Artificial intelligence psychological assessment method and system

Abstract

The application discloses an artificial intelligence psychological evaluation method and system, obtains multi-source psychological related data of a measured object, and shunts the multi-source psychological related data to a subjective data link, an objective data link and a restricted text semantic link, respectively calculates subjective risk values, objective risk values, restricted text semantic features and restricted text evidence segments, carries out time attenuation smoothing on basic features, calculates dimension risk scores of multiple psychological dimensions, and generates corrected comprehensive risk scores and dynamic psychological portraits in combination with subjective and objective consistency factors and dynamic weights; generates a structured evidence package according to the risk scores, the dynamic weights and the restricted text evidence segments, and calls a psychological evaluation large model to generate a structured output result containing risk explanation items and evidence number fields. The application can improve psychological risk identification accuracy, and enhance the explainability and reviewability of a large model evaluation report.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and mental health assessment technology, specifically to artificial intelligence psychological assessment methods and systems. Background Technology

[0002] With the increasing demand for mental health services, psychological assessment systems based on internet platforms, mobile terminals, wearable devices, and artificial intelligence algorithms are gradually being applied. Existing psychological assessment methods typically analyze the psychological state of the test subject through questionnaires, self-report texts, behavioral records, or physiological state data, and output assessment results related to anxiety, depression, stress, sleep, or emotional stability. Compared to traditional paper-based questionnaires, these methods have certain advantages in data collection efficiency, result generation speed, and service coverage.

[0003] However, existing psychological assessment methods still have significant shortcomings. On the one hand, many systems rely primarily on questionnaire scores or self-report texts for judgment, making the assessment results susceptible to subjective concealment, compliant responses, misunderstandings, or short-term emotional fluctuations by the test-taker. Even when some systems incorporate behavioral, physiological, or sleep data, they often directly mix and calculate the overall risk value, lacking independent calculation processes for subjective and objective data links, making it difficult to determine whether subjective assessment results are consistent with objective behavioral states. When a test-taker has a low subjective risk but obvious objective behavioral abnormalities, or a strong subjective sense of distress but insufficient objective evidence, existing methods struggle to provide timely risk correction, retesting prompts, or manual verification, easily leading to missed judgments, misjudgments, or unstable risk levels.

[0004] On the other hand, with the application of large language models in text generation and interpretation analysis, some psychological assessment systems have begun to utilize these models to generate assessment reports, risk interpretations, and intervention recommendations. However, existing large model report generation methods largely rely on cue word constraints, typically inputting raw text, scale results, or behavioral summaries directly into the large model, which then generates natural language conclusions. Due to the lack of structured evidence packages, evidence number binding, and post-output validation mechanisms, large models may generate inferences not supported by the underlying data, or even output diagnostic conclusions, absolute judgments, or personal information beyond the authorized scope. For sensitive applications like psychological assessment, if the risk interpretations in the report cannot be traced back to specific evidence sources, it not only reduces the efficiency of professional review but also affects the reliability and security of the assessment results.

[0005] Therefore, it is necessary to provide an AI-based psychological assessment method and system based on big data, which can perform layered processing of subjective data, objective data, and textual semantic data after multi-source data collection, improve the accuracy of risk assessment through subjective-objective consistency correction and dynamic psychological profiling, and constrain the output of the large model through structured evidence packages and evidence binding verification mechanisms, thereby generating interpretable, traceable, and verifiable psychological assessment reports. Summary of the Invention

[0006] The purpose of this invention is to provide an artificial intelligence psychological assessment method and system based on big data, which solves the problems that existing artificial intelligence psychological assessment methods have difficulty in reliably correcting when subjective assessment results are inconsistent with objective behavioral states, and that assessment reports generated based on large models lack evidence binding and post-verification mechanisms, which leads to easy misjudgment and omission in risk assessment and difficulty in retrospectively verifying the interpretation conclusions.

[0007] The first aspect of this invention provides an artificial intelligence-based psychological assessment method, comprising:

[0008] Acquire multi-source psychological data of the tested object, preprocess the multi-source psychological data, and distribute it to the subjective data link, objective data link and restricted text semantic link according to the data source attributes; Based on the subjective data of the subjective data link, a subjective feature vector is generated and a subjective risk value is calculated; based on the objective data of the objective data link, an objective feature vector is generated and an objective risk value is calculated; and based on the open text data of the restricted text semantic link, restricted text semantic features and restricted text evidence fragments are extracted. The basic features in the subjective feature vector, the objective feature vector, and the restricted text semantic features are smoothed by time decay to generate smoothed time window features, and the dimensional risk scores corresponding to multiple psychological dimensions are calculated based on the smoothed time window features. The subjective and objective risk values ​​are used to calculate a subjective-objective consistency factor to characterize the degree of subjective-objective matching, and the dynamic weights corresponding to each psychological dimension are calculated based on at least one of data quality, recent changes, historical contributions, and manual review results. A corrected comprehensive risk score is generated based on the dimensional risk score, the dynamic weight, and the subjective-objective consistency factor, and a dynamic psychological profile is constructed based on the dimensional risk score, the dynamic weight, the subjective-objective consistency factor, and the corrected comprehensive risk score. A structured evidence package is generated by encapsulating the subjective risk value, the objective risk value, the subjective-objective consistency factor, the dimensional risk score, the dynamic weight, and the restricted text evidence fragments. Based on the structured evidence package and prompt word template, the psychological assessment model is invoked to output a structured output result that conforms to the preset format and includes risk interpretation items and evidence number fields; The structured output results are subjected to evidence binding verification using post-processor software verification logic. The evidence binding verification includes verifying the subset inclusion relationship between the evidence number referenced in the structured output results and the structured evidence package. If the evidence binding verification passes, a traceable psychological assessment report is generated based on the structured output results and the structured evidence package.

[0009] In a second aspect, the present invention provides an artificial intelligence psychological assessment system based on big data, comprising: The data preprocessing module is used to preprocess multi-source psychological data and distribute it to the subjective data link, objective data link and restricted text semantic link according to the data source attributes. The feature extraction module is used to generate subjective feature vectors and calculate subjective risk values ​​based on subjective data from the subjective data link, generate objective feature vectors and calculate objective risk values ​​based on objective data from the objective data link, and extract restricted text semantic features and restricted text evidence fragments based on open text data from the restricted text semantic link. The processing module is used to perform time decay smoothing on the basic features in the subjective feature vector, the objective feature vector, and the restricted text semantic features to generate smoothed time window features; calculate the dimension risk scores corresponding to multiple psychological dimensions based on the smoothed time window features; calculate the subjective-objective consistency factor based on the subjective risk value and the objective risk value; and calculate the dynamic weights corresponding to each psychological dimension based on at least one of data quality, recent changes, historical contributions, and manual review results. The risk scoring and profile building module is used to generate a corrected comprehensive risk score based on the dimensional risk scores, the dynamic weights, and the subjective-objective consistency factors; and to build a dynamic psychological profile based on the dimensional risk scores, the dynamic weights, the subjective-objective consistency factors, and the corrected comprehensive risk score. The evidence package generation module is used to generate a structured evidence package based on the subjective risk value, the objective risk value, the subjective-objective consistency factor, the dimensional risk score, the dynamic weight, and the restricted text evidence fragment; The large model reasoning module is used to call the psychological assessment large model based on the structured evidence package and prompt word template to generate structured output results containing risk interpretation items and evidence number fields.

