Jen Hong Tan

APPLICATION OF EMPIRICAL MODE DECOMPOSITION (EMD) FOR AUTOMATED DETECTION OF EPILEPSY USING EEG SIGNALS

Epilepsy is a global disease with considerable incidence due to recurrent unprovoked seizures. These seizures can be noninvasively diagnosed using electroencephalogram (EEG), a measure of neuronal electrical activity in brain recorded along scalp. EEG is highly nonlinear, nonstationary and non-Gaussian in nature. Nonlinear adaptive models such as empirical mode decomposition (EMD) provide intuitive understanding of information present in these signals. In this study a novel methodology is proposed to automatically classify EEG of normal, inter-ictal and ictal subjects using EMD decomposition. EEG decomposition using EMD yields few intrinsic mode functions (IMF), which are amplitude and frequency modulated (AM and FM) waves. Hilbert transform of these IMF provides AM and FM frequencies. Features such as spectral peaks, spectral entropy and spectral energy in each IMF are extracted and fed to decision tree classifier for automated diagnosis. In this work, we have compared the performance of classification using two types of decision trees (i) classification and regression tree (CART) and (ii) C4.5. We have obtained the highest average accuracy of 95.33%, average sensitivity of 98%, and average specificity of 97% using C4.5 decision tree classifier. The developed methodology is ready for clinical validation on large databases and can be deployed for mass screening.

Automated diagnosis of epilepsy using CWT, HOS and texture parameters.

Epilepsy is a chronic brain disorder which manifests as recurrent seizures. Electroencephalogram (EEG) signals are generally analyzed to study the characteristics of epileptic seizures. In this work, we propose a method for the automated classification of EEG signals into normal, interictal and ictal classes using Continuous Wavelet Transform (CWT), Higher Order Spectra (HOS) and textures. First the CWT plot was obtained for the EEG signals and then the HOS and texture features were extracted from these plots. Then the statistically significant features were fed to four classifiers namely Decision Tree (DT), K-Nearest Neighbor (KNN), Probabilistic Neural Network (PNN) and Support Vector Machine (SVM) to select the best classifier. We observed that the SVM classifier with Radial Basis Function (RBF) kernel function yielded the best results with an average accuracy of 96%, average sensitivity of 96.9% and average specificity of 97% for 23.6 s duration of EEG data. Our proposed technique can be used as an automatic seizure monitoring software. It can also assist the doctors to cross check the efficacy of their prescribed drugs.

APPLICATION OF INTRINSIC TIME-SCALE DECOMPOSITION (ITD) TO EEG SIGNALS FOR AUTOMATED SEIZURE PREDICTION

Intrinsic time-scale decomposition (ITD) is a new nonlinear method of time-frequency representation which can decipher the minute changes in the nonlinear EEG signals. In this work, we have automatically classified normal, interictal and ictal EEG signals using the features derived from the ITD representation. The energy, fractal dimension and sample entropy features computed on ITD representation coupled with decision tree classifier has yielded an average classification accuracy of 95.67%, sensitivity and specificity of 99% and 99.5%, respectively using 10-fold cross validation scheme. With application of the nonlinear ITD representation, along with conceptual advancement and improvement of the accuracy, the developed system is clinically ready for mass screening in resource constrained and emerging economy scenarios.

Automated detection of coronary artery disease using different durations of ECG segments with convolutional neural network

Coronary artery disease (CAD) is caused due by the blockage of inner walls of coronary arteries by plaque. This constriction reduces the blood flow to the heart muscles resulting in myocardial infarction (MI). The electrocardiogram (ECG) is commonly used to screen the cardiac health. The ECG signals are nonstationary and nonlinear in nature whereby the transient disease indicators may appear randomly on the time scale. Therefore, the procedure to diagnose the abnormal beat is arduous, time consuming and prone to human errors. The automated diagnosis system overcomes these problems. In this study, convolutional neural network (CNN) structures comprising of four convolutional layers, four max pooling layers and three fully connected layers are proposed for the diagnosis of CAD using two and five seconds durations of ECG signal segments. Deep CNN is able to differentiate between normal and abnormal ECG with an accuracy of 94.95%, sensitivity of 93.72%, and specificity of 95.18% for Net 1 (two seconds) and accuracy of 95.11%, sensitivity of 91.13% and specificity of 95.88% for Net 2 (5 s). The proposed system can help the clinicians in their accurate and reliable decision making of CAD using ECG signals.

