Single-channel EEG Research Articles

A deep learning method for classification of steady-state visual evoked potentials in a brain-computer interface speller

ABSTRACT The main drawback of SSVEP-based BCIs is the lack of a classifier that categorizes SSVEPs with high accuracy and Information Transfer Rate (ITR). Addressing this, we proposed a deep convolutional neural network (CNN) for classifying a 40-class SSVEP. Time windows of length 2 and 3.5 seconds were used for training and testing the model by leave-one-subject-out cross-validation, using nine, three, and single-channel EEG. The proposed model reached 88.5% average accuracy for the nine-channel EEG with the mean and max ITR of 72 and 91.23 bpm, respectively. It outperformed the previous deep learning methods for SSVEP-based BCIs, in terms of accuracy and ITR. In the three-channel experiment the mean accuracy and ITR were 76.02% and 40.1 bpm. In single-channel implementation, O1 channel achieved 77.38 % average accuracy (highest) and the mean ITR was 57.51 bpm. The model showed promising performance to put this technology forward and make it more practical.

RT-Blink: A Method Toward Real-Time Blink Detection From Single Frontal EEG Signal

Eye blinks take an important role for electroencephalography (EEG) signals that on one hand, they can severely impact the EEG signals, and on the other hand, they may provide useful information for brain–computer interface (BCI) and scientific applications. In particular, it is challenging to detect blinks in real time from a single channel of EEG signals. In this work, we propose a short windowed and random forest based method toward the Real-Time Blink detection (RT-Blink), which balances the processing granularity and computation complexity. RT-Blink uses a potential blink (PB) boundary detecting algorithm and a pretrained random forest (RF) model with a set of features, including sample entropy (SampEn), standard deviation (SD), range of amplitude (RA), and rate of grade (RG), which enables fully automated identification of the duration of blinks from a single-channel EEG signal. The window size of RT-Blink can be adjusted according to the processing speed of the underlying hardware and the requirement of real time, which can be down to one-point EEG data. RT-Blink provides a scalable framework, which can be realized in software, hardware accelerator, or their mixture. Using EEG data contaminated by blinks, we show that RT-Blink achieves 96.54% and 91.25% for average sensitivity and precision, respectively. The time window is 60 ms based on our computer, with minimized overlapping for blink boundary detection. The average processing time is 5.07 ms for each time window, with average 1.65 ms of SD for all the EEG test cases. The results suggest that RT-Blink has promising potential toward real-time EEG applications.

Feature Extraction From Single-Channel EEG Using Tsfresh and Stacked Ensemble Approach for Sleep Stage Classification

The smart world under Industry 4.0 is witnessing a notable spurt in sleep disorders and sleep-related issues in patients. Artificial intelligence and IoT are taking a giant leap in connecting sleep patients remotely with healthcare providers. The contemporary single-channel-based monitoring devices play a tremendous role in predicting sleep quality and related issues. Handcrafted feature extraction is a time-consuming job in machine learning-based automatic sleep classification. The proposed single-channel work uses Tsfresh to extract features from both the EEG channels (Pz-oz and Fpz-Cz) of the SEDFEx database individually to realise a single-channel EEG. The adopted mRMR feature selection approach selected 55 features from the extracted 787 features. A stacking ensemble classifier achieved 95%, 94%, 91%, and 88% accuracy using stratified 5-fold validation in 2, 3, 4, and 5 class classification employing healthy subjects data. The outcome of the experiments indicates that Tsfresh is an excellent tool to extract standard features from EEG signals.

