Identity recognition of multi-task functional near-infrared spectroscopy signals based on group average

Abstract

The present invention discloses identity recognition of multi-task functional near-infrared spectroscopy signals based on group averaging. The main steps are as follows: preprocess the original optical signals output by the near-infrared spectroscopy instrument, and extract the preprocessed signals according to different tasks and modalities; construct a functional brain network based on the extracted signals using Pearson correlation; divide the brain network into different groups according to different ratios, and combine the averaging operation to obtain a representative brain network within the group. This method extracts features with individual specificity in a simple way, and at the same time provides a benchmark method for the identity recognition problem based on near-infrared spectroscopy data, having important theoretical significance and practical application value.

Description

Technical Field

[0001] The present invention relates to the technical field of biomedical information processing, and specifically to identity recognition of multi-task functional near-infrared spectroscopy signals based on group average. Background Art

[0002] In the rapid development process of social informatization, the problem of identity recognition has attracted our wide attention. After the 1990s, more and more biometric features have been widely used in the field of identity recognition, such as fingerprints, gait, face, iris, and veins. However, these traditional biometric features have defects to varying degrees. For example, problems such as difficult collection and easy forgery make it difficult to meet people's growing demand for information security. These demands have prompted the academic and industrial communities to explore new biometric features.

[0003] A qualified biometric feature should be universal, unique, and reliable among the population. The collection process should be relatively convenient and not easily stolen. At the same time, the ability to perform live detection has become increasingly important. Functional near-infrared spectroscopy (fNIRS), as an emerging physiological optical signal, can detect changes in blood oxygen levels during brain activity. Compared with traditional biometric features, functional near-infrared spectroscopy has its unique advantages: (1) Liveness, only living individuals can collect near-infrared signals; (2) Anti-forgery, even when signal collection is carried out under duress, it will fail due to different brain states; (3) Universality, near-infrared signals have good stability and are insensitive to movement, so they are applicable to most special populations; (4) Small data volume but rich in information. Compared with two-dimensional image data such as face and fingerprint, the data volume of near-infrared spectroscopy signal data is small, and it contains the concentration changes of two substances, oxy-hemoglobin (Oxy-Hb) and deoxy-hemoglobin (Deoxy-Hb), which also has important research value in medical diagnosis.

[0004] However, simply comparing the intensities of near-infrared spectroscopy signals is meaningless because these signals come from different subjects and the intensities are relative and non-quantitative. The brain functional network constructed based on functional near-infrared spectroscopy signals can describe the connection strength between brain regions and between brain regions in different states, depict the dependence relationship between blood oxygen signals, and at the same time provide a feasible method for exploring the functional mechanism of the brain.

[0005] As the simplest and most effective method for constructing brain functional networks, the Pearson correlation coefficient (PC) has been applied to many functional imaging techniques, such as functional magnetic resonance imaging (fMRI), magnetoencephalography (MEG), electroencephalography (EEG), etc. From a mathematical perspective, PC is used to measure the linear correlation between two variables, while in the present invention, PC is used to measure the dependence relationship between two brain regions. To explore the role of the averaging operation in the BFN for identity recognition problems, we divided the experimental signals under the same task into several groups according to different ratios, extracted the representative signals of each group, and used a support vector machine for multi-classification. This is an exploration based on near-infrared spectroscopy signals in the field of identity recognition, expanding the application of near-infrared spectroscopy in the field of biological information. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide identity recognition based on group-averaged multi-task functional near-infrared spectroscopy signals, and its brain network construction method and classification model do not involve any hyperparameters, and relatively efficient recognition tasks are achieved through the simplest method.

[0007] To solve the above technical problems, the present invention adopts the following technical means:

[0008] (1) First, read the signals collected by the near-infrared spectroscopy device; second, to remove the interference of noise on identity recognition, perform preprocessing operations: band-pass filtering of optical signals, conversion by the modified Lambert-Beer law, and band-pass filtering of oxyhemoglobin concentration signals, so as to remove low-frequency and high-frequency physiological noises such as breathing, heartbeat, and pulse, and then perform baseline processing to eliminate the influence of global signals, and finally obtain clean signals;

[0009] (2) Based on multi-task and multiple modalities, extract the signals related to each experiment from the preprocessed signals based on the execution time of each task to obtain the time series under a certain task; taking oxyhemoglobin as an example, the time series can be expressed as X = [x1, x2,..., x z ∈ R t×n×z , where x i ∈ R t×n , i = 1, 2,..., z represents the signal generated in the i-th experiment, t represents the time series of a certain channel, n represents the number of channels, and z represents the number of experiments.

