A cognitive inhibition evaluation method based on a hierarchical bayesian shift-diffusion model

Abstract

The present application relates to the technical field of data processing of cognitive psychology, in particular to a cognitive inhibition evaluation method based on a hierarchical Bayesian shift diffusion model, comprising: obtaining trial primary data of a subject under a changing interference paradigm; logically comparing the key selection with the real change state of the stimulus and generating correctness logical coding, converting the original behavior data vector into modeling data vector, and constructing a competition model space driven by cognitive hypothesis based on the shift diffusion model; using MCMC sampling algorithm to iteratively estimate the parameter posterior distribution of each competition model, and calculating the DIC value of each competition model; selecting the competition model with the optimal fitting degree by comparing the bias information standard DIC value of each competition model, and outputting the evaluation result of the cognitive inhibition site of the subject. According to the model with the optimal fitting degree with the original behavior data, the specific site of the cognitive inhibition of the subject can be identified, and the precise positioning evaluation of the cognitive inhibition site can be realized.

Description

Technical Field

[0001] This invention relates to the field of data processing technology in cognitive psychology, specifically to a cognitive inhibition assessment method based on a hierarchical Bayesian drift-diffusion model. Background Technology

[0002] Visual working memory is a core cognitive ability of an individual to temporarily store and process visual information (such as color, location, and shape) online in the brain. Because the human brain has a highly limited capacity for processing information, the cognitive system must establish efficient inhibitory mechanisms to filter out irrelevant and distracting information in order to ensure the successful execution of the current task. In cognitive psychology research, the classic method for exploring the inhibitory mechanisms of visual working memory is the change-distraction paradigm. This experimental paradigm typically includes three phases: Memory phase: Participants are presented with a set of objects containing multiple feature dimensions, such as geometric shapes with different combinations of colors and shapes. Retention phase: The memory items disappear, and the screen displays a brief blank interval (usually 200 to 1000 milliseconds), during which time the information is stored in visual working memory. Testing phase: A set of objects is presented again, and participants need to judge whether the test items have changed in terms of "target features" (such as color) compared to the memory items by pressing buttons. To study inhibitory efficacy, the state of "interference features" (such as shape) is manipulated in the experiment. If the cognitive system can successfully inhibit the interference of shape information, then whether the shape changes or not should not affect the subject's reaction speed and accuracy to the main task (color judgment).

[0003] Current implementation schemes primarily rely on the comparison of mean values ​​for macroscopic behavioral indicators. Specific implementation steps include: Experimental variable manipulation: Researchers manipulate two levels of the interfering feature according to experimental needs, namely, "conditions where the interfering feature is changed" and "conditions where the interfering feature remains unchanged." Data collection: Psychological experimental software is used to record the discrete response results of subjects in each trial, mainly including the accuracy of response selection and reaction time. Mean-based statistical processing: After data collection, the mean of data for each subject under different conditions is calculated. For data stability, the standard deviation filtering method is usually used to remove "outliers" with excessively long or short reaction times. Statistical inference logic: Paired-samples t-tests or repeated measures ANOVA are used to compare the average reaction times under the two conditions. Interfering effect determination: If the average reaction time under "conditions changed" is significantly longer than under "conditions unchanged," it is inferred that interfering information has infiltrated working memory, incurring processing costs. Suppression success determination: If there is no significant difference in the average reaction times of the two conditions, it is roughly inferred that the interfering feature was successfully suppressed, without incurring additional processing costs for the target feature. The proposed approach treats the cognitive execution process as a holistic, unanalyzable "black box," relying solely on the mean reaction time of the final output to retrospectively infer the strength of the inhibitory function. This approach struggles to reveal the internal dynamic processing. Furthermore, this method has several shortcomings, including: the cognitive "black box" nature of the approach means that existing solutions cannot identify the specific location where inhibition occurs. When dealing with visual interference, it is impossible to determine whether inhibition occurs in the "early gating interception" stage before information enters memory or in the "late decision conflict elimination" stage after information has entered memory. Summary of the Invention

[0004] To address the technical problems existing in the prior art, this invention provides a cognitive inhibition assessment method based on a hierarchical Bayesian drift-diffusion model. By recoding and mapping the original behavioral data with accuracy, the macroscopic response indicators are decomposed into independent computational parameters with clear psychological significance, such as non-decision time, drift rate, and decision boundary. Through competitive fitting of these independent computational parameters, the specific sites where cognitive inhibition occurs in the subjects can be identified, and the assessment results of the subjects' cognitive inhibition sites can be output.

[0005] The first objective of this invention is to provide a cognitive inhibition assessment method based on a hierarchical Bayesian drift-diffusion model.

[0006] A second objective of this invention is to provide a computer device.

[0007] The first objective of this invention can be achieved by adopting the following technical solution:

[0008] A cognitive inhibition assessment method based on a hierarchical Bayesian drift-diffusion model includes the following steps:

[0009] S1. Obtain the trial-second raw data of the subject under the changing interference paradigm, preprocess the trial-second raw data to obtain effective trial-second raw data, and output the raw behavioral data vector; the trial-second raw data includes reaction time, conditional encoding, actual change state of stimulus and subject key selection.

[0010] S2. Accuracy recoding mapping based on cognitive logic: logically compare the key selection in the original behavioral data vector with the actual change state of the stimulus and generate a binary correctness logic code. Based on the correctness logic code, the original behavioral data vector is converted into a modeling data vector. A modeling data table is established based on the modeling data vector.