[0010] This invention distributes multi-source psychologically relevant data into subjective, objective, and restricted text semantic links according to their source attributes, generating subjective feature vectors, objective feature vectors, restricted text semantic features, and corresponding risk variables for each. Based on this, the invention calculates a subjective-objective consistency factor based on subjective and objective risk values, and combines time decay smoothing, dimensional risk scores, and dynamic weights to generate a corrected comprehensive risk score and dynamic psychological profile. Through this approach, the invention avoids the problems of unclear data boundaries and double-counting caused by simply mixing subjective and objective data for direct scoring, and can identify inconsistencies between subjective assessment results and objective state characteristics, thereby improving the accuracy, stability, and dynamic responsiveness of psychological risk assessment.

[0011] Before invoking the large-scale psychological assessment model, this invention first encapsulates and generates a structured evidence package based on subjective risk values, objective risk values, subjective-objective consistency factors, dimensional risk scores, dynamic weights, and restricted textual evidence fragments. After the large-scale model outputs the structured results, a post-processing software verification logic performs evidence binding verification on the output results, determining whether the cited evidence numbers belong to the set of legitimate evidence numbers in the structured evidence package, and checking whether the risk interpretation items are supported by legitimate evidence. Therefore, this invention can limit the natural language risk interpretations generated by the large-scale model to the scope of verifiable evidence, ensuring that the conclusions, risk factors, and recommendations in the psychological assessment report all have clear sources of evidence. This significantly reduces the risks of large-scale model illusions, inferences without evidence, and unauthorized outputs, improving the traceability, interpretability, and efficiency of manual review of the assessment report.

[0012] The above description is merely an overview of the technical solution disclosed herein. In order to better understand the technical means of this disclosure and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0013] 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 are briefly described below. Obviously, the drawings described below are only some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0014] Figure 1 A schematic diagram of the overall architecture of the big data-based artificial intelligence psychological assessment system provided in this embodiment of the invention; Figure 2 A flowchart illustrating the big data-based artificial intelligence psychological assessment method provided in this embodiment of the invention; Figure 3 This is a schematic diagram illustrating the layered processing of the subjective data processing link, the objective data processing link, and the restricted text semantic link provided in the embodiments of the present invention. Figure 4 This is a schematic diagram illustrating the process of smoothing time decay features and constructing dynamic psychological profiles in an embodiment of the present invention. Figure 5 A schematic diagram illustrating the calculation process of dimensional risk scores, subjective-objective consistency factors, and comprehensive risk scores provided in embodiments of the present invention; Figure 6 This is a schematic diagram of the dynamic weight update process provided in an embodiment of the present invention; Figure 7 A schematic diagram illustrating the generation process of structured evidence packages and restricted text evidence fragments provided in embodiments of the present invention; Figure 8 This is a schematic diagram illustrating the process of generating a large-scale psychological assessment report based on evidence constraints, as provided in an embodiment of the present invention. Figure 9 This is a schematic diagram of the evidence binding verification process provided in an embodiment of the present invention. Detailed Implementation

[0015] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0016] This invention provides an artificial intelligence-based psychological assessment method and system based on big data, applicable to scenarios such as schools, enterprises, medical auxiliary assessment institutions, community psychological service centers, and internet-based mental health service platforms. It is used to assist in the assessment of the psychological state, emotional fluctuations, stress levels, abnormal behavioral trends, and potential psychological risks of the test subjects.

[0017] It should be noted that the psychological assessment results described in this invention are used for mental health screening, risk alerts, auxiliary assessments, retest recommendations, or manual review, and do not directly replace clinical diagnostic conclusions made by professional physicians, psychological counselors, or other professionals.

[0018] like Figure 1As shown, the psychological assessment system 100 provided in this embodiment of the invention includes a data acquisition module 101, a data preprocessing module 102, a subjective feature extraction module 103, an objective feature extraction module 104, a text semantic feature extraction module 105, a time decay processing module 106, a dimensional risk calculation module 107, a subjective-objective consistency calculation module 108, a dynamic weight calculation module 109, a comprehensive risk scoring module 110, a dynamic psychological profile construction module 111, an evidence package generation module 112, a large model reasoning module 113, an assessment report generation module 114, a manual review module 115, a feedback update module 116, a database 117, and an evidence binding and verification module 120.

[0019] Users can complete psychological assessments, authorize data collection, and view assessment results through user terminal 118; administrators or professionals can view abnormal assessment results, evidence packages, model interpretations, and manual review tasks through management terminal 119.

[0020] like Figure 2 and Figure 3 As shown, the psychological assessment method provided in this embodiment of the invention includes the following steps.

[0021] S100: Obtain multi-source psychological data related to the tested subject.

[0022] The multi-source psychological data includes one or more of the following: questionnaire assessment data 201, subjective self-assessment text 202, behavioral interaction data 203, physiological state data 204, sleep rhythm data 205, environmental scene data 206, and open text data 207.

[0023] The questionnaire assessment data 201 may include standardized or semi-standardized scale data such as anxiety scales, depression scales, stress scales, sleep quality scales, mood stability scales, job burnout scales, and student mental health scales. Subjective self-assessment text 202 may include the test subject's proactive description of their own emotions, stress, sleep, learning status, and work status. Behavioral interaction data 203 may include answering time, number of times questions were returned, number of modifications, pause duration, click frequency, number of assessment interruptions, continuous login status, and interaction frequency. Physiological state data 204 may include data collected by wearable devices such as heart rate, heart rate variability, exercise volume, and resting time. Sleep rhythm data 205 may include sleep onset time, wake-up time, sleep duration, number of nighttime awakenings, and sleep regularity. Environmental scenario data 206 may include assessment time, assessment scenario, learning or work load, and recent significant event tags. Open-ended text data 207 may include diaries, consultation messages, open-ended questions and answers, and speech-to-text transcription.

[0024] In one implementation, before collecting behavioral interaction data 203, physiological state data 204, sleep rhythm data 205, and open-ended text data 207, the data acquisition module 101 displays data authorization information through the user terminal 118 and records the authorization scope, authorization time, authorized data type, and authorization withdrawal status. For data that has not been authorized, the system does not collect or analyze it.

[0025] S200 preprocesses multi-source psychological data and categorizes it into subjective data, objective data, and restricted text data according to the data source attributes.

[0026] The data preprocessing module 102 cleans, deduplicates, desensitizes, handles outliers, corrects timestamps, and normalizes the format of the collected data. Specifically, it performs missing item detection, reverse question conversion, abnormal answer detection, and scale total score calculation on the questionnaire assessment data 201; it performs duplicate click filtering, abnormal duration removal, and interaction event merging on the behavioral interaction data 203; it performs noise removal, sampling frequency unification, outlier smoothing, and device wearing validity detection on the physiological state data 204 and sleep rhythm data 205; and it performs de-identification, sensitive personal information masking, and sentence segmentation on the open text data 207.

