Biomedical Engineering: Applications, Basis and Communications

The electrocardiogram (ECG) is a principal signal employed to automatically diagnose cardiovascular disease in shallow and deep learning models. However, ECG feature extraction is required and this may reduce diagnosis accuracy in traditional shallow learning models, while backward propagation (BP) algorithm used by the traditional deep learning models has the disadvantages of local minimization and slow convergence rate. To solve these problems, a new deep learning algorithm called deep kernel extreme learning machine (DKELM) is proposed by combining the extreme learning machine auto-encoder (ELM-AE) and kernel ELM (KELM). In the new DKELM architecture with $M+1$ hidden layers, ELM-AEs are employed by the front $M$ hidden layers for feature extraction in the unsupervised learning process, which can effectively extract abstract features from the original ECG signal. To overcome the “dimension disaster” problem, the kernel function is introduced into ELM to act as classifier by the $M+1$th hidden layer in the supervised learning process. The experiments demonstrate that DKELM outperforms the BP neural network, support vector machine (SVM), extreme learning machine (ELM), deep auto-encoder (DAE), deep belief network (DBN) in classification accuracy. Though the accuracy of convolutional neural network (CNN) is almost the same as DKELM, the computing time of CNN is much longer than DKELM.

Sleep Stage Classification (SSC) is a standard process in the Polysomnography (PSG) for studying sleep patterns and events. The SSC provides sleep stage information of a patient throughout an entire sleep test. A physician uses results from SSCs to diagnose sleep disorder symptoms. However, the SSC data processing is time-consuming and requires trained sleep technicians to complete the task. Over the years, researchers attempted to find alternative methods, which are known as Automatic Sleep Stage Classification (ASSC), to perform the task faster and more efficiently. Proposed ASSC techniques usually derived from existing statistical methods and machine learning (ML) techniques. The objective of this study is to develop a new hybrid ASSC technique, Multi-Layer Hybrid Machine Learning Model (MLHM), for classifying sleep stages. The MLHM blends two baseline ML techniques, Decision Tree (DT) and Support Vector Machine (SVM). It operates on a newly developed multi-layer architecture. The multi-layer architecture consists of three layers for classifying $W$, $R$ and $N1$, $N2$, $N3$ in different epoch lengths. Our experiment design compares MLHM and baseline ML techniques and other research works. The dataset used in this study was derived from the ISRUC-Sleep database comprising of 100 subjects. The classification performances were thoroughly reviewed using the hold-out and the 10-fold cross-validation method in both subject-specific and subject-independent classifications. The MLHM achieved a certain satisfactory classification results. It gained 0.694$\pm$0.22 of accuracy ($\mathrm{\text{AUC}}=0.822\pm 0.31$) in subject-specific classification and 0.942$\pm$0.02 of accuracy ($\mathrm{\text{AUC}}=0.920\pm 0.17$) in subject-independent classification. The pros and cons of the MLHM with the multi-layer architecture were thoroughly discussed. The effect of class imbalance was rationally discussed towards the classification results.

In the present study, a method for classifying the different ictal stages in electroencephalogram (EEG) signals is proposed. The main symptoms of epilepsy are indicated by ictal activities, which trigger widespread neurological disorders other than stroke and thus affect the world population. In this work, a novel ictal classification method that combines the spectral and temporal features of twin components in Hilbert–Huang transform is proposed. Spectral features of instantaneous amplitude (IA) function are obtained based on the power spectral density of autoregressive (AR) modeling. Here four different cases of ictal activities of EEG signal are classified. In each case first and second intrinsic mode function of Hilbert–Huang transform are tabulated. The power spectral density of AR(6) and AR(10) model are done for IA1 and IA2 components of each case. Temporal features of either instantaneous frequency (IF) function or IA are computed. The feature vectors are tested in a well-known database of different classes in interictal, ictal, and normal activities of EEG signals. The discriminating power of each vector is evaluated through one-way analysis of variance, and the classification results are verified using an artificial neural network (ANN) classifier. The performance of the classifier was assessed in term of sensitivity, specificity, and total classification accuracy. The spectral features of the AR(10) of IA and the temporal features of IA yielded 100% accuracy, 100% sensitivity, and 100% specificity in the ictal classification. By contrast, these features obtained only 83.33% of the total classification accuracy in ictal and interictal EEG signal.

Cancer causes chronic stress and is associated with impaired autonomic nervous system (ANS). Heart rate variability (HRV) has been suggested to be an important tool in the identification and prediction of performance status (PS) in cancer. Lead II surface electrocardiogram (ECG) was recorded from 24 pulmonary metastases (PM) subjects and 30 healthy controls for nonlinear HRV analysis. Artificial neural network (ANN) and support vector machine (SVM) were applied for the prediction analysis. Analysis of variance (ANOVA) along with post-hoc Tukey’s HSD test was conducted using statistical R, 64-bit, v.3.3.2, at $p\le 0.05$. The obtained results suggested lower HRV that increases with cancer severity from the Eastern Cooperative Oncology Group (ECOG)1 PS to ECOG4 PS. ANOVA results stated that approximate entropy (ApEn) ($F$-$\mathrm{\text{statistic}}=5.0821$, $p=0.001661$), detrended fluctuation analysis (DFA) ${\alpha }_2$ ($F$-$\mathrm{\text{statistic}}=3.2332$, $p=0.01969$) and correlation dimension (CD) ($F$-$\mathrm{\text{statistic}}=5.3955$, $p=0.001111$) were significant. The 13 nonlinear features were fed to ANN and SVM to obtain 82.25% and 100% accuracies, respectively. Nonlinear HRV analysis has given promising results in the prediction of diagnosis of PS in PM patients. These inputs would be very useful for clinicians to diagnose PS in their cancer patients and improve their quality of living.

