A method for objectively quantifying taste perception using electroencephalography (EEG)

Abstract

This invention belongs to the interdisciplinary field of neuroscience, sensory science and technology and food science. It provides a method for objectively quantifying taste perception using electroencephalography. The invention aims to separate and decode the neural dynamics of the brain in processing taste stimuli of different intensities, and to objectively evaluate and classify them based on individual differences in innate taste sensitivity.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of neuroscience, sensory science and technology and food science, and specifically relates to a method for objectively quantifying taste perception using electroencephalography (EEG). Background Technology

[0002] Taste perception is a complex process that depends not only on the physicochemical properties of the stimulus (such as its type and concentration) but also on an individual's internal state (such as genetic predisposition, physiological sensitivity, and past experience). Among the five basic tastes, umami, triggered by substances such as glutamate, has become a hot topic in sensory neuroscience research due to its unique neural coding mechanism.

[0003] Traditionally, it has been believed that the intensity of taste stimuli enhances activation and functional network connectivity in the cerebral cortex (especially the insula and orbitofrontal cortex) in a dose-dependent manner. However, this "universal intensity model" is increasingly challenged by studies of individual differences. Growing evidence suggests that an individual's innate taste sensitivity is a key factor in regulating neural processing strategies.

[0004] When using EEG technology to study umami perception, a core paradox emerged: on the one hand, the brain can finely distinguish different umami substances with similar subjective sensations; on the other hand, when encoding the concentration (i.e., intensity) of the most typical umami substance—monosodium glutamate (MSG)—the brain exhibits a sluggish response. This suggests that a universal intensity-response model is incomplete, and that "individual sensitivity," largely overlooked in previous studies, may be the key variable to unraveling this paradox.

[0005] Therefore, there is an urgent need in this field for an assessment method that can objectively separate the universal neural coding caused by taste intensity from the personalized neural strategies driven by individual sensitivity, in order to more comprehensively understand the individualized nature of taste perception. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method for objectively quantifying taste perception using electroencephalography (EEG), thereby resolving the issues in the prior art. The technical solution adopted by this invention is as follows:

[0007] A method for objectively quantifying taste perception using electroencephalography (EEG) includes the following steps:

[0008] Step 1, Individual sensitivity grouping and taste stimulation: First, through behavioral tests, the subjects are objectively divided into different sensitivity groups, and then at least two different concentrations of umami stimulation are applied to the subjects;

[0009] Step 2, Acquisition and processing of EEG signals: While applying taste stimulation, multi-channel EEG signals of the subject are acquired simultaneously and preprocessed according to standard procedures.

[0010] Step 3: Separate and extract two types of key neurobiological markers with different functions: Analyze the processed EEG signals, including separating the mixed neural signals into two types of biomarkers with different properties, including general intensity coding markers and personalized sensitivity strategy markers;

[0011] Step 4, Model Decoding and Evaluation of Biomarkers Based on Personalized Strategies: After the biomarkers are isolated, a multivariate pattern analysis model is constructed and trained based on the extracted personalized sensitivity strategy markers. This model is used to automatically classify and determine the sensitivity group to which the subject belongs based on the subject's EEG activity in a specific time window.

[0012] Furthermore, in step 1, the behavioral tests include: using the 3-AFC ladder method to determine the subject's detection threshold for umami substances; different sensitivity groups include: high, medium, and low sensitivity groups.

[0013] Furthermore, in step 3,

[0014] For the universal intensity-encoding marker, it showed a systematic response to umami concentration in all sensitivity groups of subjects, including:

[0015] As the umami concentration increases, the amplitude of the P200 event-related potential is systematically suppressed;

[0016] As umami concentration increases, the Beta band functional connectivity between the right orbitofrontal cortex and the right dorsolateral prefrontal cortex systematically weakens.

[0017] For personalized sensitivity strategy markers, differences exist between different sensitivity groups, reflecting different neural processing strategies, including:

[0018] The "neural efficiency" markers in the high-sensitivity group were manifested as earlier sensory processing and stronger neural gain control;

[0019] The "cognitive compensation" marker for the low-sensitivity group is manifested in mobilizing more attentional resources in the later stages of cognitive processing.

