Intelligent evaluation system and method fusing behavior data and scale

Abstract

The invention discloses an intelligent evaluation system fusing behavior data and a scale, and the system comprises a data collection module which is used for collecting multi-dimensional behavior data; the data processing and feature extraction module is used for cleaning and preprocessing the data and extracting and calculating features; the analysis and evaluation module performs analysis and evaluation based on the behavior characteristics and scale scores; and the result output and explanation module is used for displaying the psychological state evaluation result generated previously to the user and explaining the evaluation result displayed by the user. According to the method, by collecting the multi-dimensional behavior data of the user in the answering process and combining the traditional psychological scale score, the limitation that the traditional psychological scale evaluation method excessively depends on self-reporting and lacks behavior data analysis and integration is overcome, more comprehensive, accurate and objective evaluation of the psychological state of the user is achieved, and powerful support is provided for the field of psychological health.

Description

Technical Field

[0001] The present invention belongs to the technical field of psychological assessment, and in particular relates to an intelligent assessment system and method integrating behavioral data and a scale. Background Art

[0002] Psychological scale assessment is a widely used tool in psychology and psychiatry to assess an individual's mental health, personality traits, and emotional state. These assessments play an important role in many fields, such as clinical diagnosis, psychological counseling, and personnel management. Traditional psychological scale assessment methods mainly rely on self-reports from subjects, that is, subjects score the scale questions based on their own circumstances. This self-report-based assessment method has the advantages of being easy to operate and easy to standardize, and has been widely used in practical applications. Common psychological scales include depression scales (such as the Beck Depression Inventory BDI), anxiety scales (such as the Generalized Anxiety Inventory GAD-7), and personality tests (such as the Big Five Personality Inventory). These scales attempt to capture the psychological state and characteristics of the subjects through a series of carefully designed questions.

[0003] However, as psychological research deepens, researchers have gradually come to realize the limitations of traditional self-report methods. These limitations include: 1. Social desirability effect: Subjects may give "correct" answers based on social expectations rather than their true feelings. 2. Self-perception bias: Subjects may lack accurate self-perception ability. 3. Emotional fluctuations: Subjects' current emotional state may affect their answers. 4. Deliberate disguise: In some cases, subjects may deliberately conceal or distort their true situation. 5. Cultural differences: Subjects from different cultural backgrounds may have different understandings and answer tendencies for the same question. 6. Language comprehension ability: The wording of the scale may affect the subject's understanding, thereby affecting the accuracy of the answer. 7. Recall bias: Questions involving past experiences may be affected by memory bias.

[0004] To overcome these limitations, psychologists and computer scientists have begun to explore new assessment methods. Among them, using behavioral data during the answering process to assist psychological assessment has become an emerging research direction. This method is based on the assumption that the behavioral patterns of subjects when answering questions may reflect their internal psychological state and cognitive processes. Behavioral data include but are not limited to: 1. Answering time: the length of time to answer each question. 2. Mouse trajectory: the path and pattern of mouse movement during the answering process. 3. Keyboard input characteristics: key speed, rhythm, etc. 4. Page browsing mode: how to switch between different questions, review answered questions, etc. 5. Answer modification behavior: the frequency and pattern of modifying answers. The advantage of these behavioral data is that they are objective and difficult to control deliberately, so they may reflect the true state of the subjects better than self-reports. At the same time, these data can be automatically collected without increasing the burden on the subjects, providing an additional information dimension for psychological assessment.

[0005] With the development of computer technology and artificial intelligence, it has become increasingly feasible to collect, process and analyze these behavioral data. Machine learning algorithms can extract meaningful patterns from massive amounts of behavioral data, which may provide a new perspective for the assessment of psychological states. However, how to effectively combine these behavioral data with traditional psychological scales to form a comprehensive and more accurate assessment system remains a problem to be solved. This is exactly the core issue that this invention intends to explore.

[0006] In summary, the existing psychological scale assessment technology has two key shortcomings: First, the traditional psychological scale assessment method relies too much on self-reporting and does not fully consider and utilize the behavioral data during the answering process. These behavioral data, such as answering time, mouse movement trajectory, keyboard click behavior, etc., may contain valuable psychological state information, but are often ignored in existing technologies. Secondly, there is currently a lack of a complete system to integrate traditional psychological scales with behavioral data analysis. Although some studies have begun to focus on behavioral data, these attempts are often fragmented and lack a systematic method to organically combine behavioral data with traditional scales, and cannot give full play to the complementary advantages of the two. Summary of the invention

[0007] The purpose of the present invention is to address the above shortcomings. The purpose of the present invention is to provide an intelligent assessment system and method that integrates behavioral data and scales; the system aims to: (1) comprehensively collect and analyze multi-dimensional behavioral data in the answering process, including but not limited to answering time, mouse trajectory, keyboard input characteristics, etc.; (2) establish a complete assessment framework, organically combine traditional psychological scales with behavioral data analysis, and achieve multi-angle and in-depth assessment of the subject's psychological state; (3) extract meaningful features from behavioral data through real-time data processing and machine learning algorithms, and improve traditional scale scores; (4) develop a comprehensive scoring system that integrates scale scores with behavioral data analysis results to provide more comprehensive and accurate psychological assessment results. By achieving these goals. The system overcomes the limitations of existing technologies, significantly improves the accuracy, objectivity and comprehensiveness of psychological assessments, and provides more powerful tool support for research and practice in the field of mental health.

[0008] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0009] In one aspect, the present invention provides an intelligent assessment system integrating behavioral data and a scale, the system comprising the following modules:

[0010] Data collection module, used for multi-dimensional behavioral data collection;

[0011] Data processing and feature extraction module, used for data cleaning and preprocessing as well as feature extraction and calculation;

[0012] Analysis and evaluation module, which conducts analysis and evaluation based on behavioral characteristics and scale scores;

[0013] The result output and explanation module is used to display the previously generated psychological state assessment results to the user and explain the assessment results displayed to the user.

[0014] Furthermore, the data collection includes scale collection, behavioral data collection, equipment and environment information collection, and audio and video recording of the answering process.

[0015] Furthermore, the data cleaning and preprocessing include comprehensive cleaning of the collected data, processing of missing values, duplicate values ​​and outliers, converting text data into vector form using natural language processing technology for subsequent analysis, and then randomly dividing the data into training sets and test sets to ensure the randomness and representativeness of the data distribution.

[0016] Furthermore, the features in the feature extraction include time features, mouse trajectory features, hesitation time features, rhythm features, modification behavior features, keyboard input features, page interaction features, device and environment features, and the feature calculation includes calculating the hesitation-modification index, cognitive load index and attention distraction index.

[0017] Furthermore, the analysis and evaluation includes: a) feature importance calculation based on random forest, b) nonlinear feature interaction calculation, c) comprehensive evaluation index, and d) personalized evaluation.

[0018] Furthermore, the result output and explanation include: a) quantitative scoring, b) classification results, c) visual display, and d) result explanation.

[0019] On the other hand, the present invention provides an intelligent assessment method integrating behavioral data and a scale, which is based on the above system and specifically includes the following steps:

[0020] Step 1: Multi-dimensional behavioral data collection;

[0021] Step 2: Extract and calculate the behavior data features;

[0022] Step 3: Develop comprehensive evaluation indicators based on behavioral characteristics and scale scores;

[0023] Step 4: Build a personalized evaluation model;

[0024] Step 5: Dynamic evaluation and feedback;

[0025] Step 6: Output the results.

[0026] Furthermore, extracting features in step 2 includes:

[0027] a) Time characteristics: calculate the total answering time, answering time for each question, dwell time between questions, hovering time on options, and further calculate the statistical indicators of these time data, including average answering time, answering time standard deviation, answering time coefficient of variation, answering time skewness, answering time kurtosis, and relative answering time;

[0028] b) Mouse trajectory characteristics: calculate the total trajectory length, trajectory complexity, trajectory smoothness, number of direction changes, and speed distribution;

[0029] c) Hesitation characteristics: calculate the hovering ratio, option switching frequency, decision time ratio, and trajectory overlap;

[0030] d) Rhythm characteristics: calculate answer rhythm consistency, answer speed trend, rhythm fluctuation, and rhythm interruption;

[0031] e) Modification behavior characteristics: calculate the overall modification rate, statistically analyze the modification time distribution, calculate the modification decisiveness, use the Markov chain model to calculate the transition probability matrix between options, and finally normalize it to obtain the transition probability matrix;

[0032] f) Keyboard input characteristics: calculate the input speed change, use the moving average method and standard deviation to analyze the stability of the input speed, calculate the deletion rate, count the pause pattern, and calculate the usage rate of special keys.

[0033] g) Page interaction features: Calculate the scrolling frequency and average scrolling amplitude, and use the N-gram model to analyze the question switching pattern.

[0034] Furthermore, the steps of analyzing the question switching pattern include: recording the question switching sequence, building the N-gram model, calculating the conditional probability, extracting the transition probability matrix, self-loop ratio, forward and backward jump rate, average jump distance features, identifying high-frequency and low-frequency switching patterns, and reviewing the intensity of behavior.

[0035] Furthermore, in step 3, the scale score and behavioral characteristics are used for comprehensive evaluation to output a comprehensive evaluation index CEI that is comprehensive and can reflect the user's psychological state. The index considers the linear and nonlinear interactions of each feature and provides the reliability and interpretability of the evaluation through weight and uncertainty estimation; wherein the comprehensive evaluation index is:

[0036] CEI = α × L + (1-α) × NL

[0037] Where α is the parameter that balances the linear and nonlinear components, L is the linear combination, and NL is the nonlinear combination.