[0011] S3. Construct a cognitive hypothesis-driven competitive model space based on the drift-diffusion model. The parameters of the drift-diffusion model include non-decision time t, drift rate v, and decision boundary a. Four different competitive models are constructed based on the zero-effect hypothesis, early input gating hypothesis, late decision conflict hypothesis, and strategy caution hypothesis, respectively.

[0012] S4. Perform hierarchical Bayesian sampling and parameter posterior estimation, construct a hierarchical Bayesian drift-diffusion model, and use the MCMC sampling algorithm based on the data in the modeling data table to iteratively estimate the parameter posterior distribution of each competing model, and calculate the standard DIC value of the bias information of each competing model.

[0013] S5. By comparing the deviation information standard DIC values ​​of each competing model, select the competing model with the best data fit and output the assessment results of the cognitive inhibition sites of the subjects.

[0014] Specifically, step S2 includes: logically comparing the subject's key selection in each original behavioral data vector with the actual stimulus state to generate a binary correctness logic code. When the subject's key selection matches the actual stimulus state, a correctness logic code of 1 is generated, indicating a correct response; when the subject's key selection does not match the actual stimulus state, a correctness logic code of 0 is generated, indicating an incorrect response. The modeling data vector includes reaction time, conditional encoding, and correctness logic code.

[0015] Specifically, step S3 includes: using the HDDM and PyMC tools in the Python environment, based on the zero-effect hypothesis, early input gating hypothesis, late decision conflict hypothesis, and policy caution hypothesis, respectively, applying conditional dependency constraints on different parameters to construct four competing models: zero-effect model M0, early input gating model Mt, late decision conflict model Mv, and policy caution model Ma.

[0016] The zero-effect model M0 is used to characterize the absence of a stable inhibition effect, and it is set that the non-decision time t, the drift rate v, and the decision boundary a do not change with the conditional coding.

[0017] The early input gating model Mt is used to characterize that inhibition occurs in the early encoding stage, and it is set to only allow the non-decision time t to change with the conditional encoding.

[0018] The late-stage decision conflict model Mv is used to characterize the suppression that occurs during the decision processing stage, and it is set to allow only the drift rate v to change with the conditional encoding;

[0019] The strategy caution model Ma is used to characterize the inhibition that stems from the subject's subjective adjustment of decision criteria or response caution, and it is set to allow only the decision boundary a to change with the conditional encoding.

[0020] Specifically, step S4 includes:

[0021] S41. Construct a hierarchical Bayesian drift-diffusion model using open-source framework tools, establish a hierarchical structure of data layer, individual layer and group layer, set a prior distribution for hyperparameters of group layer, and obtain a complete hierarchical Bayesian drift-diffusion model by connecting data layer, individual layer and group layer.

[0022] S42. Based on the data in the modeling data table, sample each competing model using the MCMC sampling algorithm, perform posterior parameter estimation for each competing model, and calculate the standard DIC value of the deviation information for each competing model based on the posterior parameter estimation results.

[0023] Specifically, step S42 includes:

[0024] Based on the modeling data table obtained in step S2 and the four different competing models constructed in step S3, the MCMC sampling algorithm is used to sample only one competing model at a time to calculate the posterior parameter estimation results. The posterior parameter estimation results include: posterior mean bias, posterior mean, and bias at the posterior mean.

[0025] The number of effective parameters (PD) and the standard deviation information (DIC) value for each competing model are calculated based on the posterior estimation results of the parameters. The standard deviation information (DIC) value is used as an evaluation index for model fit.

[0026] Specifically, the calculation of the number of effective parameters (PD) and the standard deviation information (DIC) value for each competing model based on the posterior estimation results of the parameters includes:

[0027] The number of effective parameters PD is obtained by the difference between the posterior mean deviation and the deviation at the posterior mean. The standard deviation information (DIC) value is obtained by the sum of the posterior mean deviation and the number of effective parameters PD.

[0028] Specifically, step S5 includes:

[0029] By comparing the standard deviation information (DIC) values ​​of each competing model, the competing model with the smallest standard deviation information (DIC) value is considered to have the best fit to the behavioral data.

[0030] Based on the model with the best fit, the specific processing sites where cognitive inhibition occurs in the subjects are obtained, and the evaluation results of the subjects' cognitive inhibition sites are output.

[0031] The second objective of this invention can be achieved by adopting the following technical solution:

[0032] A computer device includes a processor and a memory for storing a processor-executable program, wherein when the processor executes the program stored in the memory, it implements the above-described cognitive inhibition assessment method based on a hierarchical Bayesian drift-diffusion model.

[0033] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0034] This invention provides a cognitive inhibition assessment method based on a hierarchical Bayesian drift-diffusion model. Different models are constructed for cognitive inhibition function, including four competing models based on the zero-effect hypothesis, early input gating hypothesis, late decision conflict hypothesis, and policy caution hypothesis. A hierarchical Bayesian drift-diffusion model is then built. Based on data from a modeling data table, the MCMC sampling algorithm is used to iteratively estimate the posterior distribution of parameters for each competing model. The model with the best fit is used as the final evaluation result, directly outputting the specific processing site where inhibition occurs, thus achieving precise localization of the inhibition site. This invention solves the technical problem of existing technologies that rely solely on average reaction time comparisons, failing to deconstruct the internal dynamic process and making it difficult to distinguish the specific cognitive stages where inhibition control occurs. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0036] Figure 1 This is an overall flowchart of the cognitive inhibition assessment method in this embodiment of the invention;

[0037] Figure 2 This is a variation interference experimental paradigm and stimulus presentation time sequence diagram in an embodiment of the present invention;