[0020] Furthermore, in step 4, a multivariate pattern analysis model is constructed and trained, including:

[0021] Step 4.1, Feature Set Construction and Input: The model's input is the personalized sensitivity strategy markers from Step 3, constructing a multimodal feature set, including:

[0022] Temporal characteristics: Extract the average amplitude of key event-related potential components within a preset time window; including: extracting the amplitude values ​​of N150 components reflecting early sensory processing, P400 components reflecting cognitive appraisal, and late positive potentials (LPP) reflecting later attentional resource allocation;

[0023] Frequency domain characteristics: The average power spectral density in different neural oscillation frequency bands was calculated using the Welch method to quantify the neural oscillation activity state in different brain regions;

[0024] Functional connectivity characteristics: The phase-locked value (PLV) reflecting the synchronization of information between brain regions is calculated using an enhanced, frequency-specific PLV calculation method: First, the signal is band-pass filtered, and then the phase synchronization of a specific electrode pair in the target frequency band is calculated within a sliding time window.

[0025] Multimodal features are integrated to construct a high-dimensional feature vector for each stimulus trial for each subject, which serves as the final input to the model.

[0026] Step 4.2, Decoding Model and Strategy: A time-resolved decoding strategy is adopted, which uses a sliding estimator to classify point by point on the entire time axis after the taste stimulus, so as to dynamically track the time window in which the sensitivity classification information is most significant;

[0027] A classifier based on Riemannian geometry is employed, which includes:

[0028] Covariance matrix estimation: Transform the multi-channel EEG data within each time window into a covariance matrix;

[0029] Minimum Mean Distance Classification: Classification is performed in the Riemannian manifold space composed of covariance matrices: First, the Riemann geometric mean covariance matrix of each sensitivity group in the training data is calculated. For a new test sample, the Riemann distance between its covariance matrix and the mean matrices of the three groups is calculated, and it is classified into the group with the closest distance.

[0030] Step 4.3, Model Training, Validation and Evaluation: A hierarchical random reshuffling cross-validation method is adopted, which randomly divides the training set and the test set in each iteration, while maintaining the sample proportion of each sensitivity group in the original dataset.

[0031] The present invention has the following beneficial effects:

[0032] This invention achieves precise separation and deconstruction of neural signals: For the first time, this invention successfully separates the two processes of taste intensity encoding and individual sensitivity strategies at the neural level, providing a novel neurobiological framework for understanding individualized differences in perception.

[0033] This invention provides a high-precision objective assessment tool based on a separation mechanism: by effectively separating and identifying specific neurobiomarkers with well-defined mechanisms, this invention can objectively and quantitatively assess an individual's taste sensitivity with significantly higher accuracy than traditional subjective methods.

[0034] This invention enables the decoding of individual sensitivities: This invention demonstrates that machine learning models can be used to successfully decode an individual's sensitivity type from early electroencephalographic activity, providing strong technical support for personalized nutrition, product development, and clinical research. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the overall process of the method of the present invention.

[0036] Figure 2 The psychophysical and hedonic responses of different umami sensitivity groups to MSG stimulation are shown; where: (a) is the average perceived intensity score of MSG stimulation at three concentrations; (b) is the average liking score of the same stimulus; and (c) is the linear regression fit between individual perceived intensity scores and MSG concentration.

[0037] Figure 3 This study demonstrates how the amplitude and latency of different ERP components (such as N150, P200, and LPP) are modulated by stimulus concentration and sensitivity group at the sensor level; where: (a) is the peak amplitude of sensor-level N150 in the frontal electrode cluster (F1 / F2 / Fz); (b) is the latency of sensor-level N150 in the frontal electrode cluster (F1 / F2 / Fz); (c) is the peak amplitude of sensor-level P200 in the frontal electrode cluster (F1 / F2 / Fz); (d) is the latency of sensor-level P200 in the frontal electrode cluster (F1 / F2 / Fz); (e) is the peak amplitude of sensor-level late positive potential (LPP) in the occipital electrode cluster (O1 / O2 / Oz); and (f) is the latency of sensor-level late positive potential (LPP) in the occipital electrode cluster (O1 / O2 / Oz).

