A method for recognizing umami-induced EEG signals in the taste perception process

Abstract

This invention belongs to the field of umami perception evaluation and provides a method for recognizing umami-induced EEG signals in the taste perception process. The method includes the following steps: acquiring taste EEG data induced by umami stimulation using an EEG signal acquisition system; establishing a umami EEG dataset using the acquired data; then cleaning the raw data using an independent component analysis algorithm, including filtering, removing electrooculograms, artifacts, and rereferences, and calculating the frequency (Hz) and power spectral density; vectorizing and regularizing the 0–13 Hz signals according to brain regions; subsequently, augmenting the dataset using minority class oversampling (MCB) and dividing it into training and test sets; constructing an ensemble model for signal classification; and finally, using the trained model for umami recognition. The ensemble model method proposed in this invention significantly improves the accuracy and stability of the umami EEG recognition model, effectively achieving umami recognition.

Description

Technical Field

[0001] This invention belongs to the field of umami perception evaluation, specifically relating to a method for recognizing umami-induced electroencephalogram (EEG) signals in the taste perception process. Background Technology

[0002] Currently, most taste sensory analyses rely on human senses and machine perception; however, these methods have numerous problems. In human sensory evaluations, taste judgments are easily influenced by the subject's subjectivity. Simultaneously, consumers tend to objectively and accurately express their preferences, leading to low reliability and poor reproducibility of experimental results. Using electroencephalography (EEG) data to assess taste has inherent advantages, combining umami perception with EEG data. When using EEG to identify umami in aqueous solutions, key parameters such as the diversity of umami stimulus samples, data processing methods, and pattern recognition strategies can affect the model's recognition results. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a method for recognizing umami-induced EEG signals in the taste perception process, thereby solving the problems in the prior art. The technical solution adopted by this invention is as follows:

[0004] A method for recognizing umami-induced EEG signals in the taste perception process, comprising the following steps:

[0005] Step S1: Configure taste stimulation samples for umami-induced EEG, and use an EEG signal acquisition system to collect EEG data induced in subjects under umami-stimulated solutions and non-umami-stimulated solutions respectively.

[0006] Step S2: The original data is cleaned sequentially using the independent component analysis algorithm, including filtering, removal of electrooculogram, artifacts and rereferences, and calculation of frequency and power spectral density. The 0-13Hz signal is vectorized and regularized according to brain regions.

[0007] Step S3: Subsequently, the dataset is augmented and divided into training and test sets using minority class oversampling. Feature filtering is then performed, and an ensemble model is constructed to classify the signals. Finally, the trained model is used for umami recognition.

[0008] Step S4: Train the classification sub-model, then fit the integrated judgment model, and finally complete the umami recognition.

[0009] Furthermore, in step S1, the umami-stimulating solution includes an aqueous solution of monosodium glutamate, an aqueous solution of inosine 5′-phosphate disodium, an aqueous solution of succinic acid, and a mixed aqueous solution of monosodium glutamate and purine nucleoside 5′-monophosphate; the non-umami-stimulating solution includes pure water.

[0010] Furthermore, in step S1, the device in the EEG Ag–AgCl 64 brain region is used to collect taste EEG data of no less than 40 subjects at a frequency of 1000 Hz. For each subject, EEG data under umami stimulation and non-umami stimulation are collected separately, with at least 3 parallel tests, and each parallel test lasting at least 30 seconds.

[0011] Furthermore, in step S2, the collected data are all vectorized into brain region signals, including: matching the relationship between the time points and frequencies of the anterior brain region F, left brain region LT, right brain region RT, central brain region C, and parieto-occipital brain region PO. The matching is performed from f1 to f25 to 0.976 to 12.695 Hz, resulting in 25 data points. Then, the data are merged in the order of anterior brain region F, left brain region LT, right brain region RT, central brain region C, and parieto-occipital brain region PO to obtain a 1*125 bit array.

[0012] Furthermore, in step S2, a percentage transformation of individualized variance PSD values ​​is performed: using the key frequency as the value, the percentage of each value among the 25 values ​​and the maximum value is calculated as the PSD percentage, which is used as the input data for the model in step S3.

