WO2024083059A1 - 一种基于机器学习的工作记忆任务脑磁图分类系统 - Google Patents

一种基于机器学习的工作记忆任务脑磁图分类系统 Download PDF

Info

Publication number
WO2024083059A1
WO2024083059A1 PCT/CN2023/124641 CN2023124641W WO2024083059A1 WO 2024083059 A1 WO2024083059 A1 WO 2024083059A1 CN 2023124641 W CN2023124641 W CN 2023124641W WO 2024083059 A1 WO2024083059 A1 WO 2024083059A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
magnetoencephalogram
module
working memory
meg
Prior art date
Application number
PCT/CN2023/124641
Other languages
English (en)
French (fr)
Inventor
张瑜
钱浩天
孙超良
王志超
张欢
蒋田仔
Original Assignee
之江实验室
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 之江实验室 filed Critical 之江实验室
Publication of WO2024083059A1 publication Critical patent/WO2024083059A1/zh

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
    • A61B5/242—Detecting biomagnetic fields, e.g. magnetic fields produced by bioelectric currents
    • A61B5/245—Detecting biomagnetic fields, e.g. magnetic fields produced by bioelectric currents specially adapted for magnetoencephalographic [MEG] signals
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235—Details of waveform analysis
    • A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
    • A61B5/7267—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems involving training the classification device
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/30—Noise filtering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects

Definitions

  • the present invention relates to the technical field of neuroimaging data analysis and machine learning, and in particular to a working memory task magnetoencephalogram classification system based on machine learning.
  • EEG EEG collects potential differences at different locations on the scalp and draws them into images. Its advantage is that it has high temporal resolution and can capture changes in the brain in a very short time.
  • the number of traditional EEG electrodes is limited, the spatial resolution is poor, and the conduction of brain waves varies greatly in different media such as brain tissue, cerebrospinal fluid, skull, and skin, making source reconstruction difficult (Roberta Grech, Cassar Tracey, Muscat Joseph, et al. Review on solving the inverse problem in EEG source analysis [J]. Journal of neuroengineering and rehabilitation, 2008, 5 (1): 25).
  • fMRI is based on the blood oxygenation level dependent (BOLD) of the brain, reflecting the changes in deoxyhemoglobin concentration caused by task-induced or spontaneous neurometabolism, thereby indirectly reflecting the functional activity of neurons in the brain (Ugurbil K. Development of functional imaging in the human brain (fMRI); the University of Minnesota experience [J]. Neuroimage, 2012, 62 (2): 613-619).
  • fMRI has the advantages of high spatial resolution and non-invasive acquisition.
  • the current temporal resolution of fMRI is generally 2s, which is still relatively low compared to the ever-changing changes in brain function and is insufficient to capture the instantaneous changes in the brain during different tasks.
  • Magnetoencephalography studies brain activity by recording the magnetic field generated by the electrical activity of neuronal groups.
  • the magnetic field signal collected by the magnetoencephalography device is the superposition of the magnetic field generated by all the neuronal activities inside the brain.
  • the complex neuronal activity is abstracted into a dipole model, and then the mathematical method is used to simulate this process according to the propagation law of the magnetic field in space, which is the so-called direct problem in magnetoencephalography research.
  • the opposite of the above process is the inverse problem in magnetoencephalography processing, that is, the internal brain signal is inferred through relevant algorithms. This process is also called MEG source reconstruction.
  • the source reconstruction step of MEG data is the key in MEG data processing (S Baillet. Magnetoencephalography for brain electrophysiology and imaging [J]. Nat Neurosci, 2017, 20 (3): 327-339).
  • MEG data processing S Baillet. Magnetoencephalography for brain electrophysiology and imaging [J]. Nat Neurosci, 2017, 20 (3): 327-339.
  • Existing methods are all based on low-level processing scripts, which often have single functions and can only implement a certain step in the MEG data processing process. The process is cumbersome and requires operators to have certain processing experience.
  • unsupervised learning methods In early studies, researchers usually used unsupervised learning methods to obtain relevant spatiotemporal data features for exploring potential explanatory factors in unlabeled data in view of the unique high-dimensional characteristics of neuroimaging data.
  • Commonly used unsupervised learning methods include K-means clustering, hierarchical clustering, and autoencoding. These methods are applied to brain science and neuropsychiatric diseases, and the main findings of the research can be roughly divided into the following aspects: differences between groups of subjects; potential spatial patterns in related fluctuations; and temporal dynamic structures under certain states.
  • the use of supervised learning techniques, especially classification techniques, for brain imaging data and predictive classification of subjects at the individual level has aroused great interest among researchers in related fields.
  • the key steps in building a classification prediction model include feature extraction and selection, model selection and training, and model effect evaluation.
  • machine learning research on neuroimaging data how to screen high-performance features from complex data and build a classification prediction model, and use it as a biomarker for further research, has always been a difficult problem that has plagued researchers. If a complete system can be built for processing neuroimaging data, extracting relevant features, and constructing machine learning classification prediction models, it will play an important role in related research on brain imaging such as magnetoencephalography and fMRI, and will also be of great significance for the study of brain-related neural mechanisms.
  • the purpose of the present invention is to address the deficiencies of the prior art and propose a working memory task magnetoencephalogram classification system based on machine learning, which can perform a complete process from preprocessing to source reconstruction analysis on magnetoencephalogram data, and use a machine learning model to classify the magnetoencephalogram data of working memory tasks.
  • a working memory task magnetoencephalogram classification system based on machine learning comprising a magnetoencephalogram data acquisition module, a magnetoencephalogram data preprocessing module, a magnetoencephalogram source reconstruction module and a machine learning classification module;
  • the magnetoencephalogram data acquisition module is used to collect magnetoencephalogram data of subjects in different working memory task states and input them into the magnetoencephalogram data preprocessing module;
  • the MEG data preprocessing module is used to preprocess the MEG data of different working memory task states, including a data quality control submodule, a low-quality channel and data segment filtering submodule, and a noise pseudo-shadow separation module;
  • the data quality control submodule is used to perform quality verification on the magnetoencephalogram data of different working memory task states; the low-quality channel and data segment filtering submodule is used to filter out channels and data segments that do not meet the requirements; the noise artifact separation module is used to perform noise removal and artifact identification;
  • the MEG source reconstruction module is used to perform sensor signal analysis and source-level traceability reconstruction analysis on the working memory MEG data after the MEG data preprocessing module to obtain power time series characteristics;
  • the machine learning classification module is used to perform dimensionality reduction using a principal component analysis method based on the power time series features obtained in the magnetoencephalogram source reconstruction module, and finally use a machine learning model to classify the working memory task category to which the subject belongs.
  • the data quality control submodule is used to perform preliminary data verification on the magnetoencephalogram data collected by the magnetoencephalogram data collection module, and output a document of data quality information.
  • the data quality information recorded by the data quality control submodule includes: MEG sampling frequency, data recording duration, number of MEG channels, number of reference channels, number of ECG channels, number of EMG channels, number of recorded events and average coil movement.
  • the filtering out low-quality channels and data segments submodule is used to detect noisy channels by checking the signal similarity between each magnetoencephalogram channel sensor and its adjacent sensors, and channels and data segments that exhibit a correlation threshold below or a variance ratio threshold above the adjacent channel will be marked as bad channels and bad data segments and removed from subsequent analysis.
  • the noise artifact separation module is used to extract independent components using an independent component analysis method and classify them into brain or noise components, and then threshold the three parameters of correlation between independent component signals, correlation between power and time series, and correlation between spectra, and perform noise removal and artifact identification by repeatedly selecting iterations with the highest brain component and lowest artifact contamination.
  • the sensor signal analysis in the magnetoencephalogram source reconstruction module includes time-locked analysis and time-frequency analysis; Frequency analysis obtains the relationship between signal frequency and time; time-locked analysis is used to obtain the brain's processing process of a certain event and the activity state before and after the event.
  • the source-level tracing reconstruction analysis goal in the magnetoencephalogram source reconstruction module is to use the magnetic field distribution around the subject's head to reversely infer the magnetic field changes inside the brain.
  • the tracing reconstruction analysis adopts beamforming technology and uses a coherent source dynamic imaging method or a linear constrained minimum variance method for source reconstruction.
  • the machine learning model in the machine learning classification module includes a support vector machine model, a logistic regression model or a random forest model.
  • the present invention Compared with the background technology, the present invention has the following beneficial effects: the present invention provides a working memory task MEG classification system based on machine learning, which first solves the problem that the current MEG data processing is based on the underlying processing script, the process is cumbersome and there is no complete automated process.
  • the system in the present invention clearly organizes the process of MEG data from preprocessing to sensor signal analysis and then to source reconstruction, which is easy to operate and meets the needs of automated analysis.
  • the system of the present invention realizes dimensionality reduction of extracted features in the machine learning classification module, and constructs a machine learning classification model for working memory MEG data, which can more accurately output the working memory task category to which the subject belongs based on the input MEG data, which plays an important role in the relevant research on MEG, and is also of great significance for the research on the neural mechanism related to the brain's working memory.