A deep convolutional neural network model to classify heartbeats

The electrocardiogram (ECG) is a standard test used to monitor the activity of the heart. Many cardiac abnormalities will be manifested in the ECG including arrhythmia which is a general term that refers to an abnormal heart rhythm. The basis of arrhythmia diagnosis is the identification of normal versus abnormal individual heart beats, and their correct classification into different diagnoses, based on ECG morphology. Heartbeats can be sub-divided into five categories namely non-ectopic, supraventricular ectopic, ventricular ectopic, fusion, and unknown beats. It is challenging and time-consuming to distinguish these heartbeats on ECG as these signals are typically corrupted by noise. We developed a 9-layer deep convolutional neural network (CNN) to automatically identify 5 different categories of heartbeats in ECG signals. Our experiment was conducted in original and noise attenuated sets of ECG signals derived from a publicly available database. This set was artificially augmented to even out the number of instances the 5 classes of heartbeats and filtered to remove high-frequency noise. The CNN was trained using the augmented data and achieved an accuracy of 94.03% and 93.47% in the diagnostic classification of heartbeats in original and noise free ECGs, respectively. When the CNN was trained with highly imbalanced data (original dataset), the accuracy of the CNN reduced to 89.07%% and 89.3% in noisy and noise-free ECGs. When properly trained, the proposed CNN model can serve as a tool for screening of ECG to quickly identify different types and frequency of arrhythmic heartbeats.

Computer aided diagnosis of atrial arrhythmia using dimensionality reduction methods on transform domain representation

Abstract Electrocardiogram (ECG) is a P-QRS-T wave, representing the depolarization and repolarization mechanism of the heart. Among different cardiac abnormalities, the atrial fibrillation (AF) and atrial flutter (AFL) are frequently encountered medical emergencies with life threatening complications. The clinical features of ECG, the amplitude and intervals of different peaks depict the functioning of the heart. The changes in the morphological features during various pathological conditions help the physician to diagnose the abnormality. These changes, however, are very subtle and difficult to correlate with the abnormalities and demand a lot of clinical acumen. Hence a computer aided diagnosis (CAD) tool can help physicians significantly. In this paper, a general methodology is presented for automatic detection of the normal, AF and AFL beats of ECG. Four different methods are investigated for feature extraction: (1) the principal components (PCs) of discrete wavelet transform (DWT) coefficients, (2) the independent components (ICs) of DWT coefficients, (3) the PCs of discrete cosine transform (DCT) coefficients, and (4) the ICs of DCT coefficients. Three different classification techniques are explored: (1) K -nearest neighbor ( K NN), (2) decision tree (DT), and (3) artificial neural network (ANN). The methodology is tested using data from MIT BIH arrhythmia and atrial fibrillation databases. DCT coupled with ICA and K NN yielded the highest average sensitivity of 99.61%, average specificity of 100%, and classification accuracy of 99.45% using ten fold cross validation. Thus, the proposed automated diagnosis system provides high reliability to be used by clinicians. The method can be extended for detection of other abnormalities of heart and to other physiological signals.

Deep convolutional neural network for the automated detection and diagnosis of seizure using EEG signals

An encephalogram (EEG) is a commonly used ancillary test to aide in the diagnosis of epilepsy. The EEG signal contains information about the electrical activity of the brain. Traditionally, neurologists employ direct visual inspection to identify epileptiform abnormalities. This technique can be time-consuming, limited by technical artifact, provides variable results secondary to reader expertise level, and is limited in identifying abnormalities. Therefore, it is essential to develop a computer-aided diagnosis (CAD) system to automatically distinguish the class of these EEG signals using machine learning techniques. This is the first study to employ the convolutional neural network (CNN) for analysis of EEG signals. In this work, a 13-layer deep convolutional neural network (CNN) algorithm is implemented to detect normal, preictal, and seizure classes. The proposed technique achieved an accuracy, specificity, and sensitivity of 88.67%, 90.00% and 95.00%, respectively.

