Speech feature learning is the core and key of speech recognition method for mental illness. Deep feature learning can automatically extract speech features, but it is limited by the problem of small samples. Traditional feature extraction (original features) can avoid the impact of small samples, but it relies heavily on experience and is poorly adaptive. To solve this problem, this paper proposes a deep embedded hybrid feature sparse stack autoencoder manifold ensemble algorithm. Firstly, based on the prior knowledge, the psychotic speech features are extracted, and the original features are constructed. Secondly, the original features are embedded in the sparse stack autoencoder (deep network), and the output of the hidden layer is filtered to enhance the complementarity between the deep features and the original features. Third, the L1 regularization feature selection mechanism is designed to compress the dimensions of the mixed feature set composed of deep features and original features. Finally, a weighted local preserving projection algorithm and an ensemble learning mechanism are designed, and a manifold projection classifier ensemble model is constructed, which further improves the classification stability of feature fusion under small samples. In addition, this paper designs a medium-to-large-scale psychotic speech collection program for the first time, collects and constructs a large-scale Chinese psychotic speech database for the verification of psychotic speech recognition algorithms. The experimental results show that the main innovation of the algorithm is effective, and the classification accuracy is better than other representative algorithms, and the maximum improvement is 3.3%. In conclusion, this paper proposes a new method of psychotic speech recognition based on embedded mixed sparse stack autoencoder and manifold ensemble, which effectively improves the recognition rate of psychotic speech.
Aiming at the problem of high-quality image reconstruction from projection data at sparse angular views, we proposed an improved fast iterative reconstruction algorithm based on the minimization of selective image total variation (TV). The new reconstruction scheme consists of two components. Firstly, the algebraic reconstruction technique (ART) algorithm was adopted to reconstruct image that met the identity and non-negativity of projection data, and then, secondly, the selective TV minimization was used to modify the above image. Two phases were alternated until it met the convergence criteria. In order to further speed up the convergence of the algorithm, we applied a fast convergence technology in the iterative process. Experiments on simulated Sheep-Logan phantom were carried out.The results demonstrated that the new method not only improved image reconstruction quality and protected the edge of the image characteristics, but also improved the convergence speed of the iterative reconstruction significantly.
Anesthesia consciousness monitoring is an important issue in basic neuroscience and clinical applications, which has received extensive attention. In this study, in order to find the indicators for monitoring the state of clinical anesthesia, a total of 14 patients undergoing general anesthesia were collected for 5 minutes resting electroencephalogram data under three states of consciousness (awake, moderate and deep anesthesia). Sparse partial least squares (SPLS) and traditional synchronized likelihood (SL) are used to calculate brain functional connectivity, and the three conscious states before and after anesthesia were distinguished by the connection features. The results show that through the whole brain network analysis, SPLS and traditional SL method have the same trend of network parameters in different states of consciousness, and the results obtained by SPLS method are statistically significant (P<0.05). The connection features obtained by the SPLS method are classified by the support vector machine, and the classification accuracy is 87.93%, which is 7.69% higher than that of the connection feature classification obtained by SL method. The results of this study show that the functional connectivity based on the SPLS method has better performance in distinguishing three kinds of consciousness states, and may provides a new idea for clinical anesthesia monitoring.
Due to the sparsity of brain encoding, the neural ensemble signals recorded by microelectrode arrays contain a lot of noise and redundant information, which could reduce the stability and precision of decoding of motion intent. To solve this problem, we proposed a decoding method based on partial least squares (PLS) feature extraction in our study. Firstly, we extracted the features of spike signals using the PLS, and then classified them with support vector machine (SVM) classifier, and decoded them for motion intent. In this study, we decoded neural ensemble signals based on plus-maze test. The results have shown that the proposed method had a better stability and higher decoding accuracy, due to the PLS combined with classification model which overcame the shortcoming of PLS regression that was easily affected by accumulated effect of noise. Meanwhile, the PLS method extracted fewer features with more useful information in comparison with common feature extraction method. The decoding accuracy of real data sets were 93.59%, 84.00% and 83.59%, respectively.
Electroencephalogram (EEG) classification for brain-computer interface (BCI) is a new way of realizing human-computer interreaction. In this paper the application of semi-supervised sparse representation classifier algorithms based on help training to EEG classification for BCI is reported. Firstly, the correlation information of the unlabeled data is obtained by sparse representation classifier and some data with high correlation selected. Secondly, the boundary information of the selected data is produced by discriminative classifier, which is the Fisher linear classifier. The final unlabeled data with high confidence are selected by a criterion containing the information of distance and direction. We applied this novel method to the three benchmark datasets, which were BCIⅠ, BCIⅡ_Ⅳ and USPS. The classification rate were 97%,82% and 84.7%, respectively. Moreover the fastest arithmetic rate was just about 0.2 s. The classification rate and efficiency results of the novel method are both better than those of S3VM and SVM, proving that the proposed method is effective.