[0038] Figure 4 The main effects of concentration and UTS are represented by time-frequency characteristics; (a), (b), (c), (d), (e), and (f) are the total average time-frequency energy (TFR) and scalp topography of the high-sensitivity and low-sensitivity groups at three MSG concentrations; (g), (h), (i), (j), and (l) are the time progression of the main power effects of stimulation concentration on the Theta and Beta bands in the frontal, central, parietal, and occipital regions; and (m) is the time progression of the main power effect of umami sensitivity on the Alpha band in the right temporal region.

[0039] Figure 5Source-level neural markers of the main effect of umami sensitivity groups show significant differences in multiple brain regions among different umami sensitivity groups; where: (a) represents the Theta band power in the precentral gyrus; (b), (c), (g), and (i) represent...

[0040] The amplitudes of the N150, N200, P350 and P400 components; (d), (e), (f), and (h) are the latency periods of the N200, P200 and P400 components;

[0041] Figure 6 The MSG concentration significantly modulates the Beta band functional connectivity between the OFC and DLPFC on the right; where: (a) is an overview of the region of interest (ROIs) and compares the average connectivity matrix at the lowest (6.25 mM) and highest (50 mM) concentrations; (b), (c), (d), and (e) show the connectivity matrix comparison between the high-sensitivity group and the low-sensitivity group at different concentrations; the boxes in all matrices indicate that connectivity weakens as the concentration increases.

[0042] Figure 7 Decoding individual umami sensitivity using time-resolved multivariate pattern analysis (MVPA); (a) decoding accuracy of the XGBoost classifier over time; (b) optimization results of the time window and number of features for the sensitive group classifier; (c) temporal generalization results of the sensitive group classifier, showing that it is a dynamic (diagonal) rather than a stable neural code. Detailed Implementation

[0043] The following will be based on embodiments of the present invention. Figures 1-7 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.

[0044] The core of this invention is the proposal of a novel scientific analytical framework that, for the first time, "separates" two distinct physiological processes at the neural level: a "universal encoding" driven by stimulus intensity and a "personalized strategy" determined by individual talent. Its contribution lies in using this newly discovered framework to decode and assess intrinsic human traits (taste sensitivity). The core elements are: signal separation and marker extraction: conceptually and operationally separating EEG signals into "universal intensity-encoding markers" and "personalized sensitivity strategy markers." This invention clarifies biomarkers with neuroscientific significance and functional orientation. Examples include: P200 amplitude, OFC-DLPFC functional connectivity (for encoding intensity); N150 latency, and alpha power (for characterizing sensitivity strategies). The markers of this invention can not only be used for classification but also explain "why," revealing the underlying neurobiological principles.

[0045] like Figure 1 This invention proposes a method for objectively quantifying taste perception using electroencephalography (EEG), comprising the following steps:

[0046] Step 1: Individual Sensitivity Grouping and Taste Stimulation. First, behavioral tests, including the 3-AFC ladder method, were used to determine the subjects' detection thresholds for umami substances, objectively dividing them into different sensitivity groups, such as high, medium, and low sensitivity groups. Then, at least two different concentrations of umami stimulation, such as different concentrations of MSG solution, were applied to the subjects.

[0047] Step 2: Acquire and process EEG signals. Simultaneously, multi-channel EEG signals from the subject are acquired while gustatory stimulation is applied, and standard preprocessing is performed.

[0048] Step 3: Separate and extract two types of key neuromarkers with distinct functions. The core of analyzing the processed EEG signals lies in separating the mixed neural signals into two different types of biomarkers:

[0049] 1. Universal intensity-encoding markers: These markers exhibit a systematic response to umami concentration in all sensitivity groups. Specifically, they may include:

[0050] As the umami concentration increases, the P200 event-related potential (ERP) amplitude is systematically suppressed.

[0051] As umami concentration increases, the functional connectivity strength in the Beta band (13-30Hz) between the right orbitofrontal cortex (OFC) and the right dorsolateral prefrontal cortex (DLPFC) systematically weakens.

[0052] 2. Personalized sensitivity strategy markers: These markers differ significantly among different sensitivity groups, reflecting different neural processing strategies. Specifically, they may include:

[0053] The markers of "neural efficiency" in the high-sensitivity group were: earlier sensory processing (e.g., shorter N150 latency) and stronger neural gain control (e.g., enhanced alpha band power in the right temporal region).

