An exposure conditioning IRT-CAT system and method for pre-adolescent school-age children behavior addiction risk early warning

Abstract

The application provides an exposure conditioning IRT-CAT system and method for prepubertal school-age children behavior addiction risk early warning, which solves the problems of long fixed questionnaire length in existing screening, misjudgment of non-exposure as low risk, and inability of direct comparison of risks in different fields, and the like. The application comprises an offline calibration stage and an online evaluation stage. The application has the advantages of reasonable early warning results, high evaluation efficiency, strong universality and the like.

Description

Technical Field

[0001] This invention belongs to the field of psychometrics and computerized adaptive testing technology, specifically relating to an exposure-conditioned IRT-CAT system and method for early warning of behavioral addiction risks in pre-adolescent children. Background Technology

[0002] The school-age period is a critical stage for the rapid development of children's eating behaviors, digital media use behaviors, self-control, and emotional regulation abilities. Addiction-like behaviors such as uncontrolled food intake, excessive internet use, continuous viewing of short videos, and excessive participation in online games may show early risk signals before puberty. If rapid, low-burden, and repeatable risk screening can be conducted on school populations before the peak of puberty, it can provide earlier warnings for families, schools, and children's health management. Currently, the field of children's behavioral health screening mainly uses fixed-length scales or electronic questionnaires for assessment. However, existing technological solutions have the following shortcomings: First, fixed questionnaire assessment methods present a trade-off between efficiency and accuracy. Fixed questionnaires require each child to answer the same set of questions, preventing dynamic adjustments based on individual responses. This results in low-risk children having to answer a large number of questions that contribute little to their risk assessment, adding unnecessary workload. Simultaneously, high-risk children may lack sufficiently informative questions in critical risk ranges, affecting the accuracy of risk estimation. Furthermore, traditional total or average scoring methods assume different items have the same measurement contribution, failing to fully utilize differences in item difficulty, discrimination, threshold positions, and information content. Additionally, assessment tools based on clinical samples are unsuitable for early warning in school populations. Most existing behavioral addiction assessment tools primarily use clinically diagnosed children as their sample source, and their norms and cutoff points are suitable for diagnostic assistance or case differentiation scenarios. However, clinical samples cannot reflect the continuous distribution characteristics of behavioral risk in the general school population; directly using clinical tools for school screening can easily lead to insufficient differentiation between low-risk and intermediate-risk children. The population database collected in school settings can cover a complete continuous spectrum from low risk to high risk in the general population, making it more suitable for establishing risk percentile references and stratified early warning thresholds. However, there is currently a lack of assessment schemes built on such data.

[0003] Meanwhile, existing technologies generally neglect the exposure prerequisites in behavioral addiction domains, leading to distorted assessments of non-exposed individuals. In multi-domain screening for childhood behavioral addictions, areas such as short video addiction and gaming addiction have clear exposure prerequisites: children who have not watched short videos or played online games do not have a realistic basis for developing the corresponding behavioral addiction risk. However, existing screening tools typically treat these non-exposed children directly as low-risk individuals, including them in total score calculations or risk rankings. This approach leads to systematic biases in item parameter estimation during the calibration phase, and in the assessment phase, it causes non-exposed individuals to be misjudged as low-risk, further distorting cross-domain risk rankings. Currently, there is a lack of a technical solution that systematically processes exposure conditions in both item parameter calibration and online assessment phases. Moreover, the assessment results for different behavioral domains lack a comparable and unified standard. The number of items, response levels, and calibration sample distributions vary across different behavioral domains, and the potential risk values ​​for each domain are defined within their respective item banks and calibration samples, resulting in inconsistent measurement scales. Therefore, the potential risk values ​​for different domains cannot be directly added, averaged, or ranked, which makes multi-domain comprehensive risk assessment lack a reasonable computational basis. How to integrate assessment results from multiple fields into a comprehensive risk indicator with interpretability is an unsolved technical problem in this field.

