View of An Optimized Prediction Of Alzheimer’s Disease Using Deep Convoltional Neural Network

(1)

17530 http://annalsofrscb.ro

An Optimized Prediction Of Alzheimer’s Disease Using Deep Convoltional Neural Network

L.Dharshana Deepthi¹

1PSNA College of Engineering and Technology,Dindigul,Tamil Nadu,India Email:*¹[email protected]

ABSTRACT

Deep Learning is a subspace of machine learning that analyzes the data with the logic of human using the algorithmic structure called Artificial Neural Network (ANN). They are highly employed in healthcare which identifies the pattern of patient’s symptoms. Alzheimer’s Disease(AD) is one of the neurogenerative disorders that is being affecting many people globally. The disease is segregated into several stages and is classified mainly as Mild, Moderate, or Severe. The symptoms includes crippled speaking and writing, inability in remembering the information and so on. Machine learning algorithm techniques such as Independent Component Analysis (ICA), Decision Tree Classifier, and Linear Discriminant Analysis (LDA), have been used to pinpoint different stages of disease, but the accuracy of identification is not considered to be good. This paper proposes a Deep Learning-based technique where the prediction is improved by employing a Convolutional Neural Network (CNN). It examines the EEG signal, features are extracted using the Fast Fourier Transform (FFT), and classification of the disease using CNN. To optimize and increase the accuracy, Adam optimization algorithm is implemented. Adam can easily handle the sparse gradients on noisy problems. It uses Momentum and Adaptive Learning rate for fast convergence..

Keywords:

Adam, Mild, Severe, Optimization

1. Introduction

Alzheimer's Disease (AD) is a neurologic disease, affects 46 million people across the world.

Forgetting recent events or conversations is an early symptom of the disorder. When the disease worsens, it causes extreme memory loss and the inability to perform daily tasks. The damage begins years before in the memory region of the brain before the first symptom appears. The other region of the brain is also affected by the loss of neurons, and eventually shrinks significantly.

Mild, Moderate, and Severe are the three levels of severity of the disease. Difficulties in using the right word, forgetting the names and misplacing or losing the valuable object is the symptom of the first stage of the disease. Then the progression goes and unable to remember own personal history and brings confusion on about what day it is and where one is. Then the severity includes physical ability changes and difficulty in communication which is considered to be the final stage. India is holding third position after China and the United States in suffering from Alzheimer’s and at the end of 2030, the disease is expected to reach 7.5 million. Various methods have been used to predict and classify the disease in advance to decrease the threat of death.

Detecting the Alzheimer's Disease at the early stage, slows down the chance of death. To perform the detection, neuroimaging such as Magnetic Resonance Imaging, Positron Emission Tomography, ElectroEncephaloGraphy are used. The electrical activity of neurons are measured by placing the terminals on the scalp. The terminal through which the electric current enters and exixts transfers the data to the machine which then records as a signal. EEG is typically divided into several regular recurrence groups, such as delta, theta, alpha, beta and gamma of 0.5-4Hz, 4- 8Hz, 8-13 Hz, 13-30 Hz and 30-40 Hz respectively. These group has a significant effect called EEG “slowing” that determines the AD. That is there will be increment in the power of low- frequency bands(delta and theta) and decrement in the higher-frequency bands(alpha and beta).

(2)

EEG produces an efficient low- cost physiological biomarker that is available at community clinics.

2.Literature Review

The supervised approach in machine learning is used to segregate different patients with Healthy Control (HC) and Mild Cognitive Impairment (MCI) or Alzheimer's disease (AD) samples. This approach is proposed by Giulia Fiscon et al. [2] which involves recording of EEG signals followed by extracting the features using Fast Fourier Transform and the Discrete Wavelet Transform. The matrix of 109 rows and 913 columns are used to organize the characteristics. The C4.5 algorithm is a classification algorithm that generates decision trees. For various databases, the result provides accurate disease stage classification.

The irregular EEG pattern "slowing" is quantified by the Relative Power (RP). Differentiation of HC and MCI is done by the relative power in deep neural network. Donghyeon Kim and Kiseon Kim [3] proposes a deep Neural Network-based classifier that calculates and normalizes the relative power of theta, alpha, and beta for discriminating the features. The number of hidden layers depends on the classifier in the classification process. Leave-One-Out cross validation is used to investigate the discussed scheme.

