Modeling method and system of cognitive ability quantitative evaluation system and evaluation method

Abstract

The invention discloses a modeling method and system of a cognitive ability quantitative evaluation system and an evaluation method. The modeling method comprises the following steps: generating an initial labyrinth, wherein the initial labyrinth is represented as a weighted graph structure; obtaining behavior data of a testee in the initial labyrinth, and extracting vector knowledge representing intuitive behavior tendency and structural knowledge representing structured learning ability; optimizing the initial labyrinth through a genetic algorithm, wherein a fitness function is combined with vector knowledge and structure knowledge; dynamically controlling the multi-level checkpoint, and generating a test number in real time based on the performance of the testee; verifying the uniqueness of the path from the starting point to the terminal point in the test times by adopting a breadth-first search algorithm, and optimizing the test times; the subject behavior is modeled based on a subsequent representation model that associates the cognitive dimension and the cross-diagnostic dimension. According to the evaluation system and the evaluation method generated by the invention, different mental disease patient groups can be obviously distinguished.

Description

Technical Field

[0001] This invention relates to a modeling method for a cognitive ability quantitative assessment system, as well as a corresponding cognitive ability quantitative assessment system and method, belonging to the field of healthcare informatics technology. Background Technology

[0002] A major challenge in the field of cognitive ability assessment is the lack of a tool that can objectively and quantitatively reflect the cognitive abilities of test subjects. Currently, assessment methods in this area largely rely on the test subject's self-report and the subjective judgment of assessors based on clinical experience. However, this approach is susceptible to interference from individual differences, cultural background, and the assessor's level of personal experience, leading to significant uncertainty and bias in the assessment results. Furthermore, descriptions of specific behavioral manifestations, such as disorganized thinking, often rely too heavily on subjective judgment, undoubtedly increasing the difficulty of accurate assessment. Simultaneously, many mental illnesses, such as anxiety disorders, depression, and schizophrenia, exhibit a high degree of overlap in cognitive manifestations, complicating accurate differentiation and identification and increasing the risk of misjudgment. Moreover, most existing assessment protocols fail to effectively assess the cognitive abilities of test subjects under natural conditions.

[0003] One common assessment method currently used is the use of standardized psychological assessment tools, such as the Minnesota Multiphasic Personality Inventory (MMPI) and the Beck Depression Inventory (BDI). Although these tools are widely used and standardized, they are primarily based on self-reports by the test-takers, making them susceptible to limitations imposed by the test-takers' honesty and insight, resulting in a relatively subjective assessment process. Furthermore, due to the relatively simple assessment models of these tools, it is difficult to capture the dynamic thinking and higher cognitive activity patterns of the assessed individuals. Clinical interviews are another assessment method, where assessors evaluate cognitive abilities through interaction with the test-takers. However, this method is highly dependent on the assessor's experience and professional judgment, easily influenced by various subjective factors, and the consistency of assessments among different assessors is difficult to guarantee, with the risk of misjudgment and omission remaining. Neuroimaging techniques, such as functional magnetic resonance imaging (fMRI) and electroencephalography (EEG), can help observe abnormalities in brain structure and function and can serve as auxiliary tools in the assessment process. However, these techniques are costly to use and, in most cases, lack stable and persistent biomarkers; therefore, they are usually only used as auxiliary tools, not the primary assessment tools.

[0004] In conclusion, the field of cognitive ability assessment urgently needs a more objective, quantitative tool that can assess patients' cognitive abilities under natural conditions to compensate for the shortcomings of existing methods and improve the accuracy and reliability of assessments. Summary of the Invention

[0005] The primary technical problem to be solved by this invention is to provide a modeling method for a cognitive ability quantitative assessment system.

[0006] Another technical problem to be solved by the present invention is to provide a cognitive ability quantitative evaluation system.

[0007] Another technical problem to be solved by the present invention is to provide a method for quantitative evaluation of cognitive ability.

[0008] To achieve the above-mentioned technical objectives, the present invention adopts the following technical solution:

[0009] According to a first aspect of the present invention, a modeling method for a cognitive ability quantitative assessment system is provided, comprising the following steps:

[0010] S1: Generate an initial maze, which is represented by a weighted graph structure;

[0011] S2: Obtain behavioral data of the test subjects in the initial maze, and extract vector knowledge representing intuitive behavioral tendencies and structural knowledge representing structured learning abilities;

[0012] S3: The initial maze is optimized using a genetic algorithm, where the fitness function combines the vector knowledge and the structural knowledge;

[0013] S4: Dynamically control multi-level levels and generate trial times in real time based on the test subject's performance;

[0014] S5: Use a breadth-first search algorithm to verify the uniqueness of the path from the starting point to the ending point in each trial and optimize the trials;

[0015] S6: Model the behavior of test subjects based on a successor representation model, wherein the successor representation model is associated with cognitive dimensions and cross-diagnostic dimensions.

[0016] Preferably, in step S2, the vector knowledge calculates the straight-line direction tendency between the current node and the destination using Euclidean distance; the structural knowledge evaluates the counterintuitive path selection ability using a metric of deviation from the shortest path.