[0038] Figure 3 This is a mapping diagram of drift diffusion model parameters and cognitive inhibition sites in an embodiment of the present invention;

[0039] Figure 4 This is a schematic diagram of the optimal model recognition result in an embodiment of the present invention;

[0040] Figure 5 This is a schematic diagram of the hardware architecture and logic modules of the computer device in an embodiment of the present invention. Detailed Implementation

[0041] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments, and the implementation of the present invention is not limited thereto. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Example 1:

[0043] This invention provides a cognitive inhibition assessment method based on a hierarchical Bayesian drift-diffusion model. By recoding and mapping the original behavioral data with accuracy, the macroscopic response indicators are decomposed into independent computational parameters with clear psychological significance, such as non-decision time, drift rate, and decision boundary. By constructing a competitive model space driven by cognitive hypotheses based on the drift-diffusion model, the competitive fitting of these independent computational parameters can accurately determine the specific stage of cognitive inhibition at the computational level, identify the specific site of cognitive inhibition, and achieve visualization and quantitative separation of the processing mechanism.

[0044] like Figure 1 The diagram shows the overall flowchart of the cognitive inhibition assessment method. This embodiment describes a cognitive inhibition assessment method based on a hierarchical Bayesian drift-diffusion model, which specifically includes the following steps:

[0045] S1. Behavioral Data Acquisition and Preprocessing: Obtain raw trial-subject data from the participants under the changing interference paradigm. Preprocess the raw trial-subject data to obtain effective raw trial-subject data, and output the raw behavioral data vector. The raw trial-subject data includes reaction time, conditional encoding, actual stimulus changes, and participant button selections.

[0046] Specifically, a variable interference paradigm task is run through a computer experimental program, collecting behavioral data from each participant in real time across hundreds of trials. In other implementations, this behavioral task can be replaced with a cognitive inhibition task involving a competitive relationship between target and interference information. The raw data for each trial includes: Reaction time: the number of milliseconds from stimulus presentation to the participant pressing a key. Conditional encoding: whether interference features change or remain unchanged. Unchanged interference features mean that the attributes of the interference stimulus (e.g., color, shape, location) remain fixed throughout the experiment or block; for example, all interference objects are the same color (e.g., blue squares), and the target object is a red square. Changed interference features mean that the attributes of the interference stimulus change randomly or switch according to rules. Actual change state of the stimulus, for example, whether the color of the main task changes. Participant key press selection: the participant's key press response in this trial. Preprocessing of the trial-level raw data includes removing premature responses (e.g., RT < 200ms) and abnormally delayed responses (e.g., RT > 3 standard deviations) to ensure data validity, representing the raw data of each trial as a raw behavioral data vector.

[0047] In this embodiment, each trial output in step S1 can be represented as a raw behavior data vector:

[0048] Xi = [RTi, Ci, Ki, Si];

[0049] Where RTi represents the reaction time of the i-th trial in milliseconds; Ci represents the conditional code; Ki represents the actual change in stimulus state; and Si represents the subject's key press response.

[0050] Specifically, the conditional encoding Ci can use a binary encoding method, for example: Ci = 0 indicates that the interference feature remains unchanged; Ci = 1 indicates that the interference feature changes. The actual change state of the stimulus Ki can also use a binary encoding method, for example: Ki = 0 indicates that the target feature has not changed; Ki = 1 indicates that the target feature has changed. The subject's key press response Si can also use a binary encoding method, for example: Si = 0 indicates that the subject pressed the "no change" key; Si = 1 indicates that the subject pressed the "change" key.

[0051] In this embodiment, the raw data collected for a subject in four consecutive trials can be as follows: Trial 1: X1 = [612, 0, 1, 1], indicating that the reaction time in this trial was 612 ms, under the condition that the interference feature remained unchanged, the target feature actually changed, and the subject pressed the "Change" button. Trial 2: X2 = [735, 1, 0, 1], indicating that the reaction time in this trial was 735 ms, under the condition that the interference feature changed, the target feature actually did not change, but the subject incorrectly pressed the "Change" button. Trial 3: X3 = [548, 0, 0, 0], indicating that the reaction time in this trial was 548 ms, under the condition that the interference feature remained unchanged, the target feature actually did not change, and the subject pressed the "No Change" button. Trial 4: X4 = [801, 1, 1, 1], indicating that the reaction time in this trial was 801 ms, under the condition that the interference feature changed, the target feature actually changed, and the subject pressed the "Change" button.

[0052] In practical applications, experimental programs (such as PsychoPy, E-Prime, or Psychtoolbox) can be run via computer terminals equipped with display and input devices, and the aforementioned fields for each trial can be written to local storage or a remote database in real time. The collected data is preprocessed, removing trials with excessively fast responses (RT < 200 ms) and trials with no button response. Abnormal trials exceeding the mean ± 3 standard deviations are then eliminated based on the participants' internal reaction time distribution, thus forming the effective input data required for subsequent modeling.

[0053] like Figure 2 As shown, the experimental paradigm of altered interference and the stimulus presentation time sequence diagram are presented. Figure 2 Each trial begins 500ms from the fixation point, followed by a 100ms memory period. Participants are tasked with memorizing only the target feature and ignoring distracting features. After the memory period ends, a blank screen is presented as a retention interval (1000ms), and finally, a test item is presented. In the test item, the target feature changes with a 50% probability, and distracting features also change independently with a 50% probability. Participants are required to quickly determine whether the target feature (e.g., color) has changed (key press response) after the test item appears, given the presence of distracting features (e.g., shape). By manipulating the stability of distracting features, a benchmark is provided for subsequent model comparisons.