  • FIG1 is a structural diagram of a working memory task magnetoencephalogram classification system based on machine learning provided by an embodiment of the present invention.
  • FIG. 2 is a schematic diagram of a magnetoencephalogram data preprocessing module provided in an embodiment of the present invention.
  • FIG. 3 is a schematic diagram of a magnetoencephalogram source reconstruction module provided in an embodiment of the present invention.
  • FIG4 is a schematic diagram of a machine learning classification module provided in an embodiment of the present invention.
  • the present invention proposes a working memory task magnetoencephalogram classification system based on machine learning, including magnetoencephalogram
  • the data acquisition module, the MEG data preprocessing module, the MEG source reconstruction module and the machine learning classification module can perform a complete process from preprocessing to source reconstruction analysis on the working memory task MEG data, and use the machine learning model to classify the working memory task MEG data.
  • the structure of a working memory task MEG classification system based on machine learning proposed by the present invention is shown in FIG1 .
  • the input system is a working memory MEG public data set collected by the MEG data acquisition module. First, it passes through the MEG data preprocessing module, which is used to preprocess the original MEG data of different working memory task states.
  • a data quality control submodule mainly includes a data quality control submodule, a low-quality channel and data segment filtering submodule, and a noise artifact separation module to perform data quality control, filter low-quality channels and data segments, and separate noise artifacts through an independent component analysis (ICA) method;
  • ICA independent component analysis
  • MEG source reconstruction module which is used to perform sensor signal analysis (Sensor-level Analysis) and source level analysis (Source Analysis) on the data after the MEG data preprocessing module to obtain a power time series;
  • the machine learning classification module using the power time series results obtained in the MEG source reconstruction module as the original features, and then the original features are subjected to feature dimension reduction through the principal component analysis (Principal Component Analysis, PCA) method, and finally the support vector machine (SVM) and other machine learning models are trained for classification.
  • PCA Principal component analysis
  • SVM support vector machine
  • the MEG data acquisition module is used to collect MEG data from the Human Connectome Project (HCP, link address: https://db.humanconnectome.org/data/projects/HCP_1200), of which 95 subjects underwent MEG data acquisition for working memory tasks, and the number of subjects available was 83.
  • the MEG system includes 248 magnetometer channels and 23 reference channels with a sampling rate of 2034.5101 Hz.
  • the ECG, EOG, and EMG signals are collected synchronously with the MEG, and the contact resistance of all electrodes is controlled within 10K ohms.
  • the magnetoencephalogram data preprocessing module is used to input the magnetoencephalogram data of the HCP into the magnetoencephalogram data preprocessing module, and the magnetoencephalogram data preprocessing module includes a data quality control submodule, a low-quality channel and data segment filtering submodule, and a noise pseudo-shadow separation module; this embodiment is described by taking the working memory magnetoencephalogram data of the HCP as an example;
  • the data quality control submodule is used to perform preliminary data verification on the MEG data collected by the MEG data acquisition module, and output a document of data quality information;
  • the recorded data quality information includes: MEG sampling frequency, data recording duration, number of MEG channels, number of reference channels, number of ECG channels, number of EMG channels, number of recorded events and average coil movement.
  • the low-quality channel and data segment filtering submodule is used to check each MEG channel sensor and its adjacent sensors. Channels that exhibit a correlation below a threshold (0.4) or a variance ratio above a threshold (0.5) with adjacent channels are marked as bad channels and bad data segments and removed from subsequent analysis.
  • the noise artifact separation module is used to separate and remove noise and artifacts using an independent component analysis method.
  • the independent component analysis method is a blind source separation method, which aims to recover the original independent signals from the mixed observation signals and find a set of linear transformations so that the unmixed signals are independent of each other.
  • the independent component analysis method is used in the noise artifact separation module to extract each independent component and classify it into brain or noise components. Then, the parameters such as the correlation between independent component signals, the correlation between power and time series, and the correlation between spectra are thresholded, and noise removal and artifact identification are performed by repeatedly selecting iterations with the highest brain component and the lowest artifact pollution.
  • the identified noise components include electrocardiogram and eye movement artifacts, power bursts, and environmental noise.
  • the preprocessed HCP working memory magnetoencephalogram data were obtained.
  • the trials of these data were split, and each task state of each subject was split into several trials. Subsequent processing was then performed based on the data split into individual trials.
  • the MEG source reconstruction module is used to input the HCP data processed by the MEG data preprocessing module, and is mainly used for MEG data sensor signal analysis and source-level traceability reconstruction analysis.
  • the embodiment of the present invention takes the processing flow of HCP working memory MEG data in the MEG source reconstruction module as an example to illustrate:
  • the MEG data sensor signal analysis is to directly analyze the signals collected by the MEG sensor and use it as the basis for source reconstruction, mainly including time-locked analysis and time-frequency analysis.
  • the time-frequency analysis method provides joint distribution information in the time domain and the frequency domain, and clearly reflects the relationship between the signal frequency and time.
  • Time-locked analysis can analyze the brain's processing of an event, that is, the activity state before and after the event, and is often used to calculate event-related fields (ERFs) and covariance matrices.
  • ERPs event-related fields
  • the system of the present invention uses the MRI data provided by HCP to construct a head model and a source model.
  • the head model is constructed based on the subject's own MRI T1 structural image.
  • a head model is constructed for each subject.
  • the individual MRI data of the subject will be aligned with the standard template, and the transformation matrix obtained by the alignment will be used to inversely transform the regular grid in the standard space to obtain the source model in the individual subject space.
  • the MEG source reconstruction module integrates two currently recognized source tracing algorithms using beamforming technology, namely the Dynamic Imaging of Coherent Sources (DICS) method and the Linear Constrained Minimum Variance (LCMV) method.
  • the DICS algorithm is based on frequency domain data, and time-frequency analysis is required before analysis to obtain the time-frequency distribution of the data.
  • the LCMV source analysis method is based on time domain data, and time-locked analysis is required to calculate the covariance matrix of the data before analysis.
  • two routes can be selected: time-frequency analysis-DICS algorithm source reconstruction or time-locked analysis-LCMV algorithm source reconstruction.
  • the sensor signal analysis adopts the time-frequency analysis method, and the source reconstruction is performed using the DICS algorithm. Finally, the power time series characteristic results in different frequency bands can be obtained.
  • the embodiment of the present invention takes the results under the alpha frequency band (8-15Hz) as an example.
  • N represents the number of defined time points and M represents the number of cortical vertices.
  • N is defined as 200, representing 200 time points from -1.5s to 2.5s, and each time point is separated by 0.02s. Its meaning is to take the moment when the subject receives the visual stimulation as the zero time point, intercept the trial time period from 1.5s before to 2.5s after the zero time point, and analyze the data of the 200 time points in this time period.
  • the system of the present invention uses the cortical space with a resolution of 4K provided by the Human Connectome Project, and the M corresponding to this space represents 8004 vertices.
  • the machine learning classification module inputs the power time series feature result data processed by the MEG data preprocessing module and the MEG source reconstruction module, which is mainly used to perform feature dimension reduction on the power time series features, and finally train a variety of machine learning models such as a support vector machine model, a logistic regression model or a random forest model for classification.
  • HCP MEG data as an example, as follows:
  • the power time series results of each subject are used as feature data sets, and the principal component analysis method is used to reduce the dimension of the feature data sets.
  • Principal component analysis is often used to reduce the dimension of a data set while retaining the features that contribute the most to the variance in the data set.
  • the specified information is retained to 95% of the original level.
  • the reduced-dimensional data set is then randomly divided into a training set, a validation set, and a test set in a ratio of 7:1:2, based on the subject.
  • the training set data is input into the machine learning model for training.
  • the machine learning models used include support vector machine model, logistic regression model or random forest model.
  • the validation set data is used to tune the model's hyperparameters.
  • the trained machine learning model can output whether the working memory task category of the subject is 0-Back or 2-Back. After testing on the test set data, the feature data of this example can obtain good classification results on all three models.