The recording of EEG are analysed using five EEG signals (delta, theta, alpha, beta). This method is proposed by Katerina D.Tzimmoutaa et al [4], who calculated 38 linear and non-linear features. Multi-regression Linear Analysis is used to investigate the magnitude of 38 extracted features and their MMSE ratings. According to the location of electrode, regression models are created for individual and cluster of channels. An ANOVA analysis has been carried out based on scrutinizing all the assumptions. Carmina Reyes-Coronel et al. [5] used test scores of neurology and Quantitative EEG (QEEG) markers to resolve whether RCD is suffered by an AD patient.

RCD and non-RCD patients are identified by the Support Vector Machine (SVM) after each function was checked individually. Using leave-one-out validation method, it yields 72.1 percent and 77.9 percent precision. The accuracy, sensitivity, and specificity of the classifier increase by 80.9 percent, 80 percent, and 81.1 percent, respectively, as a result of the neuropsychological test results.

Xin Hong et al. [6] suggest a model which pinpoints the patient affected from Mild Cognitive Impairment using Long Short Term Memory (LSTM). The data is taken as an image, and the processing starts by utilizing different stages. Initially skull stripping is carried out, then the process of normalization and registration has been carried out on the image. At the end of pre- processing smoothing and segmentation are taken to make a noise-free data. Once these process are done, the model’s training starts by providing sequential data with different time stage, and the state of succeeding sixth months is automatically predicted by the model. For example, during testing, when the 18th and 24th month's feature data are provided, then the state of the subject on the 30^th month is predicted by the model. Maryamossadat Aghili et al. [7] suggest an RNN-based approach which distinguishes the AD patients from the other healthy ones by analyzing the longitudinal data. Each subject's data related to the time point is given along with its final diagnosis mark in the LSTM and GRU models inorder to learn the pattern. The non-recurrent network’s results, such as Multi-Layer Perceptron, are compared to the results of the outcome models for all data arrangements .Each patient's data is entered into the MLP only once. Since there are huge number of trainable parameters, which makes the LSTM difficult, to train the sequential data, and susceptible to the problem of overfitting on the training data.

(3)

A machine learning method is proposed by Escudero J et al [8] for cost-effective detection of Alzheimer's disease. By utilizing ADNI data, the comparison of AD to controls, as well as MCI patients to AD is done for a year. It employs locally weighted learning to train the classifier and determines the biomarkers to find out the stage of the patient. As a result, we can customise the classifier by assigning the importance in preparation to the cases that are most similar to the AD patient. Siqi Liu et al. [9] proposes a deep learning model that uses auto- encoders and a softmax regression layer which is capable of analysing multiple classes. The

neurons are modified as a result of the dimensionality reduction. The softmax layers classifies the instance by taking the highest probabilities of each label. All of the MRI images have been nonlinearly recorded and segmented into 83 functional regions. Before the classification, the extraction of grey matter volume and by utilizing the ElasticNet, the features are selected. The characteristics are normalised to a zero mean and range from 0 to 1. Approximately 90% subjects are used for training and the remaining subjects are used for research in each fold of cross validation. The dark regions clearly show the distinction between different Regions of Interest (RoI). Darker ROIs seem to be the highest chance of progression of MCI and AD when compared to lighter ROIs.

Firouzeh Razavi et al. [10] propose a method for learning the expressive features of brain images that employs Sparse filtering. To categorise the conditions, the SoftMax regression is qualified.

The first phase consists of three stages: The training is done by Sparse Filtering, and the weight matrix W is estimated. Each sample's local features are extracted using the sparse filtering that is trained. The average of the local features are taken to get the learned features from each sample.

Padilla P et al. [11] presented a novel strategy where SPECT and PET databases are used. To examine the databases for the selection and the extraction, the Fisher Discriminant Proportion (FDP) and nonnegative grid factorization are used.