[0038] Furthermore, in step 4, a personalized evaluation model is constructed based on the fused data; the specific steps include:

[0039] 1) Data fusion: The standardized scale scores obtained in step 3 are combined with the extracted behavioral features into a feature vector;

[0040] 2) User feature vector construction: construct a feature vector for each user u;

[0041] 3) Basic model selection: select K basic models M = {M1, M2, ..., MK};

[0042] 4) Personalized adjustment function: Define a personalized adjustment function g for each basic model Mi i (U u ), used to adjust the model output according to user characteristics;

[0043] 5) Personalized model construction: For each basic model Mi, build a personalized model Fi:

[0044] F i (X,U u )=M i (X)+g i (U u )

[0045] Among them, M i (X) is the output of the basic model, g i (U u ) is a personalized adjustment item; by combining the prediction result of the basic model with the personalized adjustment function, a personalized evaluation model for each user is obtained;

[0046] 6) Model training: Using historical data {(X j ,U j ,y j )} Train each personalized model F i , where y j is the true label;

[0047] 7) Ensemble learning: Use the stacking method to integrate multiple personalized models;

[0048] 8) Dynamic update mechanism: Design an incremental learning algorithm to regularly update the model based on newly collected data;

[0049] 9) Personalized threshold adjustment: Set a personalized evaluation threshold T for each user u u :

[0050] T u =T0+h(U u )

[0051] Where T0 is the basic threshold, h(U u ) is an adjustment function based on user characteristics.

[0052] Furthermore, in step 5, based on the personalized assessment model constructed in step 4, real-time monitoring, incremental learning, and feedback generation mechanisms are introduced on this basis to achieve a more dynamic and adaptive psychological state assessment.

[0053] The beneficial effects of the present invention are:

[0054] (1) Significantly improved evaluation accuracy

[0055] The present invention provides a more comprehensive evaluation basis by combining traditional scale scores with multi-dimensional behavioral data. In particular, the introduction of behavioral data captures potential information that is difficult to obtain through self-reporting, thereby significantly improving the accuracy of the evaluation.

[0056] (2) Reduce subjective bias and improve evaluation reliability

[0057] The objective behavior data collected by the present invention, such as answering time, mouse track, etc., are not easily controlled or disguised by the subjects. The introduction of such objective data effectively reduces the influence of subjective factors such as social expectation effect and improves the reliability of the evaluation results.

[0058] (3) Personalized and dynamic evaluation

[0059] The adaptive weight adjustment mechanism of the present invention can dynamically adjust the evaluation model according to the individual's historical data to achieve personalized evaluation. At the same time, the ability of real-time data collection and analysis enables the system to perform dynamic evaluation and capture changes in psychological state in a timely manner.

[0060] (4) Improving evaluation efficiency and accessibility

[0061] The present invention adopts an automated data collection and analysis process, which does not require the full participation of professionals, greatly improving the efficiency of the assessment. At the same time, the online characteristics and cross-platform compatibility of the system make remote assessment possible, improving the accessibility of mental health services. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 It is a schematic diagram of the system composition of the present invention. DETAILED DESCRIPTION

[0063] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0064] Embodiment 1

[0065] like Figure 1 , an intelligent assessment system that integrates behavioral data and scales. Its core is to achieve a more comprehensive and accurate psychological state assessment by collecting and analyzing the user's behavioral data during the answering process and combining it with traditional psychological scale scores.

[0066] The system mainly includes the following modules:

[0067] 1. Data acquisition module

[0068] a) Scale collection

[0069] It integrates a variety of traditional psychological scales, such as the PHQ-9 scale, the Beck Depression Inventory BDI, the Generalized Anxiety Scale GAD-7, the Eysenck Personality Questionnaire (EPQ), the Symptom Checklist-90 (SCL-90), etc., covering multiple aspects of assessment, including emotional state, personality traits, and comprehensive mental health status. According to different assessment purposes and user groups, the system can intelligently select the appropriate scale to present to the user for answering questions. For example, for users who are initially screening for depression tendencies, the PHQ-9 scale is preferred; for users who need a comprehensive understanding of their mental health status, the SCL-90 scale is selected.

[0070] b) Behavioral data collection

[0071] The Microsoft Clarity tool is used to collect mouse actions, including the mouse movement trajectory, click events (including coordinate position, timestamp, press duration and click type), and wheel usage during the user's answering process. For example, the coordinate sequence of the mouse movement trajectory is recorded at a sampling frequency of 60 times per second, which can accurately capture the user's operation path on the screen.

[0072] For keyboard input, record the user's keystroke speed (characters entered per minute CPM), key interval time, deletion and modification frequency (the ratio of the number of times the backspace and delete keys are used to the total number of characters entered) and the use and duration of special keys (such as Shift, Ctrl, Alt, etc.) in the questions that need to be entered.

[0073] Comprehensively collect data related to web page switching, accurately record page scrolling patterns, including the direction (up or down), amplitude (in pixels) and timestamp (accurate to milliseconds) of each scrolling event, and determine the user's reading speed and focus areas of the question content by analyzing the scrolling pattern. Count the frequency of question switching (the number of times the user switches between different questions divided by the total answering time, in times / minute). Fast switching may mean that the user is familiar with the question or lacks patience, while slow switching may mean that the user is thinking hard or encountering difficulties. Record the number and duration of reviewing answered questions in detail, and count the cumulative number of times the user returns to review the answered questions and the cumulative time spent on these questions (accurate to seconds), from which to analyze the user's confidence in the answer and the points of struggle on certain questions.

[0074] We collect data on answer modification behavior, accurately record the number of answer modifications (count the number of times users change their answers for each question), the final answer selection time (the exact time from the first answer selection to the final confirmation of the answer, in seconds), and use these data to understand the hesitation and changes of users in the decision-making process. We record the answer modification pattern in detail, including the timestamp of each modification (accurate to milliseconds) and the content of the options before and after the modification, establish a time series model of modification behavior, and analyze the user's thinking change path and decision-making adjustment process during the answering process.

[0075] c) Equipment and environment information collection

[0076] Collect device information through device sensors to obtain the screen resolution of the user's device (in pixels, such as 1920x1080, 2560x1440, etc.). Different screen resolutions may affect the user's perception of the display of the question content and the convenience of operation; accurately identify the device type (distinguish between desktops, laptops, tablets or mobile phones). The input methods and operating habits of different devices vary greatly. For example, mobile phone users may prefer touch operations, while desktop users are better at using keyboards and mice for precise operations; obtain the operating system type and version (such as Windows 10, MacOS Monterey, iOS15, Android 12, etc.). The characteristics and functions of the operating system may affect the user's interactive experience and performance during the answering process.

[0077] The device microphone is used to record the start time of the test (accurate to seconds) and the background noise level (in decibels dB), and the device light sensor is used to collect the light intensity (in lux). At the same time, during the test, the background noise level and light intensity are collected at fixed time intervals for subsequent calculation of the environmental stability index.

[0078] d) Audio and video recording of the answering process

[0079] During the user's answering process, the entire process is recorded through the device camera and microphone to record the user's facial expression changes (captured once per second) and voice information, providing multimodal data for subsequent emotional analysis.

[0080] 2. Data processing and feature extraction module

[0081] a) Data cleaning and preprocessing

[0082] The collected data is thoroughly cleaned, and missing values ​​are carefully handled (using mean filling, median filling or model-based filling methods, and appropriate strategies are selected according to the characteristics of the data), duplicate values ​​(deleting completely duplicate data records) and outliers (identifying and correcting or deleting abnormal data points by setting a reasonable threshold range or using statistical methods). Text data (such as answers to open-ended questions entered by users, text records related to the use of special keys, etc.) are converted into vector form using natural language processing technology, such as using bag-of-words models, TF-IDF methods or pre-trained word vector models (such as Word2Vec, GloVe, etc.) to map text to low-dimensional vector space for subsequent analysis. The data is then randomly divided into a training set (usually accounting for 70%-80%) and a test set (accounting for 20%-30%) to ensure the randomness and representativeness of the data distribution, and provide a reliable data foundation for model training and evaluation.

[0083] b) Feature extraction and calculation

[0084] -Time characteristics: including the calculation of total answering time, answering time for each question, dwell time between questions, hovering time on options, etc., and further calculation of statistical indicators such as average answering time, answering time standard deviation, coefficient of variation, skewness, kurtosis, and relative answering time to analyze the time pattern and rhythm of users' answers.

[0085] - Mouse trajectory features: Calculate the total length of the mouse trajectory, trajectory complexity (using discrete Fourier transform to calculate the energy and normalization of the first k frequency components), trajectory smoothness, number of direction changes, and speed distribution to capture the behavioral patterns and hesitation of the user's mouse operation.

[0086] - Hesitation time features: Calculate the hover ratio (the ratio of total hover time to total answering time), option switching frequency (the ratio of the number of switches between options to the total answering time), decision time ratio (the ratio of the decision time before the last choice to the total time for the question), and track overlap (the number of times the mouse track crosses itself), etc., to evaluate the user's hesitation and decision-making during the answering process.

[0087] - Rhythm features: Calculate answering rhythm consistency (use autocorrelation function to calculate the autocorrelation coefficient of lag k, and take ACF(1) as the consistency index), answering speed trend (use simple linear regression method to calculate the trend of answering time with question number), rhythm volatility (calculate the mean and variance of the difference in answering time for adjacent questions), and rhythm interruption (detect abnormally long answering time) to reflect the user's thinking process, consistency of decision-making speed, and possible fatigue or attention changes.

[0088] - Modification behavior characteristics: calculate the overall modification rate (the ratio of the number of question modifications to the total number of questions), modification time distribution (calculate the proportion of modification behaviors occurring in the 25%, 50%, 75% and 100% time periods before answering the questions), modification decisiveness (the ratio of the difference between the final submission time and the last modification time to the total answering time) and option switching mode (use the Markov chain model to calculate the transition probability matrix between options), etc., to analyze the user's modification behavior and thinking process during the answering process.

[0089] -Keyboard input characteristics: calculate the input speed change (moving average and standard deviation of the number of characters per minute), deletion rate (ratio of the number of deletion operations to the total number of characters), pause pattern (calculate the number of pauses exceeding the threshold τ and the cumulative time), and special key usage rate (ratio of the number of special key usages to the total number of key presses), etc., to understand the user's input habits and psychological state.

[0090] -Page interaction features: calculate scrolling frequency (ratio of scrolling events to total answering time), average scrolling amplitude (ratio of the sum of amplitudes of each scrolling to the number of scrolling events), question switching pattern (using N-gram model to analyze the pattern of question switching sequence, including transition probability matrix, self-loop ratio, forward jump rate, backward jump rate, average jump distance, etc., and identify high-frequency and low-frequency Bigram and Trigram and calculate the intensity of review behavior), etc., to quantify users' page interaction behavior and answering strategy.