[0054] S2. Based on cognitive logic, the accuracy recoding mapping compares the key selections in the original behavioral data vector with the actual changes in the stimulus state and generates a correctness logic code. Based on the correctness logic code, the original behavioral data vector is converted into a modeling data vector, and a modeling data table is established based on the modeling data vector.

[0055] Specifically, the subject's key press selection in each raw behavioral data vector is logically compared with the actual stimulus state to generate a binary correctness logical code. When the subject's key press selection matches the actual stimulus state, a correctness logical code of 1 is generated, indicating a correct response; when the subject's key press selection does not match the actual stimulus state, a correctness logical code of 0 is generated, indicating an incorrect response. Thus, the trial-level raw data of the raw behavioral data vectors is transformed into a clean dataset containing reaction time, conditional encoding, and correctness logical code (correct / incorrect). This allows the physical response space of the raw data to be mapped to a decision logic space, laying the foundation for subsequent modeling. The recoding mapping eliminates the subject's response preference for specific physical keys (such as habitually pressing the left key), ensuring that the drift rate parameter estimated in subsequent steps only reflects the quality of target information processing.

[0056] Specifically, the original behavioral data vector Xi = [RTi, Ci, Ki, Si] is transformed into the modeling data vector Yi = [RTi, Ci, ACCi].

[0057] Where ACi represents the correctness logic code, and its generation rule is as follows:

[0058] When Si = Ki, ACci = 1;

[0059] When Si ≠ Ki, ACci = 0.

[0060] In other words, if the subject's button press response matches the actual change in the stimulus, it is recorded as correct; otherwise, it is recorded as incorrect. After this processing, the system no longer retains physical response directions such as "left button" and "right button," but instead retains the decision outcome variable of "whether it is correct," making the subsequent drift-diffusion model estimation more directly correspond to the quality of target information processing.

[0061] Taking the above four trials as an example, the data after recoding in step S2 is as follows:

[0062] First trial: The original data is [612, 0, 1, 1]. Since Si = Ki, the re-encoded data is Y1 = [612, 0, 1].

[0063] Second trial: The original data is [735, 1, 0, 1]. Since Si ≠ Ki, the re-encoded data is Y2 = [735, 1, 0].

[0064] Third trial: The original data is [548, 0, 0, 0]. Since Si = Ki, the re-encoded data is Y3 = [548, 0, 1].

[0065] Fourth trial: The original data is [801, 1, 1, 1]. Since Si = Ki, the re-encoded data is Y4 = [801, 1, 1].

[0066] A modeling data table is established based on the modeling data vectors. The resulting modeling data table should include at least three columns: reaction time (RT), condition code, and correctness logic code (ACC). This modeling data table can be directly used as the software input file for the hierarchical Bayesian drift-diffusion model. For example, each row in the CSV file can be written as: 612,0,1; 735,1,0; 548,0,1; 801,1,1. This step unifies explicit keystroke behavior into calculable cognitive decision results, avoiding systematic biases in model parameter estimation caused by differences in left-hand preference, right-hand preference, or key mapping among different participants.

[0067] S3. Construct a cognitive hypothesis-driven competitive model space based on the drift-diffusion model. The parameters of the drift-diffusion model include non-decision time t, drift rate v, and decision boundary a. Four different competitive models are constructed based on the zero-effect hypothesis, early input gating hypothesis, late decision conflict hypothesis, and policy caution hypothesis, respectively.

[0068] In this embodiment, a competitive model is constructed using a computer program and hierarchical Bayesian sampling is performed. For example, it can be implemented in a Python environment using HDDM, PyMC, or Stan. Based on the zero-effect hypothesis, early input gating hypothesis, late decision conflict hypothesis, and policy caution hypothesis, four competitive models are constructed by applying conditional constraints to different parameters: zero-effect model M0, early input gating model Mt, late decision conflict model Mv, and policy caution model Ma. All models are based on the drift-diffusion model, which assumes that decision-making is a process of accumulating evidence until a certain decision boundary is reached. According to the basic principle of the drift-diffusion model, the reaction time is determined by the following core parameters: drift rate v, representing the average rate of evidence accumulation, reflecting processing efficiency; non-decision time t, representing the time consumed by sensory encoding and motor execution that do not participate in the decision-making process; and decision boundary a, representing the total amount of evidence required for decision-making, reflecting the degree of caution. This invention constructs four competitive models based on four different inhibition control theory hypotheses by applying different experimental constraints to the above parameters: zero-effect model M0, early input gating model Mt, late decision conflict model Mv, and policy caution model Ma, where:

[0069] Zero-effect model M0: used to characterize the absence of a stable inhibition effect, where the non-decision time t, drift rate v, and decision boundary a are set to remain unchanged with conditional coding.

[0070] Early Input Gating Model Mt: This model characterizes inhibition occurring in the early encoding stage, allowing only the non-decision time t to vary with conditional encoding. The early input gating hypothesis posits that inhibition occurs in the early encoding stage, and the model sets conflict conditions to affect only the non-decision time t. If the model has the highest goodness of fit, inhibition is considered to occur in the early encoding stage, representing the time cost of inhibition occurring in this stage.

[0071] Late Decision Conflict Model Mv: This model characterizes suppression occurring during the decision processing stage, allowing only the drift rate v to vary with the conditional encoding. The late decision conflict hypothesis posits that suppression occurs at the central decision-making stage. The model assumes that conflict conditions only affect the drift rate v, as conflict information can interfere with the efficiency of evidence accumulation and slow down decision-making. By allowing only the drift rate v to vary with the interfering condition, if the model has the highest fit, suppression is considered to have occurred during the decision processing stage, indicating that incomplete suppression leads to impaired processing efficiency.