[0004] To address the aforementioned issues, this application proposes an Exposure Conditioning-based Catalysis (IRT-CAT) system and method for early warning of behavioral addiction risks in pre-adolescent children. This system utilizes a school population database, deeply integrates exposure conditioning calibration with computerized adaptive testing, and unifies risk results from multiple domains into a comparable caliber. Summary of the Invention

[0005] The purpose of this invention is to address the above-mentioned problems by providing a reasonably designed and effective exposure-conditioning IRT-CAT method for early warning of behavioral addiction risks in pre-adolescent children, which effectively reduces the risk of misjudgment.

[0006] Another objective of this invention is to address the aforementioned problems by providing an exposure-conditioned IRT-CAT system for assessing the efficiency of behavioral addiction risk warning in pre-adolescent children.

[0007] To achieve the above objectives, the present invention employs the following technical solution: an exposure-conditioned IRT-CAT method for early warning of behavioral addiction risks in pre-adolescent children, comprising: S1: Obtain multi-domain item response data and behavioral exposure variables corresponding to at least one domain; obtain multi-domain item response data from a database of pre-adolescent children collected from school settings, and record exposure variables of whether each child actually comes into contact with a specific behavior.

[0008] S2: Determine whether each domain has exposure prerequisites based on the domain configuration. For domains with exposure prerequisites, use samples that meet the exposure conditions and have a high percentage of valid responses as calibration samples for that domain. For domains without exposure prerequisites, use all samples that meet the valid response conditions as calibration samples.

[0009] S3: For the calibration samples in each domain, for each item, construct surrogate latent variables based on the responses of other items after excluding the current item, and fit a model to obtain the discrimination parameter and threshold parameter of the item, generating a parameterized item bank; fit a cumulative logistic model to obtain the discrimination parameter and threshold parameter of the item.

[0010] S4: During online assessment, applicability is assessed for each subject in each domain: for domains with exposure prerequisites, if a subject does not meet the exposure conditions for that domain, the domain is marked as inapplicable and skipped; for applicable domains, computerized adaptive testing is performed. S5: Perform computerized adaptive testing, select the item with the most information based on the current potential risk value, update the potential risk value and standard error after receiving the answer, and perform sequential double threshold stop judgment. Stop the testing of this domain when the preset accuracy condition or number of questions is met; use the expected posterior method to update the current potential risk value and standard error.

[0011] S6: After the domain testing is completed, the final potential risk value of the domain is mapped to the risk percentile in the empirical distribution of the calibration samples in that domain; S7: Using the set of applicable fields of the current subjects as the denominator, perform a weighted summation of the risk percentiles of each applicable field, and output the overall risk percentile and the highest risk field.

[0012] This method systematically embeds the domain-specific characteristics of the application scenario of early warning for childhood behavioral addiction into the technical framework of IRT-CAT, forming a complete closed loop from data preparation, offline item parameter calibration, parameterized item bank generation, online applicability determination, CAT adaptive assessment, conversion from domain theta to risk percentile, and comprehensive integration of applicable domains. The technical foundation of this method is item response theory, whose core idea is to correlate and parameterize the subject's trait level with their behavior on items. The IRT (Independent Test Response) provides a theoretical foundation with three key characteristics: First, the estimation of test-takers' abilities does not depend on the specific selection of test items, which ensures that dynamic item selection in adaptive assessment does not affect the consistency of ability estimation. Second, the estimation of item parameters in the IRT model is independent of the test-taker sample. The item characteristic curve is a regression of the probability of a test-taker's correct answer to their latent trait level. The shape and position of the regression curve do not depend on the distribution of test-takers, which provides a theoretical basis for reusing item parameters calibrated based on school population databases in different assessment scenarios. Third, the IRT accurately estimates measurement error through the model. The concept of the test information function introduced by Bernbaum makes the accuracy of measurement quantifiable at each latent trait level, which provides a computational tool for information-based item selection strategies in CAT (Computer-Aided Assessment).

[0013] By simultaneously introducing exposure conditioning during the offline calibration and online assessment phases, the fundamental assumption of IRT—parameter invariance being independent of the subject population—is strictly limited to the correct reference population, thus fundamentally resolving the issue of non-exposed individuals being incorrectly included in the risk spectrum. Building upon this, domain percentile mapping achieves unification of different scales, and through weighted synthesis of the applicable domain set, multi-domain assessment results are integrated into a comprehensive early warning indicator with public health interpretability.