[0091] -Device and environment characteristics: Encode the device type (e.g., desktop = 1, laptop = 2, tablet = 3, mobile phone = 4), standardize the screen size (ratio the width × height to a standard size such as 1920 × 1080), and calculate the coefficient of variation of background noise and light intensity (by collecting data, calculating the mean and standard deviation, and then obtaining the coefficient of variation, and combining it into a single environmental stability index, and evaluating environmental stability according to preset thresholds) to consider the impact of device and environmental factors on user answering behavior.

[0092] -Compound feature calculation: Calculate compound features such as the hesitation-modification index (combining hover ratio, modification rate, and decision time ratio), cognitive load index (combining time fluctuation, scrolling frequency, modification rate, and screen size), and attention distraction index (combining scrolling frequency, replay count, and deletion rate) to more comprehensively evaluate the user's psychological state.

[0093] 3. Evaluation model module

[0094] a) Feature importance calculation based on random forest: A random forest model is constructed, with standardized scale scores and behavioral characteristics as input, and specific evaluation indicators (such as depression, anxiety, etc.) as target variables. The mean impurity reduction method is used to calculate the importance score of each feature, and the weight wi of each feature is normalized.

[0095] b) Nonlinear feature interaction calculation: For each pair of features (Zi, Zj), their interaction term Iij = Zi × Zj is calculated, and a weight wij = (wi × wj) / 2 is assigned to each interaction term to capture the nonlinear relationship between features.

[0096] c) Comprehensive evaluation index: Calculate the linear and nonlinear combination of all features, the comprehensive evaluation index EI = α × L + (1-α) × NL, where α is a parameter that balances linearity and nonlinearity.

[0097] d) Personalized evaluation model: Combine the standardized scale scores and the extracted behavioral features into a feature vector X. For each user u, construct a feature vector Uu = [du, hu], where du is demographic information (such as age, gender, education level, etc.) and hu is a historical behavior pattern (such as past test results, behavioral feature statistics, etc.). Select K basic models (such as random forests, support vector machines, etc.), and define a personalized adjustment function gi(Uu) for each basic model to adjust the model output according to user characteristics. Construct a personalized model Fi(X,Uu) = Mi(X) + gi(Uu). Use historical data to train each personalized model.

[0098] 4. Result output and interpretation module

[0099] a) Quantitative scoring: Quantitative scores are given on multiple psychological dimensions (such as depression, anxiety, stress, etc.). Based on the user's specific psychological dimensions, the system outputs detailed numerical scores so that users or mental health professionals can clearly understand the specific evaluation results of each psychological state.

[0100] b) Classification results: According to the psychological state score, the user's emotional state is classified into "normal", "mild depression", "moderate depression", etc., providing users with more intuitive evaluation conclusions and facilitating quick understanding of their current psychological state.

[0101] c) Visual display: The assessment results of the psychological state are intuitively displayed through charts (such as bar charts, line charts, pie charts, etc.), graphs, etc., including the changing trend of the psychological state over time, the comparison of quantitative scores of various psychological dimensions, etc., to help users understand their own psychological changes more easily.

[0102] d) Result explanation: Provide a detailed explanation of the evaluation results displayed to the user, explain why the user received a certain score or classification for a certain psychological state, and provide specific suggestions. For example, if the system detects that the user may be in a state of mild depression, the system will explain which behavioral characteristics or scale scores contribute most to the evaluation results and provide corresponding suggestions (such as suggesting relaxation, seeking professional help, etc.).

[0103] Embodiment 2

[0104] An intelligent mental health assessment method integrating behavioral data and scales, including the following contents:

[0105] 1. Multi-dimensional behavioral data collection

[0106] When the user takes the psychological scale test, the system simultaneously collects the following behavioral data:

[0107] a) Time data:

[0108] -Total answering time: The total time from when the user starts answering the question to when they submit all answers, in seconds.

[0109] -Answering time for each question: The time a user spends on a single question, from the time the question is loaded to the time the user confirms the answer, in seconds.

[0110] -Time between questions: The time a user spends between two consecutive questions, from confirming the answer to the previous question to the completion of loading the next question, measured in seconds.

[0111] -Option hover time: The cumulative time the mouse hovers over each option, in milliseconds.

[0112] b) Mouse behavior data:

[0113] - Mouse movement trajectory coordinate sequence: records the X and Y coordinates of the mouse moving on the screen in pixels, with a sampling frequency of 60 times per second.

[0114] -Mouse movement speed: The speed at which the mouse moves between two consecutive sampling points, in pixels per second.

[0115] - Mouse movement acceleration: rate of change of mouse speed in pixels / second2.

[0116] - Click events: Records the coordinate position (pixels), timestamp (milliseconds), press duration (milliseconds) and click type (left click / right click / double click) of each mouse click.

[0117] - Scroll wheel usage: records the direction (up / down), amplitude (pixels) and timestamp (milliseconds) of each scroll wheel event.

[0118] c) Keyboard behavior data (applicable to questions that require input):

[0119] -Typing speed: the number of characters typed per minute (CPM).

[0120] -Key Interval: The time interval between two consecutive key presses, in milliseconds.

[0121] -Delete and modify frequency: The ratio of the number of times the backspace and delete keys are used to the total number of characters entered in an input field.

[0122] - Special key usage: records the usage and duration (in milliseconds) of special keys such as Shift, Ctrl, Alt, etc.

[0123] d) Page interaction data:

[0124] - Page scroll mode: records direction (up / down), amplitude (pixels) and timestamp (milliseconds) of each scroll event.

[0125] - Question switching frequency: The number of times a user switches between different questions divided by the total answering time, in times / minute.

[0126] -Number of times the user has reviewed answered questions: The number of times the user has reviewed answered questions.

[0127] - Duration of reviewing answered questions: The cumulative time that users stay on answered questions, in seconds.

[0128] e) Answer modification behavior:

[0129] -Answer modification count: The number of times the user changed their answer on each question.

[0130] -Final answer selection time: The time from the first answer selection to the final confirmation of the answer, measured in seconds.

[0131] -Answer modification mode: record the timestamp (in milliseconds) of each modification and the options before and after the modification.

[0132] f) Equipment information:

[0133] -Screen Resolution: The screen resolution of the user's device, in pixels (for example: 1920x1080).

[0134] -Device type: distinguish between desktop, laptop, tablet or mobile phone.

[0135] - Operating System: The operating system type and version of the user's device.

[0136] - Browser information: the type and version of browser used by the user.

[0137] g) Environmental data:

[0138] -Question answering start time: records the timestamp when the user starts answering questions, accurate to seconds.

[0139] - Background noise level: Uses the device microphone to record the ambient noise level in decibels (dB).

[0140] -Light intensity: Use the device light sensor to record the ambient light intensity in lux.

[0141] h) User status data:

[0142] -Facial expressions: Use a camera to record changes in facial expressions during the test, capturing them once per second.

[0143] -Eye movement data: Use an eye tracker to record the user's eye movement trajectory and gaze point, with a sampling frequency of 30 times per second.

[0144] By collecting these detailed and multi-dimensional behavioral data, this system can comprehensively capture various subtle behavioral characteristics of users during the psychological scale test, providing a rich source of information for subsequent data analysis and psychological state assessment.

[0145] 2. Behavioral data feature extraction and calculation

[0146] a) Time characteristics:

[0147] Accurately calculate the total answering time (the total time from the user starting to answer the question to submitting all the answers, in seconds), the answering time for each question (calculate the user's stay time for each question, from the time the question is loaded to the time the answer is confirmed, in seconds), the dwell time between questions (the precise dwell time between two consecutive questions, from the time the answer to the previous question is confirmed to the time the next question is loaded, in seconds), and the option hovering time (the cumulative time the mouse hovers over each option, in milliseconds). Further calculate the statistical indicators of these time data, such as the average answering time (the average answering time of all questions), the standard deviation of the answering time (reflecting the degree of dispersion of the answering time), the coefficient of variation (the ratio of the standard deviation to the mean, used to compare the degree of dispersion of different groups of data), skewness (describing the asymmetry of data distribution), kurtosis (measurement of the peak degree of data distribution) and relative answering time (the ratio of the answering time for each question to the difficulty coefficient of the question, the difficulty coefficient of the question can be determined in advance based on expert experience or large-scale sample testing), through which the time pattern and rhythm characteristics of the user's answering are deeply analyzed.

[0148] -Average answering time:

[0149]

[0150] Where ti is the time to answer each question, and n is the total number of questions.

[0151] -Standard deviation of answering time:

[0152]

[0153] -Coefficient of variation of answering time:

[0154]

[0155] -Answer time skewness:

[0156]

[0157] - Kurtosis of answering time:

[0158]

[0159] -Relative answering time:

[0160]

[0161] Where Di is the difficulty coefficient of the question (pre-defined).

[0162] b) Mouse track characteristics:

[0163] -Total length of track:

[0164]

[0165] Where ((x i ,y i )) is a point on the trajectory

[0166] -Trajectory complexity: Use discrete Fourier transform (DFT) to calculate the energy sum of the first k frequency components. Specifically: Represent the x and y coordinates of the mouse trajectory as time series x(t) and y(t), respectively. Make sure the number of data points is a power of 2 (such as 256, 512, etc.). If not, make it a power of 2 by interpolation or truncation. Perform discrete Fourier transform on x(t) and y(t), respectively:

[0167] [X(f)=DFT(x(t)),Y(f)=DFT(y(t))]

[0168] The calculation formula of DFT is:

[0169]

[0170] Where N is the sequence length (k = 0, 1, ..., N-1). Calculate the energy of each frequency component:

[0171] [E(f)=|X(f)| 2 +|Y(f)| 2 ]

[0172] Where |X(f)| and |Y(f)| are the moduli of X(f) and Y(f), respectively. Select the first k frequency components: Usually the low-frequency parts are selected because they contain the main features of the trajectory. The choice of k can be determined based on experience, and common choices are 8 or 16. Calculate the trajectory complexity:

[0173]

[0174] Among them, f i is the i-th frequency component. Normalization: In order to make trajectories of different lengths comparable, we usually normalize the complexity:

[0175]

[0176] Where N is the total number of points in the trajectory.

[0177] -Trajectory smoothness:

[0178]

[0179] Where L max Is the straight-line distance from the start point to the end point of the mouse track

[0180] -Number of direction changes: Count the number of times the trajectory tangent angle changes by more than a threshold value θ, where the threshold value θ is a pre-set value.