[0072] The strategic caution model Ma is used to characterize inhibition stemming from subjects' subjective adjustment of decision criteria or response caution, allowing only the decision boundary 'a' to change with condition coding. The strategic caution hypothesis posits that inhibition manifests as a strategic adjustment, allowing only the decision boundary 'a' to change with conditions. If this model has the highest goodness of fit, the inhibitory effect is determined to originate from subjects' subjective adjustment of decision criteria or response caution, serving as a control model to exclude subjects' subjective adjustment of response thresholds.

[0073] In this embodiment, the data Y = [RT, condition, ACC] output from step S2 can be imported into the HDDM tool, and the following models can be established respectively: M0: parameters do not contain condition terms; Mt: t ~ condition; Mv: v ~ condition; Ma: a ~ condition. By constructing four different competing models, multiple interpretations with clear psychological meanings can be provided for the same set of behavioral data. Subsequent steps do not directly judge based on the size of a single parameter, but compare the overall explanatory power of each competing model for the same set of data, and select the optimal model. In other words, the role of the competing models is to transform the question of "where does cognitive inhibition occur" into the computable question of "which cognitive hypothesis has the strongest ability to generate the current behavioral data". The output of this step is four predefined candidate model objects that can enter the subsequent Bayesian sampling stage, namely M0, Mt, Mv, and Ma. This output result is then sent to step S4 for parameter estimation and model comparison.

[0074] like Figure 3The diagram shows the mapping between drift-diffusion model parameters and cognitive inhibition sites. The inhibitory effect of different cognitive processing sites is reflected in the differences in drift-diffusion parameters under different experimental conditions. The diagram maps non-decision time t, drift rate v, and decision boundary a to different processing stages. Non-decision time t corresponds to the stimulus encoding and response execution stage, drift rate v corresponds to the evidence accumulation and decision processing stage, and decision boundary a corresponds to the response criterion and caution adjustment stage. By observing which type of parameter needs to change with experimental conditions, we can deduce the main cognitive sites where cognitive inhibition occurs.

[0075] S4. Perform hierarchical Bayesian sampling and parameter posterior estimation, construct a hierarchical Bayesian drift-diffusion model, and use the MCMC (Markov Chain Monte Carlo) sampling algorithm based on the data in the modeling data table to iteratively estimate the parameter posterior distribution of each competing model, and calculate the standard DIC value of the bias information of each competing model.

[0076] To robustly estimate model parameters, this embodiment employs a hierarchical Bayesian framework. This framework assumes that individual parameters originate from a group-level population distribution, thus preserving individual differences while sharing group information. Hierarchical Bayesian MCMC sampling is performed using open-source frameworks such as PyMC or Stan, iteratively estimating the posterior distribution of parameters for each competing model. MCMC sampling fully utilizes the long-tail characteristics of the total distribution of reaction time without relying on traditional mean and variance statistics. This means that seemingly "abnormal" extremely fast or slow reactions can be effectively considered by the model, resulting in more accurate parameter estimates that better reflect the stochastic nature of cognitive processes.

[0077] S41. Construct a hierarchical Bayesian drift-diffusion model using open-source framework tools, establish a hierarchical structure of data layer, individual layer and group layer, set a prior distribution for hyperparameters of group layer, and obtain a complete hierarchical Bayesian drift-diffusion model by connecting data layer, individual layer and group layer, thereby realizing the constraint and shared estimation of individual parameters.

[0078] A data layer is established. For the observation data yij of the i-th subject in the j-th trial, the observation data includes the reaction time RTij and the correctness logical code ACCij. The data layer is used to describe the probability distribution of the behavioral data in this trial under given individual parameters.

[0079] An individual layer is established, and an individual parameter set θi = [vi, ti, ai] is set for each subject. Here, vi represents the subject's drift rate parameter, ti represents the subject's non-decision time parameter, and ai represents the subject's decision boundary parameter. When using a conditional dependency model, the corresponding parameters can be further decomposed into parameters under different experimental conditions; for example, ti,0 and ti,1 represent the non-decision time under the disturbance-unchanged condition and the disturbance-changed condition, respectively.

[0080] A group hierarchy is established, where all individual parameters are not independent of each other, but rather derived from the corresponding group-level distributions. For example, the individual drift rate vi is derived from the group-level distribution N(μv, σv), the individual non-decision time ti is derived from the group-level distribution N(μt, σt), and the individual decision boundary ai is derived from the group-level distribution N(μa, σa). Here, μ represents the group mean, and σ represents the group dispersion.

[0081] A prior is defined for the group-level hyperparameters μv, μt, μa, and σv, σt, σa, which are then used to constrain the parameter search range. The prior is preferably a weak-information prior, ensuring that the model retains data-driven features while avoiding unstable convergence caused by unconstrained estimation.

[0082] A complete hierarchical model is formed by connecting the data layer, individual layer, and group layer, resulting in a complete hierarchical Bayesian drift-diffusion model. This model can estimate both the individual parameters of each subject and the overall parameters at the group level.

[0083] In this embodiment, using a dataset of 20 participants with 240 valid trials each, individual parameters are first established for each of the 20 participants. These 20 sets of individual parameters are then constrained to the same hierarchical distribution. This way, when a participant has a relatively small number of valid trials, their parameter estimates will moderately shrink towards the group mean, thereby reducing the probability of extreme estimates and improving robustness under small sample conditions. The completed hierarchical candidate model provides the parameter space, prior structure, and likelihood function basis for subsequent MCMC sampling.