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Theoretical Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • Medical Informatics (AREA)
  • Software Systems (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Multimedia (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Pathology (AREA)
  • Public Health (AREA)
  • Biophysics (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Molecular Biology (AREA)
  • Surgery (AREA)
  • Animal Behavior & Ethology (AREA)
  • Mathematical Physics (AREA)
  • Veterinary Medicine (AREA)
  • Fuzzy Systems (AREA)
  • Databases & Information Systems (AREA)
  • Physiology (AREA)
  • Psychiatry (AREA)
  • Signal Processing (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)

Abstract

本发明公开了一种基于机器学习的工作记忆任务脑磁图分类系统，包括脑磁图数据采集模块、脑磁图数据预处理模块、脑磁图源重建模块和机器学习分类模块，其中脑磁图数据采集模块用于采集被试不同工作记忆任务态的脑磁图数据，脑磁图数据预处理模块用于对不同工作记忆任务态的脑磁图数据进行质量控制并分离噪声和伪影，脑磁图源重建模块用于对经过脑磁图数据预处理模块的数据进行传感器信号分析及源重建分析，机器学习分类模块以功率时间序列作为特征，对被试所属的工作记忆任务进行分类。本发明整合了对工作记忆脑磁图数据完整的从预处理到源重建的分析流程，对工作记忆任务脑磁图数据分类，对于工作记忆解码和大脑记忆相关机制的研究有重要意义。