[0181] -Speed ​​distribution: Calculate the time proportion of different speed ranges (such as low, medium, and high).

[0182] c) Hesitation Characteristics:

[0183] Calculate the hover ratio (the ratio of total hover time to total answering time), option switching frequency (the ratio of the number of switches between options to the total answering time), decision time ratio (the ratio of the decision time before the last choice to the total time for the question), and track overlap (calculate the number of times the mouse track crosses itself, and use a complex algorithm to detect when the mouse movement path intersects with itself). These features comprehensively reflect the user's hesitation level and decision-making difficulty in the answering process.

[0184] -Hover Scale:

[0185]

[0186] , where Th is the total hovering time and Tt is the total answering time

[0187] -Option switching frequency:

[0188]

[0189] Where Ns is the number of switches between options

[0190] -Decision time ratio:

[0191]

[0192] Where Td is the decision time before the final choice, and Ti is the total time for the question.

[0193] -Trajectory overlap: Counts the number of times the mouse trajectory crosses itself, where trajectory self-crossing refers to the situation where the mouse movement path intersects with itself.

[0194] d) Rhythm characteristics:

[0195] It is mainly used to capture the time pattern of users when answering a series of questions, reflecting the user's thinking process, consistency of decision-making speed, and possible fatigue or attention changes. Consider there are n questions, and the answering time of each question is t1, t2, ..., t n .

[0196] - Answering rhythm consistency: Use the autocorrelation function (ACF) to calculate the autocorrelation coefficient of lag k (k is usually selected from 1 to 5)

[0197]

[0198] Consistency index: ACF(1) is taken as the consistency index. The closer the value is to 1, the more consistent the answering rhythm is.

[0199] -Answering speed trend: used to detect whether the user is getting faster or slower over time. Using simple linear regression method:

[0200] t i =β0+β1i+ε

[0201] Where i is the question number (1 to n), t i is the corresponding answering time; calculate β1 (slope):

[0202]

[0203] Trend indicator: If β1<0, it means the answering speed is increasing; if β1>0, it means the answering speed is slowing down; the absolute value of β1 indicates the degree of change.

[0204] - Rhythm volatility: calculate the difference in answering time for adjacent questions

[0205] Δt i =|t i+1 -t i |,

[0206] Then calculate Δt i The mean and variance of .

[0207] - Rhythm Breaks: Detects unusually long response times, which may indicate distraction or difficult questions.

[0208] e) Modify behavioral characteristics:

[0209] Calculate the overall modification rate (the ratio of the number of question modifications to the total number of questions), count the modification time distribution (calculate the proportion of modification behaviors occurring in the 25%, 50%, 75% and 100% time periods before answering the questions), calculate the modification decisiveness (the ratio of the difference between the last modification time and the final submission time to the total answering time), and use the Markov chain model to calculate the transition probability matrix between options (collect the user option switching sequence, construct the state transition matrix and count the number of transitions, and finally normalize to obtain the transition probability matrix). Through these features, we can deeply analyze the modification behavior patterns and decision-making processes of users in the process of answering questions. It is mainly used to capture the hesitation, uncertainty or reflection behavior of users when answering questions.

[0210] - Overall modification rate:

[0211]

[0212] , where N m is the number of times the question has been modified, N t Is the total number of questions.

[0213] -Revision time distribution: Calculate the proportion of revision behaviors that occur in the 25%, 50%, 75% and 100% time periods before answering the questions.

[0214] - Modification of decisiveness:

[0215]

[0216] Where T f is the final submission time, T m Is the last modification time.

[0217] -Option switching mode: Use the Markov chain model to calculate the transition probability matrix between options. Specifically, consider a question containing 4 options (A, B, C, D). The steps for calculating the transition probability matrix are as follows: First, collect the option switching sequence of the user when answering the question. For example: A->B->A->C->D; then, construct the state transition matrix to create a 4x4 matrix (because there are 4 options), and the initial values ​​are all 0. Count the number of transitions, traverse the collected option switching sequence, and count the number of each transition: A->B: 1 time B->A: 1 time A->C: 1 time C->D: 1 time. Fill the transition matrix according to the statistical results (the upper left corner is blank):

[0218]

[0219] Calculate the transition probability and normalize each row of the matrix to get the transition probability matrix (the upper left corner is blank):

[0220]

[0221] This Markov chain model can reflect the user's thinking process and hesitation pattern when choosing an answer. For example, a high self-loop probability (value on the diagonal) may indicate the user's preference or confidence in a certain option. Frequent switching back and forth (such as between A and B) may indicate that the user is hesitant between the two options. The absence of certain transition probabilities may indicate that the user believes that some options are irrelevant.

[0222] f) Keyboard input features (applicable to open-ended questions):

[0223] Calculate the change in input speed, use the moving average method (consider the window length k, calculate the moving average of the number of characters per minute CPM, the window size k can be selected according to the actual situation, such as 5, 10, etc.) and standard deviation to analyze the stability of input speed. Calculate the deletion rate (the ratio of the number of deletion operations to the total number of characters), count the pause pattern (calculate the number of pauses exceeding the threshold τ and the cumulative time, the threshold τ is determined according to the user's normal input rhythm), calculate the special key usage rate (the ratio of the number of special key usage to the total number of key presses), these features reflect the user's operating characteristics and psychological state when answering questions from the perspective of keyboard input.

[0224] - Input speed change: Calculate the moving average and standard deviation of the number of characters per minute (CPM). For the moving average of the input speed, consider the window length k, then the moving average is:

[0225]

[0226] Among them: SMA t It is the moving average at time point t. CPM t is the number of characters per minute at time point t. n is the number of periods of the moving average (window size).

[0227] - Deletion rate:

[0228]

[0229] Where N d is the number of delete operations, N c is the total number of characters.

[0230] - Pause mode: Count the number and cumulative time of pauses exceeding a threshold τ.

[0231] -Special Key Usage:

[0232]

[0233] Where N s is the number of times a special key is used, N kIt is the total number of key presses. Special keys include function keys (F1-F12), editing keys (Backspace, Delete, Insert, Home, End, Page Up, Page Down), control keys (Ctrl, Alt, Windows / Command key), modifier keys (Shift, Caps Lock), navigation keys (arrow keys), keys related to the numeric keypad, system control keys (Esc, Print Screen, Scroll Lock, Pause / Break) and combination keys (Tab, Enter / Return, Spacebar). This indicator can reflect the user's input habits, operation proficiency, and possible psychological state. For example, high frequency of Backspace use may indicate uncertainty or anxiety.

[0234] g) Page interaction features:

[0235] Calculate the scrolling frequency (the ratio of the number of scrolling events to the total answering time) and the average scrolling amplitude (the average of each scrolling amplitude), and use the N-gram model (such as 2-gram and 3-gram) to analyze the question switching pattern. The specific steps include recording the question switching sequence, building the N-gram model, calculating the conditional probability (such as P(qj|qi) and P(qk|qi,qj)), extracting features such as the transition probability matrix, self-loop ratio, forward and backward jump rate, and average jump distance, identifying high-frequency and low-frequency switching modes (setting a frequency threshold, above the threshold is high frequency, below the threshold is low frequency) and the intensity of review behavior (the ratio of the product of the number of replays and the total replay time to the total answering time). These features are helpful in analyzing the user's answering strategy, hesitation pattern, and attention distribution.

[0236] -Scrolling frequency:

[0237]

[0238] Where N s is the number of scroll events

[0239] -Average scrolling amplitude:

[0240]

[0241] where a i is the amplitude of each scroll

[0242] - Question switching pattern: Quantify the user's question switching behavior, capture the user's answering strategy, hesitation pattern, attention distribution and other information, and improve the accuracy and depth of psychological assessment. Specifically, the N-gram model is used to analyze the pattern of question switching sequence. The steps are as follows:

[0243] (1) Sequence recording: Record the user's question switching sequence during the test Q = {q1, q2, ..., qn}, where qi represents the question number of the i-th access.

[0244] (2) N-gram construction: Sequence Q constructs 2-gram (Bigram) and 3-gram (Trigram) models.

[0245] Bigram:B={(q_i,q_{i+1})|i=1,2,…,n-1}

[0246] Trigram:T={(q_i,q_{i+1},q_{i+2})|i=1,2,…,n-2}

[0247] (3) Probability calculation: Calculate the conditional probabilities P(qj|qi) and P(qk|qi,qj), where:

[0248]

[0249]

[0250] (4) Feature extraction:

[0251] (a) Transition probability matrix M, where

[0252] M ij =P(q j |q i )

[0253] (b) Self-circulation ratio

[0254]

[0255] N is the total number of questions;

[0256] (c) Forward jump rate

[0257] where j>i;

[0258] (d) Backward jump rate

[0259]

[0260] where j <i;

[0261] (e) Average jump distance

[0262]

[0263] (5) Pattern recognition:

[0264] (a) Identify the k most frequent Bigrams and Trigrams;

[0265] The purpose of identifying high-frequency Bigrams and Trigrams is to capture the common behavior patterns of users in the process of answering questions. By analyzing the high-frequency phrases in the question switching sequence, the user's answering strategy and cognitive preference can be inferred:

[0266] Answering fluency: A high-frequency question switching pattern may reflect the user's clear answering path and fluency of thinking, indicating that the user is more familiar with certain questions.

[0267] Focus: Through high-frequency switching sequences, we can locate the user's attention distribution on certain types of questions. If the switching frequency between certain types of questions is high, it may mean that these questions attract more attention from users.

[0268] Time distribution pattern: If certain questions switch frequently, it may mean that the user responds quickly to these questions. The system can combine behavior intensity analysis to determine whether the user shows high confidence on these questions.

[0269] (b) Identify the m Bigrams and Trigrams with the lowest frequency;

[0270] Low-frequency Bigram and Trigram capture the user's rare or abnormal switching patterns during the answering process. These patterns can help identify the user's hesitation and uncertainty, or detect certain special psychological states:

[0271] Hesitation and confusion: Low-frequency switching patterns may indicate that the user is confused or indecisive between these questions. These low-frequency sequences may reflect the user's cognitive load or decision-making difficulties.

[0272] Abnormal patterns: Low-frequency switching behaviors may be unexpected situations that occur during the user's answering process, which may be related to psychological states such as anxiety and distraction. The system can conduct further analysis based on these low-frequency phrases to determine whether the user exhibits atypical behavior patterns.