[0054] The marker of "cognitive compensation" in the low-sensitivity group is manifested in mobilizing more attentional resources in the later stages of cognitive processing (such as larger late positive potentials LPP or P400 amplitude).

[0055] Step 4: Model Decoding and Evaluation Based on Personalized Strategy Markers. After successfully separating the two types of markers, a Multivariate Pattern Analysis (MVPA) model is constructed and trained specifically based on the extracted personalized sensitivity strategy markers. This model is used to automatically classify and determine the sensitivity group to which the subject belongs based on their EEG activity within a specific time window (e.g., 119-320 milliseconds after stimulation).

[0056] Step four of this invention aims to accomplish two core tasks: 1) verifying the effectiveness of the "personalized sensitivity strategy markers" separated and extracted in step three; and 2) based on these markers, constructing a decoding model capable of objectively and automatically classifying individual taste sensitivity. This step employs a comprehensive technical solution integrating Multivariate Pattern Analysis (MVPA), temporally resolved decoding, multimodal feature fusion, and advanced classification algorithms, specifically including:

[0057] Step 4.1, Feature set construction and input:

[0058] The input to the decoding model consists of "personalized sensitivity strategy markers" extracted in step three, which reflect the differences in neural processing strategies among different sensitivity groups (high, medium, and low). These markers specifically refer to neural features that exhibit significant differences (e.g., p < 0.05) between different sensitivity groups in the statistical analysis (such as ANOVA or t-test) in step three. To ensure the comprehensiveness of the decoding information, this invention constructs a multimodal feature set, specifically including:

[0059] Temporal features (ERPFeatures): Extract the average amplitude of key event-related potential (ERP) components within a preset time window. For example, extract the amplitude values ​​of the N150 component (130-220ms) reflecting early sensory processing, the P400 component (380-420ms) reflecting cognitive appraisal, and the late positive potential LPP (600-800ms) reflecting later attentional resource allocation.

[0060] Frequency domain features: Welch's method was used to calculate the average power spectral density (PSD) in different neural oscillation frequency bands (e.g., Theta: 5-8 Hz, Alpha: 8-12 Hz, Beta: 12-30 Hz) to quantify the neural oscillation activity state in different brain regions.

[0061] Functional Connectivity Features: This involves calculating the Phase Locking Value (PLV), which reflects the synchronization of information between brain regions. This invention employs an enhanced, frequency-specific PLV calculation method: first, the signal is bandpass filtered, and then the phase synchronization of a specific electrode pair in a target frequency band (such as the Alpha or Beta band) is calculated within a sliding time window. To improve computational efficiency and feature validity, this method also integrates an intelligent channel pair selection strategy based on electrode spatial distance, ensuring the inclusion of representative connections at the proximal, mid-range, and long-range.

[0062] The above multimodal features are integrated to construct a high-dimensional feature vector for each stimulus trial of each subject, which serves as the final input to the decoding model.

[0063] Step 4.2, Decoding Model and Strategy:

[0064] This invention employs a time-resolved decoding strategy, which uses a sliding estimator to classify the gustatory stimulus point by point along the entire time axis to dynamically track the time window in which the sensitivity classification information is most significant.

[0065] Preferred Implementation: Riemannian Geometry-Based Classifier. This is the preferred classification model of the present invention, particularly adept at processing spatial pattern information contained in EEG signals. This model is implemented through a two-step pipeline:

[0066] 1. Covariance Matrix Estimation: The multi-channel EEG data within each time window is transformed into a covariance matrix. This matrix effectively captures the spatiotemporal correlations among all electrode channels. To improve the stability of the matrix estimation, the Ledoit-Wolf (lwf) method is used for regularization shrinkage.

[0067] 2. Minimum Mean Distance Classification (MDM): This method performs classification within a Riemannian manifold space composed of covariance matrices. First, the Riemannian geometric mean covariance matrix for each sensitivity group (high, medium, low) in the training data is calculated. For a new test sample, the Riemannian distance between its covariance matrix and the mean matrices of the three groups is calculated, and it is classified into the group with the closest distance.

