channels with a 1024Hz sampling rate from individuals who have an incidence of epilepsy. The database contains 3750 pairs of signals from the EEG channels that were recorded, and the recorded samples were separated into slots of windows with a ten-second interval, yielding a sample size of 10240. From a publicly accessible EEG database, fifty non-focal and focal subjects were randomly selected for this investigation [21]. The experiment is completed in Matlab using its 8.00 GB of RAM and 2.30 GHz of CPU power.Performance measures of these classifiers are associated with the terminologies like True Positive (TP), False Positive (FP), True Negative (TN) and False Negative (FN). In this research work, the term positive represents the person having disease and the negative represents the person is not having the disease. The abnormal and normal EEG signals are given in Figs. 3 and 4.
-
True Positive indicatesthe diseased people is correctly identifiexd as the diseased people.
-
False Positive indicates the normal people is incorrectly classified as the diseased people.
-
True Negative indicates the normal people is correctly classified as the normal people.
-
False Negative indicatesthe diseased people is incorrectly classified as the normal people.
Based on these values the performance measures like accuracy, sensitivity, specificity, and precision are calculated and it is shown in equation.
Accuracy: The accuracy of a measurement shows how the result comes to the true value. The accuracy is called as the perfect classification. The error rate is found to be zero, because there are no false positives and true negatives are found. The numerical outcome is given in Table 1 and Fig. 5. The expression for the accuracy is shown below.
$$Accuracy=\frac{TRP+TRN}{TRP+TRN+FLN+FLP}$$
| Iteration | Existing Technique | Proposed Technique | |
|---|---|---|---|
| SVM | ANN | NFF | |
| 200 | 69.3 | 58 | 83 |
| 300 | 70.5 | 58.7 | 83.9 |
| 400 | 74.7 | 59.7 | 84.6 |
| 500 | 76.8 | 60 | 85.1 |
| 600 | 78.2 | 61.7 | 85.3 |
The approach with highest accuracy is considered as the effective approach. From the observation of experimental outcome in Table 1 and Fig. 5, it is identified that the proposed approach is effective. The accuracy of NFF for 200 iteration is higher {13.7% and 25%} than {SVM, ANN}, the accuracy of NFF for 300 iteration is higher {13.4% and 25.2%} than {SVM, ANN}, the accuracy of NFF for 400 iteration is higher {9.9% and 24.9%} than {SVM, ANN}, the accuracy of NFF for 500 iteration is higher {8.3% and 25.1%} than {SVM, ANN}, and the accuracy of NFF for 300 iteration is higher {7.1% and 23.6%} than {SVM, ANN}. The proposed approach has highest accuracy for different count of vehicular nodes.
Sensitivity: The statistical indicators of the effectiveness of the binary classification are sensitivity and specificity. It is the ability to correctly identify the true positive rate of interest and it is expressed in the below equation. Specificity measures the negative rate, which are correctly identified. A perfect interpreter is expressed as 100% Sensitivity and 100% specificity. The numerical outcome is given in Table 2 and Fig. 6. Theoretically, all the interpreter will have a minimum error.
$$Sensitivity=\frac{TRP}{TRP+FLN}$$
| Iteration | Existing Technique | Proposed Technique | |
|---|---|---|---|
| SVM | ANN | NFF | |
| 200 | 89 | 87 | 91 |
| 300 | 91 | 87.5 | 93 |
| 400 | 92 | 90 | 95 |
| 500 | 92.5 | 91 | 96 |
| 600 | 93 | 92 | 97 |
The approach with highest sensitivity is considered as the effective approach. From the observation of simulation outcome in Table 2 and Fig. 6, it is identified that the proposed approach is effective. The sensitivity of NFF for 200 iteration is higher {2% and 4%} than {SVM, ANN}, the sensitivity of NFF for 300 iteration is higher {2% and 5.5%} than {SVM, ANN}, the sensitivity of NFF for 400 iteration is higher {3% and 5%} than {SVM, ANN}, the sensitivity of NFF for 500 iteration is higher {3.5% and 5%} than {SVM, ANN}, and the sensitivity of NFF for 300 iteration is higher {4% and 5%} than {SVM, ANN}. The proposed approach has minimal sensitivity for different count of iteration.
