Aiming at the problem that the small samples of critical disease in clinic may lead to prognostic models with poor performance of overfitting, large prediction error and instability, the long short-term memory transferring algorithm (transLSTM) was proposed. Based on the idea of transfer learning, the algorithm leverages the correlation between diseases to transfer information of different disease prognostic models, constructs the effictive model of target disease of small samples with the aid of large data of related diseases, hence improves the prediction performance and reduces the requirement for target training sample quantity. The transLSTM algorithm firstly uses the related disease samples to pretrain partial model parameters, and then further adjusts the whole network with the target training samples. The testing results on MIMIC-Ⅲ database showed that compared with traditional LSTM classification algorithm, the transLSTM algorithm had 0.02-0.07 higher AUROC and 0.05-0.14 larger AUPRC, while its number of training iterations was only 39%-64% of the traditional algorithm. The results of application on sepsis revealed that the transLSTM model of only 100 training samples had comparable mortality prediction performance to the traditional model of 250 training samples. In small sample situations, the transLSTM algorithm has significant advantages with higher prediciton accuracy and faster training speed. It realizes the application of transfer learning in the prognostic model of critical disease with small samples.
Heart sound is one of the common medical signals for diagnosing cardiovascular diseases. This paper studies the binary classification between normal or abnormal heart sounds, and proposes a heart sound classification algorithm based on the joint decision of extreme gradient boosting (XGBoost) and deep neural network, achieving a further improvement in feature extraction and model accuracy. First, the preprocessed heart sound recordings are segmented into four status, and five categories of features are extracted from the signals based on segmentation. The first four categories of features are sieved through recursive feature elimination, which is used as the input of the XGBoost classifier. The last category is the Mel-frequency cepstral coefficient (MFCC), which is used as the input of long short-term memory network (LSTM). Considering the imbalance of the data set, these two classifiers are both improved with weights. Finally, the heterogeneous integrated decision method is adopted to obtain the prediction. The algorithm was applied to the open heart sound database of the PhysioNet Computing in Cardiology(CINC) Challenge in 2016 on the PhysioNet website, to test the sensitivity, specificity, modified accuracy and F score. The results were 93%, 89.4%, 91.2% and 91.3% respectively. Compared with the results of machine learning, convolutional neural networks (CNN) and other methods used by other researchers, the accuracy and sensibility have been obviously improved, which proves that the method in this paper could effectively improve the accuracy of heart sound signal classification, and has great potential in the clinical auxiliary diagnosis application of some cardiovascular diseases.
Emotion plays an important role in people's cognition and communication. By analyzing electroencephalogram (EEG) signals to identify internal emotions and feedback emotional information in an active or passive way, affective brain-computer interactions can effectively promote human-computer interaction. This paper focuses on emotion recognition using EEG. We systematically evaluate the performance of state-of-the-art feature extraction and classification methods with a public-available dataset for emotion analysis using physiological signals (DEAP). The common random split method will lead to high correlation between training and testing samples. Thus, we use block-wise K fold cross validation. Moreover, we compare the accuracy of emotion recognition with different time window length. The experimental results indicate that 4 s time window is appropriate for sampling. Filter-bank long short-term memory networks (FBLSTM) using differential entropy features as input was proposed. The average accuracy of low and high in valance dimension, arousal dimension and combination of the four in valance-arousal plane is 78.8%, 78.4% and 70.3%, respectively. These results demonstrate the advantage of our emotion recognition model over the current studies in terms of classification accuracy. Our model might provide a novel method for emotion recognition in affective brain-computer interactions.
