Individual differences of P300 potentials lead to that a large amount of training data must be collected to construct pattern recognition models in P300-based brain-computer interface system, which may cause subjects’ fatigue and degrade the system performance. TrAdaBoost is a method that transfers the knowledge from source area to target area, which improves learning effect in the target area. Our research purposed a TrAdaBoost-based linear discriminant analysis and a TrAdaBoost-based support vector machine to recognize the P300 potentials across multiple subjects. This method first trains two kinds of classifiers separately by using the data deriving from a small amount of data from same subject and a large amount of data from different subjects. Then it combines all the classifiers with different weights. Compared with traditional training methods that use only a small amount of data from same subject or mixed different subjects’ data to directly train, our algorithm improved the accuracies by 19.56% and 22.25% respectively, and improved the information transfer rate of 14.69 bits/min and 15.76 bits/min respectively. The results indicate that the TrAdaBoost-based method has the potential to enhance the generalization ability of brain-computer interface on the individual differences.
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.
Affective brain-computer interfaces (aBCIs) has important application value in the field of human-computer interaction. Electroencephalogram (EEG) has been widely concerned in the field of emotion recognition due to its advantages in time resolution, reliability and accuracy. However, the non-stationary characteristics and individual differences of EEG limit the generalization of emotion recognition model in different time and different subjects. In this paper, in order to realize the recognition of emotional states across different subjects and sessions, we proposed a new domain adaptation method, the maximum classifier difference for domain adversarial neural networks (MCD_DA). By establishing a neural network emotion recognition model, the shallow feature extractor was used to resist the domain classifier and the emotion classifier, respectively, so that the feature extractor could produce domain invariant expression, and train the decision boundary of classifier learning task specificity while realizing approximate joint distribution adaptation. The experimental results showed that the average classification accuracy of this method was 88.33% compared with 58.23% of the traditional general classifier. It improves the generalization ability of emotion brain-computer interface in practical application, and provides a new method for aBCIs to be used in practice.
Transfer learning is provided with potential research value and application prospect in motor imagery electroencephalography (MI-EEG)-based brain-computer interface (BCI) rehabilitation system, and the source domain classification model and transfer strategy are the two important aspects that directly affect the performance and transfer efficiency of the target domain model. Therefore, we propose a parameter transfer learning method based on shallow visual geometry group network (PTL-sVGG). First, Pearson correlation coefficient is used to screen the subjects of the source domain, and the short-time Fourier transform is performed on the MI-EEG data of each selected subject to acquire the time-frequency spectrogram images (TFSI). Then, the architecture of VGG-16 is simplified and the block design is carried out, and the modified sVGG model is pre-trained with TFSI of source domain. Furthermore, a block-based frozen-fine-tuning transfer strategy is designed to quickly find and freeze the block with the greatest contribution to sVGG model, and the remaining blocks are fine-tuned by using TFSI of target subjects to obtain the target domain classification model. Extensive experiments are conducted based on public MI-EEG datasets, the average recognition rate and Kappa value of PTL-sVGG are 94.9% and 0.898, respectively. The results show that the subjects’ optimization is beneficial to improve the model performance in source domain, and the block-based transfer strategy can enhance the transfer efficiency, realizing the rapid and effective transfer of model parameters across subjects on the datasets with different number of channels. It is beneficial to reduce the calibration time of BCI system, which promote the application of BCI technology in rehabilitation engineering.
In the field of brain-computer interfaces (BCIs) based on functional near-infrared spectroscopy (fNIRS), traditional subject-specific decoding methods suffer from the limitations of long calibration time and low cross-subject generalizability, which restricts the promotion and application of BCI systems in daily life and clinic. To address the above dilemma, this study proposes a novel deep transfer learning approach that combines the revised inception-residual network (rIRN) model and the model-based transfer learning (TL) strategy, referred to as TL-rIRN. This study performed cross-subject recognition experiments on mental arithmetic (MA) and mental singing (MS) tasks to validate the effectiveness and superiority of the TL-rIRN approach. The results show that the TL-rIRN significantly shortens the calibration time, reduces the training time of the target model and the consumption of computational resources, and dramatically enhances the cross-subject decoding performance compared to subject-specific decoding methods and other deep transfer learning methods. To sum up, this study provides a basis for the selection of cross-subject, cross-task, and real-time decoding algorithms for fNIRS-BCI systems, which has potential applications in constructing a convenient and universal BCI system.
Sudden cardiac arrest (SCA) is a lethal cardiac arrhythmia that poses a serious threat to human life and health. However, clinical records of sudden cardiac death (SCD) electrocardiogram (ECG) data are extremely limited. This paper proposes an early prediction and classification algorithm for SCA based on deep transfer learning. With limited ECG data, it extracts heart rate variability features before the onset of SCA and utilizes a lightweight convolutional neural network model for pre-training and fine-tuning in two stages of deep transfer learning. This achieves early classification, recognition and prediction of high-risk ECG signals for SCA by neural network models. Based on 16 788 30-second heart rate feature segments from 20 SCA patients and 18 sinus rhythm patients in the international publicly available ECG database, the algorithm performance evaluation through ten-fold cross-validation shows that the average accuracy (Acc), sensitivity (Sen), and specificity (Spe) for predicting the onset of SCA in the 30 minutes prior to the event are 91.79%, 87.00%, and 96.63%, respectively. The average estimation accuracy for different patients reaches 96.58%. Compared to traditional machine learning algorithms reported in existing literatures, the method proposed in this paper helps address the requirement of large training datasets for deep learning models and enables early and accurate detection and identification of high-risk ECG signs before the onset of SCA.