[0017] Preferably, in step S4, the multi-level checkpoints include:

[0018] Level 1: Fixed start and end points, providing locally visible edges, and learning repeated paths;

[0019] Level 2: Fixed endpoint, random starting point, with locally visible edges;

[0020] Level 3: Random start and end points, with locally visible edges provided;

[0021] Level 4: Random start and end points, hidden path visibility.

[0022] Preferably, in step S6, the cognitive dimensions include short-term memory, long-term memory, planning ability, planning flexibility, and abstract integration ability; the cross-diagnostic dimensions include depression and anxiety, compulsive behavior, social withdrawal, and schizophrenia-like symptoms.

[0023] According to a second aspect of the present invention, a cognitive ability quantitative assessment system is provided, including a main control module, a task flow control module, a generation module, and a clinical impression assessment module; wherein, the main control module is connected to the task flow control module, the generation module, and the clinical impression assessment module, and is used for resource loading and logical control of role behavior;

[0024] The task flow control module is used to dynamically generate the aforementioned multi-level levels and trials.

[0025] The generation module is used to generate maze structures that can effectively assess the cognitive abilities of test subjects;

[0026] The clinical impression assessment module is used to capture the test subject's performance in the task. Based on the test subject's performance, the discount factor and state transition probability are calculated using the successor representation model, and the test subject's disease and condition are predicted accordingly.

[0027] Preferably, the generation module uses the Prim algorithm to generate the initial maze structure and combines it with a network library in Python to generate the corresponding weighted graph structure; then, it combines a genetic algorithm to optimize the maze structure to ensure that the maze can distinguish between intuitive behavior and structured learning behavior.

[0028] Preferably, in the clinical impression assessment module, the discount factor is adjusted by the cognitive dimension parameter, and the state transition probability is adjusted by the cross-diagnostic dimension parameter.

[0029] According to a third aspect of the present invention, a method for quantitatively assessing cognitive ability is provided, comprising the following steps:

[0030] Based on the patient's condition configuration parameters, the above initial maze is generated;

[0031] Capture behavioral data of test subjects and calculate cognitive dimension parameters and cross-diagnostic dimension parameters through a subsequent representation model;

[0032] The disease classification and severity assessment results are output based on the parameters.

[0033] Preferably, the disease classification includes:

[0034] When the depression and anxiety parameters are above the threshold and the schizophrenia-like parameters are below the threshold, an assessment result suggesting suspected depression is output.

[0035] When the schizophrenia-like parameter is higher than the threshold and the path randomness index is significant, the assessment result of suspected schizophrenia is output.

[0036] Compared with existing technologies, this invention achieves objective quantitative assessment of cognitive abilities through gamified tasks, completely eliminating the bias caused by the reliance on subjective judgment in traditional methods. Based on a clinically validated disease-adapted scenario library and a genetic algorithm-optimized structure, it significantly improves diagnostic accuracy for different mental illness groups. Utilizing a successor representation model to dynamically link cognitive dimensions and cross-diagnostic dimensions, it comprehensively captures cognitive characteristics such as short-term memory and planning abilities, as well as clinical symptoms such as depression and anxiety, under natural conditions. This invention combines low cost and high ease of use, and experimental verification has shown that it can effectively distinguish the core behavioral patterns of patients with depression and schizophrenia. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the cognitive ability quantitative assessment system provided in the first embodiment of the present invention;

[0038] Figure 2 This is a schematic diagram of the graphical interface of the computer or mobile device used in an embodiment of the present invention;

[0039] Figure 3 This is a comparison chart of the average scores of the depression group and the schizophrenia group in the standardized maze task in an embodiment of the present invention;

[0040] Figure 4 This is a schematic diagram illustrating the comparison results of "depression-anxiety" cross-diagnostic dimension parameters between two groups of patients in an embodiment of the present invention;

[0041] Figure 5 This is a schematic diagram comparing the parameters of the "schizophrenia-like symptoms" dimension in an embodiment of the present invention. Detailed Implementation

[0042] The technical content of the present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0043] The cognitive ability quantitative assessment system and method provided in this invention are illustrated in the following embodiments using a maze treasure hunt game as an example, but the scope of protection of this invention is not limited to this form. This technical solution can be applied to any task operable on electronic devices such as computers, mobile phones, and VR (virtual reality) that meets the following specific conditions: the task structure can be represented in tabular form; the task environment model is partially or completely invisible (i.e., the participant cannot fully understand the entire environment); and the state in the task can be clearly distinguished and defined. Task forms that meet these conditions include, but are not limited to, flying games, indoor escape games, and cargo handling games. In different types of games, the "maze" concept used in this invention can be adjusted accordingly to spatial elements with equivalent functions, such as rooms (e.g., indoor escape games) and roads (e.g., cargo handling games), depending on the characteristics of the specific game.