Segmentation of optic disc, fovea and retinal vasculature using a single convolutional neural network

We have developed and trained a convolutional neural network to automatically and simultaneously segment optic disc, fovea and blood vessels. Fundus images were normalized before segmentation was performed to enforce consistency in background lighting and contrast. For every effective point in the fundus image, our algorithm extracted three channels of input from the point’s neighbourhood and forwarded the response across the 7-layer network. The output layer consists of four neurons, representing background, optic disc, fovea and blood vessels. In average, our segmentation correctly classified 92.68% of the ground truths (on the testing set from Drive database). The highest accuracy achieved on a single image was 94.54%, the lowest 88.85%. A single convolutional neural network can be used not just to segment blood vessels, but also optic disc and fovea with good accuracy.

Automated screening system for retinal health using bi-dimensional empirical mode decomposition and integrated index

Posterior Segment Eye Diseases (PSED) namely Diabetic Retinopathy (DR), glaucoma and Age-related Macular Degeneration (AMD) are the prime causes of vision loss globally. Vision loss can be prevented, if these diseases are detected at an early stage. Structural abnormalities such as changes in cup-to-disc ratio, Hard Exudates (HE), drusen, Microaneurysms (MA), Cotton Wool Spots (CWS), Haemorrhages (HA), Geographic Atrophy (GA) and Choroidal Neovascularization (CNV) in PSED can be identified by manual examination of fundus images by clinicians. However, manual screening is labour-intensive, tiresome and time consuming. Hence, there is a need to automate the eye screening. In this work Bi-dimensional Empirical Mode Decomposition (BEMD) technique is used to decompose fundus images into 2D Intrinsic Mode Functions (IMFs) to capture variations in the pixels due to morphological changes. Further, various entropy namely Renyi, Fuzzy, Shannon, Vajda, Kapur and Yager and energy features are extracted from IMFs. These extracted features are ranked using Chernoff Bound and Bhattacharyya Distance (CBBD), Kullback-Leibler Divergence (KLD), Fuzzy-minimum Redundancy Maximum Relevance (FmRMR), Wilcoxon, Receiver Operating Characteristics Curve (ROC) and t-test methods. Further, these ranked features are fed to Support Vector Machine (SVM) classifier to classify normal and abnormal (DR, AMD and glaucoma) classes. The performance of the proposed eye screening system is evaluated using 800 (Normal=400 and Abnormal=400) digital fundus images and 10-fold cross validation method. Our proposed system automatically identifies normal and abnormal classes with an average accuracy of 88.63%, sensitivity of 86.25% and specificity of 91% using 17 optimal features ranked using CBBD and SVM-Radial Basis Function (RBF) classifier. Moreover, a novel Retinal Risk Index (RRI) is developed using two significant features to distinguish two classes using single number. Such a system helps to reduce eye screening time in polyclinics or community-based mass screening. They will refer the patients to main hospitals only if the diagnosis belong to the abnormal class. Hence, the main hospitals will not be unnecessarily crowded and doctors can devote their time for other urgent cases.

Novel risk index for the identification of age-related macular degeneration using radon transform and DWT features

Age-related Macular Degeneration (AMD) affects the central vision of aged people. It can be diagnosed due to the presence of drusen, Geographic Atrophy (GA) and Choroidal Neovascularization (CNV) in the fundus images. It is labor intensive and time-consuming for the ophthalmologists to screen these images. An automated digital fundus photography based screening system can overcome these drawbacks. Such a safe, non-contact and cost-effective platform can be used as a screening system for dry AMD. In this paper, we are proposing a novel algorithm using Radon Transform (RT), Discrete Wavelet Transform (DWT) coupled with Locality Sensitive Discriminant Analysis (LSDA) for automated diagnosis of AMD. First the image is subjected to RT followed by DWT. The extracted features are subjected to dimension reduction using LSDA and ranked using t-test. The performance of various supervised classifiers namely Decision Tree (DT), Support Vector Machine (SVM), Probabilistic Neural Network (PNN) and k-Nearest Neighbor (k-NN) are compared to automatically discriminate to normal and AMD classes using ranked LSDA components. The proposed approach is evaluated using private and public datasets such as ARIA and STARE. The highest classification accuracy of 99.49%, 96.89% and 100% are reported for private, ARIA and STARE datasets. Also, AMD index is devised using two LSDA components to distinguish two classes accurately. Hence, this proposed system can be extended for mass AMD screening.