[0013]

[0014] Among them, F i Represents the percentage of the current data point; PSD i Power spectral density representing the current data point; PSD1~PSD 25 The maximum power spectral density of 25 data points (PSD1~PSD) 25 ) max This represents the maximum value among them.

[0015] Furthermore, in step S2, the collected data is standardized, and the data is converted to absolute values ​​and then to percentages.

[0016] Furthermore, in step S3, the Kolmogorov-Smirnov test and the T-test are used to screen features, and features with a significance P≤0.05 are selected as the feature values ​​for modeling in step S4.

[0017] Furthermore, in step S4, a sub-model is trained based on at least five binary classification algorithms, and the best model is selected using accuracy and receiver operating characteristic (ROC) curves as criteria. The evaluation index is calculated using the following formula:

[0018]

[0019] Where ACC represents accuracy, PRE represents precision, MCC represents Matthews coefficient, AP represents average precision, TP represents true positives, TN represents true negatives, FP represents false positives, and FN represents false negatives.

[0020] Furthermore, in step S4, the ensemble model is fitted using the SVC algorithm.

[0021] The present invention has the following beneficial effects:

[0022] (1) This invention establishes a paradigm for umami EEG experiments and a method for acquiring umami EEG signals, including a process for acquiring EEG signals under umami stimulation and a method for preprocessing EEG signals.

[0023] (2) This invention proposes a novel ensemble model that enables umami-based EEG signal recognition. This model identifies characteristic brain regions and eliminates the need for highly accurate classification.

[0024] (3) The umami EEG signal recognition strategy proposed in this invention can be used in taste-related brain-computer interface technology and has important value in sensory evaluation, auxiliary judgment of taste disorder-related diseases and virtual taste application. Attached Figure Description

[0025] Figure 1 This is a flowchart of the umami signal recognition method based on electroencephalography in an embodiment of the present invention;

[0026] Figure 2 This is a schematic diagram showing the details of the sample stimulus recording;

[0027] Figure 3 This is a graph showing the parameters and evaluation results of the sub-model;

[0028] Figure 4 This is a schematic diagram showing the distribution of PSD percentage before and after EEG readings;

[0029] Figure 5 This is a schematic diagram illustrating the effect of data augmentation sampling methods;

[0030] Figure 6 This is a diagram showing the model's integration mode and evaluation results. Detailed Implementation

[0031] The following will be based on embodiments of the present invention. Figures 1-6 The technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Unless otherwise specified, the technical means used in the embodiments are conventional means well known to those skilled in the art.

[0032] A method for recognizing umami-induced EEG signals in the umami perception process includes the following specific steps:

[0033] S1, configured with taste stimulation samples for umami-induced EEG, and using an EEG signal acquisition system to collect EEG data induced by subjects under umami / non-umami solution stimulation respectively.

[0034] S2, the original data is cleaned sequentially using independent component analysis algorithms, including filtering, removal of electrooculograms, artifacts and rereferences, and calculation of frequency (Hz) and power spectral density. The signals from 0 to 13 Hz are then vectorized and regularized according to brain regions.

[0035] S3. Subsequently, the dataset was augmented using minority class oversampling and divided into training and test sets. Feature filtering was performed, and an ensemble model was constructed using the idea of ​​an ensemble model for signal classification. Finally, the trained model was used for taste recognition.

[0036] S4. Train the classification sub-model, then fit the integrated judgment model, and finally complete the umami taste recognition.

[0037] In step S1, 40 participants were recruited. These participants had all undergone rigorous EEG and sensory training, were familiar with umami, were in good physical and mental health, and could accurately distinguish umami. The experiment was conducted in a safe, quiet, and odor-free EEG testing room at a suitable temperature of 21±1℃. To avoid the influence of negative emotions or feelings of fullness during the testing process, EEG measurements were performed in the morning (10:00-11:00 AM) and afternoon (3:00-4:00 PM).