[0027] To avoid data duplication during subsequent subjective-objective consistency calculations, this invention divides the data into three calculation chains after preprocessing: First, the subjective data link. The subjective data link includes questionnaire assessment data 201 and subjective self-assessment text 202, which are used to form subjective feature vector 301 and subjective risk value 501.

[0028] Second, the objective data link. The objective data link includes behavioral interaction data 203, physiological state data 204, sleep rhythm data 205, and environmental scene data 206, which are used to form objective feature vector 302 and objective risk value 502.

[0029] Third, the restricted text semantic link. The restricted text semantic link includes open text data 207. This link does not directly input the original full text into the psychological assessment model 803, but instead generates restricted text semantic features 303 and restricted text evidence fragments 702 by the text semantic feature extraction module 105. The restricted text semantic features 303 are used for risk calculation, and the restricted text evidence fragments 702 are used to provide necessary context when interpreting the report.

[0030] Through the above-described hierarchical processing, both subjective risk value 501 and objective risk value 502 have clear data source boundaries, and they will not be calculated repeatedly when calculating the subjective-objective consistency factor 503. The subsequent fusion feature vector 304 is generated only after the calculation of subjective risk value 501, objective risk value 502, and subjective-objective consistency factor 503 are completed, and is used for dynamic psychological profile construction and comprehensive risk scoring.

[0031] S300 generates subjective feature vectors, objective feature vectors, and restricted text semantic features, respectively.

[0032] The subjective feature extraction module 103 generates scale features based on the questionnaire assessment data 201, including the total score of each scale, scores of each dimension, abnormal response indicators, consistency of reverse questions, consistency of homogeneous questions, and intensity of subjective stress description. For scores of different scales, the system uses scale rule conversion or norm normalization methods to map them to a unified risk range of 0 to 100 points.

[0033] For example, the raw score of a certain scale is The norm mean is The standard deviation is Then the standard score can be calculated first. :

[0034] Then, it is mapped to a risk score using the Sigmoid function:

[0035] in, This represents the risk score for the corresponding subjective indicator. This represents the raw score of the scale. This represents the norm mean of the corresponding scale or population. This represents the standard deviation of the corresponding scale or the corresponding population. Represents the natural constant.

[0036] For scales without normative data, the minimum value of the scale can be used. and maximum value Perform range normalization:

[0037] in, This indicates the preset minimum value of the scale or indicator. This indicates the preset maximum value of the scale or indicator.

[0038] The objective feature extraction module 104 generates an objective feature vector 302 based on behavioral interaction data 203, physiological state data 204, sleep rhythm data 205, and environmental scene data 206. For heterogeneous data such as click frequency, pause duration in answering questions, number of question modifications, sleep duration, sleep onset time deviation, and heart rate variability, the system uses threshold mapping, statistical normalization, or abnormal deviation mapping to convert them into risk indicators ranging from 0 to 100 points.

[0039] For example, the degree of deviation can be calculated based on the historical average or group average of the test subjects regarding the duration of pauses during answering questions.

[0040] in, Indicates the degree of deviation. This indicates the average pause duration in the current assessment. This indicates the duration of historical benchmark pauses or the duration of pauses within the same group of benchmarks. This represents the standard deviation of the historical baseline pause duration or the pause duration within the same group. The system then... The risk score is mapped to a range of 0 to 100. For sleep duration, if it falls below a preset lower limit or decreases by more than a preset percentage relative to the individual's historical mean, the risk score increases with the magnitude of the decrease. For heart rate variability, if it falls below the individual's baseline level for multiple consecutive time windows, the stress-related risk score increases.

[0041] After de-identifying the open text data 207, the text semantic feature extraction module 105 extracts restricted text semantic features 303 using dictionary rules, emotion classification models, or semantic embedding models. The restricted text semantic features 303 include negative emotion intensity, stress semantic intensity, helplessness semantic intensity, sleep disturbance semantic intensity, avoidance expression intensity, and positive support expression intensity.

[0042] In one implementation, the text semantic feature extraction module 105 segments the open text data 207 into several candidate text fragments; it performs semantic classification on each candidate text fragment to identify its psychosemantic category; it masks content containing identity information such as name, contact information, address, and company name; and it retains only restricted text evidence fragments 702 that are no longer than a preset number of characters and are directly related to risk interpretation. These restricted text evidence fragments 702 are not used to freely generate conclusions but are instead used as citationable evidence fields in the evidence package 705.

[0043] For example, when open-ended text data contains the phrase "I've been feeling very tired lately, and I can't sleep at night," the system can generate the following restricted text semantic features: fatigue semantic intensity of 0.82, sleep disturbance semantic intensity of 0.76, and negative emotion intensity of 0.64; simultaneously, it generates a restricted text evidence fragment: "I've been feeling very tired lately, and I can't sleep at night." This fragment, after being desensitized and length-limited, is entered into evidence package 705, instead of directly inputting the entire diary or message into the psychological assessment model 803.

[0044] S400 performs time decay smoothing on the basic features to generate smoothed time window features.

[0045] like Figure 4 As shown, the time decay processing module 106 operates on the feature layer, not the final report layer. That is, the system first performs time decay smoothing on the basic features within each time window to obtain smoothed time window features 402, and then calculates the risk scores 504 for each dimension based on the smoothed features.

[0046] For any fundamental feature In the The original values ​​within each time window are denoted as follows: The smoothed eigenvalues ​​are denoted as ,but:

[0047] in, This is the time update coefficient, with a value ranging from 0 to 1; Indicates the first The first basic feature in the The original values ​​within a time window; Indicates the first The smoothed values ​​of the basic features within the previous time window; Indicates the first The smoothed values ​​of the basic features within the current time window. This applies when the data quality in the current time window is high, the data collection is continuous, and the recent changes are significant. Take the larger value; when there are many missing data points, high noise levels, or discontinuous data collection, Take the smaller value.

[0048] In one implementation, It can be represented as:

[0049] in, Based on the updated coefficients; Representation of features In the Data quality within each time window; Representation of features In the The magnitude of recent changes within each time window; Representation of features In the Noise level for each time window; , , These are preset coefficients; This represents the preset lower limit of the time update coefficient; This indicates the preset upper limit of the time update coefficient; This is a truncation function used to restrict input values ​​to a preset range; that is, when the input value is less than a certain value... Time output When the input value is greater than Time output When the input value is located to The input value is output in between.

[0050] Therefore, the level of action and input-output relationship of the time decay factor are both limited to the basic feature layer, avoiding a disconnect between the subsequent risk scoring formula and the time smoothing formula.

[0051] S500 calculates the dimensional risk score based on the smoothed time window features.

[0052] The dimension risk calculation module 107 maps the smoothed time window feature 402 to the emotion dimension 404, behavior dimension 405, cognition dimension 406 and social adaptation dimension 407, and calculates the dimension risk score 504 respectively.