[0044] During the learning process of these tasks, participants need to gradually master the transition matrix between states (i.e., state transition probabilities) and infer the task reward probability associated with each state. Currently, this task is implemented through a web application, so any device capable of connecting to the internet and browsing web pages can use the system and method provided by this invention. The core advantage of this method lies in its applicability to various dynamic environments. Through continuous interaction between the participant and the system, the exploration strategy is constantly optimized, ultimately achieving the exploration and determination of the optimal solution or a near-optimal solution.

[0045] First Embodiment

[0046] like Figure 1 As shown, the first embodiment of the present invention provides a cognitive ability quantitative assessment system, including a main control module, a task flow control module, a generation module, and a clinical impression assessment module. More preferably, it also includes a score calculation module. The design of each module follows modular programming principles to ensure the functional independence and high cohesion of each module.

[0047] The main control module controls the task flow control module, generation module, clinical impression assessment module, and score calculation module. The main control module has resource loading and character behavior logic control functions. In this embodiment, the main control module serves as the game's main control module.

[0048] During resource loading, the game's main control module first reads the various material resources required by the game from a predefined resource folder asynchronously to generate the game scene. This process uses the fetch API in JavaScript for resource loading and management, and is based on Phaser3, combined with HTML5. <canvas>Elements are used to render game content. Resources include, but are not limited to: map textures, character sprite images, target object images, user interface button icons, and path icons. All image resources are managed through XML-based configuration files, which define the metadata of each image (such as coordinates, size, scaling ratio, etc.). During gameplay, mazes are also loaded based on mazes generated / optimized by the generation module (referred to as the "maze generation module" in this embodiment).

[0049] The character's behavior logic is implemented in JavaScript by binding event listeners. When controlling the character's behavior, the game's main control module uses EventListener to capture user key input and controls the character's movement by handling keyboard events. Each time a key is pressed, the system first calls the checkMove() function to check the traversability of the current character's position and adjacent paths. Specifically, the traversability of a path is determined by the edge weights in the graph structure generated by the maze module.

[0050] When the `checkMove()` function returns that the path is passable, the system updates the character's 2D coordinates. This operation is accomplished through the `updatePosition(x,y)` function, which modifies the character's current position and calls the `render()` function to update the character's display on the canvas. Updating the character's position at the data level is achieved by updating a `position` object (containing x and y coordinates). If an impassable path is detected (edge ​​weight is infinity or a specific impassable marker is used), collision behavior is triggered. Collision behavior is handled by the `triggerCollision()` function, which invokes the character sprite's knockdown animation. The animation implementation relies on the CSS3 `@keyframes` rule and the JavaScript `requestAnimationFrame` function for smooth frame-by-frame rendering.

[0051] The core task of the maze generation module is to generate maze structures that can effectively assess the cognitive abilities of test subjects. This module uses the Prim algorithm to generate the initial maze structure and combines it with Python network libraries (such as NetworkX) to generate the corresponding weighted graph structure. Subsequently, a genetic algorithm is used to optimize the maze structure to ensure that the maze can distinguish between intuitive and structured learning behaviors.

[0052] The task flow control module is responsible for the overall management and control of the gamification testing process, and generates and optimizes the test sessions in real time based on the test subjects' performance during the evaluation.

[0053] The clinical impression assessment module is responsible for capturing the test subject's performance in the game task, and using the successor representation model to calculate the discount factor and state transition probability based on the test subject's performance, and predicting the test subject's disease and condition accordingly.

[0054] The scoring module provides real-time feedback to test subjects during gameplay, encouraging them to optimize their exploration strategies and improve their learning efficiency.

[0055] Second Embodiment

[0056] Based on the cognitive ability quantitative assessment system in the first embodiment, the second embodiment of the present invention provides a modeling method for a cognitive ability quantitative assessment system, which includes at least the following steps:

[0057] S1: Generate an initial maze, which is represented as a weighted graph structure.

[0058] Prim's algorithm, a classic minimum spanning tree algorithm, is particularly suitable for generating mazes. The initial maze is generated starting from a randomly selected node, and nodes are gradually added to the existing tree. The tree is expanded by selecting adjacent nodes with the smallest edge weight.

[0059] The maze is represented as a weighted graph G = (V, E), where V is the set of nodes, representing each cell in the maze; E is the set of edges, representing the paths between adjacent cells. The edge weight w(e) represents the cost of the path between two nodes and can be assigned a value according to specific design requirements. A random node in the maze is selected as the starting point and added to the minimum spanning tree set T. The node v with the minimum edge weight is selected from all its neighbors in T and added to T, with the corresponding edge added to the spanning tree. This process is repeated until all nodes are included in T. The generated weighted graph structure is stored and processed using the Network X library in Python. To accommodate JavaScript processing requirements, the graph structure is converted to JSON format, where the coordinates of each node and the connection information of its neighbors are encoded as key-value pairs.