Measuring Brain Activation Patterns from Raw Single-Channel EEG during Exergaming: A Pilot Study

Physical and cognitive rehabilitation is deemed crucial to attenuate symptoms and to improve the quality of life in people with neurodegenerative disorders, such as Parkinson’s Disease. Among rehabilitation strategies, a novel and popular approach relies on exergaming: the patient performs a motor or cognitive task within an interactive videogame in a virtual environment. These strategies may widely benefit from being tailored to the patient’s needs and engagement patterns. In this pilot study, we investigated the ability of a low-cost BCI based on single-channel EEG to measure the user’s engagement during an exergame. As a first step, healthy subjects were recruited to assess the system’s capability to distinguish between (1) rest and gaming conditions and (2) gaming at different complexity levels, through Machine Learning supervised models. Both EEG and eye-blink features were employed. The results indicate the ability of the exergame to stimulate engagement and the capability of the supervised classification models to distinguish resting stage from game-play (accuracy > 95%). Finally, different clusters of subject responses throughout the game were identified, which could help define models of engagement trends. This result is a starting point in developing an effectively subject-tailored exergaming system.

A Siamese Network-Based Method for Improving the Performance of Sleep Staging with Single-Channel EEG

Sleep staging is of critical significance to the diagnosis of sleep disorders, and the electroencephalogram (EEG), which is used for monitoring brain activity, is commonly employed in sleep staging. In this paper, we propose a novel method for improving the performance of sleep staging models based on Siamese networks, based on single-channel EEG. Our proposed method consists of a Siamese network architecture and a redesigned loss with distance metrics. Two encoders are used in the Siamese network to generate latent features of the EEG epochs, and the contrastive loss, which is also a distance metric, is used to compare the similarity or differences between EEG epochs from the same or different sleep stages. We evaluated our method on single-channel EEGs from different channels (Fpz-Cz and F4-EOG (left)) from two public datasets SleepEDF and MASS-SS3 and achieved the overall accuracies MF1 and Cohen's kappa coefficient of 85.2%, 78.3% and 0.79 on SleepEDF and 87.2%, 82.1% and 0.81 on MASS-SS3. The results show that our method can significantly improve the performance of sleep staging models and outperform the state-of-the-art sleep staging methods. The performance of our method also confirms that the features captured by Siamese networks and distance metrics are useful for sleep staging.

Prediction of Increased Intracranial Pressure in Traumatic Brain Injury Using Quantitative Electroencephalogram in a Porcine Experimental Model.

Continuous and non-invasive measurement of intracranial pressure (ICP) in traumatic brain injury (TBI) is important to recognize increased ICP (IICP), which can reduce treatment delays. The purpose of this study was to develop an electroencephalogram (EEG)-based prediction model for IICP in a porcine TBI model. Thirty swine were anaesthetized and underwent IICP by inflating a Foley catheter in the intracranial space. Single-channel EEG data were collected every 6 min in 10 mmHg increments in the ICP from baseline to 50 mmHg. We developed EEG-based models to predict the IICP (equal or over 25 mmHg) using four algorithms: logistic regression (LR), naive Bayes (NB), support vector machine (SVM), and random forest (RF). We assessed the performance of each model based on the accuracy, sensitivity, specificity, and AUC values. The accuracy of each prediction model for IICP was 0.773 for SVM, 0.749 for NB, 0.746 for RF, and 0.706 for LR. The AUC of each model was 0.860 for SVM, 0.824 for NB, 0.802 for RF, and 0.748 for LR. We developed a machine learning prediction model for IICP using single-channel EEG signals in a swine TBI experimental model. The SVM model showed good predictive power with the highest AUC value.

LWSleepNet: A lightweight attention-based deep learning model for sleep staging with singlechannel EEG.