[0084] S42. Based on the data in the modeling data table, sample each competing model using the MCMC sampling algorithm, perform posterior parameter estimation for each competing model, and calculate the standard DIC value of the deviation information for each competing model based on the posterior parameter estimation results.

[0085] S421. Input data and competition model: Based on the modeling data table obtained in step S2 and the four different competition models constructed in step S3, the MCMC sampling algorithm is used to sample only one competition model at a time, and the posterior parameter estimation results are calculated. The posterior parameter estimation results include: posterior mean deviation, posterior mean, and deviation at the posterior mean.

[0086] Specifically, sampling parameters were set, and four independent sampling chains were established, each iterating 4000 times to obtain the posterior distribution of the parameters. The first 1000 iterations were discarded as warm-up samples, and the remaining samples were used for posterior estimation. This setting can be adjusted according to the sample size and computational resources. Assume that S effective samples were sampled, with S = 8000 sample points. The MCMC sampling algorithm was used to repeatedly sample the parameters of the competing model to obtain the posterior distribution samples of the drift rate v, non-decision time t, decision boundary a, and its group hyperparameters. This posterior distribution is not a single point estimate, but a set of samples reflecting the possible range and uncertainty of the parameters. This distribution comprehensively characterizes the uncertainty of all possible parameter values ​​given the data and model.

[0087] Parameter samples θ obtained by MCMC sampling (1) ,…,θ (S) The parameter θ for the s-th MCMC sample is (s) , parameter θ (s) Includes: v (s) , a (s) , t (s) The bias D of this sample is calculated as a transformation of the logarithmic form of the likelihood function:

[0088] ;

[0089] Among them, D (s) Let θ represent the bias of the S-th sample, y represent the observed data (including reaction time (RT) and accuracy), and θ represent the bias of the S-th sample. (s) Represents model parameters (v) (s) , a (s) , t (s) ), p(y∣θ (s) ) indicates that the given parameter θ (s) At time t=0.3, the likelihood value of the observed data y (RT and accuracy). Given a set of parameters (e.g., v=0.5, a=1.2, t=0.3), generate the probability of all currently observed trial responses (those specific RT and choices).

[0090] Specifically, the posterior mean deviation is calculated by averaging the deviations D of all S samples obtained from the MCMC sampling. , can be represented as:

[0091] ;

[0092] Among them, D (s) This represents the bias of the S-th sample, where S is the number of valid samples.

[0093] Specifically, the posterior mean of the parameters is calculated. It can be represented as:

[0094] ;

[0095] Where, θ (S) It is the parameter value obtained from the s-th sample.

[0096] Specifically, calculate the deviation at the posterior mean. It can be represented as:

[0097] ;

[0098] in, It is the likelihood value of the observed data y when the parameter takes the posterior mean. The larger the likelihood value, the more likely the parameter is. The better the model fits the data, the better.

[0099] S422. Output model fit index: Calculate the number of effective parameters PD (posterior deviance penalty for complexity) and the deviation information standard DIC value of each competing model based on the posterior estimation results. The deviation information standard DIC value is used as a model fit evaluation index for subsequent optimal model selection.

[0100] The number of effective parameters is obtained by calculating the difference between the deviation at the posterior mean and the deviation at the posterior mean. It can be represented as:

[0101] ;

[0102] The standard deviation information (DIC) value is obtained by summing the posterior mean deviation and the number of valid parameters. The calculation formula is as follows:

[0103] DIC = Posterior mean deviation + number of effective parameters;

[0104] The posterior mean bias measures how "bad" the model fits the data; a smaller posterior mean bias indicates a better fit. The number of effective parameters measures the model's complexity. It is not necessarily an integer and reflects the "degrees of freedom" of the parameters in Bayesian estimation. The DIC value penalizes complex models; therefore, a simpler model with a similar fit will have a lower DIC. The absolute value of DIC itself has no specific meaning; its value lies in the relative comparison between models. Generally, a DIC difference (ΔDIC) greater than 10 is considered to indicate a substantial difference between models.

[0105] In this embodiment, hierarchical Bayesian MCMC sampling is performed using the open-source framework tools PyMC or Stan. After running the tool pm.sample(), the relevant fields dic or deviance can be directly found in the posterior summary (az.summary) or trajectory diagnosis. The standard deviation information (DIC) value of each model is recorded, and the DIC values ​​of the zero-effect model M0, the early input gating model Mt, the late decision conflict model Mv, and the policy-prudent model Ma are compared. In this example, 5000 MCMC samples were performed on the Mv model. The specific calculation process of its standard deviation information (DIC) value is shown in Table 1 below:

[0106] Table 1 Example of the calculation process for the standard DIC value of deviation information.

[0107]

[0108] S5. Output the mechanism judgment results based on model evidence. By comparing the bias information standard DIC value of each competing model, select the competing model with the best data fit and output the cognitive inhibition site assessment results of the subjects.

[0109] Specifically, the bias information standard (DIC) values ​​of each competing model are compared, and the competing model with the smallest bias information standard (DIC) value is selected as the model with the best fit to the behavioral data. Based on the model with the best fit, the specific processing sites where cognitive inhibition occurs in the subjects are obtained, and the evaluation results of the subjects' cognitive inhibition sites are output.