Description

一种基于机器学习的工作记忆任务脑磁图分类系统 技术领域
本发明涉及神经影像数据分析和机器学习技术领域，尤其涉及一种基于机器学习的工作记忆任务脑磁图分类系统。
背景技术
大脑作为人体中最为神秘的器官，人们不断的在探索大脑的奥秘，试图了解大脑的内部机理和工作机制。近年来，随着功能磁共振成像(functional magnetic resonance imaging,fMRI)、脑电图(Electroencephalogram,EEG)、脑磁图(Magnetoencephalography,MEG)、正电子发射断层扫描(Positron Emission Computed Tomography,PET)等影像技术的产生与应用，人们对于大脑的研究有了新的发展，这些技术也在探索人类大脑机理和相关认知活动方面发挥越来越大的作用。其中，目前利用EEG、fMRI和MEG等功能成像技术探索大脑已成为热门研究之一(He B,Liu Z.Multimodal Functional Neuroimaging:Integrating Functional MRI and EEG/MEG[J].IEEE reviews in biomedical engineering,2008,1(2008):23-40)。EEG通过采集头皮上不同位置电位差绘制成图像，其优势在于时间分辨率高，可以捕捉大脑在极短时间内的变化，但传统的EEG电极数量有限，空间分辨率差，且脑电传导在脑组织、脑脊液、颅骨、皮肤等不同介质中差异较大，源重建较为困难(Roberta Grech,Cassar Tracey,Muscat Joseph,et al.Review on solving the inverse problem in EEG source analysis[J].Journal of neuroengineering and rehabilitation,2008,5(1):25)。fMRI基于大脑的血氧水平依赖(blood oxygenation level dependent，BOLD)，反映了由任务诱发或自发神经代谢活动而引起的脱氧血红蛋白浓度变化，从而间接反映大脑的神经元功能活动(Ugurbil K.Development of functional imaging in the human brain(fMRI)；the University of Minnesota experience[J].Neuroimage,2012,62(2):613-619)。fMRI具有空间分辨率高、采集无创等优点，然而现有的fMRI的时间分辨率一般为2s，相比于瞬息万变的脑功能变化来说，仍然较低，不足以捕捉大脑处理不同任务过程中的瞬间变化。脑磁图(MEG)通过记录神经元群电活动产生的磁场来研究大脑活动，它以非常高(毫秒内)的时间分辨率直接、无创地记录大脑活动，并生成动态、信息丰富的大规模大脑活动表示(J Gross.Magnetoencephalography in Cognitive Neuroscience:A Primer[J].Neuron,2019,104(2):189-204)。其具有极高的空间分辨率(1-3mm)和时间分辨率(1-2ms)，且磁场受脑脊液、颅骨等干扰小，检查无创、无放射性，在研究脑功能方面有其独特的优势。利用MEG技术 对人体大脑功能进行研究具有重要意义，有望为揭示大脑工作机制提供新的资料。
其中，对于脑磁图技术，其收集的是被试头部周围的磁场分布，而不是大脑内部的磁场分布，脑磁图设备采集到的磁场信号是由大脑内部所有神经元活动产生磁场叠加而成的。将复杂的神经元活动抽象成偶极子模型，再使用数学方法按照磁场在空间中的传播规律对这一过程进行模拟，即脑磁图研究中所谓的正问题。上述过程相反的是脑磁图处理中的逆问题，即通过相关算法对大脑内部信号进行推算，这个过程也被称作MEG源重建。由于大脑内部神经元数远大于脑磁图设备的传感器数，故逆问题会有无穷多个解，使得MEG的源重建问题存在一定不确定性。因此，对于MEG数据的源重建步骤是MEG数据处理中的关键(S Baillet.Magnetoencephalography for brain electrophysiology and imaging[J].Nat Neurosci,2017,20(3):327-339)。目前，关于脑磁图数据的从预处理到传感器信号分析再到溯源重建的流程，还没有工具将其完整的整合起来，实现较为自动的处理MEG数据。现有的方法都是基于较底层的处理脚本，这些脚本往往功能单一，只能实现MEG数据处理流程中的某一步骤，并且过程繁琐，需要操作人员有一定处理经验。
随着人工智能技术正在迅速发展，越来越多的研究人员都开始将机器学习技术引入认知神经科学和脑科学的探索中来(Khosla M,Jamison K,Ngo G H,et al.Machine learning in resting-state fMRI analysis[J].Magnetic Resonance Imaging,2019,64:101-121)。在脑科学研究中，应用最广泛的机器学习方法有分类、回归和聚类等。随着数据质量规模的提升以及机器学习研究方法的不断发展，机器学习方法在fMRI和EEG的研究中已经取得了较为突出的成果，但在MEG领域相关的研究还很少。
早期的研究中，针对神经影像数据独特的高维特性，研究人员通常采用无监督学习的方法，以获得相关的时空数据特征，用于探索未标记数据中潜在的解释因素。常用的无监督学习方法有K-means聚类、分层聚类、自编码等。这些方法应用于脑科学和神经精神疾病中，研究的主要发现可大致划分为以下几方面：被试者组间的差异性；体现相关波动中潜在的空间模式；某些状态下的时间动态结构。近年来，使用监督学习技术，尤其是分类技术，用于脑影像数据并对被试进行个体水平上的预测分类，在相关领域引起了研究者们极大的兴趣。构建分类预测模型的关键步骤包括特征提取和选择，模型选择和训练，模型效果评估等。在对神经影像数据的机器学习研究中，如何从繁杂的数据中筛选出高性能特征并构建分类预测模型，并以此作为进一步研究的生物标志物，一直是困扰研究者的难题。若能构建出完整的用于神经影像数据处理，相关特征提取，并构建机器学习分类预测模型的系统，将会对脑磁图、fMRI等脑影像的相关研究起到重要的作用，对于大脑相关神经机制的研究也有着重要意义。
发明内容
本发明的目的在于针对现有技术的不足，提出一种基于机器学习的工作记忆任务脑磁图分类系统，可以对脑磁图数据进行完整的从预处理到源重建分析的流程，并使用机器学习模型对工作记忆任务的脑磁图数据进行分类。
本发明是通过以下技术方案来实现的：一种基于机器学习的工作记忆任务脑磁图分类系统，包括脑磁图数据采集模块、脑磁图数据预处理模块、脑磁图源重建模块和机器学习分类模块；
所述脑磁图数据采集模块用于采集被试不同工作记忆任务态的脑磁图数据，并输入至脑磁图数据预处理模块；
所述脑磁图数据预处理模块用于对不同工作记忆任务态的脑磁图数据进行预处理，包括数据质量控制子模块、滤除低质量信道与数据段子模块和分离噪声伪影子模块；
所述数据质量控制子模块用于对不同工作记忆任务态的脑磁图数据进行质量校验；所述滤除低质量信道与数据段子模块用于滤除不满足需求的通道和数据段；所述分离噪声伪影子模块用于进行噪声去除和伪影识别；
所述脑磁图源重建模块用于对经过脑磁图数据预处理模块后的工作记忆脑磁图数据进行传感器信号分析及源级别的溯源重建分析，得到功率时间序列特征；
所述机器学习分类模块用于根据脑磁图源重建模块中得到的功率时间序列特征通过主成分分析方法进行降维，最后采用机器学习模型，对被试所属的工作记忆任务类别进行分类。
进一步地，所述数据质量控制子模块用于对脑磁图数据采集模块采集的脑磁图数据进行初步的数据校验，并输出数据质量信息的文档。
进一步地，所述数据质量控制子模块记录的数据质量信息包括：脑磁图采样频率、记录数据时长、脑磁图通道数、参考通道数、心电图通道数、肌电图通道数、记录的事件数和平均线圈移动。
进一步地，所述滤除低质量信道与数据段子模块用于通过检查每个脑磁图通道传感器与其相邻传感器之间的信号相似性来检测噪声通道，与相邻通道表现出低于相关性阈值或高于方差比阈值的通道和数据段将被标记为坏通道和坏数据段并从后续分析中移除。
进一步地，所述分离噪声伪影子模块用于使用独立成分分析方法提取出各独立成分，并将其分类为大脑或噪声成分，再通过对独立成分信号之间的相关性、功率与时间序列之间的相关性以及频谱之间的相关性这三个参数进行阈值化，通过多次选择具有最高大脑成分和最低伪影污染的迭代来进行噪声去除和伪影识别。
进一步地，所述脑磁图源重建模块中传感器信号分析包括锁时分析和时频分析；所述时 频分析得到信号频率随时间变化的关系；锁时分析用于得到大脑对某一事件的处理过程，得到事件前后的活动状态。
进一步地，所述脑磁图源重建模块中源级别的溯源重建分析目标是利用被试头部周围的磁场分布情况逆向推算出大脑内部的磁场变化，溯源重建分析采用波束形成技术，使用基于相干源动态成像方法或基于线性约束最小方差方法进行溯源重建。
进一步地，所述机器学习分类模块中机器学习模型包括支持向量机模型、逻辑回归模型或随机森林模型。