Sleep is vital to human health, and sleep staging is an essential process in sleep assessment. However, manual classification is an inefficient task. Along with the increased demand for portable sleep quality detection devices, lightweight automatic sleep staging needs to be developed. This study proposes a novel attention-based lightweight deep learning model called LWSleepNet. A depthwise separable multi-resolution convolutional neural network is introduced to analyze the input feature map and captures features at multiple frequencies using two different sized convolutional kernels. The temporal feature extraction module divides the input into patches and feeds them into a multi-head attention block to extract time-dependent information from sleep recordings. The model's convolution operations are replaced with depthwise separable convolutions to minimize its number of parameters and computational cost. The model's performance on two public datasets (Sleep-EDF-20 and Sleep-EDF-78) was evaluated and compared with those of previous studies. Then, an ablation study and sensitivity analysis were performed to evaluate further each module. LWSleepNet achieves an accuracy of 86.6% and Macro-F1 score of 79.2% for the Sleep-EDF-20 dataset and an accuracy of 81.5% and Macro-F1 score of 74.3% for the Sleep-EDF-78 dataset with only 55.3 million floating-point operations per second and 180 K parameters. On two public datasets, LWSleepNet maintains excellent prediction performance while substantially reducing the number of parameters, demonstrating that our proposed Light multiresolution convolutional neural network and temporal feature extraction modules can provide excellent portability and accuracy and can be easily integrated into portable sleep monitoring devices.

MS-HNN: Multi-Scale Hierarchical Neural Network With Squeeze and Excitation Block for Neonatal Sleep Staging Using a Single-Channel EEG.

Most existing neonatal sleep staging appro- aches applied multiple EEG channels to obtain good performance. However, it potentially increased the computational complexity and led to an increased risk of skin disruption to neonates during data acquisition. In this paper, a multi-scale hierarchical neural network (MS-HNN) with a squeeze and excitation (SE) block for neonatal sleep staging is presented in this study on the basis of a single EEG channel. MS-HNN composes of multi-scale convolutional neural network (MSCNN), temporal information learning (TIL) module, and squeeze and excitation (SE) block. MSCNN can extract features from different scales and frequencies, and TIL module can acquire the transition information among adjacent stages. In addition, for these extracted features, SE block can selectively concentrate on informative features and weaken redundant features for achieving better performance. The proposed approach was validated on a clinical dataset involving 64 neonates from the Children's Hospital of Fudan University (CHFU). The proposed network achieves an accuracy of 75.4% and 76.5% for three-class automatic neonatal sleep staging with the single-EEG channel and the eight-EEG channels, respectively. The experimental results show that the proposed method can maintain good performance by making full use of the information in the single channel while reducing the channels to control the computational overhead.

Fusion of EEG and Eye Blink Analysis for Detection of Driver Fatigue.

The driver fatigue detection using multi-channel electroencephalography (EEG) has been extensively addressed in the literature. However, the employment of a single prefrontal EEG channel should be prioritized as it provides users with more comfort. Furthermore, eye blinks from such channel can be analyzed as the complementary information. Here, we present a new driver fatigue detection method based on simultaneous EEG and eye blinks analysis using an Fp1 EEG channel. First, the moving standard deviation algorithm identifies eye blink intervals (EBIs) to extract blink-related features. Second, the discrete wavelet transform filters the EBIs from the EEG signal. Third, the filtered EEG signal is decomposed into sub-bands, and various linear and nonlinear features are extracted. Finally, the prominent features are selected by the neighbourhood components analysis and fed to a classifier to discriminate between fatigue and alert driving. In this paper, two different databases are investigated. The first one is used for parameters' tuning of proposed method for the eye blink detection and filtering, nonlinear EEG measures, and feature selection. The second one is solely used for testing the robustness of the tuned parameters. The comparison between the obtained results from both databases by the AdaBoost classifier in terms of sensitivity (90.2% vs. 87.4%), specificity (87.7% vs. 85.5%), and accuracy (88.4% vs. 86.8%) indicates the reliability of the proposed method for the driver fatigue detection. Considering the existence of commercial single prefrontal channel EEG headbands, the proposed method can be used to detect the driver fatigue in real-world scenarios.

Channel Contribution In Deep Learning Based Automatic Sleep Scoring - How Many Channels Do We Need?