[0068] Other alternative embodiments: The decoding framework of this invention is also compatible with other advanced classification algorithms, such as Linear Discriminant Analysis (LDA) with automatic shrinkage, Support Vector Machine (SVC), or Logistic Regression. Before using these classifiers, the features usually need to be standardized (StandardScaler).

[0069] Step 4.3, Model Training, Validation, and Evaluation: To ensure the objectivity and reliability of model performance evaluation, this invention employs a Stratified Shuffle Split cross-validation method. This method randomly divides the training and test sets in each iteration while strictly maintaining the sample proportion of each sensitivity group in the original dataset, effectively avoiding evaluation bias that may be caused by imbalanced samples.

[0070] Analysis of participants' neural activity into taste sensitivity groups (high, medium, or low) showed that the XGBoost classifier exhibited the highest performance, achieving a peak accuracy of 59.9%. As shown in the figure, individual taste sensitivity can be significantly decoded early after stimulus presentation (particularly within a 50–100 ms time window) through patterns of neural activity in the brain, with decoding accuracy far exceeding random levels. This supports the view that taste sensitivity is a "trait" or "intrinsic" characteristic that can be identified and distinguished in early neural processing stages. Examination of the optimal model parameters revealed a significant time dependence, with the most discriminative information for classifying sensitivity groups occurring within a 119–320 ms time window. Figure 7 b). Furthermore, temporal generalization analysis revealed that the sensitive neural coding exhibits high dynamics and limited cross-temporal generalization ability, manifested as a strong diagonal pattern in the generalization matrix ( Figure 7 c). These results confirm the existence of a transient neural feature with an early to mid-latency latency that can effectively predict an individual's innate taste sensitivity. This result not only strongly confirms the effectiveness and specificity of the "personalized sensitivity strategy markers" extracted in step three, but also demonstrates that the entire method provided by this invention can achieve an objective and accurate assessment of an individual's taste sensitivity, and has significant application value.

[0071] A specific embodiment of the present invention is as follows:

[0072] Step 1: Determining Individual Sensitivity and Applying Stimuli. Forty healthy adult subjects were recruited. Their detection thresholds for MSG were determined, and they were divided into high-sensitivity (n=13), medium-sensitivity (n=14), and low-sensitivity (n=11) groups using the interquartile range. In the EEG experiment, subjects were randomly presented with three concentrations of MSG solution (6.25mM, 25mM, 50mM) as taste stimuli.

[0073] Step 2: Extract universal intensity markers and analyze the EEG data of all subjects. The results showed two universal intensity coding patterns independent of sensitivity groups:

[0074] P200 amplitude suppression: such as Figure 3 As shown in c, at the recording sites in the frontal, central, and parietal lobes, the amplitude of P200 decreased significantly with increasing MSG concentration (p<.001). This reflects that as the stimulus signal becomes clearer, the brain shifts from an initial detection state to a more complex analysis state, which is a manifestation of neural efficiency rather than weaker processing.

[0075] OFC-DLPFC functional connectivity weakens: such as Figure 6 As shown, as MSG concentration increased from 6.25 mM to 50 mM, the Beta band functional connectivity between the right orbitofrontal cortex (OFC) and the right dorsolateral prefrontal cortex (DLPFC) decreased significantly and linearly (p < .001). This reflects that the brain requires stronger cognitive control (i.e., stronger connectivity) when processing vague, weak stimuli, while it is "easier" to process clear, strong stimuli, requiring less network communication.

[0076] Step 3: Extract personalized sensitivity strategy markers and compare the EEG activity of different sensitivity groups to find significant differences in their neural strategies.

[0077] High sensitivity group (neural efficiency strategy): such as Figure 3 As shown in b, the N150 latency in the frontal lobe of this group was the fastest among the three groups. Figure 4 As shown in m, this group exhibited significantly higher alpha band power in the right temporal region. This indicates that they were able to classify stimuli earlier and effectively suppress irrelevant neural activity through alpha oscillations, achieving efficient processing.

[0078] Low-sensitivity group (cognitive compensation strategy): This group showed signs of greater "effort" in the later stages of processing. For example, their P400 amplitude was larger in source space analysis, reflecting a more strenuous cognitive reappraisal process. MVPA analysis also supported this, finding that the low-sensitivity group had the highest decoding accuracy (65.2%) for the three different concentrations, indicating that they generated more stable and clearer neural representations to distinguish intensity—neural traces left by their "strenuous" processing.