[0010] (3) Based on the segmented signals, regarding the channels as brain regions and the signals as time series, use Pearson correlation to obtain the connection weights between every two channels in the signals generated in the i-th experiment, and finally obtain a functional brain network matrix W ∈ R i such that each element in the matrix n ×n represents the weight between the i-th channel and the j-th channel;

[0011] (4) Taking all the signals related to one task of one subject as an example, considering the individual specificity of the representative brain functional network after averaging multiple functional brain networks, divide all the signals related to the experiments into several groups, define the number of brain functional networks within each group, and then perform the within-group averaging operation to obtain the within-group representative features, and straighten them into vectors;

[0012] The brain functional networks of all experiments under one task W = [W1, W2,..., W z ∈ R n×n×z ; Considering the utility of the above-constructed brain functional network in the identity recognition problem, the specific process is as follows: Remove the elements on the diagonal of each generated brain functional network to exclude the influence of the self-brain region on itself, and then straighten the brain functional network into a row vector.

[0013] (5) Perform the same operations on the deoxyhemoglobin and oxyhemoglobin obtained in the same experiment respectively, and perform a concatenation operation on the two obtained row vectors to obtain a comprehensive feature A of the hybrid modality ∈ R (n×n)×t where n represents the number of channels and t represents the dimension of the comprehensive feature.

[0014] (6) Next, consider the influence of the within-group averaging method on the identity recognition accuracy. Divide the mixed features of all experiments under the same task into several groups according to different ratios. To maximize the utilization rate of the data, consider three cases of group division: Group 1 (one experiment under the same task is divided into one group), Group 2 (five experiments under the same task are divided into one group), and Group 3 (all experiments under the same task are divided into one group). Then perform an averaging operation on the features in the same group to obtain the comprehensive features of the hybrid modality for each group. The identity recognition task is essentially a classification task, and the most commonly used and simplest support vector machine classifier is used for classification to obtain the final classification result.

[0015] Use the classification accuracy of the classification result to evaluate the influence of the averaging method on the identity recognition accuracy.

[0016] The beneficial technical effects of the present invention are: The brain functional network construction method and the classification model involved in the present invention do not involve any hyperparameters, and strive to achieve a relatively efficient recognition task with the simplest method, which is a bold attempt in the field of biometrics.​ Description of the Drawings

[0017] Figure 1 It is a flowchart of the implementation of the present invention for identity recognition by combining a brain functional network constructed based on near-infrared spectroscopy data with a support vector machine classifier. Detailed Implementation Manner

[0018] The present invention will be further described below in conjunction with embodiments.

[0019] (1) Obtain near-infrared spectroscopy data: In this example, we select a publicly available near-infrared spectroscopy dataset under multi-task, including left-hand unilateral finger-tapping (LHT), right-hand unilateral finger-tapping (RHT), and unilateral foot-tapping (FT). 25 experiments were conducted for each task, and each experiment lasted for an average of 30 s. The first two seconds were the introduction period, and then for the next ten seconds, the subjects performed corresponding actions according to different instructions. Subsequently, the 17th - 19th seconds were the rest state. The dataset contains 30 subjects (17 males and 13 females), and they were all clearly informed of the experimental process before the experiment, signed the experimental consent form, and none of them had mental illnesses that could affect the experimental results.

[0020] (2) Preprocess the near-infrared spectroscopy data: First, read the signals collected by the near-infrared spectroscopy device; secondly, to remove the interference of noise on identity recognition, preprocessing operations are performed: band-pass filtering of the optical signal, conversion by the modified Lambert-Beer law, and band-pass filtering of the oxyhemoglobin concentration signal, so as to remove low-frequency and high-frequency physiological noises such as respiration, heartbeat, and pulse. After that, baseline processing is performed to eliminate the influence of the global signal, and finally clean signals are obtained. In this example, the band-pass filtering range of the optical signal is 0.01 - 0.2 HZ, and the band-pass filtering range of the oxyhemoglobin concentration signal is 0.01 - 0.2 HZ. It should be noted that these preprocessing processes are processed using the nirsLAB of MATLAB and the BBCI toolbox.