Machine learning based sleep scoring methods aim to automate the process of annotating polysomnograms with sleep stages. Although sleep signals of multiple modalities and channels should contain more information according to sleep guidelines, most multi-channel multi-modal models in the literature showed only a little performance improvement compared to single-channel EEG models and sometimes even failed to outperform them. In this paper, we investigate whether the high performance of single-channel EEG models can be attributed to specific model features in their deep learning architectures and to which extent multi-channel multi-modal models take the information from different channels of modalities into account. First, we transfer the model features from single-channel EEG models, such as combinations of small and large filters in CNNs, to multi-channel multi-modal models and measure their impacts. Second, we employ two explainability methods, the layer-wise relevance propagation as post-hoc and the embedded channel attention network as intrinsic interpretability methods, to measure the contribution of different channels on predictive performance. We find that i) single-channel model features can improve the performance of multi-channel multi-modal models and ii) multi-channel multi-modal models focus on one important channel per modality and use the remaining channels to complement the information of the focused channels. Our results suggest that more advanced methods for aggregating channel information using complementary information from other channels may improve sleep scoring performance for multi-channel multi-modal models.

Extracting Multi-Scale and Salient Features by MSE Based U-Structure and CBAM for Sleep Staging.

According to the World Health Organization, more and more people in the world are suffering from somnipathy. Automatic sleep staging is critical for assessing sleep quality and assisting in the diagnosis of psychiatric and neurological disorders caused by somnipathy. Many researchers employ deep learning methods for sleep stage classification and have achieved high performance. However, there are still no effective methods to modeling intrinsic characteristics of salient wave in different sleep stages from physiological signals. And transition rules hidden in signals from one to another sleep stage cannot be identified and captured. In addition, class imbalance problem in dataset is not conducive to building a robust classification model. To solve these problems, we construct a deep neural network combining MSE(Multi-Scale Extraction) based U-structure and CBAM (Convolutional Block Attention Module) to extract the multi-scale salient waves from single-channel EEG signals. The U-structured convolutional network with MSE is utilized to extract multi-scale features from raw EEG signals. After that, the CBAM is used to focus more on salient variation and then learn transition rules between successive sleep stages. Further, a class adaptive weight cross entropy loss function is proposed to solve the class imbalance problem. Experiments in three public datasets show that our model greatly outperform the state-of-the-art results compared with existing methods. The overall accuracy and macro F1-score (Sleep-EDF-39: 90.3%-86.2, Sleep-EDF-153: 89.7%-85.2, SHHS: 86.8%-83.5) on three public datasets suggest that the proposed model is very promising to completely take place of human experts for sleep staging.

EEG Signal Epilepsy Detection with a Weighted Neighbour Graph Representation and Two-stream Graph-based Framework.

Epilepsy is one of the most common neurological diseases. Clinically, epileptic seizure detection is usually performed by analyzing electroencephalography (EEG) signals. At present, deep learning models have been widely used for single-channel EEG signal epilepsy detection, but this method is difficult to explain the classification results. Researchers have attempted to solve interpretive problems by combining graph representation of EEG signals with graph neural network models. Recently, the combination of graph representations and graph neural network (GNN) models has been increasingly applied to single-channel epilepsy detection. By this methodology, the raw EEG signal is transformed to its graph representation, and a GNN model is used to learn latent features and classify whether the data indicates an epileptic seizure episode. However, existing methods are faced with two major challenges. First, existing graph representations tend to have high time complexity as they generally require each vertex to traverse all other vertices to construct a graph structure. Some of them also have high space complexity for being dense. Second, while separate graph representations can be derived from a single-channel EEG signal in both time and frequency domains, existing GNN models for epilepsy detection can learn from a single graph representation, which makes it hard to let the information from the two domains complement each other. For addressing these challenges, we propose a Weighted Neighbour Graph (WNG) representation for EEG signals. Reducing the redundant edges of the existing graph, WNG can be both time and space-efficient, and as informative as its less efficient counterparts. We then propose a two-stream graph-based framework to simultaneously learn features from WNG in both time and frequency domain. Extensive experiments demonstrate the effectiveness and efficiency of the proposed methods.