[0053] Specifically, the emotion dimension 404 can be calculated from features such as scale emotion score, negative emotion semantic intensity, degree of positive emotion deficiency, and emotion fluctuation amplitude; the behavior dimension 405 can be calculated from features such as sleep rhythm deviation, activity frequency changes, number of test interruptions, and stability of answering behavior; the cognitive dimension 406 can be calculated from features such as answering reaction time, number of revisions, hesitation duration, attention-related questionnaire score, and cognitive stress semantic intensity; and the social adaptation dimension 407 can be calculated from features such as social avoidance expression, interaction frequency changes, participation in group activities, and environmental adaptation labels.

[0054] In one implementation, the risk score for each dimension is calculated using a linear weighting function:

[0055] in, Indicates the first Within the first time window Risk scores in each dimension; Represents the smoothed first... One basic feature; Indicates the first The first basic feature in the Feature weights in each dimension; This indicates a summation operation; the sum of the feature weights under the same dimension is 1.

[0056] In another implementation, dimensional risk scores can also be calculated using a logistic regression model, a gradient boosting tree model, a support vector machine model, or a neural network model. In this case, the input is the smoothed time window feature 402, and the output is the risk score of 0 to 100 for each dimension. Model training samples can come from historical assessment data, manual review results, and subsequent follow-up results.

[0057] For example, behavioral dimension risk score The risk score can be calculated by weighting the risk values ​​for sleep duration, sleep onset time deviation, nighttime awakening, assessment interruption, and decreased interaction frequency. If the above indicators are 70, 65, 60, 40, and 55, with corresponding weights of 0.30, 0.25, 0.15, 0.10, and 0.20, respectively, then the behavioral dimension risk score is:

[0058] Thus, heterogeneous raw data such as heart rate variability, click frequency, and pause duration in answering questions are all transformed into calculable scalar risk scores through explicit normalization, smoothing, and weighting processes.

[0059] S600 calculates the subjective risk value and the objective risk value respectively.

[0060] like Figure 5 As shown, the subjective-objective consistency calculation module 108 does not separate the subjective risk value and the objective risk value from the already integrated dynamic psychological profile 408, but instead calculates them based on the subjective data link and the objective data link respectively.

[0061] Subjective risk value 501 is denoted as Its inputs include the smoothed subjective features corresponding to questionnaire assessment data 201 and subjective self-assessment text 202. The objective risk value 502 is denoted as... The input includes smoothed objective features corresponding to behavioral interaction data 203, physiological state data 204, sleep rhythm data 205, and environmental scene data 206. The restricted text semantic features 303 extracted from the open text data 207 can be configured as subjective auxiliary features based on their source attributes, or they can be used solely as evidence interpretation features and not participate in the basic difference calculation of the subjective-objective consistency factor. Specific configurations can be preset by the system according to the application scenario.

[0062] In one implementation, subjective risk value Calculated based on subjective feature vector 301:

[0063] in, Indicates the first Subjective risk value within a time window; Indicates the first One subjective risk indicator; Indicates the first The weight of each subjective risk indicator; This indicates a summation operation; the sum of the weights of each subjective risk indicator is 1.

[0064] Objective risk value Calculated based on objective feature vector 302:

[0065] in, Indicates the first Objective risk values ​​within a time window; Indicates the first One objective risk indicator; Indicates the first The weight of each objective risk indicator; This indicates a summation operation; the sum of the weights of each objective risk indicator is 1.

[0066] Subjective risk index weights and objective risk indicator weights It can be determined based on scale rules, expert experience, training sample fitting results, or manual review results.

[0067] S700 calculates the subjective-objective consistency factor based on subjective risk value and objective risk value.

[0068] The subjective-objective consistency factor 503 is used to represent the degree of consistency between the subjective expression results and the objective behavioral state of the tested object. Its calculation formula is as follows:

[0069] in, Indicates the first The subjective-objective consistency factor within a time window is 503; This indicates a subjective risk value of 501. This indicates an objective risk value of 502. This indicates the preset maximum risk difference; This represents the absolute value operation. When... and When all scores are in the range of 0 to 100, You can take 100.

[0070] when When the value is close to 1, it indicates that the subjective risk value is close to the objective risk value, and the consistency between subjective and objective risk is relatively high; when... When the value is close to 0, it indicates that there is a large difference between the subjective risk value and the objective risk value, and the system needs to reduce the certainty of automatic judgment or trigger manual review.

[0071] In one implementation, the system further calculates the trend consistency factor. If recently If both subjective and objective risk values ​​show an upward trend within a given time window, the trend consistency factor is high; if the two values ​​change in opposite directions, the trend consistency factor is low. (Comprehensive consistency factor) It can be represented as:

[0072] in, Indicates the first The overall consistency factor within a time window; Indicates the consistency factor between subjective and objective factors; Indicates the trend consistency factor; The preset weight has a value range of 0 to 1.

[0073] Through the above design, the present invention retains independent subjective risk value 501 and objective risk value 502 before calculating dynamic psychological profile 408, avoiding the logical conflict of "integration first, separation later" and also avoiding the repeated calculation of objective behavioral characteristics.

[0074] S800 calculates dynamic weights based on data quality, recent changes, historical contributions, and manual review results.

[0075] like Figure 6 As shown, the dynamic weight calculation module 109 calculates the data quality factor 601, the recent change factor 602, the historical contribution factor 603, and the manual review and correction factor 604 respectively, and obtains the normalized dynamic weight 605 based on the above factors.

[0076] For the The dimension, in the first Data quality factor within a time window It can be calculated based on data completeness, acquisition continuity, noise level, and equipment effectiveness:

[0077] in, Indicates the first The dimension in the first Data quality factor within a time window; Indicates data integrity; Indicates continuous data collection; Indicates noise level; Indicates equipment availability; , , , As preset weights, and .

[0078] Recent change factor It can be calculated based on the degree of deviation of the current dimension risk score from the historical baseline:

[0079] in, Indicates the first The dimension in the first Recent change factors within a time window; This indicates the current risk score for that dimension. Indicates the first The dimension in the first The time window to the A set of historical risk scores within a time window; This is an arithmetic mean function used to calculate the average of a set of historical risk scores; Indicates the first Historical standard deviation or preset standardization coefficient for each dimension; This is a minimum value function used to output the smaller value from two input values; This represents absolute value operations.

[0080] Historical contribution factor The contribution factor can be determined based on the degree to which this dimension contributes to the final risk assessment in historical samples. In one implementation, for systems that use machine learning models to calculate risk scores, the historical contribution factor can be calculated using feature importance, model coefficients, or SHAP contribution values; for systems that use rule-based models, the historical contribution factor can be calculated based on the proportion of samples manually reviewed where this dimension is confirmed to be valid.

[0081] Manual review correction factor The correction factor is determined based on the results of manual review. For example, if manual review repeatedly confirms that the risk warning for a certain dimension is accurate, the system increases the correction factor for that dimension; if manual review repeatedly confirms that a certain dimension contains false alarms, the system decreases the correction factor for that dimension.

[0082] No. Unnormalized weight scores for each dimension It can be represented as:

[0083] in, Indicates the first The dimension in the first Unnormalized weighted scores within each time window; , , , The preset coefficients, and .