This paper is proposed by Hongming Li et al. [12] to determine when and where a Mild Cognitive Impaired individual can turn to an Alzheimer's patient. On comparision with cross- sectional data, the pattern classifier works well on longitudinal data. To construct a prognostic model for AD progression, the Recurrent Neural Nwteork is used to obtain insightful representations about individual’s cognitive measures and combine them with the baseline of hippocampus. To identify the transformation of Cognitively Impaired patient to the AD patients, the LSTM encoder is used to learn detailed representations about cognitive tests. The results show that separating MCI individuals from AD patients has a promising prognostic performance. Similarly, a grading biomarker suggested by Tong Tong et al. [13] for predicting the conversion of AD from MCI.

The ADNI pipeline pre-processes MRI brain images. The function is chosen using a sparse regression technique after pre-processing. For an input of millions of features, an elastic net is used in this study. The collection of features in this study is based on AD and NC topics. For each MCI subject, a grading value is determined and used as a biomarker for classification. A weighting function is used to model the relationship between the MCI subject and the training population. The Elastic Net technique is used for Sparse Representation. SVM with a linear kernel was used to train a classifier with a single type of function. The SVM is implemented with the aid of liblinear.

3.Methodology

The work uses EEG signals as an input dataset. A deep CNN network architecture is used to segregate different stages of the disease by considering the input signal. The following modules are included in this work:

(4)

𝑛=0

3.1.Data Preprocess

The ElectroEncephaloGram signal is obtained from the medical research data repository physionet and the signal is in the form of European Data Format (EDF). Preprocess the signal with a Low Pass Filter if it includes any noise or artifacts. The signal's noise involves eyeball moments, muscle contractions, and so on. To make the low frequency signal and exclude certain noises, the threshold value of Low Pass Filter is set to high.

3.2.Feature Extraction

FFT is used to extract the features from preprocessed data. It selects the feature subset from the data in the model’s construction. The time domain function is converted to the frequency domain using FFT. The following is a description of the Discrete Fourier Transform:

𝑋𝑘 = ∑𝑁−1 𝑥𝑛𝑒−𝑖𝜋2𝑘𝑛/𝑁

(1) where k, n, N represents the frequency domain, time domain and the length of sequence respectively. The FFT decreases the time of computation from O(N²) to O(N) as compared to the Discrete Fourier Transform O(NlogN). Different relative band powers such as delta, theta, alpha, and beta are produced using FFT. Then the features such as power spectrum's magnitude, frequency vector, signal order, signal coefficients, and the average of the band signals are extracted. These characteristics allows a more accurate classification of the dataset.

3.3. Classification

Classification is a method of selecting the group to which a new observation belongs, which is based on the determined training set. Convolutional Neural Networks (CNN) are used to classify the data. Image and Signals are easily analysed using the deep learning neural network known as CNN. The features are provided to the CNN, which automatically train the given dataset.

Figure 1 Convolutional Neural Network

The signal is accepted as input by the input layer, as shown in Figure 1. The Hidden Layers extract features by conducting calculations and manipulation of data. The function extraction is

(5)

𝐧=𝟎

performed by several hidden layers such as Convolution Layer, ReLU Layer, Pooling Layer, Flattening layer and so on. The Fully Connected Layer in the Output Layer determines which group the input belongs to.

The proposed CNN's building block is depicted in Figure 2.The first layer in the CNN is Convolution Layer, where the convolution is done using a filter (2).

𝐂𝐦 = ∑𝐍−𝟏 𝐟𝐧𝐤𝐦−𝐧

(2) where k is the signal, f is the filter, N is the number of datapoints in k and C denotes the output.

Figure 2 Building Block of CNN

The rectified feature map is obtained by applying the activation function to the convolution layer.

The number of parameters and computation is reduced using Pooling Layer. The signal is recognised and classified by the Fully Connected Layer (3).

𝑥𝑖 = ∑𝑗𝑤𝑗𝑖𝑦𝑗 + 𝑏𝑖 (3)

where w,y,b and x denotes weights, previous layer's output, bias and current layer's output respectively. The softmax function, which defines the output group, receives the output from the last completely connected layer. In training, the model takes the input signals and the disease is classified to different stages using the classifier as described in the architecture.

Figure 3 Training phase of the prediction system

The training process of the prediction system is depicted by the architecture in Figure 3. The dataset is given to the model, where the Low Pass Filter processes the input signal, FFT extracts

(6)

the features and feed the CNN with the extracted features to make a prediction of classifying the disease. Different datasets are taken to train the model which makes them more effective.