[0273] Focus on the topic: Some topics may induce less switching behavior due to their complexity or challenge, which can be used as a basis for evaluating the user's focus or engagement on a specific topic.

[0274] -Review Behavior Intensity:

[0275]

[0276] Where N r is the number of replays, T ris the total review time. The behavior intensity value can be used to evaluate the subject's interest and concentration in the content of a question during the answering process. It further quantifies the user's attention distribution and decision weight during the answering process. Questions or switching modes with high behavior intensity values ​​may reflect that the user has invested more time and energy in these parts. By combining these high-frequency or low-frequency N-gram patterns with behavior intensity values, the system can achieve a deeper assessment of the user's psychological state, thereby improving the accuracy and depth of psychological assessments.

[0277] (c) Equipment and environmental characteristics:

[0278] The collection and processing of device and environmental features indirectly affects the pattern recognition in the N-gram model by affecting the user's answering behavior. Factors such as device type, screen size, and environmental stability are important external variables that affect user behavior patterns. Pattern recognition algorithms need to consider these features to make a more comprehensive and accurate psychological state assessment. Different devices may cause changes in the user's answering behavior pattern; the screen size directly affects the way users view and operate questions; the coefficient of variation of background noise and light intensity is an important feature to measure the stability of the user's answering environment. These environmental factors may affect the user's answering behavior and directly affect the results of pattern recognition.

[0279] -Device type encoding: By encoding device types as numerical features, pattern recognition algorithms can better adjust the analysis and classification of behavior patterns and improve the accuracy of understanding user behavior under different devices. Convert device types to numerical features (such as desktop = 1, laptop = 2, tablet = 3, mobile phone = 4).

[0280] - Screen size standardization: After standardization, the screen sizes of all devices are mapped to a unified scale, which helps the pattern recognition algorithm to make fair comparisons between different devices and reduce the interference caused by device differences.

[0281]

[0282] -Environmental stability: By calculating the coefficient of variation of noise and light intensity, the pattern recognition algorithm can capture the difference in user behavior in stable and unstable environments. The specific steps for calculating the coefficient of variation of background noise and light intensity are as follows:

[0283] 1. Data collection: During the test, the background noise level N = {n1, n2, ..., nk} and light intensity L = {l1, l2, ..., lk} are collected at fixed time intervals Δt, where ni and li represent the noise level (unit: dB) and light intensity (unit: lux) at the i-th time point, respectively.

[0284] 2. Mean calculation: Calculate the arithmetic mean of the noise level and the light intensity:

[0285]

[0286]

[0287] 3. Standard deviation calculation: Calculate the sample standard deviation of the noise level and the light intensity:

[0288]

[0289]

[0290] 4. Coefficient of variation calculation: Calculate the coefficient of variation of the noise level and the light intensity:

[0291]

[0292]

[0293] 5. Environmental stability index: Combine the coefficients of variation of noise and light into a single environmental stability index:

[0294] ESI = w1 × CV N + w2 × CV L ;

[0295] where w1 and w2 are weight coefficients, and w1 + w2 = 1;

[0296] 6. Stability assessment: Evaluate the environmental stability according to the preset threshold T:

[0297] If ESI < T, the environment is considered stable;

[0298] If ESI ≥ T, the environment is considered unstable.

[0299] Through the above steps, the stability of the test environment can be quantified, providing an objective basis for the reliability of the psychological assessment results.

[0300] h) Composite features:

[0301] - Hesitation - modification index:

[0302]

[0303] Combined with the hover ratio, modification rate, and decision - time ratio

[0304] - Cognitive load index:

[0305]

[0306] Take into account time fluctuations, scrolling frequency, modification rate and screen size

[0307] -Distraction Index:

[0308]

[0309] Combine scroll frequency, replay count, and deletion rate.

[0310] 3. Comprehensive evaluation indicators based on behavioral characteristics and scale scores

[0311] Develop a new comprehensive evaluation index that can effectively integrate traditional psychological scale scores and extracted behavioral features to provide a more comprehensive and accurate psychological state assessment. Use scale scores and behavioral features for comprehensive evaluation, and output a comprehensive comprehensive evaluation index CEI that can reflect the user's psychological state. This index takes into account the linear and nonlinear interactions of each feature, and provides the reliability and interpretability of the evaluation through weight and uncertainty estimation.

[0312] 3.1 Data Preprocessing and Standardization

[0313] a) Standardization of scale scores:

[0314] The score S for each scale dimension i Perform z-score standardization:

[0315]

[0316] b) Behavioral characteristics standardization:

[0317] For each behavioral feature F j Perform z-score standardization:

[0318]

[0319] 3.2 Feature Importance Calculation

[0320] The importance of each feature is calculated using the random forest algorithm:

[0321] a) Construct a random forest model, using the standardized scale scores and the above-extracted multidimensional behavioral characteristics as input variables, and specific evaluation indicators (such as the classification or quantitative evaluation of psychological states such as depression and anxiety) as target variables. For example, if evaluating depression, the target variable can be the depression level classified according to clinical diagnostic criteria (such as no depression, mild depression, moderate depression, severe depression), or the depression quantitative score obtained through professional psychological assessment tools.

[0322] b) Use the Mean Decrease Impurity method to calculate the importance score of each feature in the model. In random forests, impurity is usually measured by Gini impurity or information entropy. For each decision tree, calculate the contribution of each feature to the reduction of impurity when the node is split, and then average the contributions of all decision trees to get the importance score of each feature. For example, for a data set containing multiple behavioral features and scale scores, after training the random forest model, it was found that the answer time feature made a greater contribution to reducing impurity, and its average impurity reduction value was higher, indicating that this feature plays an important role in distinguishing users with different psychological states.

[0323] c) Normalize the calculated feature importance scores so that the sum of the importance weights of all features is 1. The normalized weight wi represents the relative importance of each feature in the overall evaluation model and can be used in subsequent model construction and feature selection processes. For example, suppose that after normalization, the weight of mouse trajectory complexity is 0.15, the weight of answering speed trend is 0.12, etc. This means that when comprehensively evaluating the user's psychological state, mouse trajectory complexity has a higher importance in the model than other features such as answering speed trend.

[0324] 3.3 Nonlinear feature interactions: To capture the nonlinear relationship between features, second-order feature interactions are introduced to capture the additional information provided by the combination of two features. For example, taking a long time to answer questions (feature A) and frequently revising answers (feature B) may mean different things when viewed separately, but if they occur at the same time, they may more strongly suggest an anxious state. Specifically,

[0325] a) For each pair of features (Z i ,Z j ), and calculate their interaction terms:

[0326] I ij =Z i ×Z j

[0327] b) Assign weights to each interaction term:

[0328]

[0329] 3.4 Calculation of comprehensive evaluation index: Combine linear terms and nonlinear interaction terms to calculate comprehensive evaluation index:

[0330] a) Linear combination:

[0331] L=Σ(w i ×Z i ).

[0332] This step weights and sums each feature according to its importance to obtain a linear comprehensive value that reflects the user's psychological state. For example, for a data set containing multiple features such as answering time, mouse track complexity, and scale score, a linear combination is calculated based on their respective weights. If the answering time feature weight is 0.1, its standardized value is 0.5, the mouse track complexity weight is 0.08, and the standardized value is 0.6, etc., the value of the linear combination is obtained by weighted summation, which integrates the contribution of each feature to the psychological state at the linear level.

[0333] b) Nonlinear combination:

[0334] NL=∑(w ij ×I ij ).

[0335] The nonlinear combination takes into account the interaction between features, further enriching the evaluation model's ability to capture complex data relationships. For example, by combining the interaction term and weight of the above-mentioned answering time and mouse hover ratio, as well as the interaction terms of other feature pairs, the value of the nonlinear combination NL is calculated, which reflects the impact of the nonlinear relationship between features on the psychological state evaluation.

[0336] c) Comprehensive evaluation indicators:

[0337] CEI (ComprehensiveEvaluationIndex)=α×L+(1-α)×NL

[0338] Among them, α is a parameter that balances the linear and nonlinear components. By adjusting the value, the relative importance of the linear and nonlinear parts in the comprehensive evaluation can be flexibly determined according to the actual data and evaluation needs. For example, when α = 0.7, the model focuses more on the linear combination part, that is, it pays more attention to the independent contribution of each feature; when α = 0.4, the weight of the nonlinear combination part increases relatively, and more emphasis is placed on the impact of the interaction between features on the psychological state assessment. The comprehensive evaluation index EI combines linear and nonlinear feature information, and can more comprehensively and accurately reflect the user's psychological state.

[0339] 3.5 Adaptive Weight Adjustment

[0340] Calculate the performance stability of each feature in historical data. Performance stability can be calculated by a variety of methods, such as calculating the variance or coefficient of variation of feature values ​​at different time points or in different samples. Features with smaller variance or lower coefficient of variation are considered to have higher stability, indicating that their contribution to psychological state assessment in different situations is relatively consistent. For example, for the answering time feature, calculate its variance in multiple historical test data sets. If the variance is small, it means that the answering time feature has less variation in different user groups or different test scenarios and is more stable.

[0341] Introduce adaptive mechanisms to enable evaluation indicators to adapt to different individuals and situations:

[0342] a) Calculate the performance stability s_i of each feature in historical data.

[0343] b) Update feature weights regularly:

[0344] w i,new =β×w i +(1-β)×s i

[0345] Where β is the update rate, which controls the balance between new and old information.

[0346] 3.6 Uncertainty Estimation

[0347] In order to provide reliable evaluation results, uncertainty estimation is introduced:

[0348] a) Use the bootstrap method to sample multiple times with replacement from the original data to generate multiple sample data sets of the same size as the original data set. For example, from an original data set containing 1,000 user answer data, 100 sample data sets are generated by sampling with replacement. Each sample data set also contains 1,000 data points, but some of the data points may appear repeatedly, while some data points may not appear.

[0349] b) For each sample data set obtained by sampling, the comprehensive evaluation index CEI is calculated to obtain a series of estimated values. Due to the randomness of sampling, the CEI value calculated for each sample data set may be different, and these estimated values ​​constitute a distribution about .