[0014] The database used in this method is a behavioral addiction item response database built from data collected in school settings targeting pre-adolescent children. Unlike clinical-based data, which primarily consists of outpatient cases or clinically diagnosed children, this database is population-based data from a school population. It preserves a continuous distribution from low to high risk among children in the general population, making it suitable for establishing risk percentile references and early warning thresholds. This database is used to estimate item discrimination, thresholds, and information functions, and also to generate domain experience distributions and risk percentile lookup tables. Because the database originates from a school population, the percentile results output by the system can be interpreted as the subject's risk position relative to a similar school population.

[0015] In the above technical solution, step S2, for areas with exposure prerequisites, uses samples that meet the exposure conditions and have a qualified response rate as calibration samples, specifically including: For the default applicable domain, the calibration sample set C_s is all samples that satisfy the requirement that the effective response ratio is not less than the preset threshold τ_s; For domains with exposure prerequisites, the calibration sample set C_s is a set of samples that simultaneously satisfy the condition that the exposure variable value is equal to the preset exposure value and the effective response ratio is not lower than the preset threshold τ_s. Samples that do not meet the exposure conditions will not be included in the calibration sample set for the corresponding field.

[0016] This step addresses the determinism of the reference population in IRT calibration. A domain configuration is established for each behavioral domain *s*. This configuration includes at least the domain identifier, item list, response level range, presence of exposure prerequisites, exposure variable name, exposure value, minimum effective response ratio, and CAT stopping parameters. Let *D_s* be the original sample set for domain *s*, *E_s* be the exposure variable for that domain, and *v_s* be the exposure value. Then, the calibration sample set for the default applicable domain is *C_s = {r in D_s : valid_ratio(r,s)>= τ_s}, and the calibration sample set for domains with exposure prerequisites is *C_s = {r in D_s : E_s(r)=v_s and valid_ratio(r,s)>= τ_s}. Here, *valid_ratio(r,s)* represents the proportion of valid responses to *r* in domain *s* out of the total number of items in that domain.

[0017] In classic IRT theory, item parameters (discrimination a, difficulty b) are estimated through samples, and the parameter values ​​themselves have no absolute meaning; their interpretation depends entirely on the group represented by the calibration sample. Including non-exposed individuals in the calibration sample is tantamount to forcibly placing two fundamentally different groups—those with no exposure opportunity and those with exposure opportunity and potential risk—on the same scale for parameter estimation. This leads to distortions on two levels: first, a systematic bias in parameter estimation because the response patterns of non-exposed individuals reflect "inapplicability" rather than "low risk"; second, the resulting percentile reference table anchors non-exposed individuals at the low end, causing a systematic overestimation of the risk of actual exposed individuals in subsequent assessments. This step, through exposure conditional screening during the calibration phase, defines the correct reference population boundaries for each domain. For domains without exposure preconditions, the reference population is the valid school sample; for domains with exposure preconditions, the reference population is the children who have actually been exposed to the behavior. Only by first determining the correct reference population can the subsequently calculated item difficulty, discrimination, risk percentiles, and overall risk be interpretable.

[0018] In the above technical solution, step S3 involves constructing proxy latent variables based on the responses of other entries after excluding the current entry, and fitting a model to obtain the discrimination parameter and threshold parameter. Specifically, this includes: For item i in domain s, remove the current item i from the domain response matrix, construct the item-rest surrogate score using the mean of the valid responses of the remaining items and standardize it to obtain the surrogate latent variable θ_-i; For each ordered response threshold k of an item, construct a binary variable, fit a cumulative logistic model, and obtain the discrimination parameter a_i and the multi-level threshold parameter b_ik from the model coefficients; During the fitting process, at least one robust processing is also performed: ridge regularization of the slope term, deletion of threshold fitting with insufficient sample size, limiting the threshold to a preset range, monotonically sorting the thresholds, or generating a backoff threshold based on the empirical cumulative ratio when the threshold fitting fails.