The testing process of the model is depicted in Figure 4. The input samples are used to train the qualified classifier, and the Low Pass Filter pre-process the testing data before the FFT extracts the features. The extracted features are then fed into a qualified classifier, which predicts the result as Mild or Severe.

Figure 4 Testing phase of the prediction system 4.Results and Discussions

The work uses the Matlab software in which the EEG signal is taken in the form of EDF, then the signal is pre-processed and plotted as a graph which represents the time in x-axis and amplitude in y-axis as shown in Figure 5.

Figure 5 Representaion of EEG input signal

The Fast Fourier Transform extracts the features like power spectrum magnitude, coefficient of the signal, order of the signal and average of the band power as shown in Figure 6. It can be represented as a frequency vs amplitude graph.

Figure 6 Representaion of Power Spectrum

Different frequency relative band power signals are depicted is given. The x-axis shows the number of samples, and the y-axis is the amplitude. The lower band signals like delta and theta

(7)

gains an increment as shown in Figure 7.1 and Figure 7.2 and on the other hand the higher band signals like alpha and beta faces decrement as shown in the Figure 7.3, Figure 7.4.

Figure 7.1

Figure 7.2

Figure 7.3

Figure 7.4

The Convolutional Neural Network uses these signals and their average relative band power to distinguish disease stages and accuracy is taken as shown in Figure 7.5 and Figure 7.6.

Figure 7.5 Classification Result

(8)

Figure 7.6 Accuracy of the classification

In EEG classification, the band waves delta, theta, alpha and beta waves gets fluctuated in some cases like sleepy and drowsy in the given dataset. The accuracy and loss of such waves are predicted as depicted in Figure 7.7.

Figure 7.7 Accuracy and Loss of classification

To improve the accuracy Adaptive Moment Estimation(Adam) is used. It collobrates the RMSprop(Root-Mean-Square prop) and momentum based Gradient Descent. In Adam, the momentum GD holds the history of updates that is being taken and the learning rate provided by RMSprop makes Adam a powerful since it combines the Adagrad and RMSprop. It is a first order gradient based algorithm which is computationally efficient, need little memory,is invariant to diagonal rescale of the gradients and suits best for problrms that contains large data or parameters.

Adaptive learning rates, also known as Learning Rate Schedules, are changes to the learning rate in the training process by reducing the learning rate to a pre-defined schedule, as seen in AdaGrad, RMSprop, Adam, and AdaDelta.Gradient Descent Optimization resolves the AdaGrad’s diminishing learmimg rates. The gradient of the accuracy and loss are taken in Adam Optimizer to improve the accuracy and decrease the loss.

(9)

Figure 7.8 Accuracy and Loss of the classification after using Adam

Figure 7.8 depicts the accuracy of nearly 90% and above after using Adam which is very efficient than the learning without using the optimizer. The loss function is also reduced to below 0.5%.

The Adam works perfectly for large amount of data with gradient.

5.Conclusion

This work uses the EEG signal which is less expensive on comparing with MRI and all the other scans. From the pre-processed signal, the required features are extracted by the Fast Fourier Transform. The classification operation is carried out by the Convolutional Neural Network using the extracted features. The CNN which divides the signal into the classes called Mild and Severe.

The CNN in deep learning is more feature compatible and can be utilized for different spatial data. Then Adam Optimizer is used which is memory efficient, improves the accuracy and reduces the loss.

Limitations and Future Studies

The work that is proposed above can be extended to predict and analyze the disease classification in image processing of various datasets. The described method can also be used to pinpoint various Parkinson’s and other Seizure related diseases that uses the EEG signal

References

[1] L.Dharshana Deepthi, Dr.D.Shanthi, Dr.M.Buvana, “An Intelligent Alzheimer’s Disease Prediction using Convolutional Neural Network” International Journal of Advanced Research in Engineering and Technology (IJARET),Volume 11, Issue 4, April 2020, pp. 12-22, Article ID: IJARET_11_04_003.

[2] Giulia Fiscon,Emanuel Weitschek, Alessio Cialini, Giovanni Felici, ”Combining EEG signal processing with supervised methods for Alzheimer patient’s classification”, BMC Medical Informatics and Decision Making, pp.18-35,doi:

https://doi.org/10.1186/s12911-018-0613-y, 2018.