[0350] c) Calculate the mean and 95% confidence interval of the CEI estimate. The mean, as a point estimate of the comprehensive evaluation indicator, reflects the average level of the user's psychological state; the 95% confidence interval provides an uncertainty range, indicating the range in which the true psychological state evaluation value may be at a certain confidence level. For example, the calculated CEI mean is 0.65, and the 95% confidence interval is [0.58, 0.72], which means that we are 95% sure that the user's true psychological state evaluation value is between 0.58 and 0.72. Uncertainty estimation provides a reliability assessment for the psychological state assessment results, helping users and professionals better understand the accuracy and stability of the assessment results.

[0351] 3.7 Results calibration and interpretation

[0352] a) Use quantile transformation to map the CEI to a score range of 0-100 for easier understanding and comparison.

[0353] b) Calculate the contribution of each original feature to the final CEI using the SHAP (SHapley AdditiveexPlanations) value.

[0354] c) Generate an interpretation report that explains which scale scores and behavioral characteristics have the greatest impact on the assessment results.

[0355] Through this approach, a new comprehensive evaluation index (CEI) is created, which has the following characteristics:

[0356] Comprehensiveness: Both traditional scale scores and behavioral characteristics are considered.

[0357] Nonlinearity: Capturing complex nonlinear relationships through feature interactions.

[0358] Adaptive: Ability to adjust feature weights based on historical data.

[0359] Interpretable: Provides the contribution of each feature to the final evaluation.

[0360] Reliability: Provides confidence in the assessment results through uncertainty estimates.

[0361] 4. Construction of personalized evaluation model: Build a personalized evaluation model based on the fused data.

[0362] The output of step 3 is a comprehensive evaluation index CEI that can reflect the user's psychological state. This index takes into account the linear and nonlinear interactions of each feature and provides the reliability and interpretability of the evaluation through weight and uncertainty estimation. However, this output is relatively static and only analyzes the feature set shared by all users. Therefore, based on the data in step 3, the data is fused to further build a more personalized evaluation model. The goal of this step is to use the user's individual information (such as demographic information and historical behavior patterns) and the fused features obtained from step 3 to build a personalized prediction model, so that the evaluation results are more adapted to the unique situation of each user. The specific steps are as follows:

[0363] 4.1 Data Fusion:

[0364] The standardized scale score S = {s1, s2, ..., sn} obtained in step 3 is combined with the extracted behavioral features F = {f1, f2, ..., fm} to form a feature vector X = [S, F]. For example, the scale score vector may contain scores of multiple dimensions such as depression scale scores and anxiety scale scores, while the behavioral feature vector includes feature values ​​of multiple behavioral data dimensions such as answering time features, mouse trajectory features, and keyboard input features. Through data fusion, information from different data sources is integrated to provide a more comprehensive data foundation for building a personalized assessment model.

[0365] 4.2 User feature vector construction:

[0366] Construct a feature vector Uu=[du,hu] for each user u, where:

[0367] du is demographic information (such as age, gender, education level, etc.)

[0368] hu is the historical behavior pattern (such as past test results, behavior feature statistics, etc.)

[0369] Demographic information can provide basic background characteristics of users, which may be related to their mental state. For example, users of different age groups or genders may have differences in their mental state performance and coping methods; education level may affect users' understanding and expression of questions, which in turn affects their answering behavior and mental state assessment. Historical behavior patterns record users' performance and behavioral characteristics in past mental health assessments, such as previous answering time distribution, modification behavior frequency, etc. This historical information can help the model better understand the user's mental state change trends and individual characteristics, thereby achieving more accurate personalized assessments.

[0370] 4.3 Basic model selection:

[0371] Select K basic models M = {M1, M2, ..., MK}, such as random forest, support vector machine, gradient boosting tree, etc. These basic models have good performance and wide application in the field of machine learning, and are suitable for processing different types of data and prediction tasks. For example, the random forest model can process high-dimensional data, has a good ability to capture complex relationships between features, and has good anti-overfitting performance; the support vector machine has high accuracy when processing classification problems, especially for linear or nonlinear separable data; the gradient boosting tree improves the performance of the model by gradually iterating, and has good adaptability to the distribution characteristics of the data. Selecting multiple different types of basic models can give full play to their respective advantages and provide a variety of options for building personalized evaluation models.

[0372] 4.4 Personalized adjustment function:

[0373] Define a personalized adjustment function g for each base model Mi i (U u) is used to adjust the model output according to user characteristics. The personalized adjustment function can modify or optimize the model's prediction results based on the user's demographic information and historical behavior patterns. For example, for young and well-educated users, the model's weight on their psychological state assessment may be adjusted to make them more inclined to positive assessment results based on their historical faster answering speed and higher answering accuracy; while for users who have a history of specific psychological problems and have shown more hesitation and modification in their recent answering behavior, the adjustment function may increase the weight of their current negative psychological state. Specifically, the adjustment function may dynamically adjust the model output based on factors such as the type, severity, and frequency of historical psychological problems of the user, combined with the hesitation and modification behavior characteristics reflected in the current answering process, such as high frequency of answer modification and long decision time. For example, if the user has suffered from moderate anxiety, and in this answering, the average number of modifications per question exceeds a certain threshold, and the mouse hovers over the option for a long time, indicating that they have difficulty in making decisions, the adjustment function may correspondingly increase the possibility of anxiety or other related negative psychological state assessments. This can be achieved by modifying the weight coefficients of anxiety-related traits in the model or by adding penalty terms related to these traits in the calculation of the final assessment results.

[0374] 4.5 Personalized model construction:

[0375] For each base model Mi, build a personalized model Fi:

[0376] F i (X,U u )=M i (X)+g i (U u )

[0377] Among them, M i (X) is the output of the basic model, g i (U u ) is a personalized adjustment item. By combining the prediction results of the basic model with the personalized adjustment function, a personalized evaluation model for each user is obtained. In this way, the model not only takes into account the user's current answering behavior data and scale score, but also incorporates the user's personal characteristics and historical behavior patterns, so that the user's psychological state can be evaluated more accurately, improving the accuracy and pertinence of the evaluation.

[0378] 4.6 Model training:

[0379] Using historical data {(X j ,U j ,y j )} Train each personalized model F i , where y j is the true label.

[0380] Optimization goal:

[0381] min∑L(F i (X j ,U j ),y j ),

[0382] L is the loss function. The loss function can be selected in an appropriate form according to the specific evaluation task and data characteristics, such as mean square error (MSE) for quantifying the difference between the predicted value and the true value, and cross entropy loss for classification problems. By continuously adjusting the model parameters, the model's ability to fit historical data is improved, thereby enhancing the model's prediction accuracy in practical applications.

[0383] 4.7 Ensemble Learning:

[0384] Use the Stacking method to integrate multiple personalized models:

[0385] a) Take the output of each Fi as a new feature.

[0386] b) Train the meta-model G to combine these features:

[0387] y pred =G(F1(X,U),F2(X,U),…,F K (X,U)).

[0388] 4.8 Dynamic update mechanism:

[0389] Design an incremental learning algorithm to regularly update the model based on newly collected data. When new user answer data and related information (including new scale scores, behavioral data, demographic information, etc.) enter the system, first update the user feature vector U to incorporate the new information into the user's historical features. For example, update the new answer time features, modified behavior features, etc. to the historical behavior pattern part, and update the demographic information (if there are changes). Then, use the new data to fine-tune the personalized model Fi and meta-model G, and recalculate the model parameters so that the model can adapt to new data patterns and changes in the user's psychological state. In this way, the model can be continuously optimized over time and with the accumulation of data, maintaining the timeliness and accuracy of the assessment of the user's psychological state.

[0390] 4.9 Personalized threshold adjustment:

[0391] Set a personalized evaluation threshold Tu for each user u:

[0392] T u =T0+h(U u )

[0393] Where T0 is the basic threshold, h(U u ) is an adjustment function based on user characteristics.

[0394] This approach enables the construction of an individualized assessment model for each user, taking into account not only the scale scores and behavioral characteristics, but also the user's personal characteristics and historical behavior patterns. This personalized approach significantly improves the accuracy and applicability of psychological assessments, capturing individual differences and providing mental health professionals with more precise and insightful assessment results.

[0395] 5. Dynamic evaluation and feedback: Implement dynamic evaluation and feedback mechanism of psychological state.

[0396] Based on the personalized evaluation model constructed in step 4, real-time monitoring, incremental learning, feedback generation and other mechanisms are introduced on this basis to achieve a more dynamic and adaptive psychological state evaluation. In step 4, multiple personalized models F(X,U) are constructed, which are trained based on the user's historical data, individual information, behavioral characteristics and scale scores. They provide a personalized static evaluation of the user's psychological state. On the basis of step 4, through real-time updated data stream processing and incremental learning, the model can dynamically adapt to user changes and provide more real-time evaluation and feedback. The introduction of the time dimension enables the model to be continuously updated and adjusted according to new user data, so that the evaluation results not only reflect the user's static psychological state, but also dynamically track its changes. The specific steps are as follows:

[0397] 5.1 Real-time data stream processing:

[0398] a) Establish a data stream pipeline D(t) = {X(t), U(t)}, where X(t) is the feature vector at time t and U(t) is the user feature.

[0399] b) Use sliding window technology with a window size of w and a step size of Δt to update data:

[0400] W(t)={D(t-w+1),D(t-w+2),…,D(t)}.

[0401] 5.2 Incremental Learning Algorithm:

[0402] a) Define the incremental update function I(M,W(t)) to update the model M.

[0403] b) Every Δt time, perform model update:

[0404] M(t)=I(M(t-Δt),W(t)).

[0405] 5.3 Real-time evaluation calculation:

[0406] a) Define the evaluation function E(X,U,M).

[0407] b) At each time point t, calculate the real-time evaluation score:

[0408] S(t)=E(X(t),U(t),M(t)).

[0409] 5.4 State change detection:

[0410] a) Calculate the moving average MA(t) and moving standard deviation MSD(t) of the score.

[0411] b) Define the change detection function:

[0412]

[0413] c) When C(t)>θ, a state change alarm is triggered, and θ is the preset threshold.

[0414] 5.5 Adaptive Question Selection:

[0415] a) Maintain a pool of questions Q = {q1, q2, ..., qn}, where each question qi has a difficulty di and a discrimination ai.