[0084] To ensure that the sum of dynamic weights is 1, the system uses a normalization function to calculate the first... Dynamic weights of each dimension :

[0085] in, Indicates the first The dimension in the first Dynamic weights within a time window; This represents the sum of the unnormalized weight scores for all dimensions within the same time window.

[0086] Alternatively, the Softmax function can be used for calculation:

[0087] in, Represented by natural constant The base is an exponential function; the Softmax function is used to convert multiple unnormalized weight scores into normalized weights with positive values ​​and a sum of 1.

[0088] Therefore, dynamic weights are no longer qualitative descriptions, but computable variables driven by data quality, recent changes, historical contributions, and manual review results.

[0089] S900 calculates the comprehensive basic risk score and the corrected comprehensive risk score.

[0090] The comprehensive risk scoring module 110 calculates a comprehensive basic risk score of 505 based on the risk scores of each dimension (504) and the normalized dynamic weight (605).

[0091] in, Indicates the first The comprehensive basic risk score within the time window is 505; This indicates the risk score in the emotional dimension; This indicates the risk score in the behavioral dimension; This indicates the risk score in the cognitive dimension; This indicates the risk score for the social adaptation dimension; , , , These represent the dynamic weights of the corresponding dimensions, and the sum of the four is 1.

[0092] The comprehensive risk scoring module 110 is based on the comprehensive consistency factor. The overall basic risk score of 505 is corrected. Specifically, when the subjective risk value is low but the objective risk value is high, the system believes that there may be hidden risks or underestimated risks, and the overall risk score of 506 is appropriately increased after correction; when the subjective risk value is high but the objective risk value is low, the system retains the risk warning, but reduces the certainty of the conclusion and suggests retesting or manual consultation.

[0093] In one implementation, the corrected comprehensive risk score It can be represented as:

[0094] in, Indicates the first The corrected comprehensive risk score within the time window is 506; This represents the overall basic risk score; Indicates the correction factor; Indicates the objective risk value; Indicates the subjective risk value; This indicates the preset maximum risk difference; This is a maximum value function used to output the maximum value from multiple input values. This formula is used to handle situations where the objective risk is higher than the subjective risk.

[0095] In another implementation, if the subjective risk value is significantly higher than the objective risk value, the system does not directly lower the overall risk score. Instead, it generates a prompt stating "strong subjective distress but insufficient objective behavioral evidence" and triggers a retest or manual review.

[0096] in, To preset the difference threshold, Marked for manual review.

[0097] The comprehensive risk scoring module 110 determines the risk level 507 based on the corrected comprehensive risk score 506. For example, when When it is less than 30, it is considered low risk; when When the value is between 30 and 50, it is judged as low to medium risk; when When the value is between 50 and 70, it is considered a medium risk; when A score of 70 to 85 indicates a medium-to-high risk level; when A value greater than 85 is considered high-risk. These thresholds can be adjusted based on the application scenario, population norms, and manual review results.

[0098] S1000 constructs dynamic psychological profiles based on dimensional risk scores, dynamic weights, and time trends.

[0099] The dynamic psychological profile construction module 111 constructs a dynamic psychological profile based on the emotional dimension 404, behavioral dimension 405, cognitive dimension 406, social adaptation dimension 407, risk scores of each dimension 504, normalized dynamic weights 605, time trend and subjective-objective consistency factor 503.

[0100] Dynamic psychological profile 408 can be represented as:

[0101] in, Indicates the first Dynamic psychological profile within a time window; Indicates the emotional dimension state; Indicates the state of the behavioral dimension; Indicates the state of cognitive dimensions; Indicates the state of social adaptation; Represents a dynamic set of weights; Indicates the overall consistency factor; This represents the adjusted overall risk score.

[0102] Unlike existing methods that generate static assessment conclusions based on only a single questionnaire, the dynamic psychological profile 408 of this invention updates over a time window and can reflect the direction of risk changes in the tested subject. For example, the system can record whether a tested subject's risk scores in the emotional dimension, behavioral dimension, and sleep rhythm have been continuously rising over the past four weeks, in order to determine whether there is a trend of gradual accumulation of psychological risk.

[0103] S1100 generates an evidence package based on the risk scoring process.

[0104] like Figure 7 As shown, the evidence package generation module 112 generates evidence package 705 based on subjective risk value 501, objective risk value 502, subjective-objective consistency factor 503, dimensional risk score 504, dynamic weight 605 and restricted text semantic features 303.

[0105] Evidence package 705 includes structured evidence fields 701, restricted text evidence fragments 702, evidence number 703, and evidence confidence level 704. Each piece of evidence has a unique evidence number and records the evidence source, evidence type, evidence value, time window, confidence level, and citationable content.

[0106] In one implementation, evidence package 705 is represented in JSON format. In evidence package 705, the open-ended text data is not entered into the large model inference module 113 in its original full-text form. Instead, it is provided as restricted text evidence fragments 702 after text semantic feature extraction, anonymization, length limitation, and evidence number binding. This preserves the text semantic context while preventing the large model from directly accessing a large amount of irrelevant or sensitive original text.

[0107] S1200 uses evidence packages and prompt word templates to call up a large psychological assessment model and generate structured output results.

[0108] like Figure 8 As shown, the large model reasoning module 113 receives the evidence package 705 and the prompt template 801. The prompt template 801 includes role constraints, task constraints, evidence citation constraints, output format constraints, and security constraints.

[0109] In one implementation, the prompt word template 801 includes the following: "You are the mental health auxiliary assessment report generation module. You can only generate explanations based on the input evidence_list and must not use any information not provided. Each risk explanation must cite at least one evidence_id. Medical diagnostic conclusions must not be output. For parts with insufficient evidence, output 'Insufficient evidence' or 'Retest / Manual review recommended'. The output must be in JSON format and include the fields summary, risk_factors, protective_factors, suggestions, and review_flag. Each item in risk_factors must contain description and evidence_ids fields." The large model reasoning module 113 requires the large psychological assessment model 803 to output structured results according to a preset JSON structure 804. For example: json {"summary":"This assessment shows that the subjects have a moderate level of psychological stress risk, mainly manifested as increased subjective stress, insufficient sleep, and increased expression of fatigue."} "risk_factors":[ {"description":"Recent stress levels have increased","evidence_ids":["E001"]}, {"description":"Sleep rhythm shows a downward trend","evidence_ids":["E002"]}, {"description":"Open-ended expressions containing content related to fatigue and sleep problems","evidence_ids":["E003"]} ], "protective_factors":[], "suggestions": ["It is recommended to retest in one week", "It is recommended that a psychologist or professional conduct a further interview"] "review_flag":true} Here, summary represents the assessment summary, risk_factors represents the set of risk factors, protective_factors represents the set of protective factors, suggestions represents the set of suggestions, review_flag represents the flag indicating whether manual review is required, and evidence_ids represents the set of evidence numbers cited in the corresponding risk interpretation.

[0110] S1300 performs evidence binding verification on the structured output results of the large psychological assessment model.