[0060] It should be noted that the embodiments of the present invention provide multiple optional schemes when generating the initial scenario. The initial scenario can be a default scenario or it can include multiple specific scenarios: for test subjects without a clinical diagnosis, the default scenario is used as the initial scenario; for test subjects with a clinical diagnosis (e.g., patients with depression), a specific scenario that has been previously validated and is suitable for such patients is selected as the initial scenario. It should be noted that the "initial scenario" mentioned here is a comprehensive concept, referring to the basic environmental framework of the evaluation task; in the specific embodiment of the present invention, which takes a maze treasure hunt game as an example, the initial scenario is implemented as an "initial maze" (i.e., a maze environment generated by the Prim algorithm and represented as a weighted graph structure). In other application scenarios (such as flight games, indoor escape games, and cargo handling games), the concept of "initial scenario" can be correspondingly reflected as structural elements with equivalent functions, such as route networks, room layouts, or road systems.

[0061] S2: Based on the test subject's performance in the initial maze, obtain the test subject's behavioral data in the initial maze, and extract vector knowledge representing intuitive behavioral tendencies and structural knowledge representing structured learning abilities.

[0062] To differentiate between intuitive (moving towards the goal) and structured learning (relying on the understanding of the maze's structure) behaviors in mazes by test subjects (including healthy individuals and patients), the concepts of vector knowledge and structural knowledge were introduced during maze generation. These two concepts were optimized using a genetic algorithm, ensuring that the generated mazes could assess both the test subjects' intuitive reaction abilities and their learning and adaptive capabilities.

[0063] Vector knowledge represents a test subject's tendency to act intuitively in a maze, namely, to move in a straight line toward the destination. This behavior is measured by Euclidean distance.

[0064] Assuming the current node is v and the endpoint is t, the formula for measuring vector knowledge is:

[0065]

[0066] Where (x) v y v ) and (x t y t ) are the coordinates of node v and endpoint t, respectively.

[0067] Structural knowledge measures a test subject's learning and mastery of the complex structure of a maze. To enhance the complexity of the maze and assess the test subject's cognitive abilities, structural knowledge is defined as a counterintuitive path-choice strategy that requires, in certain situations, test subjects to first move away from the endpoint and then gradually approach it.

[0068] The measurement of structural knowledge is achieved by measuring deviations from the shortest path. Assuming the shortest path for the full path of node v is P_shortest, while the actual generated path is P_actual, the optimization objective of structural knowledge is to increase the deviation between these two:

[0069] D eviation =|P actual (v,t)-P shortest (v,t)|.

[0070] S3: Optimize the initial maze using a genetic algorithm, where the fitness function combines the vector knowledge and the structural knowledge.

[0071] The maze is optimized using a genetic algorithm to ensure that there are both intuitive paths based on vector knowledge and complex paths that require structural knowledge to navigate.

[0072] Multiple initial mazes are generated, and the structure of each maze is initialized randomly. The fitness function is defined by combining vector knowledge and structural knowledge as a metric, and the specific formula is as follows:

[0073] Fitness (M) = α·D vector +β·D eviation

[0074] Here, α and β are weighting coefficients that control the influence of vector knowledge and structural knowledge in maze generation. Selection, crossover, and mutation: Classic operations from genetic algorithms (selection, crossover, and mutation) are used to iteratively update the population, gradually generating maze structures that meet the requirements (unique shortest path, and effective separation of structural and vector knowledge). Through these techniques, the maze generation module can generate complex maze structures that test both intuitive responses and cognitive learning abilities, providing a reliable foundation for subsequent cognitive ability assessments.

[0075] S4: Dynamically controls multi-level levels, generating trials in real time based on the test subject's performance.

[0076] In this embodiment, the level flow control includes the control of multiple levels and the control of the number of trials in each level. For example, this embodiment designs four levels, each designed to progressively increase the cognitive challenge of the test subjects to the maze structure, thereby accurately assessing their cognitive abilities and learning strategies.

[0077] Level 1 is designed to help test subjects quickly familiarize themselves with the maze structure and environment through repeated trials. In this level, the start and end points remain unchanged to ensure that test subjects can learn the shortest path through repeated practice. A total of three trials are conducted, with the start point S_1 and end point G_1 remaining constant in each trial. The system displays the visible edges around the character for one step, allowing test subjects to gradually grasp the overall layout of the maze during exploration.

[0078] Level 2 further enhances the test takers' understanding and familiarity with the maze structure by fixing the endpoint and randomizing the starting position. A total of five trials are conducted. The endpoint G_2 is fixed, and the starting point S_i (i = 1, 2, ..., 5) is randomly assigned in each trial. Visible edges around the character for one step are still provided to help the test takers gradually deduce the optimal path from different starting points to the fixed endpoint.

[0079] Level 3 tests the participant's exploration and learning abilities in a completely unknown situation by randomizing the start and end points. The start point S_i and end point G_i are randomly generated for each trial, and the visible edges around the character for one step are still visible, allowing the participant to gradually deduce the global maze structure from local information.

[0080] Level 4 removes visual support for the paths around the character, testing the test taker's planning and decision-making abilities and cognitive strategies in a completely hidden path situation. The starting point S_i and ending point G_i are randomly generated for each trial. The system no longer displays visible edges around the character; the test taker must make path choices based on structural knowledge accumulated from previous levels.