Lie Detection: Truth Identification from EEG Signal Using Frequency and Time Features with SVM Classifier

This study investigated the approach of extracting features from single EEG channels when the minimum number of features in Electroencephalogram (EEG) channels, hence the visibility of using sets of features extracted from a single channel. The feature sets considered in previous studies are utilized to establish a combined set of features extracted from one channel. The feature is the set of statistical moments. Publicly available EEG datasets like the Dryad dataset, obtained from 15 participants, are tested into a support vector machine classifier. The 12 channels were trained separately, where each channel was divided into a different number of blocks, and the results indicated that some channels were bad. Some were very encouraging, reaching 100% in the number of blocks 16 in channels 8 and 12. In this article, the comparison of ANN algorithm test results published in a previous article with SVM algorithm test results for the same tested features and channels will be presented.

Automatic Sleep Staging Based on Single-Channel EEG Signal Using Null Space Pursuit Decomposition Algorithm

Sleep quality is related to people’s physical and mental health, so an accurate assessment of sleep quality is key to recognizing sleep disorders and taking effective interventions. To address the shortcomings of traditional manual and automatic staging methods, such as being time-consuming and having low classification accuracy, an automatic sleep staging method based on the null space pursuit (NSP) decomposition algorithm of single-channel electroencephalographic (EEG) signals is proposed, which provides a new way for EEG signal decomposition and automatic identification of sleep stages. First, the single-channel EEG signal data from the Sleep-EDF database, DREAMS Subject database, and Sleep Heart Health Study database (SHHS), available on PhysioNet, were preprocessed, respectively. Second, the preprocessed single-channel EEG signals were decomposed by the NSP algorithm. Third, we extracted nine features in the time domain of the nonlinear dynamics and statistics from the original EEG signal and the six simple signals that were decomposed. Finally, the extreme gradient boosting (XGBOOST) algorithm was used to construct a classification model to classify and identify the 63 extracted EEG signal features for automatic sleep staging. The experimental results showed that, on the Sleep-EDF database, the accuracy of four and five categories were 93.59% and 92.89%, respectively; on the DREAMS Subject database, the accuracy rates of four and five categories were 91.32% and 90.01%, respectively; on the SHHS database, the accuracy rates of four and five categories were 90.25% and 88.37%, respectively. The experimental results show that the automatic sleep staging model proposed in this work has high classification accuracy and efficiency, as well as strong applicability and robustness.

Single-channel electroencephalogram signal used for sleep state recognition based on one-dimensional width kernel convolutional neural networks and long-short-term memory networks

Aiming at the problem that the unbalanced distribution of data in sleep electroencephalogram(EEG) signals and poor comfort in the process of polysomnography information collection will reduce the model's classification ability, this paper proposed a sleep state recognition method using single-channel EEG signals (WKCNN-LSTM) based on one-dimensional width kernel convolutional neural networks(WKCNN) and long-short-term memory networks (LSTM). Firstly, the wavelet denoising and synthetic minority over-sampling technique-Tomek link (SMOTE-Tomek) algorithm were used to preprocess the original sleep EEG signals. Secondly, one-dimensional sleep EEG signals were used as the input of the model, and WKCNN was used to extract frequency-domain features and suppress high-frequency noise. Then, the LSTM layer was used to learn the time-domain features. Finally, normalized exponential function was used on the full connection layer to realize sleep state. The experimental results showed that the classification accuracy of the one-dimensional WKCNN-LSTM model was 91.80% in this paper, which was better than that of similar studies in recent years, and the model had good generalization ability. This study improved classification accuracy of single-channel sleep EEG signals that can be easily utilized in portable sleep monitoring devices.