[0019] This step addresses the problem of how to stably and unbiasedly estimate item parameters in a school census scenario where a pre-calibrated item bank is unavailable. Let R_s be the set of items in domain s, Y_rj be the response score of sample r on item j, and the formula for calculating the item-rest surrogate score is M_ri = mean{Y_rj : j in R_s, j != i, Y_rj valid}. The surrogate latent variable is θ_ri(-i) = [M_ri - mean(M_i)] / sd(M_i), where mean(M_i) and sd(M_i) are calculated within the calibrated samples C_s of domain s. The core principle of using item-rest surrogate latent variables is to avoid homology bias: if the total score or factor score containing item i itself is used as the independent variable to predict the response of item i, circular reasoning will occur, as the item parameters include the correlation between the item and itself, leading to a systematic overestimation of the discrimination parameter. Deleting the current entry and then constructing a proxy variable is equivalent to providing each entry with a potential risk reference that does not contain information about the entry itself, so that the estimation of the discrimination can better reflect the entry's true ability to distinguish the level of risk in the domain.

[0020] For a multi-level item with m+1 ordered response levels, the system constructs a binary variable Z_rik = 1 (Y_ri>= k), k=1,...,m, for each threshold k, and fits a cumulative logistic model: logit P(Z_rik=1 | θ_ri(-i)) = α_ik + β_ik θ_ri(-i), thus obtaining b_ik = -α_ik / β_ik, a_i = median{β_ik : β_ik satisfies the validity condition}, and the cumulative probability is P_i(Y_i>= k | θ) = 1 / [1 + exp{-a_i(θ - b_ik)}]. a_i represents the item's ability to distinguish children at different risk levels, and b_ik represents the potential risk position corresponding to a child reaching response level k or higher. For binary items, the above model degenerates into a single-threshold two-parameter logistic model. The introduction of multiple robust handling methods is to address the unavoidable issues of sparse categories and skewed response distributions in school census data. For example, the logistic model uses IRLS iteration and applies a ridge penalty to the slope term. If the number of samples in each category on both sides of a certain threshold is not less than a preset lower limit and the slope is greater than the lower limit, the fitting result of that threshold is retained; otherwise, empirical proportion backoff is implemented. The validity condition refers to the number of samples in each category on both sides of the threshold being not less than a preset lower limit and β_ik being greater than a preset lower limit for the slope.

[0021] After obtaining the item discrimination parameter a_i and the threshold parameter b_ik, the system calculates the probability of each response category based on these parameters. For items at level m+1, the cumulative probability S_k(θ) = 1 / [1 + exp{-a_i(θ - b_ik)}] is calculated first, and then the category probability is obtained from the difference between adjacent cumulative probabilities: P_i(Y=0|θ) = 1 - S_1(θ), P_i(Y=k|θ) = S_k(θ) -S_{k+1}(θ), k=1,...,m-1, P_i(Y=m|θ) = S_m(θ). The information function of an item is I_i(θ) = sum_k[P'_ik(θ)]^2 / P_ik(θ). The category probability is used to calculate the likelihood of a child choosing each response level at a given risk level, and the information function is used to determine how much measurement information a question can provide near the current risk level. The higher the information content, the more suitable the question is for judging the current child's risk level. The system calculates the information function and domain test information curve for each item on a preset theta grid, and writes the item stem, option label, original code, discrimination index, threshold, calibration sample size, missing rate, information peak, domain applicability rule, CAT stopping parameter, risk threshold, percentile lookup table, and version information into a machine-readable item database file. The preset theta grid ranges from -4 to +4, with appropriate step sizes such as 0.01 or 0.05, to achieve a balance between computational accuracy and efficiency.

[0022] In the above technical solution, in step S5, the sequential dual-threshold stop judgment has the following priority: The first priority is to determine if the number of questions answered is not less than the minimum precision judgment number and the standard error is not greater than the preset precision threshold. If these conditions are met, the process stops when the precision target is met. Only when the aforementioned conditions are not met will the second priority check be performed to determine whether the number of questions answered has reached half of the total number of questions in the question bank for that field and rounded up. If the condition is met, the process will be forcibly stopped at the maximum number of questions. If neither of the two conditions is met, continue selecting a topic.

[0023] This step addresses the trade-off between accuracy and efficiency in CAT assessments. Classic CAT typically employs a single stopping rule, such as a fixed number of questions or a fixed standard error threshold. However, in large-scale school screening scenarios, the risk levels of subjects vary significantly: low-risk children may achieve sufficiently accurate risk estimates after a few questions, while high-risk or intermediate-risk children require more questions to achieve the same level of accuracy.