[3] Donghyeon Kim, Kiseon Kim, ”Detection of Early Stage Alzheimer’s Disease using EEG Relative Power with Deep Neural Network”, Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), 2018.

(10)

[4] Katerina D. Tzimourta, Theodora Afrantou , Panagiotis Ioannidis, ”Analysis of electroencephalographic signals complexity regarding Alzheimer’s Disease ”, Computers and Electrical Engineering Vol.76, pp.198–

212,doi:https://doi.org/10.1016/j.compeleceng.2019.03.018, 2019.

[5] Carmina Reyes-Coronel, Markus Waser, Heinrich Garn, ” Predicting Rapid Cognitive Decline In Alzheimer’s Disease Patients Using Quantitative Eeg Markers And Neuropsychological Test Scores”,IEEE Engineering in Medicine and Biology Society, 2016.

[6] Xin Hong, Rongjie Lin, Chenhui Yang,”Predicting Alzheimer's Disease Using LSTM”,IEEE Access, Vol.7, pp. 2169-3536, 2019.

[7] Maryamossadat Aghili, Solale Tabarestani, Malek Adjouadi,” Predictive Modeling of Longitudinal Data for Alzheimer’s Disease Diagnosis Using RNNs ”, pp. 112–

119,https://doi.org/10.1007/978-3-030-00320-3_14, 2018.

[8] Escudero, J., Ifeachor, E., Zajicek, J. P., Green, C., Shearer, J, & Pearson, S.,“Machine Learning-Based Method for Personalized and Cost-Effective Detection of Alzheimer's Disease”,IEEE Transactions on Biomedical Engineering, Vol.60(1), pp.164-168, doi:10.1109/tbme.2012, 2013.

[9] Siqi Liu , Sidong Liu, Ron Kikinis , Dagan Feng ,” Early Diagnosis Of Alzheimer’s Disease With Deep Learning”,IEEE 11th International Symposium on Biomedical Imaging (ISBI), 2014 .

[10] Firouzeh Razavi, Mohammad Jafar Tarokh and Mahmood Alborzi,” An intelligent Alzheimer’s disease diagnosis method using unsupervised feature learning”, doi:

https://doi.org/10.1186/s40537-019-0190-7, 2019.

[11] Padilla, P., Lopez, M., Gorriz, J. M., Ramirez, J., Salas-Gonzalez, D., & Alvarez, I ,

“NMF- SVM Based CAD Tool Applied to Functional Brain Images for the Diagnosis of Alzheimer’s Disease”, IEEE Transactions on Medical Imaging, Vol.31(2), doi:10.1109/tmi.2011.2167628,pp.207–216, 2012.

[12] Hongming Li anad Yong Fan,” Early Prediction Of Alzheimer’s Disease Dementia Based On Baseline Hippocampal MRI And 1-Year Follow-Up Cognitive Measures Using Deep Recurrent Neural Networks”, IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019), 2019.

[13] Tong Tong, Qinquan Gao, Ricardo Guerrero, Christian Ledig, Liang Chen Daniel Rueckert and the Alzheimer’s Disease Neuroimaging Initiative(ADNI),”A Novel Gradding Biomarker for the Prediction of Conversion from Mild Cognitive Impairment to Alzheimer’s Diseasse”, IEEE Transactions on Biomedical Engineering, pp. 0018- 9294,doi: 10.1109/TBME.2016.2549363, 2015.

[14] Charlotte Cecere, Christen Corrado, Robi Polikar ,”Diagnostic Utility of EEG Based Biomarkers for Alzheimer’s Disease”,IEEE Annual Northeast Bioengineering Conference, 2014.

[15] Somenath Bera & Vimal K. Shrivastava,” Analysis of various optimizers on deep convolutional neural network model in the application of hyperspectral remote sensing image classification”, International Journal of Remote Sensing, ISSN: 0143-1161,

(11)

2020.

[16] K. Kranthi Kumar, D.N.V.S.L.S. Indira, Brahmaiah Madamanchi, Aravinda Kasukurthi, and Vinay Kumar Dasari,” An Efficient Image Classification of Malaria Parasite Using Convolutional Neural Network and ADAM Optimizer”, Turkish Journal of Computer and Mathematics Education, Vol.12 No 2(2021), 3376-3384,2021.