MLMRS-Net: Electroencephalography (EEG) motion artifacts removal using a multi-layer multi-resolution spatially pooled 1D signal reconstruction network

Electroencephalogram (EEG) signals suffer substantially from motion artifacts when recorded in ambulatory settings utilizing wearable sensors. Because the diagnosis of many neurological diseases is heavily reliant on clean EEG data, it is critical to eliminate motion artifacts from motion-corrupted EEG signals using reliable and robust algorithms. Although a few deep learning-based models have been proposed for the removal of ocular, muscle, and cardiac artifacts from EEG data to the best of our knowledge, there is no attempt has been made in removing motion artifacts from motion-corrupted EEG signals: In this paper, a novel 1D convolutional neural network (CNN) called multi-layer multi-resolution spatially pooled (MLMRS) network for signal reconstruction is proposed for EEG motion artifact removal. The performance of the proposed model was compared with ten other 1D CNN models: FPN, LinkNet, UNet, UNet+, UNetPP, UNet3+, AttentionUNet, MultiResUNet, DenseInceptionUNet, and AttentionUNet++ in removing motion artifacts from motion-contaminated single-channel EEG signal. All the eleven deep CNN models are trained and tested using a single-channel benchmark EEG dataset containing 23 sets of motion-corrupted and reference ground truth EEG signals from PhysioNet. Leave-one-out cross-validation method was used in this work. The performance of the deep learning models is measured using three well-known performance matrices viz. mean absolute error (MAE)-based construction error, the difference in the signal-to-noise ratio (ΔSNR), and percentage reduction in motion artifacts (η). The proposed MLMRS-Net model has shown the best denoising performance, producing an average ΔSNR, η, and MAE values of 26.64 dB, 90.52%, and 0.056, respectively, for all 23 sets of EEG recordings. The results reported using the proposed model outperformed all the existing state-of-the-art techniques in terms of average η improvement.

A novel moving window-based power spectrum features for single-channel EEG classification using machine learning

Electroencephalogram (EEG) signal classification is a crucial and very difficult task. Meanwhile, extracting features that are representative and able to discriminate different types of EEG signals is a complex task. Such features are usually fed to machine learning algorithms to classify the EEG signals based on the extracted features. This paper proposed a highly accurate and real-time features extraction method that can be used to help physicians in detecting different types of seizures and states in EEG signals characterized by a set of features extracted from the power spectrum of the EEG window. This is achieved by applying the following four steps. First, the EEG signals dataset contains different classes of EEG signals: Normal Eye Closed, Normal Eye Opened, Focal Seizure, Non-Focal Seizure, and Ictal Seizure activities. Second, each EEG signal has a length of 4097 samples sampled with a sampling frequency of 173.6 Hz which resulted in 23.6 seconds in length, this signal will be truncated into windows (Sub-signals) with a length of 349 samples (Approximately 2 seconds) with a total number of 12 windows for each signal. Afterward, the Fourier Transform (FT) based power spectrum will be computed for each window, then a set of different features are extracted from each window's FT power spectrum, and these features are classified using different Machine Learning (ML) algorithms. The results showed that the proposed methodology yields around 98% accuracy for the five different classification scenarios using different ML algorithms. The suggested method is hence robust, fast, real-time, accurate, and simple.

Fast Sleep Stage Classification Using Cascaded Support Vector Machines with Single-Channel EEG Signals.

Long-term sleep stage monitoring is very important for the diagnosis and treatment of insomnia. With the development of wearable electroencephalogram (EEG) devices, we developed a fast and accurate sleep stage classification method in this study with single-channel EEG signals for practical applications. The original sleep recordings were collected from the Sleep-EDF database. The wavelet threshold denoising (WTD) method and wavelet packet transformation (WPT) method were applied as signal preprocessing to extract six kinds of characteristic waves. With a comprehensive feature system including time, frequency, and nonlinear dynamics, we obtained the sleep stage classification results with different Support Vector Machine (SVM) models. We proposed a novel classification method based on cascaded SVM models with various features extracted from denoised EEG signals. To enhance the accuracy and generalization performance of this method, nonlinear dynamics features were taken into consideration. With nonlinear dynamics features included, the average classification accuracy was up to 88.11% using this method. In addition, with cascaded SVM models, the classification accuracy of the non-rapid eye movement sleep stage 1 (N1) was enhanced from 41.5% to 55.65% compared with the single SVM model, and the overall classification time for each epoch was less than 1.7 s. Moreover, we demonstrated that it was possible to apply this method for long-term sleep stage monitor applications.