Herniation in the lumbar area is one of the most common diseases which results in lower back pain (LBP) causing discomfort and inconvenience in the patients’ daily lives. A computer aided diagnosis (CAD) system can be of immense benefit as it generates diagnostic results within a short time while increasing precision of diagnosis and eliminating human errors. We have proposed a new method for automatic diagnosis of lumbar disc herniation based on clinical MRI data. We use T2-W sagittal and myelograph images. The presented method has been applied on 30 clinical cases, each containing 7 discs (210 lumbar discs) for the herniation diagnosis. We employ Otsu thresholding method to extract the spinal cord from MR images of lumbar disc. A third order polynomial is then aligned on the extracted spinal cords, and by the end of preprocessing stage, all the T2-W sagittal images will have been prepared for specifying disc boundary and labeling. Having extracted an ROI for each disc, we proceed to use intensity and shape features for classification. The extracted features have been selected by Local Subset Feature Selection. The results demonstrated 91.90%, 92.38% and 95.23% accuracy for artificial neural network, K-nearest neighbor and support vector machine (SVM) classifiers respectively, indicating the superiority of the proposed method to those mentioned in similar studies.

Alginate hydrogel fibers are receiving a great attention for tissue engineering applications. However, an important limitation of alginate is that it does not provide cell adhesion motifs. In this work, albumin was blended with alginate to improve the compatibility of alginate fibers with cells. Cell laden alginate/albumin fibers, potentially usable for skin regeneration, were obtained through a spinning process, by extruding an alginate/albumin solution containing cells into a calcium chloride solution. Cell laden pure alginate fibers were prepared for comparison. Plain alginate and alginate/albumin fibers were also produced. Morphological, mechanical and functional properties of the produced fibers were investigated. In addition, the ability of the fibers to release albumin and to support the viability and growth of A549 cells embedded into them was studied. Fibers with a uniform shape and an average diameter within the range 550–570$\mu$m were produced. The water content was $94.6\pm 0.4$% for alginate fibers, and $94.9\pm 0.3$% for alginate/albumin fibers. Stress–strain tests showed, up to a strain value of 20%, the same Young’s modulus for the produced fibers, regardless of the presence of albumin. Overall, obtained results demonstrated that morphology, size, hydrophilicity and mechanical properties were not affected by albumin. Albumin was gradually released over a period of 4 days, with a residual amount (13%) remaining into the fibers. Viability test was carried out on A549 cells, laden inside alginate and alginate/albumin fibers, to evaluate cell proliferation ability. A favorable effect of albumin on the loaded cells was evidenced by a faster kinetics of growth.

There is a growing tendency for the concealment of secure information into electrocardiogram (ECG) signals in a way that the embedded ECGs still remain diagnosable. The average length of ECG recording for a primary diagnosis takes no longer than 1min yielding to limit its concealment capacity. To overcome this drawback, we enhanced both concealment capacity and embedding quality by: (I) using 12-lead ECGs to span more embedding space, (II) shuffling input message bits via nonlinear feedback shift register (NLFSR) method, (III) inserting the selected bits of each channel into the high-frequency wavelet coefficients of non-QRS parts. Inserting the message bits into high frequency coefficients of less important ECG parts leads to preserve the quality watermarked ECGs. To assess the proposed method, a text containing different letters (changes with size of both non-QRS segments and high-frequency sub-band) was hidden through 12-lead ECG signals of 56 randomly selected subjects of PTB database, where each signal length is 10s. The performance was compared to state-of-the-art ECG-based steganography schemes in terms of the following criteria: percentage residual difference (PRD), peak signal to noise ratio (PSNR), structural similarity index measure (SSIM) and bit error rate (BER). Our results showed that the proposed scheme has benefits of fast computing along with secure embedding, providing high capacity of data hiding.

The firing patterns and recruitment behavior of motor units (MUs) during a muscle contraction can be used in diagnosing neuromuscular disorder, studying motor control, improving the performance of electromyography (EMG) signal decomposition, and assessing the validity of MU potential trains extracted by an EMG decomposition algorithm. However, MU firing patterns extracted via EMG decomposition might contain several missed or erroneous that can lead to misleading conclusion. In this paper, we presented a multiple model estimation system (MMES) for estimating the mean ($\mu$) and standard deviation ($\sigma$) of inter-discharge intervals (IDIs) of a MU. The presented MMES aggregates an existing error-filtered estimation (EFE) algorithm and a multiple linear regression model to estimate both $\mu$ and $\sigma$. The MMES estimates these two parameters using 10 features extracted from given IDIs and initial estimation of $\mu$ and $\sigma$ values provided by EFE algorithm. Evaluation results using both simulated and real IDIs revealed that MMES performed better than EFE algorithm in estimating both $\mu$ and $\sigma$ values in terms of root mean square error (RMSE), estimating variance and the range of estimated values. The RMSE values for MMES in estimating $\mu$ for simulated and real IDIs, respectively, were $6.2\pm 0.4$ and $2.2\pm 0.6$ that are statistically lower than that of EFE algorithm with RMSE $9.3\pm 1.1$ and $4.1\pm 3.2$. In estimating $\sigma$ in simulated and real IDIs, RMSE values of MMES, respectively were $4.7\pm 0.2$ and $1.9\pm 0.4$ that was significantly lower than those values obtained for EFE $6.3\pm 0.4$ (for simulated IDIs) and $3.7\pm 3.0$ (for real IDIs). More importantly, MMES outperformed EFE algorithm for real train with right-skewed IDI distribution. Consequently, MMES is more accurate, reliable and consistent than EFE algorithm for estimating IDI mean and standard deviation.