The medical magnetic resonance (MR) image reconstruction is one of the key technologies in the field of magnetic resonance imaging (MRI). The compressed sensing (CS) theory indicates that the image can be reconstructed accurately from highly undersampled measurements by using the sparsity of the MR image. However, how to improve the image reconstruction quality by employing more sparse priors of the image becomes a crucial issue for MRI. In this paper, an adaptive image reconstruction model fusing the double dictionary learning is proposed by exploiting sparse priors of the MR image in the image domain and transform domain. The double sparse model which combines synthesis sparse model with sparse transform model is applied to the CS MR image reconstruction according to the complementarity of synthesis sparse and sparse transform model. Making full use of the two sparse priors of the image under the synthesis dictionary and transform dictionary learning, the proposed model is tackled in stages by the iterative alternating minimization algorithm. The solution procedure needs to utilize the synthesis and transform K-singular value decomposition (K-SVD) algorithms. Compared with the existing MRI models, the experimental results show that the proposed model can more efficiently improve the quality of the image reconstruction, and has faster convergence speed and better robustness to noise.
Current epilepsy prediction methods are not effective in characterizing the multi-domain features of complex long-term electroencephalogram (EEG) data, leading to suboptimal prediction performance. Therefore, this paper proposes a novel multi-scale sparse adaptive convolutional network based on multi-head attention mechanism (MS-SACN-MM) model to effectively characterize the multi-domain features. The model first preprocesses the EEG data, constructs multiple convolutional layers to effectively avoid information overload, and uses a multi-layer perceptron and multi-head attention mechanism to focus the network on critical pre-seizure features. Then, it adopts a focal loss training strategy to alleviate class imbalance and enhance the model's robustness. Experimental results show that on the publicly created dataset (CHB-MIT) by MIT and Boston Children's Hospital, the MS-SACN-MM model achieves a maximum accuracy of 0.999 for seizure prediction 10-15 minutes in advance. This demonstrates good predictive performance and holds significant importance for early intervention and intelligent clinical management of epilepsy patients.
This paper aims to realize the decoding of single trial motor imagery electroencephalogram (EEG) signal by extracting and classifying the optimized features of EEG signal. In the classification and recognition of multi-channel EEG signals, there is often a lack of effective feature selection strategies in the selection of the data of each channel and the dimension of spatial filters. In view of this problem, a method combining sparse idea and greedy search (GS) was proposed to improve the feature extraction of common spatial pattern (CSP). The improved common spatial pattern could effectively overcome the problem of repeated selection of feature patterns in the feature vector space extracted by the traditional method, and make the extracted features have more obvious characteristic differences. Then the extracted features were classified by Fisher linear discriminant analysis (FLDA). The experimental results showed that the classification accuracy obtained by proposed method was 19% higher on average than that of traditional common spatial pattern. And high classification accuracy could be obtained by selecting feature set with small size. The research results obtained in the feature extraction of EEG signals lay the foundation for the realization of motor imagery EEG decoding.
In order to solve the problem that the early onset of paroxysmal atrial fibrillation is very short and difficult to detect, a detection algorithm based on sparse coding of Riemannian manifolds is proposed. The proposed method takes into account that the nonlinear manifold geometry is closer to the real feature space structure, and the computational covariance matrix is used to characterize the heart rate variability (RR interval variation), so that the data is in the Riemannian manifold space. Sparse coding is applied to the manifold, and each covariance matrix is represented as a sparse linear combination of Riemann dictionary atoms. The sparse reconstruction loss is defined by the affine invariant Riemannian metric, and the Riemann dictionary is learned by iterative method. Compared with the existing methods, this method used shorter heart rate variability signal, the calculation was simple and had no dependence on the parameters, and the better prediction accuracy was obtained. The final classification on MIT-BIH AF database resulted in a sensitivity of 99.34%, a specificity of 95.41% and an accuracy of 97.45%. At the same time, a specificity of 95.18% was realized in MIT-BIH NSR database. The high precision paroxysmal atrial fibrillation detection algorithm proposed in this paper has a potential application prospect in the long-term monitoring of wearable devices.
Exploring the functional network during the interaction between emotion and cognition is an important way to reveal the underlying neural connections in the brain. Sparse Bayesian network (SBN) has been used to analyze causal characteristics of brain regions and has gradually been applied to the research of brain network. In this study, we got theta band and alpha band from emotion electroencephalogram (EEG) of 22 subjects, constructed effective networks of different arousal, and analyzed measurements of complex network including degree, average clustering coefficient and characteristic path length. We found that: ① compared with EEG signal of low arousal, left middle temporal extensively interacted with other regions in high arousal, while right superior frontal interacted less; ② average clustering coefficient was higher in high arousal and characteristic path length was shorter in low arousal.