[0081] Short-term memory (STM) refers to the brain's ability to temporarily store information. STM has a limited capacity and can typically only process a few pieces of information simultaneously. It is the cornerstone of performing complex cognitive tasks such as learning, reasoning, and planning. The basic STM index extraction is based on Level 1, where three consecutive trials remain unchanged, for a total of 17 sets and 51 trials. Ideally, participants' performance scores would continuously improve across the three levels. A linear regression analysis of the average scores from the three consecutive levels yields a slope index that reflects their short-term memory ability.

[0082] Long-term memory refers to the brain's ability to store information or knowledge over a long period. It is considered to have an unlimited capacity. Long-term memory is divided into declarative memory and procedural memory. Declarative memory is further divided into episodic memory and semantic memory, while procedural memory includes acquired skills, habits, and other information. Long-term memory supports activities or behaviors such as recognizing familiar people and places, and mastering language and knowledge. Long-term memory ability reflects whether an individual's performance improves significantly when similar trials occur across all levels of a task. It is calculated by determining the proportion of trials with overlapping paths (greater than 70%) where the subject has good long-term memory; in such trials, performance will be significantly improved if the subject has good long-term memory. This is calculated by determining the proportion of trials where the individual chooses structure-based behaviors on overlapping paths.

[0083] Planning ability refers to the capacity to select a single action or a series of actions based on their utility; it is the ability to plan action strategies using structural knowledge. It requires individuals to predict the outcomes of a specific strategy (action or sequence of actions) and assess the utility of those outcomes. This ability helps individuals utilize time and resources efficiently, reduce risks and uncertainties in task execution, and increase the likelihood of successfully achieving goals. Planning flexibility refers to an individual's ability to adjust strategies promptly in the face of constantly changing goals. It requires individuals to quickly and appropriately evaluate strategies based on past experience and generate new actions or sequences of actions when faced with new goals. This ability enables individuals to effectively cope with unexpected events. This indicator is extracted through modeling. Specifically, the model will fit participant data and set a planning phase before each trial. In this phase, individuals will perform n-step planning through a planning mechanism. These n-step planning steps can be updated in a model-free manner to improve model performance (the number of planning steps n depends on the participant's "planning time" between task presentation and button press). For further reference, see the paper "A recurrent network model of planning explains hippocampal replay and human behavior" by Kristopher T. Jensen et al. (published in *Nature Neuroscience*, volume 27, 1340–1348 (2024)). In one embodiment of this invention, planning flexibility is reflected in the second task trials, with five consecutive trials starting at different points and ending at the same point, for a total of 50 trials. Participants are required to fully utilize the environmental model and flexibly switch goals according to task requirements to achieve better behavioral performance. This is reflected in behavioral indicators, where individuals will perform more structure-based behaviors as they get closer to the finish line. The calculated indicator is the proportion of structure-based behaviors performed at different distances from the finish line.

[0084] Abstract integration ability refers to the capacity to quickly adjust and update existing global concepts or frameworks in a mental model when local information or the environment changes. It requires a series of higher-level cognitive processes, such as logical thinking, reasoning, and innovation. This ability allows individuals to adapt cognitively and behaviorally to changes in the environment in a timely manner, thus processing new information and challenges more effectively. In this embodiment, abstract integration ability is mainly reflected in the transition from level three to level four. Due to the disappearance of visible edges, the maze no longer provides participants with any hints, and individual performance declines. However, since the map settings remain consistent, participants need to transfer the structural knowledge integrated from the first three levels to the task of the fourth level. The performance from level three to level four can be used to calculate the slope, thereby obtaining an index reflecting abstract integration ability.

[0085] S5: Use a breadth-first search algorithm to verify the uniqueness of the path from the starting point to the ending point in each trial and optimize the trials.

[0086] In embodiments of the present invention, since the path in each trial is automatically generated, furthermore, to ensure that the path calculation in each trial meets the requirements—that is, there is only one unique shortest path between the start and end points—the task flow control module employs a breadth-first search algorithm for path calculation to optimize the trials. The breadth-first search algorithm is a classic algorithm for finding the shortest path in an unweighted graph. Given a start point S_i and an end point G_i, the main steps of the breadth-first search algorithm are as follows:

[0087] Initialization: Starting from the starting node S_i, create an empty queue and add the starting node to the queue. Initialize the distance vector dist(v), setting the distance of the starting node to 0 and the distances of other nodes to infinity.

[0088] Expanding nodes: Take the first node v from the queue. For all adjacent nodes u, if dist(u) is greater than dist(v)+1, update dist(u) and add node u to the queue.

[0089] Path backtracking: After the destination node G_i is visited, the shortest path from the starting point to the destination is constructed using the backtracking method.

[0090] Path uniqueness verification: Verify whether, given a maze structure, there exists only one unique shortest path between the starting point and the ending point. Path uniqueness can be verified by the following conditions:

[0091] For a node v, if there are multiple child nodes in its child node set C(v) that satisfy dist(u) = dist(v) + 1, then there are multiple shortest paths, which does not meet the trial condition. If all nodes satisfy the unique child node condition, then the path is unique.