[0021] (3) Obtain the time series of oxyhemoglobin and deoxyhemoglobin: According to the experimental paradigm in (1), the preprocessed signals are segmented according to the three tasks, and the signals corresponding to the states are extracted. Taking the change in the oxyhemoglobin concentration of the right-hand unilateral finger-tapping experiment as an example, the extracted signal sequence is expressed as X = [x1, x2,..., x z ∈ R t×n×z , where x i ∈ R t ×nDenote the time series obtained from the \(i\)-th experiment. Let \(t\) represent the length of the time series, \(n\) represent the number of channels, and \(z\) represent the number of experiments. In this example, \(t = 160\), \(n = 20\), and \(z = 25\). Since each subject has three tasks, and each task is repeated 25 times, and during each experiment, two substances, oxygenated hemoglobin and deoxygenated hemoglobin, are measured, each subject thus contains 6 signal data, with the dimension being the same as that of \(X\).

[0022] (4) Construct a brain functional network: Taking the oxygenated hemoglobin signal \(X\) obtained from the right - hand unilateral tapping of a subject as an example, based on the signal \(x\) in each experiment i , construct a functional brain network \(W\in\mathbb{R}^{n\times n}\) using Pearson correlation, where each element \(w_{ij}\) n×n represents the weight element between channel \(i\) and channel \(j\). Calculate the corresponding brain functional network for each experiment, obtaining \(W\in\mathbb{R}^{n\times n}\). ij In this example, \(n = 20\) and \(z = 25\). n×n×z

[0023] (5) Since the obtained brain functional network is a diagonal matrix with all diagonal elements being 1, representing the correlation within the brain regions themselves, subtract an identity matrix to make all the main - diagonal elements become 0. Then, straighten the matrix into a row - vector and arrange it by rows. Then \(W\in\mathbb{R}^{n\times n}\) n ×n×z is transformed into \(W'\in\mathbb{R}^{z\times(n\times n)}\), changing from a third - order tensor to a matrix form. In this example, the dimension of this matrix is \(25\times400\). z×(n×n)

[0024] (6) Repeat the operations in (4) - (5) for the deoxygenated hemoglobin signal obtained from the right - hand unilateral tapping of the subject. Similarly, obtain a feature matrix with a dimension of \(25\times400\). Concatenate the oxygenated hemoglobin and deoxygenated hemoglobin feature matrices by rows to obtain a comprehensive feature under the mixed modality, with a dimension of \(25\times800\).

[0025] (7) Repeat the operations in (3) - (6) for the oxygenated hemoglobin and deoxygenated hemoglobin signals of the right - hand unilateral tapping of all subjects. Then, concatenate them by columns to obtain a comprehensive feature of the mixed modality under the right - hand unilateral tapping task, with a dimension of \(750\times800\).

[0026] (8) Since there are three different tasks in the experimental paradigm, following the principle of maximizing data utilization, repeat the operations in (3) - (7) for both the left - hand unilateral tapping and single - foot stepping tasks. Similarly, obtain comprehensive features of the mixed modality for the corresponding tasks, with dimensions all being \(750\times800\).

[0027] ​​(9)Based on the above steps, the comprehensive features of the mixed modality of all subjects are obtained. The comprehensive features of the mixed modality of all subjects are input into a support vector machine classifier, and a relatively stable classification result is obtained by combining ten-fold cross-validation. The performance index is the classification accuracy (ACC), that is, the ratio of the number of correctly classified samples to the total number of samples, which is defined as:

[0028]

[0029] where FP and TP represent false positives and true positives.

[0030] (10)To verify the influence of the group averaging method on the accuracy of identity recognition, the functional brain networks generated in multiple experiments of the same task are grouped according to a certain proportion and averaged within the group. Similarly, taking the feature matrix W ∈ R under the unilateral tapping of the right hand of a certain subject as an example, in order to make full use of the data, the representative feature matrices within the group are obtained by averaging in the ways of group two and group three respectively, and then steps (5)-(7) are repeated. At this time, for the feature matrices of unilateral tapping of the right hand generated under the group averaging method, the dimensions are 150×800 and 30×800 respectively. Then steps (8)-(9) are repeated to obtain the specific accuracy. 20×20×25 For example, in order to make full use of the data, the representative feature matrices within the group are obtained by averaging in the ways of group two and group three respectively, and then steps (5)-(7) are repeated. At this time, for the feature matrices of unilateral tapping of the right hand generated under the group averaging method, the dimensions are 150×800 and 30×800 respectively. Then steps (8)-(9) are repeated to obtain the specific accuracy.