Electrocardiogram (ECG) can visually reflect the physiological electrical activity of human heart, which is important in the field of arrhythmia detection and classification. To address the negative effect of label imbalance in ECG data on arrhythmia classification, this paper proposes a nested long short-term memory network (NLSTM) model for unbalanced ECG signal classification. The NLSTM is built to learn and memorize the temporal characteristics in complex signals, and the focal loss function is used to reduce the weights of easily identifiable samples. Then the residual attention mechanism is used to modify the assigned weights according to the importance of sample characteristic to solve the sample imbalance problem. Then the synthetic minority over-sampling technique is used to perform a simple manual oversampling process on the Massachusetts institute of technology and Beth Israel hospital arrhythmia (MIT-BIH-AR) database to further increase the classification accuracy of the model. Finally, the MIT-BIH arrhythmia database is applied to experimentally verify the above algorithms. The experimental results show that the proposed method can effectively solve the issues of imbalanced samples and unremarkable features in ECG signals, and the overall accuracy of the model reaches 98.34%. It also significantly improves the recognition and classification of minority samples and has provided a new feasible method for ECG-assisted diagnosis, which has practical application significance.
Objective To recognize the different phases of Korotkoff sounds through deep learning technology, so as to improve the accuracy of blood pressure measurement in different populations. Methods A classification model of the Korotkoff sounds phases was designed, which fused attention mechanism (Attention), residual network (ResNet) and bidirectional long short-term memory (BiLSTM). First, a single Korotkoff sound signal was extracted from the whole Korotkoff sounds signals beat by beat, and each Korotkoff sound signal was converted into a Mel spectrogram. Then, the local feature extraction of Mel spectrogram was processed by using the Attention mechanism and ResNet network, and BiLSTM network was used to deal with the temporal relations between features, and full-connection layer network was applied in reducing the dimension of features. Finally, the classification was completed by SoftMax function. The dataset used in this study was collected from 44 volunteers (24 females, 20 males with an average age of 36 years), and the model performance was verified using 10-fold cross-validation. Results The classification accuracy of the established model for the 5 types of Korotkoff sounds phases was 93.4%, which was higher than that of other models. Conclusion This study proves that the deep learning method can accurately classify Korotkoff sounds phases, which lays a strong technical foundation for the subsequent design of automatic blood pressure measurement methods based on the classification of the Korotkoff sounds phases.
The recurrent neural network architecture improves the processing ability of time-series data. However, issues such as exploding gradients and poor feature extraction limit its application in the automatic diagnosis of mild cognitive impairment (MCI). This paper proposed a research approach for building an MCI diagnostic model using a Bayesian-optimized bidirectional long short-term memory network (BO-BiLSTM) to address this problem. The diagnostic model was based on a Bayesian algorithm and combined prior distribution and posterior probability results to optimize the BO-BiLSTM network hyperparameters. It also used multiple feature quantities that fully reflected the cognitive state of the MCI brain, such as power spectral density, fuzzy entropy, and multifractal spectrum, as the input of the diagnostic model to achieve automatic MCI diagnosis. The results showed that the feature-fused Bayesian-optimized BiLSTM network model achieved an MCI diagnostic accuracy of 98.64% and effectively completed the diagnostic assessment of MCI. In conclusion, based on this optimization, the long short-term neural network model has achieved automatic diagnostic assessment of MCI, providing a new diagnostic model for intelligent diagnosis of MCI.
Sleep staging is the basis for solving sleep problems. There’s an upper limit for the classification accuracy of sleep staging models based on single-channel electroencephalogram (EEG) data and features. To address this problem, this paper proposed an automatic sleep staging model that mixes deep convolutional neural network (DCNN) and bi-directional long short-term memory network (BiLSTM). The model used DCNN to automatically learn the time-frequency domain features of EEG signals, and used BiLSTM to extract the temporal features between the data, fully exploiting the feature information contained in the data to improve the accuracy of automatic sleep staging. At the same time, noise reduction techniques and adaptive synthetic sampling were used to reduce the impact of signal noise and unbalanced data sets on model performance. In this paper, experiments were conducted using the Sleep-European Data Format Database Expanded and the Shanghai Mental Health Center Sleep Database, and achieved an overall accuracy rate of 86.9% and 88.9% respectively. When compared with the basic network model, all the experimental results outperformed the basic network, further demonstrating the validity of this paper's model, which can provide a reference for the construction of a home sleep monitoring system based on single-channel EEG signals.