[0092] S6: Model the behavior of test subjects based on a successor representation model, wherein the successor representation model is associated with cognitive dimensions and cross-diagnostic dimensions.

[0093] This module uses successor representation (SR) models from reinforcement learning to model the cognitive processes of test subjects in game tasks. Successor representation models are a type of reinforcement learning model that makes behavioral decisions based on predictions of future states. It not only focuses on immediate rewards but also learns the transition relationships between states, using future states to estimate current behavior. This model is better able to capture the long-term strategic planning and flexibility of test subjects in game tasks.

[0094] Successor representation models combine the advantages of temporal difference learning (TD learning) with the flexibility of model-based algorithms, and are known as the third type of reinforcement learning algorithm besides model-based and model-free.

[0095] Modeling methods based on successor representation models include:

[0096] 1) Value function decomposition: The successor representation model decomposes the value function into two parts: a reward predictor and an estimate of the number of state visits, which is the successor map. Thus, the value function can be represented as the inner product of these two parts.

[0097]

[0098] Where 1[.] = 1 when the parameter is true.

[0099]

[0100] Wherein, the value function Q π (s, a) represents the expected reward after taking action a from state s under policy π. R(s) represents the expected value. γ represents the discount factor used to weigh the importance of immediate rewards against future rewards; the smaller the value, the lower the weight of future rewards. t ) represents the state s at time step t. t The reward obtained under the given policy. M(s, s′, a) represents the estimated probability of reaching state s′ from state s through action a given policy.

[0101] 2) Rapid adaptation to change: If the distant reward changes (such as the immediate reward of the final goal), the value function estimate of the successor representation model can be adjusted quickly.

[0102] 3) Obtaining bottleneck states: By estimating the number of times a state is accessed, the successor representation model can easily obtain the bottleneck state, which is the sub-goal. This makes it easier for the successor representation model to be combined with hierarchical reinforcement learning methods.

[0103] 4) Applications of deep learning: With the development of deep learning technology, successor representation models have been generalized to large-scale environments, enabling end-to-end estimation of successor representation models.

[0104] As mentioned earlier, the Q-value is calculated as the sum of the rewards at each time step with a discount factor γ. The successor representation model calculates the reward by first calculating the expected future occupancy of a state, multiplying it by R(s), and then summing up the rewards of all states.

[0105] In this way, by associating the test subject's decision-making process with different cognitive ability dimensions and cross-diagnostic indicators, the successor representation model can extract the structured information that the test subject continuously learns in the task, thereby reflecting their cognitive ability and clinical symptoms.

[0106] The core idea of ​​the successor representation model is to represent the state s t and the possible future states s t+k The association is established through the future state transition matrix M:

[0107]

[0108] Where γ is the discount factor, representing the discount rate of future rewards, P(s) t+k |s k ) indicates starting from the current state s t Transition to future state s t+k The probability of this transition matrix. The successor representation model optimizes policy π by learning this transition matrix.

[0109] To correlate the successor representation model with different cognitive and cross-diagnostic dimensions, the model will use a set of free parameters θ, which characterize the test subjects' performance across each dimension. Different weights are applied to the five cognitive dimensions and four cross-diagnostic dimensions to modulate the learning and decision-making processes of the successor representation model.

[0110] Cognitive dimensions are expressed by a set of free parameters θ cognitive = [θ1, θ2, θ3, θ4, θ5] to represent the following cognitive abilities: short-term memory parameter θ1, long-term memory parameter θ2, planning ability parameter θ3, planning flexibility parameter θ4, and abstract integration ability parameter θ5.

[0111] In the successor representation model, these dimensions influence the model's behavior by adjusting the discount factor γ and the state transition matrix M.

[0112]

[0113] Among them, w i These are the weights of the cognitive dimensions, used to adjust the impact of each dimension on the discount to future states. For example, higher planning ability and long-term memory levels will cause the model to focus more on long-term rewards and states, resulting in a higher γ.

[0114] Cross-diagnostic dimensions through a set of independent parameters θ diagnostic =[θ d1 θ d2 θ d3 θ d4 The learning mechanism of the successor representation model is independently adjusted for each dimension, corresponding to the depression and anxiety dimensions θ respectively. d1 Compulsive behavior dimension θ d2 Social withdrawal dimension θ d3 and the schizophrenia-like symptoms dimension θ d4 .

[0115] Across diagnostic dimensions, test subject performance is influenced by modulating exploration parameters and behavioral decision rules in the successor representation model:

[0116] Depression and anxiety tendencies lead to a lower exploratory tendency, making the model too conservative. This manifests as participants adopting fixed strategies, which assign high weights to behaviors with the highest action value and low weights to behaviors with lower action value, thus avoiding the use of exploratory strategies.