[0079] Step 4: Construct the model and decode the sensitivity. Personalized markers capable of distinguishing different groups (such as early ERP features, spectral features, functional connectivity features, etc.) are used as a feature set and input into an XGBoost classifier. The results are as follows: Figure 7 As shown in Figure ac, the model can successfully classify subjects into their respective sensitivity groups with an accuracy of 59.9% based on EEG activity within a time window of 119-320 milliseconds after stimulation. This demonstrates that the method proposed in this invention can objectively and accurately decode an individual's taste sensitivity traits from EEG signals.

[0080] 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 objective quantification of taste perception using electroencephalography, characterized in that, Includes the following steps: Step 1, Individual sensitivity grouping and taste stimulation: First, through behavioral tests, the subjects are objectively divided into different sensitivity groups, and then at least two different concentrations of umami stimulation are applied to the subjects; Step 2, Acquisition and processing of EEG signals: While applying taste stimulation, multi-channel EEG signals of the subject are acquired simultaneously and preprocessed according to standard procedures. Step 3: Separate and extract two types of key neurobiological markers with different functions: Analyze the processed EEG signals, including separating the mixed neural signals into two types of biomarkers with different properties, including general intensity coding markers and personalized sensitivity strategy markers; Step 4, Model Decoding and Evaluation of Biomarkers Based on Personalized Strategies: After separating the biomarkers, a multivariate pattern analysis model is constructed and trained based on the extracted general intensity coding markers and personalized sensitivity strategy markers. This model is used to automatically classify and determine the sensitivity group to which the subject belongs based on the subject's EEG activity in a specific time window. In step 1, the behavioral tests include: using the 3-AFC ladder method to determine the subject's detection threshold for umami substances; different sensitivity groups include: high, medium, and low sensitivity groups; In step 3, For the universal intensity-encoded markers, they showed a systematic response to umami concentration in all sensitivity groups, including: P200 event-related potentials and Beta band functional connectivity strength between the right orbitofrontal cortex and the right dorsolateral prefrontal cortex. The markers for personalized sensitivity strategies differed among different sensitivity groups, reflecting different neural processing strategies, including: Alpha band power in the right temporal region, late positive potential (LPP), and P400 amplitude.

2. The method for objective quantification of taste perception using electroencephalography according to claim 1, wherein, Step 4 involves constructing and training a multivariate pattern analysis model, including: Step 4.1, Feature Set Construction and Input: The model's input consists of the general intensity encoding markers and personalized sensitivity strategy markers from Step 3, constructing a multimodal feature set, including: Temporal characteristics: Extract the amplitude values ​​of N150 components reflecting early sensory processing, P400 components reflecting cognitive appraisal, and late positive potentials (LPP) reflecting later attentional resource allocation. Frequency domain characteristics: The average power spectral density in different neural oscillation frequency bands was calculated using the Welch method; Functional connectivity characteristics: The phase-locked value (PLV) reflecting the synchronization of information between brain regions is calculated using an enhanced, frequency-specific PLV calculation method: First, the signal is band-pass filtered, and then the phase synchronization of a specific electrode pair in the target frequency band is calculated within a sliding time window. Multimodal features are integrated to construct a high-dimensional feature vector for each stimulus trial for each subject, which serves as the final input to the model. Step 4.2, Decoding Model and Strategy: A time-resolved decoding strategy is adopted, which uses a sliding estimator to classify point by point along the entire time axis after the taste stimulus; A classifier based on Riemannian geometry is employed, which includes: Covariance matrix estimation: Transform the multi-channel EEG data within each time window into a covariance matrix; Minimum Mean Distance Classification: Classification is performed in the Riemannian manifold space composed of covariance matrices: First, the Riemann geometric mean covariance matrix of each sensitivity group in the training data is calculated. For a new test sample, the Riemann distance between its covariance matrix and the mean matrices of the three groups is calculated, and it is classified into the group with the closest distance. Step 4.3, Model Training, Validation and Evaluation: A hierarchical random reshuffling cross-validation method is adopted, which randomly divides the training set and the test set in each iteration, while maintaining the sample proportion of each sensitivity group in the original dataset.

Images