[0111] like Figure 9 As shown, the evidence binding verification module 120 is used to perform post-verification on the structured output result 804 after the psychological assessment model 803 generates the structured output result 804 and before the assessment report generation module 114 generates the formal assessment report. This post-verification is used to prevent the psychological assessment model 803 from generating risk interpretations not supported by the evidence package 705, citing non-existent evidence numbers, outputting content beyond the authorized scope, or outputting diagnostic conclusions.

[0112] The evidence binding verification module 120 includes an output format verification unit 901, an evidence number legality verification unit 902, an evidence coverage verification unit 903, an unauthorized content verification unit 904, and a downgrade processing unit 907.

[0113] S1310, the output format verification unit 901 verifies the format of the structured output result 804.

[0114] Specifically, the output format validation unit 901 checks whether the structured output result 804 conforms to the preset JSON structure and whether it contains the fields summary, risk_factors, protective_factors, suggestions, and review_flag; it further checks whether each item in risk_factors contains the description field and the evidence_ids field. If the structured output result 804 is not a valid JSON format or is missing necessary fields, a validation failure result 906 is generated.

[0115] S1320, the evidence number legality verification unit 902 performs legality verification on the evidence number in the structured output result 804.

[0116] Specifically, the evidence number validity verification unit 902 reads all valid evidence numbers from the evidence package 705 to form a set of valid evidence numbers. Simultaneously, all cited evidence numbers are read from the structured output result 804 to form a set of cited evidence numbers. Subsequently, a judgment was made. Is it subset of:

[0117] in, This represents the set of evidence numbers cited in the structured output result 804; This represents the set of legally existing evidence numbers in evidence package 705; This indicates a subset relationship. If... There are those not included in The evidence number, for example, if the large model outputs evidence number E999, but E999 does not exist in evidence package 705, then a verification failure result 906 is generated.

[0118] S1330, Evidence Coverage Verification Unit 903 verifies the coverage relationship between risk interpretation and evidence number.

[0119] Specifically, the evidence coverage verification unit 903 checks whether each risk interpretation in risk_factors is bound to at least one valid evidence number. If a risk interpretation is not bound to an evidence number, or the evidence number is empty, the risk interpretation is determined to lack evidentiary support, and a verification failure result 906 is generated.

[0120] In one implementation, the evidence coverage verification unit 903 can also match keywords in the risk interpretation with evidence types, field names, or descriptions in the evidence package 705. For example, when the risk interpretation is "recent decline in sleep rhythm," the system checks whether the associated evidence originates from sleep rhythm data 205 or expressions of sleep disturbance in open text data 207. If the risk interpretation content does not clearly match the associated evidence type, a verification failure result 906 is generated or the confidence level of the interpretation is reduced.

[0121] S1340, the unauthorized content verification unit 904 performs unauthorized content verification on the structured output result 804.

[0122] Specifically, the unauthorized content verification unit 904 checks whether the structured output result 804 contains medical diagnostic conclusions, unauthorized personal identification information, factual descriptions not appearing in the evidence package 705, absolute risk judgments, or statements with discriminatory or stigmatizing tendencies. If any of the above content is found, a verification failure result 906 is generated.

[0123] For example, if the psychological assessment model 803 outputs "the subject has depression" or "the subject has clear self-harm behavior", but the evidence package 705 only contains general stress and decreased sleep evidence, then the overreach content verification unit 904 determines that the output exceeds the scope of evidence support and generates a verification failure result 906.

[0124] S1350, perform report generation or downgrade processing based on the verification results.

[0125] When the output format verification, evidence number legality verification, evidence coverage verification, and unauthorized content verification all pass, the evidence binding verification module 120 generates a verification pass result 905 and sends the structured output result 804 to the evaluation report generation module 114.

[0126] When any verification fails, the evidence binding verification module 120 generates a verification failure result 906, and the degradation processing unit 907 performs at least one of the following processes: First, we refuse to generate a formal evaluation report based on the current structured output result (804 error). Second, delete or block risk explanations that failed verification, and only retain content that passed verification; Third, re-launch the psychological assessment model 803 and add the reason for the failure to the prompt word template 801; Fourth, transfer the assessment task to the manual review module 115; Fifth, generate a prompt on the user or management end that says "The model interpretation failed the evidence verification and manual review is recommended".

[0127] Through the aforementioned evidence binding and verification mechanism, this invention does not rely solely on prompt words to constrain the large psychological assessment model 803. Instead, after the large model outputs, software verification logic performs structured constraints and evidence consistency reviews on the output results, ensuring that every risk interpretation generated by the large model can be traced back to a legitimate evidence number in evidence package 705. This mechanism effectively reduces the risks of large model illusions, unauthorized inferences, and interpretations without evidence, improving the reliability, verifiability, and engineering controllability of psychological assessment reports.

[0128] S1400 generates traceable psychological assessment reports.

[0129] The assessment report generation module 114 generates a traceable psychological assessment report 806 based on the corrected comprehensive risk score 506, risk level 507, dynamic psychological profile 408, evidence package 705, structured output result 804, and the verification pass result 905 generated by the evidence binding verification module 120.

[0130] The traceable psychological assessment report 806 includes basic assessment information, comprehensive risk level, risk analysis by dimension, consistency analysis of subjective and objective factors, main risk evidence, explanation of trend changes, recommended measures, explanation of uncertainty, and prompts for manual review.

[0131] Each major risk explanation in the report is accompanied by a corresponding evidence number. For example, the report shows "Recently increased stress levels" and is associated with evidence number E001; it shows "A downward trend in sleep rhythm" and is associated with evidence number E002; it shows "Fatigue and sleep disturbances appear in open-ended expressions" and is associated with evidence number E003. When managers or professionals view the report through management terminal 119, they can click on the evidence number to view the corresponding evidence source, time window, data value, and confidence level.

[0132] In one implementation, if the evidence binding verification module 120 generates a verification failure result 906, the evaluation report generation module 114 will not generate a formal evaluation report, or will only generate a basic risk warning report that does not contain a large model natural language interpretation, and will push the evaluation result to the manual review module 115.

[0133] S1500 performs manual review of abnormal results and updates system parameters based on the review results.

[0134] The manual review module 115 is used to receive the following evaluation results: results with a risk level of medium-high risk or high risk; results with a subjective-objective consistency factor of 503 lower than the preset threshold; results with an objective risk value that is significantly higher than the subjective risk value; results whose large model output fails the evidence binding verification; and results with data quality lower than the preset threshold but with a high risk score.

[0135] Professionals can view dynamic psychological profiles (408), subjective and objective risk differences, evidence packages (705), structured output results (804), evidence binding verification results, and historical trends through management terminal (119), and provide review conclusions. Review conclusions may include risk confirmation, risk reduction, recommendations for retesting, recommendations for interviews, and recommendations for referrals.

[0136] The feedback update module 116 updates the manual review correction factor 604 in the dynamic weight calculation module 109 based on the manual review results, updates the feature weights in the dimensional risk calculation module 107, or updates the risk level threshold in the comprehensive risk scoring module 110.