本发明与背景技术相比，具有的有益效果是：本发明提供了一种基于机器学习的工作记忆任务脑磁图分类系统，首先解决了目前对于脑磁图数据处理都基于底层的处理脚本，过程繁琐且没有自动化完整流程的问题，本发明中的系统清晰的整理出了脑磁图数据从预处理到传感器信号分析再到溯源重建的流程，便于操作和满足自动化分析的需要。其次，本发明系统基于脑磁图数据预处理模块和脑磁图源重建模块得到的脑磁图功率时间序列结果，在机器学习分类模块中实现对提取特征的降维，并构建了工作记忆脑磁图数据的机器学习分类模型，可以较为准确的根据输入的脑磁图数据输出被试所属的工作记忆任务类别，这对于脑磁图的相关研究起到重要的作用，对于大脑工作记忆相关神经机制的研究也有着重要意义。
附图说明
为了更清楚地说明本发明的技术方案，下面将对实施例描述中所需要使用的附图作简单地介绍。显而易见的是，下面描述中的附图仅仅是本申请中记载的特定实施例，其不是对本发明的保护范围的限制。对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，当然还可以根据本发明的如下实施例及其附图获得一些其它的实施例和附图。
图1为本发明实施例提供的一种基于机器学习的工作记忆任务脑磁图分类系统结构图。
图2为本发明实施例提供的脑磁图数据预处理模块示意图。
图3为本发明实施例提供的脑磁图源重建模块示意图。
图4为本发明实施例提供的机器学习分类模块示意图。
具体实施方式
为了使本领域的人员更好地理解本申请中的技术方案，下面将结合附图对本发明作进一步的说明。但这仅仅是本申请的一部分实施例，而不是全部的实施例。基于本申请所述的具体实施例，本领域的其他人员在没有做出创造性劳动的前提下所获得的其他实施例，都应当落在本发明的构思范围之内。
以下参考附图描述本发明的优选实施例。
总体而言，本发明提出一种基于机器学习的工作记忆任务脑磁图分类系统，包括脑磁图 数据采集模块、脑磁图数据预处理模块、脑磁图源重建模块和机器学习分类模块，可以对工作记忆任务脑磁图数据进行完整的从预处理到源重建分析的流程，并使用机器学习模型对工作记忆任务的脑磁图数据进行分类。本发明所提出的一种基于机器学习的工作记忆任务脑磁图分类系统的结构如图1所示。输入系统的为脑磁图数据采集模块采集的工作记忆脑磁图公开数据集，首先经过脑磁图数据预处理模块，用于对不同工作记忆任务态的原始脑磁图数据进行预处理，主要包括数据质量控制子模块、滤除低质量信道与数据段子模块和分离噪声伪影子模块，进行数据质量控制、滤除低质量信道与数据段和通过独立成分分析(Independent Component Analysis,ICA)方法分离噪声伪影；其次经过脑磁图源重建模块，用于对经过脑磁图数据预处理模块后的数据进行传感器信号分析(Sensor-level Analysis)及源级别分析(Source Analysis)，得到功率时间序列；最后经过机器学习分类模块，采用脑磁图源重建模块中得到的功率时间序列结果作为原始特征，再对原始特征通过主成分分析(Principal Component Analysis,PCA)方法进行特征降维，最后训练支持向量机(Support Vector Machine,SVM)等机器学习模型进行分类。输出结果为被试所属的任务态。本发明所提出系统的各模块具体实现如下：
所述脑磁图数据采集模块用于收集来自人类连接组计划(Human Connectome Project,HCP,链接地址：https://db.humanconnectome.org/data/projects/HCP_1200)中公开的脑磁图数据，其中有95名被试进行了工作记忆任务的脑磁图数据采集，可获取的被试数目为83。每个被试工作记忆的数据共进行两次采集，每次采集包含160次经典的N-Back(N＝0或2)工作记忆任务试验，被试需要不断记忆包含人脸和常见工具的图片，并进行判断该图片是否与之前的第N幅图片相同。所有的数据都利用一台位于圣路易斯大学的全脑MAGNES 3600(4D Neuroimaging，San Diego)系统在磁屏蔽室内采集。该MEG系统包括248个磁力计通道和23个参考通道，采样率为2034.5101Hz，心电、眼电、肌电信号与MEG同步采集，所有电极的接触电阻控制在10K欧姆以内。
如图2所示，所述脑磁图数据预处理模块用于将上述HCP的脑磁图数据输入脑磁图数据预处理模块，所述脑磁图数据预处理模块包括数据质量控制子模块、滤除低质量信道与数据段子模块和分离噪声伪影子模块；本实施例以HCP的工作记忆脑磁图数据为例进行说明；
所述数据质量控制子模块用于对脑磁图数据采集模块采集的脑磁图数据进行初步的数据校验，并输出数据质量信息的文档；记录的数据质量信息包括：脑磁图采样频率、记录数据时长、脑磁图通道数、参考通道数、心电图通道数、肌电图通道数、记录的事件数和平均线圈移动。
所述滤除低质量信道与数据段子模块用于通过检查每个脑磁图通道传感器与其相邻传感 器之间的信号相似性来检测噪声通道，与相邻通道表现出低于相关性阈值(0.4)或高于方差比阈值(0.5)的通道将被标记为坏通道和坏数据段并从后续分析中移除。
所述分离噪声伪影子模块用于使用独立成分分析方法分离并去除噪声和伪影。独立成分分析方法是一种盲源分离方法，目的是从混合的观察信号中恢复原始的独立信号，找到一组线性变换使得解混出来的信号互不相关。所述分离噪声伪影子模块中使用独立成分分析方法提取出各独立成分，并将其分类为大脑或噪声成分，再通过对独立成分信号之间的相关性、功率与时间序列之间的相关性和频谱之间的相关性等参数进行阈值化，通过多次选择具有最高大脑成分和最低伪影污染的迭代来进行噪声去除和伪影识别。识别的噪声成分包括心电和眼球运动伪迹、电源爆裂和环境噪声。
经过上述数据质量控制子模块、滤除低质量信道与数据段子模块和分离噪声伪影子模块处理后，得到预处理完成的HCP工作记忆脑磁图数据。对这些数据的试验(trial)进行拆分，每位被试的每个任务态都被拆分为若干个试验。之后基于拆分成单个试验的数据进行后续处理。
如图3所示，所述脑磁图源重建模块用于输入经过了脑磁图数据预处理模块处理后的HCP数据，主要用于脑磁图数据传感器信号分析及源级别的溯源重建分析。本发明实施例以HCP工作记忆脑磁图数据在脑磁图源重建模块的处理流程为例进行说明：
所述脑磁图数据传感器信号分析是对脑磁图传感器采集到的信号直接进行分析并作为源重建时的依据，主要包括锁时分析和时频分析两部分。时频分析方法提供了时间域与频率域的联合分布信息，清楚的体现了信号频率随时间变化的关系。锁时分析可以分析大脑对某一事件的处理过程，即事件前后的活动状态，常用来用于计算事件相关场(Event-related field,ERF)和协方差矩阵。
参考传感器信号分析的结果，进一步进行源级别的溯源重建分析，其目标是利用被试头部周围的磁场分布情况逆向推算出大脑内部的磁场变化。在进行源重建时，需要清晰的定义传感器与大脑的相对空间位置以及大脑的空间范围，即通常所说的头部模型和源模型。本发明系统使用HCP提供的MRI数据构建头部模型和源模型。头部模型是根据被试自身的MRI T1结构像来构建的，参考FieldTrip工具包中的ft_prepareheadmodel函数，对每个被试都构建了头部模型。在构建源模型的过程中，会将被试个体的MRI数据与标准模板进行配准，利用配准的得到的变换矩阵对标准空间上的规则的网格进行逆变换，得到被试个体空间上的源模型。
在溯源重建分析过程中，脑磁图源重建模块集成了目前认可度较高的两种使用波束形成技术的溯源算法，分别是基于相干源动态成像(DICS)方法和基于线性约束最小方差(LCMV) 方法。DICS算法是基于频域数据的，在分析之前需要进行时频分析来获取数据的时频分布。LCMV这种源分析方法是基于时域数据的，在分析之前需要对数据进行锁时分析计算数据的协方差矩阵。在脑磁图源重建模块中，可以选择时频分析-DICS算法溯源重建或是锁时分析-LCMV算法溯源重建两种路线。为能清晰的说明后续流程，本发明实施例中传感器信号分析采用时频分析方法，溯源重建采用DICS算法进行，最后可以得到不同频段下的功率时间序列特征结果。
本发明实施例以alpha频段(8-15Hz)下的结果为例，对于每个被试，分别可以得到N*M的功率序列，其中N表示定义的时间点数目，M表示皮层顶点数。N在本实施例中定义为200，代表-1.5s到2.5s中的200个时间点，每个时间点间隔0.02s，其含义是以被试接受到视觉刺激的时刻为零时间点，截取零时间点的前1.5s到后2.5s内的trial时间段，分析该时间段内200个时间点的数据。本发明系统中采用人类连接组计划提供的4K分辨率的皮层空间，该空间对应的M代表8004个顶点。
如图4所示，所述机器学习分类模块输入经过了脑磁图数据预处理模块和脑磁图源重建模块处理后的功率时间序列特征结果数据，其主要用于对功率时间序列特征进行特征降维，最后训练支持向量机模型、逻辑回归模型或随机森林模型等多种机器学习模型进行分类。本实施例以HCP脑磁图数据为例进行说明，具体如下：
首先将每个被试的功率时间序列结果作为特征数据集，使用主成分分析方法对特征数据集进行降维。主成分分析经常用于减少数据集的维数，同时保持数据集中的对方差贡献最大的特征，本实施例中将指定信息保留到原来的95％的程度。再将降维后的数据集以被试为单位，按7:1:2的比例随机分成训练集、验证集和测试集。
然后将训练集数据输入机器学习模型进行训练，采用的机器学习模型包括支持向量机模型，逻辑回归模型或随机森林模型，采用验证集的数据对模型的超参数进行调优。训练完成的机器学习模型可以输出被试所属的工作记忆任务类别为0-Back还是2-Back。经过在测试集数据上的测试，本实例的特征数据在这三种模型上都可以得到良好的分类结果。
以上所述仅是本申请的优选实施方式。本申请不会被限制于本文所述的这些具体实施例，而是可以覆盖与本文所公开的原理和新颖特点相一致的最宽的范围。