[0110] By using the Discrete Informative Value (DIC) to assess the generalization ability of a model in a balance between model fit and complexity, the model with the smallest DIC value is considered the best approximation of the data generation process and thus the model with the best fit. In this embodiment, after completing hierarchical Bayesian MCMC sampling, the DIC of each model was calculated, and the DIC values ​​of each competing model were compared with the model with the best fit to the behavioral data. The DIC values ​​of each model are shown in Table 2 below:

[0111] Table 2. Examples of DIC value results for each model.

[0112]

[0113] As shown in the table above, the early input gating model Mt has the lowest DIC value (4289.1), significantly lower than the other three models. This indicates that, in terms of fitting the observed data and controlling model complexity, the early input gating model Mt is the best model in the current candidate model space. The model with the best fit to the behavioral data directly points to the specific cognitive processing site where inhibition occurs:

[0114] When the early input gating model Mt wins, it indicates that inhibition mainly acts on the early stages of perception or attention.

[0115] When the late-stage decision conflict model Mv wins, it indicates that inhibition mainly plays a role in the core process of accumulating decision evidence.

[0116] When the cautious model Ma wins, it indicates that inhibition is primarily manifested as an active, strategic adjustment of the response criteria.

[0117] When the zero-effect model M0 wins the comparison, it indicates that no effective inhibition mechanism was detected in this paradigm.

[0118] Based on the comparison of the DIC values ​​above, we can draw a clear conclusion on the specific processing sites where cognitive inhibition occurs: the early input gating model Mt is the best model for fitting behavioral data under the "variable interference" paradigm. In the "variable interference" task, cognitive inhibition mainly acts on the early gating stage.

[0119] like Figure 4 As shown in the diagram, the optimal model identification results are illustrated. Based on the DIC value of the optimal model (Mt), the DIC difference (ΔDIC) between the other models and the optimal model is calculated. The competing models are ranked according to the DIC difference, and the evidence strength of each competing model is evaluated.

[0120] Mt: ΔDIC = 0.0 (Optimal model);

[0121] Mv: ΔDIC =4346.7 - 4289.1 = 57.6;

[0122] Ma: ΔDIC = 4345.4 - 4289.1 = 56.3;

[0123] M0: ΔDIC = 4309.3 - 4289.1 = 20.2;

[0124] According to conventional standards (Burnham & Anderson, 2002), a ΔDIC > 10 indicates that the supporting evidence for the compared model is severely insufficient. In this embodiment, the ΔDIC of all competing models (Mv, Ma, M0) and the optimal model Mt is much greater than 10 (the smallest difference is 56.3), indicating that the Mt model has overwhelmingly strong supporting evidence compared to the other models.

[0125] In this embodiment, 24 participants were included in a set of experimental data with varying interference paradigms, and each participant completed 240 valid trials. After completing steps S1 to S4, four candidate models were compared, and the Mt model was ultimately identified as the model that best fit the behavioral data. Therefore, the final output is: "Optimal Model: Early Input Gating Model Mt; Evaluation Conclusion: Cognitive inhibition in this group of participants mainly occurred in the early input gating stage; Application Explanation: Interference information is mainly intercepted before entering subsequent decision processing, and the additional processing cost is mainly reflected in stimulus encoding or input gating." Furthermore, in a single-participant application scenario, if a participant's data ultimately generates a late decision conflict model Mv, the output is: "Optimal Model: Late Decision Conflict Model Mv; Evaluation Conclusion: Cognitive inhibition in this participant mainly occurred in the late decision conflict stage; Application Explanation: Interference information has entered the evidence accumulation process and generates conflict costs in the decision-making stage." "The final output of step S5 includes two parts: the name of the optimal model and the assessment conclusion of the cognitive inhibition sites corresponding to the optimal model. This result can be directly used for interpreting scientific research experimental results, evaluating the effectiveness of cognitive training, analyzing individual cognitive function status, and for automated behavioral data processing platforms."

[0126] Example 2:

[0127] This embodiment provides a computer device, which can be a server, a personal computer, or a dedicated psychological experiment data processing terminal. Its internal logical architecture is as follows: Figure 5 As shown, the computer device includes a processor, memory, input device, display, and network interface connected via a system bus. The memory stores a computer program, which the processor executes to implement a cognitive inhibition assessment method based on a hierarchical Bayesian drift-diffusion model as described in Example 1. The specific execution process includes:

[0128] S1. Obtain the trial-second raw data of the subject under the changing interference paradigm, preprocess the trial-second raw data to obtain effective trial-second raw data, and output the raw behavioral data vector; the trial-second raw data includes reaction time, conditional encoding, actual change state of stimulus and subject key selection.

[0129] S2. Accuracy recoding mapping based on cognitive logic: logically compare the key selection in the original behavioral data vector with the actual change state of the stimulus and generate a binary correctness logic code. Based on the correctness logic code, the original behavioral data vector is converted into a modeling data vector. A modeling data table is established based on the modeling data vector.

[0130] S3. Construct a cognitive hypothesis-driven competitive model space based on the drift-diffusion model. The parameters of the drift-diffusion model include non-decision time t, drift rate v, and decision boundary a. Four different competitive models are constructed based on the zero-effect hypothesis, early input gating hypothesis, late decision conflict hypothesis, and strategy caution hypothesis, respectively.

[0131] S4. Perform hierarchical Bayesian sampling and parameter posterior estimation, construct a hierarchical Bayesian drift-diffusion model, and use the MCMC sampling algorithm based on the data in the modeling data table to iteratively estimate the parameter posterior distribution of each competing model, and calculate the standard DIC value of the bias information of each competing model.

[0132] S5. By comparing the deviation information standard DIC values ​​of each competing model, select the competing model with the best data fit and output the assessment results of the cognitive inhibition sites of the subjects.