[0038] In step S1, the EEG signals are acquired using a device with adhesive-free electrode caps and EEG Ag-AgCl 64 brain regions, such as a 64-channel non-invasive EEG device from Brain Products GmbH, Munich, Germany. The Fcz channel is used as a reference recording. The resistance is 5kΩ, and the sampling frequency is 1000Hz.

[0039] During the preparation phase, subjects wore caps and prepared to rest, ensuring a good fit between the cap and scalp. When the instructor gave the "start" command, the subject placed the sample in their mouth (holding it for 10 seconds to taste), and the instructor began recording EEG data. During the experiment, subjects closed their eyes and kept their heads in one position to minimize motion artifacts. Afterwards, subjects were asked to empty their mouths and rinse with ultrapure water to restore the electrical signal to a nearby baseline. See details below. Figure 2 .

[0040] In step S2, the brain signals are processed using the Neuracle system software. Through data processing, the waveform of the raw EEG signal is cleaned using the Independent Component Analysis (ICA) algorithm with EEGLAB software, including filtering, removal of electrooculogram (EOG), artifacts, and re-referencing. Frequency (Hz) and power spectral density (PSD) are also calculated. This invention selects signals in the 0-13Hz range related to taste perception; the specific matching relationships are shown in Table 1.

[0041] Table 1 Point-Time Relationship Table

[0042]

[0043] Furthermore, in step S2, a percentage transformation of individualized variance PSD values ​​is performed: using the key frequency as the value, the percentage of each value among the 25 values ​​and the maximum value is calculated as the PSD percentage, which is used as the input data for the model in step S3.

[0044]

[0045] Among them, F i Represents the percentage of the current data point; PSD i Power spectral density representing the current data point; PSD1~PSD 25 The maximum power spectral density of 25 data points (PSD1~PSD) 25 ) max This represents the maximum value among them.

[0046] Furthermore, the collected data is standardized, and then converted to absolute values ​​and percentages.

[0047] Furthermore, the Kolmogorov-Smirnov test and the T-test were used for feature selection, and features with a significance P≤0.05 were selected as the feature values ​​for modeling in step S4.

[0048] Furthermore, in step S4, a sub-model is trained based on at least five binary classification algorithms, and the best model is selected using accuracy and receiver operating characteristic (ROC) curves as criteria. The evaluation index is calculated using the following formula:

[0049]

[0050] Where ACC represents accuracy, PRE represents precision, MCC represents Matthews coefficient, AP represents average precision, TP represents true positives, TN represents true negatives, FP represents false positives, and FN represents false negatives.

[0051] Furthermore, the ensemble model is fitted using the SVC algorithm.

[0052] In step S3, considering that input data with over two hundred samples but 125 features is prone to overfitting, and that low-quality features are detrimental to model training, this invention uses the Kolmogorov-Smirnov test and the T-test for feature selection. Under the significance selection threshold P≤0.05, out of a total of 125 features, 52 and 19 features remain through the KS-Test and T-Test, respectively. Taking the intersection of the results of the two selection conditions yields 14 common features. These features (f2-F, f18-F, f22-F, f23-F, f2-LT, f11-LT, f2-RT, f18-RT, f2-C, f17-C, f23-C, f2-PO, f18-PO, f20-PO) will be used for model training.

[0053] In step S3, data augmentation sampling takes into account the difference in the number of positive and negative data samples, and oversampling of a minority of samples is necessary. Figure 5 The results of benchmark modeling with the original data and after data augmentation using support vector machine / minority class oversampling techniques are presented separately. The results show that the data augmentation method is minority class oversampling with a K-Neighbour of 4, and the results of several evaluation metrics are all at a high level.

[0054] In step S4, the sub-model is constructed using the algorithm shown in Table 2. Figure 3 The evaluation of the top four sub-models with the highest scores is displayed.