[0117] Compulsive behaviors will cause test subjects to repeatedly choose specific behavioral paths, making the model's strategy more fixed;

[0118] Social withdrawal may lead to lower decision-making activity, causing the model to make fewer attempts in the state space;

[0119] Schizophrenia-like symptoms introduce more randomness, leading to policy instability in the model.

[0120] Therefore, the model's policy π is modulated by these dimensions:

[0121] π(s t |θ diagnostic ) = softmax(Q(s) t a t θ diagnostic ))

[0122] Where Q(s) t a t θd iagnostic ) is a state behavior value function that includes diagnostic dimension information.

[0123] To combine the five cognitive dimensions with the four cross-diagnostic dimensions, the calculation formula for the entire model is as follows:

[0124]

[0125] γ(θ cognitive The discount factor is adjusted by the parameters θ1, θ2, θ3, θ4, and θ5 of the cognitive dimension.

[0126] P( st+ k|s t θ diagnostic ) is affected by the cross-diagnostic dimension parameter θ d1 θ d2 θ d3 θ d4 Adjusted state transition probability.

[0127] This module compares the overlap between the test subject's actual travel path and the shortest path. To calculate the shortest path, a breadth-first search algorithm is first used, as described in detail in the previous section. The score is calculated based on the following formula to evaluate the optimality of the test subject's path selection in a given number of trials:

[0128]

[0129] Where N_overlap represents the number of steps that overlap with the shortest path actually traversed by the test subject. N_actual represents the total number of steps actually traversed by the test subject from the starting point to the ending point. The score ranges from 0 to 100 points. When the path chosen by the test subject completely overlaps with the shortest path, i.e., N_overlap = N_actual, the score is 100 points. When the path chosen by the test subject deviates from the shortest path, containing redundant or incorrect steps, the score will decrease accordingly.

[0130] To preliminarily verify the discriminant validity of the embodiments of the present invention for different mental illness groups, behavioral data from 34 clinically diagnosed patients (depression group: n=16; schizophrenia group: n=18) were collected and analyzed. All participants completed a standardized maze task (containing four progressive cognitive challenge levels), and cognitive ability dimension parameters and cross-diagnostic symptom dimension parameters were extracted using a subsequent representation model. The analysis results are as follows: Figures 3 to 5 As shown.

[0131] 1. Task performance indicators: differences in the degree of cognitive impairment

[0132] Figure 3 The average standardized scores (out of 100) of the two groups of patients in the task are presented. The results show that the depression group scored significantly higher than the schizophrenia group in the first three tasks, indicating that schizophrenia patients have more severe functional impairment in core cognitive areas such as spatial navigation, path planning, and environmental modeling, which is consistent with the characteristics of clinical cognitive impairment.

[0133] 2. Depression-Anxiety Dimension Parameters: Disease-Specific Behavioral Patterns

[0134] Figure 4 The results of the intergroup comparison of the "depression-anxiety" parameter across diagnostic dimensions are presented: the parameter value of the depression group is significantly higher than that of the schizophrenia group. The high parameter value corresponds to a reduction in exploratory behavior and rigidity of strategies, which is consistent with theoretical expectations—depression and anxiety symptoms are manifested as decision-making conservatism and exploratory inhibition, verifying the specific response of this parameter to affective disorders.

[0135] 3. Schizophrenia-like symptoms dimension: Differences in strategy stability

[0136] Figure 5 Key findings revealing the "schizophrenia-like symptoms" dimension: The parameter values ​​were significantly higher in the schizophrenia group, representing an increase in the path randomness index and the entropy of the state transition matrix. This indicates that the parameter successfully captures the characteristics of decision-making randomness and strategy instability, which is consistent with the cognitive disintegration theory model of schizophrenia.

[0137] In summary, the technical features of the embodiments of the present invention include:

[0138] 1. Objectivity and Quantification: Computerized testing provides a standardized and repeatable assessment process, reducing human error and subjective bias, making diagnostic results more objective and reliable. It can more accurately distinguish subtle differences between different mental illnesses, especially when there is significant overlap in symptoms, helping doctors make more accurate diagnoses.

[0139] 2. Ease of use and low cost: This system is designed with gamification in mind, and can conduct preliminary cognitive function assessments without the guidance of professional medical personnel. It is suitable for use in medical institutions at all levels and in home environments, and can be used to regularly monitor cognitive and mental health.

[0140] 3. Natural and comprehensive: Gamified assessment methods enable a more comprehensive understanding of patients' cognitive functions, extending to all aspects of social functioning. The assessment is conducted under natural conditions, providing diagnostic information that traditional methods cannot cover.

[0141] Third Embodiment

[0142] Based on the modeling method of the above-mentioned cognitive ability quantitative assessment system, the third embodiment of the present invention provides a cognitive ability quantitative assessment method, which includes at least the following steps:

[0143] Based on the patient's condition configuration parameters, the generation module generates the initial maze described above;

[0144] Capture behavioral data of test subjects and calculate cognitive dimension parameters and cross-diagnostic dimension parameters through a subsequent representation model;

[0145] The disease classification and severity assessment results are output based on the parameters.