[0133] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A cognitive inhibition assessment method based on a hierarchical Bayesian drift-diffusion model, characterized in that, Including the following steps: S1. Obtain the trial-second raw data of the subject under the changing interference paradigm, preprocess the trial-second raw data to obtain effective trial-second raw data, and output the raw behavioral data vector; the trial-second raw data includes reaction time, conditional encoding, actual change state of stimulus and subject key selection. S2. Accuracy recoding mapping based on cognitive logic: logically compare the key selection in the original behavioral data vector with the actual change state of the stimulus and generate a binary correctness logic code. Based on the correctness logic code, the original behavioral data vector is converted into a modeling data vector. A modeling data table is established based on the modeling data vector. S3. Construct a cognitive hypothesis-driven competitive model space based on the drift-diffusion model. The parameters of the drift-diffusion model include non-decision time t, drift rate v, and decision boundary a. Four different competitive models are constructed based on the zero-effect hypothesis, early input gating hypothesis, late decision conflict hypothesis, and strategy caution hypothesis, respectively. Using the HDDM and PyMC tools in the Python environment, four competing models were constructed by applying conditional dependencies on different parameters based on the zero-effect hypothesis, the early input gating hypothesis, the late decision conflict hypothesis, and the policy caution hypothesis: zero-effect model M0, early input gating model Mt, late decision conflict model Mv, and policy caution model Ma. The zero-effect model M0 is used to characterize the absence of a stable inhibition effect, and it is set that the non-decision time t, the drift rate v, and the decision boundary a do not change with the conditional coding. The early input gating model Mt is used to characterize that inhibition occurs in the early encoding stage, and it is set to only allow the non-decision time t to change with the conditional encoding. The late-stage decision conflict model Mv is used to characterize the suppression that occurs during the decision processing stage, and it is set to allow only the drift rate v to change with the conditional encoding; The strategic caution model Ma is used to characterize the inhibition that stems from the subject's subjective adjustment of decision criteria or response caution. It is set to allow only the decision boundary a to change with the conditional encoding. S4. Perform hierarchical Bayesian sampling and parameter posterior estimation, construct a hierarchical Bayesian drift-diffusion model, and use the MCMC sampling algorithm based on the data in the modeling data table to iteratively estimate the parameter posterior distribution of each competing model, and calculate the standard DIC value of the bias information of each competing model. S5. By comparing the deviation information standard DIC values ​​of each competing model, select the competing model with the best data fit and output the assessment results of the cognitive inhibition sites of the subjects.

2. The cognitive inhibition assessment method based on a hierarchical Bayesian drift-diffusion model according to claim 1, characterized in that, Step S2 includes: logically comparing the subject's key selection in each original behavioral data vector with the actual stimulus state to generate a binary correctness logic code. When the subject's key selection matches the actual stimulus state, a correctness logic code of 1 is generated, indicating a correct response; when the subject's key selection does not match the actual stimulus state, a correctness logic code of 0 is generated, indicating an incorrect response. The modeling data vector includes reaction time, conditional encoding, and correctness logic code.

3. The cognitive inhibition assessment method based on a hierarchical Bayesian drift-diffusion model according to claim 1, characterized in that, Step S4 includes: S41. Construct a hierarchical Bayesian drift-diffusion model using open-source framework tools, establish a hierarchical structure of data layer, individual layer and group layer, set a prior distribution for hyperparameters of group layer, and obtain a complete hierarchical Bayesian drift-diffusion model by connecting data layer, individual layer and group layer. S42. Based on the data in the modeling data table, sample each competing model using the MCMC sampling algorithm, perform posterior parameter estimation for each competing model, and calculate the standard DIC value of the deviation information for each competing model based on the posterior parameter estimation results.

4. The cognitive inhibition assessment method based on a hierarchical Bayesian drift-diffusion model according to claim 3, characterized in that, Step S42 includes: Based on the modeling data table obtained in step S2 and the four different competing models constructed in step S3, the MCMC sampling algorithm is used to sample only one competing model at a time to calculate the posterior parameter estimation results. The posterior parameter estimation results include: posterior mean bias, posterior mean, and bias at the posterior mean. The effective number of parameters (PD) and the bias information standard (DIC) value for each competing model are calculated based on the posterior estimation results of the parameters. The bias information standard (DIC) value is used as the evaluation index for model fit.

5. The cognitive inhibition assessment method based on a hierarchical Bayesian drift-diffusion model according to claim 4, characterized in that, The calculation of the effective parameter count (PD) and the standard deviation information (DIC) value for each competing model based on the posterior estimation results includes: The number of effective parameters PD is obtained by the difference between the posterior mean deviation and the deviation at the posterior mean. The standard deviation information (DIC) value is obtained by the sum of the posterior mean deviation and the number of effective parameters PD.

6. The cognitive inhibition assessment method based on a hierarchical Bayesian drift-diffusion model according to claim 5, characterized in that, Step S5 includes: By comparing the standard deviation information (DIC) values ​​of each competing model, the competing model with the smallest standard deviation information (DIC) value is considered to have the best fit to the behavioral data. Based on the model with the best fit, the specific processing sites where cognitive inhibition occurs in the subjects are obtained, and the evaluation results of the subjects' cognitive inhibition sites are output.

7. A computer device comprising a processor and a memory for storing a processor-executable program, characterized in that, When the processor executes the program stored in the memory, it implements the cognitive inhibition assessment method based on the hierarchical Bayesian drift-diffusion model as described in any one of claims 1-6.

Images