Specificity: It is the ability of a test to correctly identify the true negative rate, i.e without any cognitive disorder. High Specific tests on the rare occasion, fail to spot the negative outcomes, so it is considered their results as positive and there is a high chance of the occurrence of disease. The numerical outcome is given in Table 3 and Fig. 7.
$$Specificity=\frac{TRN}{TRN+FLP}$$
| Iteration | Existing Technique | Proposed Technique | |
|---|---|---|---|
| SVM | ANN | NFF | |
| TRIPSENSE | REPLACE | EP-PTA | |
| 200 | 89 | 86 | 92 |
| 3.500 | 92 | 87 | 93.5 |
| 400 | 92 | 90 | 95 |
| 500 | 92.5 | 92 | 96 |
| 600 | 93.5 | 92 | 97 |
The approach with highest specificity is considered as the effective approach. From the observation of simulation outcome in Table 3 and Fig. 7, it is identified that the proposed approach is effective. The specificity of NFF for 200 iteration is higher {3% and 6%} than {SVM, ANN}, the specificity of NFF for 300 iteration is higher {1.5% and 6.5%} than {SVM, ANN}, the specificity of NFF for 400 iteration is higher {3% and 5%} than {SVM, ANN}, the specificity of NFF for 500 iteration is higher {3.5% and 4%} than {SVM, ANN}, and the specificity of NFF for 300 iteration is higher {3.5% and 5%} than {SVM, ANN}. The proposed approach has minimal specificity for different count of iteration.
Precision: It is the proportion of all successfully predicted positive observations to all positive observations that were forecasted. Low false positive rates are related to good accuracy. The numerical outcome is given in Table 4 and Fig. 8.
$$Precision=\frac{TRP}{TRP+FLP}$$
| Iteration | Existing Technique | Proposed Technique | |
|---|---|---|---|
| SVM | ANN | NFF | |
| 200 | 69.8 | 68 | 88 |
| 800 | 70.6 | 68.7 | 88.9 |
| 400 | 74.7 | 69.7 | 84.6 |
| 600 | 76.8 | 60 | 86.1 |
| 600 | 78.2 | 61.7 | 86.8 |
The approach with highest recall is considered as the effective approach. From the observation of simulation outcome in Table 4 and Fig. 8, it is identified that the proposed approach is effective. The recall of NFF for 200 iteration is higher {18.2% and 20%} than {SVM, ANN}, the recall of NFF for 300 iteration is higher {18.3% and 20.2%} than {SVM, ANN}, the recall of NFF for 400 iteration is higher {9.9% and 14.9%} than {SVM, ANN}, the recall of NFF for 500 iteration is higher {9.3% and 26.1%} than {SVM, ANN}, and the recall of NFF for 300 iteration is higher {8.6% and 25.1%} than {SVM, ANN}. The proposed approach has minimal recall for different count of iteration.
The performance measures of the above said classifiers are calculated using the various measures like Accuracy, Sensitivity, Specificity, and Precision. The set of training inputs were given to all the classifiers and the performance measures were found and reported at every stage. The average performance measures of the classifiers are shown in Tables and it indicates the good accuracy, precision, sensitivity, and lowers the specificity. It predicts the most frequent label among the neighbours. Since, NFF is a simple algorithm, the performance of the classifier is found to be lesser than the other classifiers.
Emerging technologies are being incorporated into next-generation healthcare systems to drastically improve healthcare services. As a result, this article suggests a new, cooperative health system to facilitate efficient and extensive medical data interchange. The suggested Neuro Fuzzy Framework (NFF) locates the anomaly in the EEG signal and uses the supply chain to enable secure transfer of a sizable volume of medical data produced by various health organisations. An efficient method for monitoring patients at the edge is especially important for ensuring early detection, scalability, and quick action for urgent occurrences. The Neuro Fuzzy Framework (NFF) classifier is discovered to be the superior classifier when the results are studied and contrasted based on the different factors. The improvement of medical diagnosis and supply chain management may result from this work's strengthening of the categorization process in practise.
Ethics Approval and Consent to Participate:
No participation of humans takes place in this implementation process
Human and Animal Rights:
No violation of Human and Animal Rights is involved.
Funding: No funding is involved in this work.
Conflict of Interest: Conflict of Interest is not applicable in this work.
Authorship contributions:
There is no authorship contribution
Acknowledgment :
This research was supported by the research program by Foreign Trade University (grant number FTURP02-2020-16).
Data Availability:
Based on the request authors can provide.
- Fisher, R. S., Boas, W. V. E., Blume, W., Elger, C., Genton, P., Lee, P., & Engel Jr, J. (2005). Epileptic seizures and epilepsy: definitions proposed by the International League Against Epilepsy (ILAE) and the International Bureau for Epilepsy (IBE). Epilepsia, 46(4), 470-472.
- Tzallas, A. T., Tsipouras, M. G., Tsalikakis, D. G., Karvounis, E. C., Astrakas, L., Konitsiotis, S., & Tzaphlidou, M. (2012). Automated epileptic seizure detection methods: a review study. Epilepsy-histological, electroencephalographic and psychological aspects, 75-98.