Claims (6)

  1. 一种基于机器学习的工作记忆任务脑磁图分类系统，其特征在于，包括脑磁图数据采集模块、脑磁图数据预处理模块、脑磁图源重建模块和机器学习分类模块；
    所述脑磁图数据采集模块用于采集被试不同工作记忆任务态的脑磁图数据，并输入至脑磁图数据预处理模块；
    所述脑磁图数据预处理模块用于对不同工作记忆任务态的脑磁图数据进行预处理，包括数据质量控制子模块、滤除低质量信道与数据段子模块和分离噪声伪影子模块；
    所述数据质量控制子模块用于对不同工作记忆任务态的脑磁图数据进行质量校验；所述滤除低质量信道与数据段子模块用于滤除不满足需求的通道和数据段；所述分离噪声伪影子模块用于进行噪声去除和伪影识别；
    所述脑磁图源重建模块用于对经过脑磁图数据预处理模块后的工作记忆脑磁图数据进行传感器信号分析及源级别的溯源重建分析，得到功率时间序列特征；
    具体地，所述脑磁图源重建模块中传感器信号分析包括锁时分析和时频分析；所述时频分析得到信号频率随时间变化的关系；锁时分析用于得到大脑对某一事件的处理过程，得到事件前后的活动状态；所述脑磁图源重建模块中源级别的溯源重建分析目标是利用被试头部周围的磁场分布情况逆向推算出大脑内部的磁场变化，溯源重建分析采用波束形成技术，使用基于相干源动态成像方法或基于线性约束最小方差方法进行溯源重建；
    所述机器学习分类模块用于根据脑磁图源重建模块中得到的功率时间序列特征通过主成分分析方法进行降维，最后采用机器学习模型，对被试所属的工作记忆任务进行分类。
  2. 根据权利要求1所述的一种基于机器学习的工作记忆任务脑磁图分类系统，其特征在于，所述数据质量控制子模块用于对脑磁图数据采集模块采集的脑磁图数据进行初步的数据校验，并输出数据质量信息的文档。
  3. 根据权利要求2所述的一种基于机器学习的工作记忆任务脑磁图分类系统，其特征在于，所述数据质量控制子模块记录的数据质量信息包括：脑磁图采样频率、记录数据时长、脑磁图通道数、参考通道数、心电图通道数、肌电图通道数、记录的事件数和平均线圈移动。
  4. 根据权利要求1所述的一种基于机器学习的工作记忆任务脑磁图分类系统，其特征在于，所述滤除低质量信道与数据段子模块用于通过检查每个脑磁图通道传感器与其相邻传感器之间的信号相似性来检测噪声通道，与相邻通道表现出低于相关性阈值或高于方差比阈值的通道和数据段将被标记为坏通道和坏数据段并从后续分析中移除。
  5. 根据权利要求1所述的一种基于机器学习的工作记忆任务脑磁图分类系统，其特征在于，所述分离噪声伪影子模块用于使用独立成分分析方法提取出各独立成分，并将其分类为大脑或噪声成分，再通过对独立成分信号之间的相关性、功率与时间序列之间的相关性以及频谱之间的相关性这三个参数进行阈值化，通过多次选择具有最高大脑成分和最低伪影污染的迭代来进行噪声去除和伪影识别。
  6. 根据权利要求1所述的一种基于机器学习的工作记忆任务脑磁图分类系统，其特征在于，所述机器学习分类模块中机器学习模型包括支持向量机模型、逻辑回归模型或随机森林模型。
PCT/CN2023/124641 2022-10-19 2023-10-16 一种基于机器学习的工作记忆任务脑磁图分类系统 WO2024083059A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202211276176.0A CN115359236B (zh) 2022-10-19 2022-10-19 一种基于机器学习的工作记忆任务脑磁图分类系统
CN202211276176.0 2022-10-19

Publications (1)

Publication Number Publication Date
WO2024083059A1 true WO2024083059A1 (zh) 2024-04-25