[0416] b) Define the question selection function T(S(t),Q) based on the Item Response Theory (IRT):

[0417]

[0418] c) Choose the next question with the most information:

[0419] q next = argmax(I i (θ)),

[0420] Among them I i (θ) is the question information function.

[0421] 5.6 Personalized Feedback Generation:

[0422] a) Establish a feedback rule base R = {r1, r2, ..., rm}.

[0423] b) Define the feedback selection function F(S(t), C(t), R).

[0424] c) Generate personalized feedback:

[0425] feedback(t)=F(S(t),C(t),R).

[0426] 5.7 Multi-level early warning mechanism:

[0427] a) Define multiple warning thresholds Θ = {θ1, θ2, ..., θk}, corresponding to different severity levels.

[0428] b) Real-time calculation of warning level:

[0429] L(t)=max{i|C(t)>θ i ,θ i ∈Θ}

[0430] c) Trigger an early warning response of the corresponding level.

[0431] 5.8 Long-term trend analysis:

[0432] a) Use time series analysis methods, such as ARIMA model:

[0433] ARIMA(p,d,q):(1-φ1B-…-φ p B p )(1-B) d X t =(1+θ1B+…+θ q B q )∈ t

[0434] b) Predict the psychological state at n time points in the future:

[0435] S(t+1),S(t+2),…,S(t+n).

[0436] 5.9 Feedback loop optimization:

[0437] a) Record the user’s response to the feedback Response(t).

[0438] b) Define the optimization function O(F,Response).

[0439] c) Update feedback strategy regularly:

[0440] F′=O(F,{Response(tk),…,Response(t)})

[0441] Through this dynamic assessment and feedback mechanism, the present invention can monitor the user's psychological state changes in real time, provide timely and personalized feedback, and predict long-term trends. This method significantly improves the timeliness and pertinence of psychological assessment, can identify potential mental health risks at an early stage, and provide strong support for clinical intervention and self-management.

[0442] 6. Result Output

[0443] The final stage of the entire assessment process is responsible for presenting the psychological state assessment results generated in the previous steps to the user and ensuring that the results are presented in an easy-to-understand and intuitive way:

[0444] 6.1 Quantitative scoring: quantitative scores are given on multiple psychological dimensions (such as depression, anxiety, stress, etc.). Based on the user's specific psychological dimensions, the system will output detailed numerical scores so that users or mental health professionals can clearly understand the specific assessment results of each psychological state. For example, in the depression dimension, the score range may be set from 0 to 100 points, with 0 points indicating no symptoms of depression and 100 points indicating extreme depression. Based on the user's answer data, behavioral characteristics, and model calculation results, the system gives a specific depression quantitative score, such as 65 points, which means that the user is at a certain risk level for depression and requires further attention or intervention. Quantitative scoring can provide users with an intuitive measure of their psychological state, helping them better understand their mental health.

[0445] 6.2 Classification results: In addition to giving a quantitative score, the system also needs to classify the user's emotional state according to the psychological state score. For example, the user is classified as "normal", "mild depression", "moderate depression", "severe depression", "mild anxiety", "moderate anxiety", "severe anxiety", etc. The classification results provide users with a more intuitive evaluation conclusion, which is convenient for quickly understanding the current psychological state. The classification criteria can be determined based on professional psychological diagnostic criteria and a large amount of clinical data. For example, when the depression quantitative score is between 0 and 30 points, it is judged as "normal"; between 31 and 50 points is "mild depression"; between 51 and 70 points is "moderate depression"; 71 points and above is "severe depression". Through the classification results, users can quickly know the approximate range of their psychological state, which provides a reference for subsequent decisions (such as whether to seek professional help).

[0446] 6.3 Visual display: The results of the psychological state assessment are intuitively displayed through charts (such as bar charts, line charts, pie charts, etc.), graphs, etc., to help users understand their psychological changes more easily. For example, a bar chart is used to display the quantitative score comparison of users in different psychological dimensions such as depression, anxiety, and stress, so that users can see their performance in various aspects at a glance; a line chart is used to present the trend of changes in the user's psychological state in multiple assessments, helping users observe the dynamic evolution of their own psychological state; a pie chart is used to show the proportion of different emotional states (such as positive, negative, and neutral) of the user during the answering process, so that users have a more intuitive understanding of the distribution of their emotions when answering questions. Visual display can enhance users' understanding and acceptance of the assessment results, and it also facilitates mental health professionals to conduct more in-depth analysis and diagnosis.

[0447] 6.4 Result explanation: Provide a detailed explanation of the evaluation results displayed to the user, explain why the user received a score or classification of a certain mental state, and provide specific suggestions. For example, if the system detects that the user may be in a state of mild depression, the system will explain which behavioral characteristics or scale scores contribute most to the evaluation results. It may be pointed out that the user's longer answering time, more answer modifications, lower mouse movement speed, and higher scores on certain specific questions on the depression scale during the answering process have combined to lead to the evaluation result of mild depression. At the same time, provide corresponding suggestions based on the evaluation results, such as suggesting that the user conduct appropriate relaxation training (such as deep breathing, meditation, yoga, etc.), increase social activities, maintain a regular schedule, seek psychological counseling or professional help, etc. Result explanations and suggestions can help users better understand their own mental state and guide them to take positive measures to improve their mental health.

[0448] The following is a further introduction to the mental health survey of college students combined with the application:

[0449] System deployment and preparation

[0450] Deploy this mental health assessment system on the server within the university campus network to ensure that the system can run stably and is compatible with the school's information infrastructure. Provide system usage training to students participating in the assessment, inform them of the purpose, process and precautions of the assessment, and emphasize data privacy protection measures to improve student participation and cooperation.

[0451] Data collection process

[0452] Students log in to the system through the campus network and first register their personal information, including name, student number, age, gender, major, grade, and other demographic information. Then, students choose the type of psychological scale assessment to be conducted. The system provides a variety of scales suitable for student groups, such as the Symptom Checklist-90 (SCL-90) and the College Students' Mental Health Survey Questionnaire.

[0453] During the answering process, the system uses the Clarity tool to comprehensively collect students' behavioral data. In terms of mouse movements, the system accurately records the mouse movement trajectory, click events, and scroll wheel usage when students select answers, browse questions, view prompts, and other operations. For example, by analyzing the mouse trajectory, it is found that some students have more complex mouse movement trajectories and frequently wander between options on some questions involving interpersonal relationships or academic pressure, which may indicate that they are more confused or hesitant about these issues. Keyboard input data collection focuses on students' input speed, key interval, deletion and modification frequency, and special key usage when answering open-ended questions. For example, it is observed that some students have a slow input speed and a high deletion rate when describing their feelings of learning pressure, which may indicate that they have some difficulties in organizing language to express their inner feelings or are not confident in their own expressions. At the same time, the system records students' switching behavior between different question pages, including question switching frequency, the number and duration of reviewing answered questions, and other information, in order to understand students' answering strategies and their confidence in knowledge mastery. For answer modification behavior, record the number of modifications, the final answer selection time and the modification mode in detail. If it is found that a student has frequently modified multiple questions and the modification time is long, it may indicate that his thinking is not clear enough during the answering process or that there are large fluctuations in his cognition of his own mental state.

[0454] In terms of device and environmental information collection, the screen resolution, device type (such as laptop, tablet or mobile phone), operating system type and version, and browser type and version of the device used by the student are obtained. At the same time, the device microphone and light sensor are used to collect the start time of answering questions, background noise level and light intensity. For example, when conducting an assessment in a dormitory environment, a relatively high background noise level may be recorded, which may cause some interference to the answering of some students. Subsequent analysis can explore the impact of environmental factors on the assessment results of students' psychological state. In addition, the recording and video recording function during the answering process continuously records the changes in students' facial expressions and voice information, providing multimodal data support for sentiment analysis.

[0455] Data processing and analysis

[0456] After data collection is completed, the system automatically preprocesses the data. During the data cleaning process, missing values, duplicate values, and outliers are processed. For individual missing items in the scale answers caused by network problems or misoperation, a reasonable filling strategy is adopted according to the question type and other relevant information. For example, for multiple-choice questions, they can be filled according to the distribution of options, and for numerical data, the mean or median can be used for filling. For outliers in behavioral data, such as extremely short or extremely long answering times, judgments and corrections are made in combination with other behavioral data and device environment information of students. For example, if a student's answer time for a certain question is extremely short, but subsequent analysis finds that his mouse clicks and keyboard input behaviors on the question are normal, it may be an outlier caused by system timing errors, which can be reasonably corrected based on the answer time of adjacent questions.

[0457] In the data conversion phase, the text answers of students in open-ended questions and the text options in the scale questions are converted into numerical vectors through the word vector model for subsequent feature extraction and model analysis. At the same time, some data are normalized or standardized, such as normalizing the device screen resolution data to between 0 and 1, so that the data of different devices are comparable. Then, the cleaned and converted data are randomly divided into training sets and test sets in a ratio of 70%:30%, ensuring that the training set and the test set have similar data distribution and avoiding the impact of data skew on model training and evaluation.

[0458] In the data processing and feature extraction stage, various behavioral data features are calculated. In terms of time features, the total answering time, the answering time for each question, the dwell time between questions, and the hovering time of options are calculated, and further statistical indicators such as the average answering time, the standard deviation of the answering time, the coefficient of variation, the skewness, the kurtosis, and the relative answering time are obtained. For example, the analysis found that the average answering time of a certain student group is longer and the standard deviation of the answering time is larger, which may indicate that there are large individual differences in the answering process of this group, and some students may have difficulty understanding the questions or be more cautious when answering questions. The calculation of mouse trajectory features includes the total length of the trajectory, the complexity of the trajectory, the smoothness, the number of direction changes, and the speed distribution. Through these features, the operation behavior patterns of students in the answering process are analyzed. For example, students with higher trajectory complexity may have more active thinking when thinking about problems, but they may also have more hesitation and uncertainty. The hesitation features calculate the hovering ratio, the option switching frequency, the decision time ratio, and the trajectory overlap, etc., to evaluate the students' hesitation degree and decision-making process in the answering process. Rhythm features calculate the consistency of answering rhythm, answering speed trend, rhythm fluctuation and rhythm interruption. For example, if a student's answering speed trend is gradually slowing down, it may indicate that he is tired or encounters difficult problems during the answering process, which increases the difficulty of thinking. Modification behavior features calculate the overall modification rate, modification time distribution, modification decisiveness and option switching mode (using the Markov chain model to calculate the transition probability matrix between options), etc., to gain a deep understanding of students' modification behavior and thinking process for answers. Keyboard input features calculate (applicable to open-ended questions) input speed changes, deletion rate, pause mode and special key usage rate, etc., to obtain clues about students' behavioral habits and psychological state during text input. Page interaction features calculate scrolling frequency, average scrolling amplitude, question switching mode (using the N-gram model to analyze the pattern of question switching sequence) and review behavior intensity, etc., and analyze students' interactive behavior during page browsing and answering through these features. Device and environment feature processing includes device type encoding, screen size standardization and environmental stability calculation, considering the impact of device and environmental factors on students' answering behavior. Finally, composite features such as the hesitation-revision index, cognitive load index, and attention distraction index were calculated to comprehensively evaluate students’ psychological states from multiple dimensions.