MRASleepNet: a multi-resolution attention network for sleep stage classification using single-channel EEG

Objective. Computerized classification of sleep stages based on single-lead electroencephalography (EEG) signals is important, but still challenging. In this paper, we proposed a deep neural network called MRASleepNet for automatic sleep stage classification using single-channel EEG signals. Approach. The proposed MRASleepNet model consisted of a feature extraction (FE) module, a multi-resolution attention (MRA) module, and a gated multilayer perceptron (gMLP) module, as well as a direct pathway for computing statistical features. The FE, MRA, and gMLP modules were used to extract features, establish feature attention, and obtain temporal relationships between features, respectively. EEG signals were normalized and cut into 30 s segments, and enhanced by incorporating contextual information from adjacent data segments. After data enhancement, the 40 s data segments were input to the MRASleepNet model. The model was evaluated on the SleepEDF and the cyclic alternating pattern (CAP) databases, using such metrics as the accuracy, Kappa, and macro-F1 (MF1). Main results. For the SleepEDF-20 database, the proposed model had an accuracy of 84.5%, an MF1 of 0.789, and a Kappa of 0.786. For the SleepEDF-78 database, the model had an accuracy of 81.4%, an MF1 of 0.754, and a Kappa of 0.743. For the CAP database, the model had an accuracy of 74.3%, an MF1 of 0.656, and a Kappa of 0.652. The proposed model achieved satisfactory performance in automatic sleep stage classification tasks. Significance. The time- and frequency-domain features extracted by the FE module and filtered by the MRA module, together with the temporal features extracted by the gMLP module and the statistical features extracted by the statistical highway, enabled the proposed model to obtain a satisfying performance in sleep staging. The proposed MRASleepNet model may be used as a new deep learning method for automatic sleep stage classification. The code of MRASleepNet will be made available publicly on https://github.com/YuRui8879/.

A time-frequency denoising method for single-channel event-related EEG.

Electroencephalogram (EEG) acquisition is easily affected by various noises, including those from electrocardiogram (ECG), electrooculogram (EOG), and electromyogram (EMG). Because noise interference can significantly limit the study and analysis of brain signals, there is a significant need for the development of improved methods to remove this interference for more accurate measurement of EEG signals. Based on the non-linear and non-stationary characteristics of brain signals, a strategy was developed to denoise brain signals using a time-frequency denoising algorithm framework of short-time Fourier transform (STFT), bidimensional empirical mode decomposition (BEMD), and non-local means (NLM). Time-frequency analysis can reveal the signal frequency component and its evolution process, allowing the elimination of noise according to the signal and noise distribution. BEMD can be used to decompose the time-frequency signals into sub-time-frequency signals for noise removal at different scales. NLM relies on structural self-similarity to locally smooth an image to remove noise and restore its main geometric structure, making this method appropriate for time-frequency signal denoising. The experimental results show that the proposed method can effectively suppress the high-frequency components of brain signals, resulting in a smoother brain signal waveform after denoising. The correlation coefficient of the reference signal, a superposition average of multiple trial signals, and the original single trial signal was determined, and then correlation coefficients were calculated between the reference signal and single trial signals processed by time-frequency denoising, ensemble empirical mode decomposition (EEMD)-independent component analysis (ICA), EEMD-canonical correlation analysis (CCA), and wavelet threshold denoising methods. The correlation coefficient was highest for the signal processed by the time-frequency denoising method and the reference signal, indicating that the single trial signal after time-frequency denoising was most similar to the waveform of the reference signal and suggesting this is a feasible strategy to effectively reduce noise and more accurately determine signals. The proposed time-frequency denoising method exhibits excellent performance with promising potential for practical application.