[0137] For example, when manual review results repeatedly confirm that sleep rhythm data contributes significantly to the psychological risk assessment of a certain group, the system increases the weight of sleep rhythm-related features in behavioral dimension 405; when manual review results show that a certain type of text expression is prone to false alarms due to contextual differences, the system reduces the historical contribution factor of the semantic features of that text, or requires the large model inference module 113 to add an "uncertainty explanation" to the report.

[0138] In a school mental health screening scenario, students fill out a psychological assessment questionnaire via a user terminal (118) and authorize the system to collect assessment behavior data and sleep rhythm data from the past two weeks. The system first calculates a subjective risk value through a subjective data link. The score is 48, and the objective risk value is calculated through an objective data link. The score is 72. Because the objective risk value is significantly higher than the subjective risk value, and the consistency factor between objective and subjective risk is low, the system's judgment may contain subjective underestimation or concealment of risk.

[0139] Subsequently, the system applied time-decay smoothing to basic characteristics such as sleep duration, sleep onset time, pause duration during quizzes, and number of quiz modifications over the past two weeks, and calculated risk scores for emotional, behavioral, cognitive, and social adaptation dimensions. The behavioral dimension showed a high risk score, primarily evidenced by a decrease in average sleep duration, delayed nighttime sleep onset, and an increase in the number of assessment interruptions. The system further calculated dynamic weights based on data quality factors, recent change factors, historical contribution factors, and manual review and correction factors, resulting in a corrected overall risk score of 68.5, classifying it as medium risk.

[0140] For open-ended text data, the system does not directly input students' complete diaries or comments into the psychological assessment model 803. Instead, it extracts limited text evidence fragments that have undergone de-identification and length restrictions. For example, the system identifies the fragment "I've been feeling very tired lately, and I can't sleep at night" and writes it as evidence number E003 into evidence package 705. The psychological assessment model 803 can only generate interpretations based on evidence package 705, and each interpretation must cite the evidence number.

[0141] After the large psychological assessment model 803 outputs the structured output result 804, the evidence binding verification module 120 performs a post-verification. For example, if the risk interpretation output by the large model references evidence number E003, and E003 exists in evidence package 705, then the evidence number is valid; if the large model outputs evidence number E999, but E999 does not exist in evidence package 705, then the validity verification of the evidence number fails. As another example, if the large model outputs "the subject has depression," but there is no evidence in evidence package 705 to support this diagnostic conclusion, then the overreach content verification fails. For output results that fail verification, the system refuses to generate a formal assessment report or transfers the result to the manual review module 115.

[0142] Ultimately, the assessment report generation module 114 generates a traceable psychological assessment report 806 only after the evidence binding verification is passed. The report indicates that the student is in a medium-risk state, mainly based on elevated stress scale scores, decreased sleep rhythm, abnormal assessment behavior, and fatigue and sleep disturbances expressed in open-ended statements. Since the subjective risk value is lower than the objective risk value, the system also suggests retesting or further interviews by a psychological counselor.

[0143] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Any modifications, equivalent substitutions, combinations, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An artificial intelligence-based psychological assessment method, characterized in that, include: Acquire multi-source psychological data of the tested object, preprocess the multi-source psychological data, and distribute it to the subjective data link, objective data link and restricted text semantic link according to the data source attributes; Based on the subjective data of the subjective data link, a subjective feature vector is generated and a subjective risk value is calculated; based on the objective data of the objective data link, an objective feature vector is generated and an objective risk value is calculated; and based on the open text data of the restricted text semantic link, restricted text semantic features and restricted text evidence fragments are extracted. The basic features in the subjective feature vector, the objective feature vector, and the restricted text semantic features are smoothed by time decay to generate smoothed time window features, and the dimensional risk scores corresponding to multiple psychological dimensions are calculated based on the smoothed time window features. The subjective and objective risk values ​​are used to calculate a subjective-objective consistency factor to characterize the degree of subjective-objective matching, and the dynamic weights corresponding to each psychological dimension are calculated based on at least one of data quality, recent changes, historical contributions, and manual review results. A corrected comprehensive risk score is generated based on the dimensional risk score, the dynamic weight, and the subjective-objective consistency factor, and a dynamic psychological profile is constructed based on the dimensional risk score, the dynamic weight, the subjective-objective consistency factor, and the corrected comprehensive risk score. A structured evidence package is generated by encapsulating the subjective risk value, the objective risk value, the subjective-objective consistency factor, the dimensional risk score, the dynamic weight, and the restricted text evidence fragments. Based on the structured evidence package and prompt word template, the psychological assessment model is invoked to output a structured output result that conforms to the preset format and includes risk interpretation items and evidence number fields; The structured output results are subjected to evidence binding verification using post-processor software verification logic. The evidence binding verification includes verifying the subset inclusion relationship between the evidence number referenced in the structured output results and the structured evidence package. If the evidence binding verification passes, a traceable psychological assessment report is generated based on the structured output results and the structured evidence package.

2. The method according to claim 1, characterized in that, The specific steps of routing data according to its source attribute to the subjective data link, objective data link, and restricted text semantic link include: The questionnaire assessment data and subjective self-assessment texts in the multi-source psychological data are categorized into the subjective data link; One or more of the behavioral interaction data, physiological state data, sleep rhythm data, and environmental scene data from the multi-source psychological data are classified into the objective data link; The open text data in the multi-source psychologically relevant data are classified into the restricted text semantic link.

3. The method according to claim 1, characterized in that, Extracting restricted text semantic features and restricted text evidence fragments from open text data based on the aforementioned restricted text semantic links specifically includes: The open text data is de-identified and segmented to obtain multiple candidate text fragments; The candidate text fragments are semantically classified to obtain restricted text semantic features, including one or more of the following: negative emotion intensity, stress semantic intensity, helplessness semantic intensity, sleep disturbance semantic intensity, avoidance expression intensity, and positive support expression intensity. Select a target text fragment that matches a preset psychosemantic category from the candidate text fragments, and perform length trimming on the target text fragment to obtain the restricted text evidence fragment; The restricted text evidence fragments are input into the psychological assessment model as a citationable evidence field in the structured evidence package, instead of directly inputting the original full text of the open text data into the psychological assessment model.

4. The method according to claim 1, characterized in that, The process of performing time decay smoothing on the basic features to generate smoothed time window features specifically includes: For any fundamental feature Calculate its number of seconds according to the following formula. Smoothed feature values ​​within a time window: in, Indicates the first The first basic feature in the The original values ​​within a time window Indicates the first The smoothed values ​​of the basic features within the previous time window Indicates the first The smoothed values ​​of the basic features within the current time window This represents the time update coefficient.

5. The method according to claim 4, characterized in that: The time update coefficient Determined according to the following formula: in, Based on the updated coefficients, Indicates that the basic feature is in the first Data quality within a time window Indicates that the basic feature is in the first The recent changes within each time window, Indicates that the basic feature is in the first Noise level for each time window, , and For preset coefficients, This represents the preset lower limit of the time update coefficient. This indicates the preset upper limit of the time update coefficient. This is a truncation function used to restrict the input value to a specific range. to between.