[0031] (11)Table 1 shows the recognition accuracies based on oxyhemoglobin, deoxyhemoglobin, and mixed hemoglobin in three groups (that is, the representative functional brain networks in three groups obtained after averaging the functional brain networks of each subject according to different proportions). Based on Table 1, the following observation results are obtained. At the same grouping ratio, the classification accuracy of mixed blood oxygen protein is always the highest, which is predictable because mixed blood oxygen protein combines all the information of oxyhemoglobin and deoxyhemoglobin, resulting in an increase in the specificity of each sample; taking mixed blood oxygen protein as an example, the fewer the representative functional brain networks, the higher the recognition accuracy. To a certain extent, this indicates that although the averaging operation is simple, the discriminability of the representative brain networks obtained after averaging is gradually increasing and can better highlight the specificity of the subjects.

[0032] (12)For the richness of the experiment, we consider the recognition accuracy of functional brain networks under different sparsities of mixed blood oxygen protein. It can be seen that as the sparsity increases, the specificity of the functional brain networks generally shows a trend of first increasing and then decreasing; it is worth noting that when using the representative brain networks obtained by grouping and averaging with the largest proportion for identity recognition, the accuracy is the highest, and the classification accuracy of the brain networks obtained by sparsifying according to different proportions decreases significantly because this brain network has the best identity specificity, and after sparsification, some of the specificity is lost instead. In summary, it can be concluded that the group averaging operation is indeed effective for the problem of identity recognition based on near-infrared spectroscopy functional brain networks.

[0033] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and partial supplements made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. Identity recognition of multi-task functional near-infrared spectroscopy signals based on group average, characterized in that It includes the following steps: (1) Obtain near-infrared spectroscopy data: Design an experimental paradigm for signal acquisition, conduct multitask experiments, and use an infrared spectroscopy device to collect signals; (2) Preprocess the near-infrared spectroscopy data: First, read the signals collected by the near-infrared spectroscopy device; Second, perform preprocessing operations, including band-pass filtering of the optical signal, conversion using the modified Lambert-Beer law, and band-pass filtering of the blood oxygen protein concentration signal; After that, perform baseline processing to obtain clean signals; (3) Obtain the time series of oxyhemoglobin and deoxyhemoglobin: First, according to the experimental paradigm design of signal acquisition, extract the signals related to each experiment from the signals based on the execution time of each task. Then, segment the signals according to different tasks. After that, based on the segmented signals, extract them separately for deoxyhemoglobin and oxyhemoglobin. The extracted signal sequence is expressed as X = [x1, x2,..., x z ∈ R t×n×z , where x i ∈ R t×n represents the time series obtained from the i-th experiment, t represents the length of the time series, n represents the number of channels, and z represents the number of experiments; (4) Construct a brain functional network: Based on the segmented signals, regarding the channels as brain regions and the signals as time series, based on the signal x in each experiment i , use Pearson correlation to construct the functional brain network W ∈ R n×n , where each element w ij represents the weight element between channel i and channel j. For each experiment, calculate the corresponding brain functional network to obtain W ∈ R n×n×z ; (5) Obtain representative features within a group using the group averaging method: Taking all the signals related to one task of one subject as an example, considering the individual specificity of the representative brain functional network after averaging multiple brain functional networks, divide all the signals related to the experiment into several groups, define the number of brain functional networks within the group, and then perform the group averaging operation to obtain the representative features within the group, and straighten them into row vectors; (6) Obtain the comprehensive features of the hybrid modality: splice the corresponding representative features of oxyhemoglobin and deoxyhemoglobin to obtain the comprehensive features A of the hybrid modality, where A ∈ R (n×n)×t ; where n represents the number of channels and t represents the dimension of the comprehensive features; (7) Input the comprehensive features of the mixed modality of all subjects into a support vector machine classifier, and combine ten-fold cross-validation to obtain a relatively stable classification result.

Images