[0055] Table 2. Parameters of the Binary Classification Algorithm

[0056]

[0057]

[0058]

[0059] Step S4: Model integration stage Figure 6 The accuracy of the model was recorded when the voting decision threshold was 1-4 votes. It can be seen that the optimal case is N=2 (N=1 means that if one of the four sub-models considers it umami, then it is umami; this works well when the number of positive samples in the test set is dominant, but therefore has poor generalization performance). The optimal voting method (N=2) achieved accuracies of 0.9 in Validation and 0.75 in Test Set. Although the voting method achieved good results, the ensemble model based on the support vector machine algorithm to fit the confidence probabilities output by the four sub-models achieved an accuracy of 0.78 and an MCC of 0.39, demonstrating even better performance.

[0060] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Any modifications, alterations, substitutions, or variations made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for recognizing umami-induced EEG signals in the taste perception process, characterized in that, Includes the following steps: Step S1: Configure taste stimulation samples for umami-induced EEG, and use an EEG signal acquisition system to collect EEG data induced in subjects under umami-stimulated solutions and non-umami-stimulated solutions respectively. Step S2: The original data is cleaned sequentially using the independent component analysis algorithm, including filtering, removal of electrooculogram, artifacts and rereferences, and calculation of frequency and power spectral density. The 0~13Hz signal is vectorized and regularized according to brain regions. Step S3: Subsequently, the dataset is augmented and divided into training and test sets using minority class oversampling. Feature filtering is then performed, and an ensemble model is constructed to classify the signals. Finally, the trained model is used for umami recognition. Step S4: Train the classification sub-model, then fit the integrated judgment model, and finally complete the umami recognition; In step S2, the collected data are all vectorized into brain region signals, including: matching the relationship between the time points and frequencies of the anterior brain region F, left brain region LT, right brain region RT, central brain region C, and parieto-occipital brain region PO. The matching is performed from f1 to f25 to 0.976 to 12.695 Hz, resulting in 25 data points. Then, the data are merged in the order of anterior brain region F, left brain region LT, right brain region RT, central brain region C, and parieto-occipital brain region PO to obtain a 1*125 bit array. In step S2, the individualized variance PSD values ​​are converted to percentages: using the key frequency as the value, the percentage of each value among the 25 values ​​and the maximum value is calculated as the PSD percentage, which is used as the input data for the model in step S3. ； Among them, F i Represents the percentage of the current data point; PSD i The power spectral density represents the current data point; (PSD1~PSD) 25 ) max The maximum value of the power spectral density represents the power of the 25 data points. In step S4, a sub-model is trained based on at least five binary classification algorithms. Accuracy and receiver operating characteristic (ROC) curves are used as criteria for optimal selection. The evaluation index is calculated using the following formula: ； Where ACC represents accuracy, PRE represents precision, MCC represents Matthews coefficient, AP represents average precision, TP represents true positives, TN represents true negatives, FP represents false positives, and FN represents false negatives.

2. The method for recognizing umami-induced EEG signals in the taste perception process as described in claim 1, characterized in that, In step S1, the umami-stimulating solution includes an aqueous solution of monosodium glutamate, an aqueous solution of inosine 5′-phosphate disodium, an aqueous solution of succinic acid, and a mixed aqueous solution of monosodium glutamate and purine nucleoside 5′-monophosphate; the non-umami-stimulating solution includes pure water.

3. The method for recognizing umami-induced EEG signals in the taste perception process as described in claim 1, characterized in that, In step S1, the EEG device for the Ag–AgCl 64 brain region was used to collect taste EEG data from no less than 40 subjects at a frequency of 1000 Hz. For each subject, EEG data were collected under umami stimulation and non-umami stimulation, with at least 3 parallel tests and each parallel test lasting at least 30 seconds.

4. The method for recognizing umami-induced EEG signals in the taste perception process as described in claim 1, characterized in that, In step S2, the collected data is standardized, and then the data is converted to absolute values ​​and then to percentages.

5. The method for recognizing umami-induced EEG signals in the taste perception process as described in claim 1, characterized in that, In step S3, the Kolmogorov-Smirnov test and the T-test are used to screen features, and features with a significance P ≤ 0.05 are selected as the feature values ​​for modeling in step S4.

6. The method for recognizing umami-induced EEG signals in the taste perception process as described in claim 1, characterized in that, In step S4, the ensemble model is fitted using the SVC algorithm.

Images