[0146] In a preferred embodiment of the present invention, the disease classification includes:

[0147] When the depression and anxiety parameters are above the threshold and the schizophrenia-like parameters are below the threshold, an assessment result suggesting suspected depression is output.

[0148] When the schizophrenia-like parameter is higher than the threshold and the path randomness index is significant, the assessment result of suspected schizophrenia is output.

[0149] Therefore, by utilizing the cognitive ability quantitative assessment method provided in the embodiments of the present invention, task indicators are calculated and aligned with cross-diagnostic symptoms of mental illness, thereby providing more objective and quantifiable assessment results for mental illness.

[0150] It should be noted that the above embodiments are merely illustrative examples. The technical solutions of the various embodiments can be combined, and the order of the steps can be changed, all of which are within the protection scope of this invention.

[0151] The modeling method, system, and evaluation method of the cognitive ability quantitative evaluation system provided by this invention have been described in detail above. Any obvious modifications made to this invention by those skilled in the art without departing from its essential content will constitute an infringement of the patent rights of this invention and will incur corresponding legal liability.< / canvas>

Claims

1. A modeling method for a cognitive ability quantitative assessment system, characterized in that... Includes the following steps: S1: Generate an initial maze, which is represented by a weighted graph structure; S2: Obtain behavioral data of the test subjects in the initial maze, and extract vector knowledge representing intuitive behavioral tendencies and structural knowledge representing structured learning abilities; S3: The initial maze is optimized using a genetic algorithm, where the fitness function combines the vector knowledge and the structural knowledge; S4: Dynamically control multi-level levels and generate trial times in real time based on the test subject's performance; S5: Use a breadth-first search algorithm to verify the uniqueness of the path from the starting point to the ending point in each trial and optimize the trials; S6: Model the behavior of test subjects based on a successor representation model, wherein the successor representation model is associated with cognitive dimensions and cross-diagnostic dimensions.

2. The modeling method as described in claim 1, characterized in that... In step S2, the vector knowledge calculates the straight-line direction tendency between the current node and the destination using Euclidean distance; the structural knowledge evaluates the counterintuitive path selection ability using a metric of deviation from the shortest path.

3. The modeling method as described in claim 1, characterized in that... In step S4, the multi-level checkpoints include: Level 1: Fixed start and end points, providing locally visible edges, and learning repeated paths; Level 2: Fixed endpoint, random starting point, with locally visible edges; Level 3: Random start and end points, with locally visible edges provided; Level 4: Random start and end points, hidden path visibility.

4. The modeling method as described in claim 1, characterized in that... In step S6, the cognitive dimensions include short-term memory, long-term memory, planning ability, planning flexibility, and abstract integration ability; the cross-diagnostic dimensions include depression and anxiety, compulsive behavior, social withdrawal, and schizophrenia-like symptoms.

5. A cognitive ability quantitative assessment system, characterized in that... It includes a main control module, a task flow control module, a generation module, and a clinical impression assessment module; among which, the main control module connects the task flow control module, the generation module, and the clinical impression assessment module, and is used for the logical control of resource loading and role behavior; The task flow control module is used to dynamically generate the multi-level levels and trials as described in any one of claims 1 to 4; The generation module is used to generate maze structures that can effectively assess the cognitive abilities of test subjects; The clinical impression assessment module is used to capture the test subject's performance in the task. Based on the test subject's performance, the discount factor and state transition probability are calculated using the successor representation model, and the test subject's disease and condition are predicted accordingly.

6. The cognitive ability quantitative assessment system as described in claim 5, characterized in that... It also includes a scoring module to provide real-time feedback to test subjects during gameplay, encouraging them to optimize their exploration strategies and improve their learning efficiency.

7. The cognitive ability quantitative assessment system as described in claim 5, characterized in that... The generation module uses the Prim algorithm to generate the initial maze structure and combines it with a network library in Python to generate the corresponding weighted graph structure. Then, it combines a genetic algorithm to optimize the maze structure to ensure that the maze can distinguish between intuitive behavior and structured learning behavior.

8. The cognitive ability quantitative assessment system as described in claim 5, characterized in that... In the clinical impression assessment module, the discount factor is adjusted by the cognitive dimension parameter, and the state transition probability is adjusted by the cross-diagnostic dimension parameter.

9. A method for quantitatively assessing cognitive ability, characterized in that... Includes the following steps: Based on the patient's condition configuration parameters, an initial maze as described in any one of claims 1 to 4 is generated; Capture behavioral data of test subjects and calculate cognitive dimension parameters and cross-diagnostic dimension parameters through a subsequent representation model; The disease classification and severity assessment results are output based on the parameters.

10. The cognitive ability quantitative assessment method as described in claim 9, characterized in that, The disease classification includes: When the depression and anxiety parameters are above the threshold and the schizophrenia-like parameters are below the threshold, an assessment result suggesting suspected depression is output. When the schizophrenia-like parameter is higher than the threshold and the path randomness index is significant, the assessment result of suspected schizophrenia is output.

Images