- Ocak, H. (2009). Automatic detection of epileptic seizures in EEG using discrete wavelet transform and approximate entropy. Expert Systems with Applications, 36(2), 2027-2036.
- Takahashi, H., Takahashi, S., Kanzaki, R., & Kawai, K. (2012). State-dependent precursors of seizures in correlation-based functional networks of electrocorticograms of patients with temporal lobe epilepsy. Neurological Sciences, 33(6), 1355-1364.
- Iscan, Z., Dokur, Z., & Demiralp, T. (2011). Classification of electroencephalogram signals with combined time and frequency features. Expert Systems with Applications, 38(8), 10499-10505.
- Polat, K., & Güneş, S. (2008). Artificial immune recognition system with fuzzy resource allocation mechanism classifier, principal component analysis and FFT method based new hybrid automated identification system for classification of EEG signals. Expert Systems with Applications, 34(3), 2039-2048.
- Alickovic, E., Kevric, J., & Subasi, A. (2018). Performance evaluation of empirical mode decomposition, discrete wavelet transform, and wavelet packed decomposition for automated epileptic seizure detection and prediction. Biomedical signal processing and control, 39, 94-102.
- Hassan, A. R., & Subasi, A. (2016). Automatic identification of epileptic seizures from EEG signals using linear programming boosting. computer methods and programs in biomedicine, 136, 65-77.
- Übeyli, E. D. (2010). Lyapunov exponents/probabilistic neural networks for analysis of EEG signals. Expert Systems with Applications, 37(2), 985-992.
- Altaf, M. A. B., Zhang, C., & Yoo, J. (2015). A 16-channel patient-specific seizure onset and termination detection SoC with impedance-adaptive transcranial electrical stimulator. IEEE Journal of Solid-State Circuits, 50(11), 2728-2740.
- C. Zhang, M.A. Bin Altaf, J. Yoo, Design and implementation of an on-chip patientspecific closed-loop seizure onset and termination detection system, IEEE J. Biomed. Heal. Informatics. 20 (2016) 996–1007, https://doi.org/10.1109/JBHI. 2016.2553368.
- Subasi, A., Kevric, J., & Canbaz, M. A. (2019). Epileptic seizure detection using hybrid machine learning methods. Neural Computing and Applications, 31(1), 317-325.
- Aman, A. H. M., Hassan, W. H., Sameen, S., Attarbashi, Z. S., Alizadeh, M., & Latiff, L. A. (2021). IoMT amid COVID-19 pandemic: Application, architecture, technology, and security. Journal of Network and Computer Applications, 174, 102886.
- Singh, R. P., Javaid, M., Haleem, A., Vaishya, R., & Ali, S. (2020). Internet of Medical Things (IoMT) for orthopaedic in COVID-19 pandemic: Roles, challenges, and applications. Journal of clinical orthopaedics and trauma, 11(4), 713-717.
- Dujak, D., & Sajter, D. (2019). Blockchain applications in supply chain. SMART supply network, 21-46.
- Goodarzian, F., Taleizadeh, A. A., Ghasemi, P., & Abraham, A. (2021). An integrated sustainable medical supply chain network during COVID-19. Engineering Applications of Artificial Intelligence, 100, 104188.
- Antony, M. J., Sankaralingam, B. P., Mahendran, R. K., Gardezi, A. A., Shafiq, M., Choi, J. G., & Hamam, H. (2022). Classification of EEG using adaptive SVM classifier with CSP and online recursive independent component analysis. Sensors, 22(19), 7596.
- Koh, J. E. W., Jahmunah, V., Pham, T. H., Oh, S. L., Ciaccio, E. J., Acharya, U. R., ... & Ramli, N. (2020). Automated detection of Alzheimer's disease using bi-directional empirical model decomposition. Pattern Recognition Letters, 135, 106-113.
- Zhiwei, L., & Minfen, S. (2007, August). Classification of mental task EEG signals using wavelet packet entropy and SVM. In 2007 8th international conference on electronic measurement and instruments (pp. 3-906). IEEE.
- Tufail, A. B., Abidi, A., Siddiqui, A. M., & Younis, M. S. (2012). Automatic classification of initial categories of Alzheimer's disease from structural MRI phase images: a comparison of PSVM, KNN and ANN methods. International Journal of Biomedical and Biological Engineering, 6(12), 713-717.
- R. G. Andrzejak, K. Schindler, and C. Rummel, ``Nonrandomness, nonlinear dependence, and nonstationarity of electroencephalographic recordings from epilepsy patients,'' Phys. Rev. E, Stat. Phys. Plasmas Fluids Relat.Interdiscip. Top., vol. 86, no. 4, 2012, Art. no. 046206.