Family

ID=84008847

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/124641 WO2024083059A1 (zh) 2022-10-19 2023-10-16 一种基于机器学习的工作记忆任务脑磁图分类系统

Country Status (2)

Country Link
CN (1) CN115359236B (zh)
WO (1) WO2024083059A1 (zh)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115359236B (zh) * 2022-10-19 2023-03-28 之江实验室 一种基于机器学习的工作记忆任务脑磁图分类系统
CN116863025B (zh) * 2023-09-05 2023-12-05 之江实验室 脑磁图数据的溯源重建方法、装置、电子装置和介质

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112446264A (zh) * 2019-09-03 2021-03-05 西安慧脑智能科技有限公司 一种基于图像特征的脑磁图解码方法
CN113160975A (zh) * 2021-04-25 2021-07-23 宁波磁波智能科技有限公司 基于原子磁强计的高精度多通道脑磁图系统
US20210346096A1 (en) * 2020-05-11 2021-11-11 Carnegie Mellon University Methods and apparatus for electromagnetic source imaging using deep neural networks
CN114065825A (zh) * 2022-01-17 2022-02-18 北京航空航天大学杭州创新研究院 一种基于结构相似性的脑磁meg源定位方法
CN115132349A (zh) * 2022-06-10 2022-09-30 华中师范大学 一种事件相关脑活动时空频动态定位系统及方法
CN115359236A (zh) * 2022-10-19 2022-11-18 之江实验室 一种基于机器学习的工作记忆任务脑磁图分类系统

Family Cites Families (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP5245063B2 (ja) * 2008-08-22 2013-07-24 敏明 松井 脳磁界計測装置の位置合わせ装置
CN110393525B (zh) * 2019-06-18 2020-12-15 浙江大学 一种基于深度循环自编码器的大脑活动检测方法
CN110604565A (zh) * 2019-08-02 2019-12-24 北京脑陆科技有限公司 一种基于便携式eeg设备的脑部健康筛查方法
CN111340142B (zh) * 2020-05-14 2020-08-14 南京慧脑云计算有限公司 一种癫痫脑磁图棘波自动检测方法与溯源定位系统
CN112220485A (zh) * 2020-09-03 2021-01-15 北京慧脑云计算有限公司 一种癫痫脑电图/脑磁图异常波自动检测方法与定位系统
CN114052668B (zh) * 2022-01-17 2022-06-17 北京航空航天大学杭州创新研究院 一种基于脑磁图数据的脑功能分析方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112446264A (zh) * 2019-09-03 2021-03-05 西安慧脑智能科技有限公司 一种基于图像特征的脑磁图解码方法
US20210346096A1 (en) * 2020-05-11 2021-11-11 Carnegie Mellon University Methods and apparatus for electromagnetic source imaging using deep neural networks
CN113160975A (zh) * 2021-04-25 2021-07-23 宁波磁波智能科技有限公司 基于原子磁强计的高精度多通道脑磁图系统
CN114065825A (zh) * 2022-01-17 2022-02-18 北京航空航天大学杭州创新研究院 一种基于结构相似性的脑磁meg源定位方法
CN115132349A (zh) * 2022-06-10 2022-09-30 华中师范大学 一种事件相关脑活动时空频动态定位系统及方法
CN115359236A (zh) * 2022-10-19 2022-11-18 之江实验室 一种基于机器学习的工作记忆任务脑磁图分类系统

Also Published As

Publication number Publication date
CN115359236B (zh) 2023-03-28
CN115359236A (zh) 2022-11-18

Similar Documents

Publication Publication Date Title
Khosla et al. A comparative analysis of signal processing and classification methods for different applications based on EEG signals
Becker et al. Emotion recognition based on high-resolution EEG recordings and reconstructed brain sources
Fatourechi et al. EMG and EOG artifacts in brain computer interface systems: A survey
WO2024083059A1 (zh) 一种基于机器学习的工作记忆任务脑磁图分类系统
Hsu Single-trial motor imagery classification using asymmetry ratio, phase relation, wavelet-based fractal, and their selected combination
CN106407733A (zh) 基于虚拟现实场景脑电信号的抑郁症风险筛查系统和方法
Zhao et al. Classification of five emotions from EEG and eye movement signals: complementary representation properties
Islam et al. Probability mapping based artifact detection and removal from single-channel EEG signals for brain–computer interface applications
Yi et al. A novel method for constructing EEG large-scale cortical dynamical functional network connectivity (dFNC): WTCS
Rommers et al. Electrophysiological methods
Shi et al. Categorizing objects from MEG signals using EEGNet
Miladinović et al. Performance of dual-augmented Lagrangian method and common spatial patterns applied in classification of motor-imagery BCI
Torse et al. Nonlinear blind source separation for EEG signal pre-processing in brain-computer interface system for epilepsy
Vadivu et al. An Novel Versatile Inspiring Wavelet Transform and Resilient Direct Neural Network Classification Techniques for Monitoring Brain Activity System Based on EEG Signal
Qayyum et al. Assessment of cognitive load using multimedia learning and resting states with deep learning perspective
Shen et al. Contrastive Learning of Shared Spatiotemporal EEG Representations Across Individuals for Naturalistic Neuroscience
Sun et al. MEEG-Transformer: transformer Network based on Multi-domain EEG for emotion recognition
Borhani et al. EEG-based visual attentional state decoding using convolutional neural network
Aljalal et al. Mild Cognitive Impairment Detection from EEG Signals Using Combination of EMD Decomposition and Machine Learning
Liu et al. Mutual information in the frequency domain for the study of biological systems
Li CNN-Based Image Analysis for EEG Signal Characterization
Liu Blink-related oscillations: neurotechnology advances that open a new window into brain function
Bashivan Commonality and singularity in working memory network predicting performance and individual differences
Dou Artifact Analysis and Removal of Electroencephalographic (EEG) Recordings
Li A Dual-Modality Emotion Recognition System of EEG and Facial Images and its

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 23879048

Country of ref document: EP

Kind code of ref document: A1