6. The method according to claim 1, characterized in that: The multiple psychological dimensions include emotional dimension, behavioral dimension, cognitive dimension and social adaptation dimension; The step of calculating the dimensional risk scores corresponding to multiple psychological dimensions based on the smoothed time window features includes: Based on one or more of the following features in the smoothed time window: scale emotion score, negative emotion semantic intensity, degree of absence of positive emotion, and emotion fluctuation amplitude, calculate the risk score corresponding to the emotion dimension. Based on one or more of the following characteristics of the smoothed time window: sleep rhythm shift, activity frequency change, number of assessment interruptions, and stability of answering behavior, calculate the risk score corresponding to the behavioral dimension. Based on one or more of the smoothed time window features, such as response time, number of modifications, hesitation duration, attention-related questionnaire score, and cognitive stress semantic intensity, calculate the risk score corresponding to the cognitive dimension. Based on one or more of the following features in the smoothed time window: social avoidance expression, changes in interaction frequency, participation in group activities, and environmental adaptation labels, calculate the risk score corresponding to the social adaptation dimension.

7. The method according to claim 1, characterized in that, The calculation of the subjective-objective consistency factor based on the subjective risk value and the objective risk value specifically includes: Calculate the first according to the following formula. Subjective and objective consistency factors within a time window: in, This represents the subjective-objective consistency factor. This represents the subjective risk value. This represents the objective risk value. This indicates the preset maximum risk difference.

8. The method according to claim 1, characterized in that, The calculation of the dynamic weights corresponding to each psychological dimension based on at least one of data quality, recent changes, historical contributions, and manual review results specifically includes: Calculate the data quality factor, recent change factor, historical contribution factor, and manual review and correction factor for each psychological dimension. The unnormalized weight score is calculated by linear weighting the data quality factor, the recent change factor, the historical contribution factor, and the manual review and correction factor. The unnormalized weight scores are normalized or subjected to Softmax mapping to obtain the dynamic weights corresponding to each psychological dimension.

9. The method according to claim 8, characterized in that: The unnormalized weight score is calculated according to the following formula: in, Indicates the first The psychological dimension in the first Unnormalized weighted scores within each time window Indicates the data quality factor. Indicates the factor of recent change. Indicates historical contribution factor, This indicates a manual review correction factor. , , and These are preset coefficients; The dynamic weights are calculated according to the following formula: or, in, Indicates the first The psychological dimension in the first Dynamic weights within a time window This represents the sum of the unnormalized weighted scores of each psychological dimension within the same time window. Represented by natural constant An exponential function with base 0.

10. The method according to claim 1, characterized in that, The specific steps of performing evidence binding verification on the structured output result include: Verify whether the structured output results conform to the preset structured JSON or XML format; Read the set of legal evidence numbers from the structured evidence package, and read the set of cited evidence numbers from the structured output result, and determine whether the set of cited evidence numbers is a subset of the set of legal evidence numbers; Check whether each risk interpretation item in the structured output is bound to at least one valid evidence number; Based on preset sensitive word rules, diagnostic conclusion rules, or authorized scope rules, check whether the structured output results contain medical diagnostic conclusions, unauthorized personal identification information, factual descriptions not appearing in the structured evidence package, absolute risk judgments, or stigmatizing expressions. In cases where the evidence binding verification fails, at least one of the following downgraded actions will be taken: refusing to generate a formal assessment report, deleting the risk explanation items that failed the verification, re-calling the large psychological assessment model, transferring to manual review, or generating a verification failure message.

11. An artificial intelligence psychological assessment system, characterized in that, include: The data preprocessing module is used to preprocess multi-source psychological data and distribute it to the subjective data link, objective data link and restricted text semantic link according to the data source attributes. The feature extraction module is used to generate subjective feature vectors and calculate subjective risk values ​​based on subjective data from the subjective data link, generate objective feature vectors and calculate objective risk values ​​based on objective data from the objective data link, and extract restricted text semantic features and restricted text evidence fragments based on open text data from the restricted text semantic link. The processing module is used to perform time decay smoothing on the basic features in the subjective feature vector, the objective feature vector, and the restricted text semantic features to generate smoothed time window features; calculate the dimension risk scores corresponding to multiple psychological dimensions based on the smoothed time window features; calculate the subjective-objective consistency factor based on the subjective risk value and the objective risk value; and calculate the dynamic weights corresponding to each psychological dimension based on at least one of data quality, recent changes, historical contributions, and manual review results. The risk scoring and profile building module is used to generate a corrected comprehensive risk score based on the risk scores of the dimensions, the dynamic weights, and the subjective-objective consistency factors. A dynamic psychological profile is constructed based on the risk scores of the aforementioned dimensions, the dynamic weights, the subjective-objective consistency factors, and the corrected comprehensive risk scores. The evidence package generation module is used to generate a structured evidence package based on the subjective risk value, the objective risk value, the subjective-objective consistency factor, the dimensional risk score, the dynamic weight, and the restricted text evidence fragment; The large model reasoning module is used to call the psychological assessment large model based on the structured evidence package and prompt word template to generate structured output results containing risk interpretation items and evidence number fields.

12. The system according to claim 11, characterized in that, The evidence binding verification module includes: An output format verification unit is used to verify whether the structured output result conforms to a preset structured format. The evidence number validity verification unit is used to determine whether the evidence number referenced in the structured output result exists in the structured evidence package by using a set subset determination method; An evidence coverage verification unit is used to determine whether each risk interpretation item in the structured output is bound to at least one legitimate evidence number; The unauthorized content verification unit is used to determine whether the structured output results contain medical diagnostic conclusions, unauthorized personal identification information, factual descriptions not appearing in the structured evidence package, absolute risk judgments, or stigmatizing statements. The downgrade processing unit is used to perform downgrade processing when any of the following checks fail: the output format verification unit, the evidence number legality verification unit, the evidence coverage verification unit, and the unauthorized content verification unit.

13. The system according to claim 12, characterized in that, The evidence number validity verification unit is specifically used for: Read the set of legal evidence numbers from the structured evidence package; Read the set of cited evidence numbers from the structured output; Determine whether the set of cited evidence numbers is a subset of the set of legitimate evidence numbers; If there is an evidence number in the set of cited evidence numbers that is not included in the set of legal evidence numbers, a verification failure result will be generated.

14. The system according to claim 11, characterized in that, The system also includes; The evidence binding verification module is used to perform evidence binding verification on the structured output results after the structured output results are generated and before the traceable psychological assessment report is generated, so as to verify the subset inclusion relationship between the evidence number referenced in the structured output results and the structured evidence package. The assessment report generation module is used to generate a traceable psychological assessment report based on the structured output results and the structured evidence package, provided that the evidence binding verification is passed.

15. The system according to claim 11, characterized in that, The system also includes a manual review module and a feedback update module; The manual review module is used to receive evaluation results for risk levels reaching a preset level, subjective and objective consistency factors falling below a preset threshold, objective risk values ​​exceeding subjective risk values ​​by a preset difference, evidence binding verification failing, or data quality falling below a preset quality threshold. The feedback update module is used to update the manual review correction factor in the dynamic weight calculation module, the feature weight in the dimensional risk calculation module, or the risk level threshold in the comprehensive risk scoring module based on the manual review results.

Images