[0459] In the evaluation model construction and training phase, the importance of each feature is calculated based on the random forest algorithm, and the feature importance score is determined by the average impurity reduction method and normalized to obtain the weight. For example, in the evaluation of students' mental health, it was found that the scores of questions about depression and insomnia in the scale, as well as the mouse hover time and modification behavior during the answering process, have high importance weights. Calculate nonlinear feature interactions to capture nonlinear relationships between features. For example, the interaction between the two features of long answering time and frequent answer modification may more strongly suggest that students are in an anxious state. Calculate comprehensive evaluation indicators, combine linear combinations and nonlinear combinations, and determine the relative importance of linear and nonlinear parts by adjusting the balance parameters to obtain comprehensive evaluation indicators. At the same time, perform adaptive weight adjustment, calculate the performance stability of each feature based on historical data, and regularly update the feature weights so that the model can adapt to the changing trend of the psychological state of the student group. Uncertainty estimation is performed through the bootstrap method, and the calculated mean and 95% confidence interval provide reliability evaluation for the evaluation results.

[0460] In terms of building a personalized assessment model, the standardized scale scores and behavioral characteristics are combined into a feature vector, and the user feature vector is constructed by combining the student's demographic information and historical behavior patterns. Random forests, support vector machines, etc. are selected as basic models, and personalized adjustment functions are defined for each basic model to adjust the model output according to the individual characteristics of the students. For example, for students with excellent grades and stable performance in previous psychological assessments, the adjustment function may appropriately reduce the assessment weight of their psychological problem risk in the model prediction; while for students who have recently experienced major life events (such as family changes, broken love, etc.) and whose answering behavior has changed significantly (such as extended answering time, frequent revisions, etc.), the adjustment function will increase the sensitivity to changes in their psychological state and increase the weight of psychological problem risk assessment. By training personalized models and integrating multiple models using the Stacking method, and finally setting personalized thresholds, the assessment thresholds are dynamically adjusted according to the individual characteristics of the students to improve the accuracy and pertinence of the assessment.

[0461] Result output and feedback

[0462] The system generates a detailed mental health assessment report for each student based on the assessment results. The report includes quantitative scores, such as giving specific numerical scores on psychological dimensions such as depression, anxiety, and stress; classification results, classifying students' mental states as "normal", "mild problems", "moderate problems" or "severe problems", etc.; visual display, using bar charts to show the comparison of quantitative scores of different psychological dimensions, line charts to show the trend of students' mental state changes in multiple assessments (if there is historical assessment data), and pie charts to show the proportion of different emotional states during the answering process, etc., to help students intuitively understand their mental state. At the same time, the report provides an explanation of the results, detailing the basis of the assessment results, such as which behavioral characteristics or scale scores contribute more to the results, and providing personalized suggestions based on the assessment results. For example, for students who are assessed as mildly anxious, it is recommended that they participate in mental health lectures organized by the school, learn relaxation techniques (such as progressive muscle relaxation), and increase physical exercise; for students who are at risk of moderate depression, it is recommended that they seek professional help from the school's psychological counseling center in a timely manner, and inform their classmates and teachers around them to give appropriate attention and support.

[0463] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the protection scope of the present invention in any form, and all technical solutions obtained by equivalent replacement and the like fall within the protection scope of the present invention. The parts not involved in the present invention are the same as the prior art or can be implemented by the prior art.

Claims

1. An intelligent assessment system integrating behavioral data and scales, characterized in that: The system includes the following modules: Data collection module, used for multi-dimensional behavioral data collection; Data processing and feature extraction module, used for data cleaning and preprocessing as well as feature extraction and calculation; Analysis and evaluation module, which conducts analysis and evaluation based on behavioral characteristics and scale scores; The result output and explanation module is used to display the previously generated psychological state assessment results to the user and explain the assessment results displayed to the user.

2. The intelligent assessment system integrating behavioral data and scale according to claim 1, characterized in that: The data collection includes scale collection, behavioral data collection, equipment and environment information collection, and audio and video recording of the answering process.

3. The intelligent assessment system integrating behavioral data and scale according to claim 1, characterized in that: The data cleaning and preprocessing include comprehensive cleaning of the collected data, processing of missing values, duplicate values ​​and outliers, converting text data into vector form using natural language processing technology for subsequent analysis, and then randomly dividing the data into training sets and test sets to ensure the randomness and representativeness of data distribution.

4. The intelligent assessment system integrating behavioral data and scale according to claim 1, characterized in that: The features in the feature extraction include time features, mouse trajectory features, hesitation time features, rhythm features, modification behavior features, keyboard input features, page interaction features, device and environment features, and the feature calculation includes calculating the hesitation-modification index, cognitive load index and attention distraction index.

5. The intelligent assessment system integrating behavioral data and scales according to claim 1, characterized in that: The analysis and evaluation includes: a) feature importance calculation based on random forest, b) non-linear feature interaction calculation, c) comprehensive evaluation index, and d) personalized evaluation.

6. The intelligent assessment system integrating behavioral data and scales according to claim 1, characterized in that: The result output and explanation include: a) quantitative scoring, b) classification results, c) visual display, and d) result explanation.

7. An intelligent assessment method integrating behavioral data and a scale, characterized in that: The method is carried out based on the system according to any one of claims 1 to 6, and specifically comprises the following steps: Step 1: Multi-dimensional behavioral data collection; Step 2: Extract and calculate the behavior data features; Step 3: Develop comprehensive evaluation indicators based on behavioral characteristics and scale scores; Step 4: Build a personalized evaluation model; Step 5: Dynamic evaluation and feedback; Step 6: Output the results.

8. The intelligent assessment method integrating behavioral data and a scale according to claim 7, characterized in that: The feature extraction in step 2 includes: a) Time characteristics: calculate the total answering time, answering time for each question, dwell time between questions, hovering time on options, and further calculate the statistical indicators of these time data, including average answering time, answering time standard deviation, answering time coefficient of variation, answering time skewness, answering time kurtosis, and relative answering time; b) Mouse trajectory characteristics: calculate the total trajectory length, trajectory complexity, trajectory smoothness, number of direction changes, and speed distribution; c) Hesitation characteristics: calculate the hovering ratio, option switching frequency, decision time ratio, and trajectory overlap; d) Rhythm characteristics: calculate answer rhythm consistency, answer speed trend, rhythm fluctuation, and rhythm interruption; e) Modification behavior characteristics: calculate the overall modification rate, statistically analyze the modification time distribution, calculate the modification decisiveness, use the Markov chain model to calculate the transition probability matrix between options, and finally normalize it to obtain the transition probability matrix; f) Keyboard input characteristics: calculate the input speed change, use the moving average method and standard deviation to analyze the stability of the input speed, calculate the deletion rate, count the pause pattern, and calculate the usage rate of special keys. g) Page interaction features: Calculate the scrolling frequency and average scrolling amplitude, and use the N-gram model to analyze the question switching pattern.

9. The intelligent assessment method integrating behavioral data and a scale according to claim 8, characterized in that: The steps for analyzing the question switching pattern include: recording the question switching sequence, building an N-gram model, calculating the conditional probability, extracting the transition probability matrix, self-loop ratio, forward and backward jump rate, average jump distance features, identifying high-frequency and low-frequency switching patterns, and reviewing the intensity of behavior.

10. The intelligent assessment method integrating behavioral data and a scale according to claim 7, characterized in that: In step 3, a comprehensive evaluation is performed using the scale score and behavioral characteristics to output a comprehensive evaluation index CEI that is comprehensive and can reflect the user's psychological state. The index considers the linear and nonlinear interactions of each feature and provides the reliability and interpretability of the evaluation through weight and uncertainty estimation; wherein the comprehensive evaluation index is: CEI = α × L + (1-α) × NL Where α is the parameter that balances the linear and nonlinear components, L is the linear combination, and NL is the nonlinear combination.

11. The intelligent assessment method integrating behavioral data and a scale according to claim 7, characterized in that: In step 4, a personalized evaluation model is constructed based on the fused data; the specific steps include: 1) Data fusion: The standardized scale scores obtained in step 3 are combined with the extracted behavioral features into a feature vector; 2) User feature vector construction: construct a feature vector for each user u; 3) Basic model selection: select K basic models M = {M1, M2, ..., MK}; 4) Personalized adjustment function: Define a personalized adjustment function g for each basic model Mi i (U u ), used to adjust the model output according to user characteristics; 5) Personalized model construction: For each basic model Mi, build a personalized model Fi: F i (X,U u )=M i (X)+g i (U u ) Among them, M i (X) is the output of the basic model, g i (U u ) is a personalized adjustment item; by combining the prediction result of the basic model with the personalized adjustment function, a personalized evaluation model for each user is obtained; 6) Model training: Using historical data {(X j ,U j ,y j )} Train each personalized model F i , where y j is the true label; 7) Ensemble learning: Use the stacking method to integrate multiple personalized models; 8) Dynamic update mechanism: Design an incremental learning algorithm to regularly update the model based on newly collected data; 9) Personalized threshold adjustment: Set a personalized evaluation threshold T for each user u u : T u =T0+h(U u ) Where T0 is the basic threshold, h(U u ) is an adjustment function based on user characteristics.

12. The intelligent assessment method integrating behavioral data and a scale according to claim 7, characterized in that: In step 5, based on the personalized assessment model constructed in step 4, real-time monitoring, incremental learning, and feedback generation mechanisms are introduced to achieve a more dynamic and adaptive psychological state assessment.

Images