[0024] In a single domain, the online CAT assessment takes "read the current theta and SE - calculate the information amount of all unanswered items - present the item with the maximum information amount at the current theta - receive and record the response - update theta and SE via EAP - sequentially determine the precision stopping condition and the forced item count stopping condition" as one iteration unit. Only when both stopping conditions are not satisfied, the system returns to the information amount calculation step and enters the next round. Let n_s be the number of responded items in the current domain, SE_s be the posterior standard deviation after this round of EAP, N_s be the total number of items in the parameterized item bank of this domain, and M_s=ceil(N_s / 2) be the forced stopping upper limit of this domain. After each response and completion of EAP update, the system performs the judgments strictly in accordance with the priority. The first priority is precision stopping: if n_s is not less than the preset minimum item count and SE_s is not greater than the preset precision threshold, it is considered that the preset measurement precision has been reached, and the CAT of the current domain is immediately terminated. This minimum item count constraint avoids premature stopping caused by accidental responses. The second priority is forced item count stopping: only when the aforementioned precision stopping condition is not satisfied, if n_s≥M_s, the CAT of the current domain is terminated regardless of whether SE_s has reached the precision threshold, thereby ensuring that the assessment burden does not exceed half of the total item count of the domain's item bank, and the ceiling rounding is adopted for odd item counts. The continuation condition is: if n_s is less than the minimum item count, or n_s reaches the minimum item count but SE_s is greater than the precision threshold, and n_s<M_s at the same time, the system calculates the information amount I_i(θ_hat) of all unanswered items at the current θ_hat, selects the item with the highest information amount as the next item, and enters the next iteration. The above judgment sequence cannot be swapped, so as to ensure that those who meet the precision requirement can finish in advance, and at the same time all subjects are constrained by the upper limit of half the item count of the domain. If SE_s is still greater than the precision threshold when the process is forcibly stopped due to reaching M_s, the system retains the final theta and SE, and simultaneously marks that the estimation uncertainty is relatively high, which is used to prompt subsequent retest or supplementary assessment.

[0025] During online assessment, the system executes CAT (Adaptive Testing) sequentially by domain. For domains without exposure prerequisites, it directly enters the item adaptation process; for domains with exposure prerequisites, it first reads existing exposure variables or poses applicability questions to the subject. If the domain is not applicable, the system records the inapplicability and moves to the next domain. For applicable domains, the system calculates the posterior distribution on the theta grid. Assuming the prior π(θ) can be a standard normal or an empirical prior determined by calibration samples, for the set of answered items H, the system accumulates the category probability of each answered response in log space to obtain the posterior weight, using EAP (Experience-Based Advancement) as the theta estimate and the posterior standard deviation as the SE (Self-Adjustment). Based on a small number of questions already answered by the child, the system estimates the potential behavioral addiction risk level and estimation uncertainty of the current domain in real time. θ_hat represents the current risk position, and SE represents whether the estimate is stable. For applicable domains, the system first initializes the current potential risk value θ_hat to 0 and the standard error SE to 1, or determines the initial values ​​based on the subject's historical assessment records.

[0026] After each response, the system calculates the information content of all unanswered items at the current θ_hat and selects the item with the highest information content as the next question. If multiple items have the same or similar information content, they can be grouped together based on symptom coverage, question exposure frequency, question length, historical usage rate, or random rules. In this way, the system no longer distributes all questions but prioritizes presenting the items that are most valuable for judging the current risk level of children, thereby maximizing the accuracy of risk estimation with a shorter number of questions, making it suitable for large-scale rapid screening in school settings.

[0027] The design principle of the sequential dual-threshold stopping rule lies in assigning different technical roles and priorities to the two stopping conditions: the accuracy condition acts as a quality gatekeeper, ensuring that the estimation error of theta in each domain is controlled within an acceptable range; the question load limit condition acts as a burden controller, ensuring that even in difficult estimation situations where it is difficult to reach the accuracy threshold, there is a hard upper limit to the children's answering burden. This design technically achieves personalized allocation of the test-taker's assessment burden, with those requiring high accuracy doing more and those requiring low accuracy doing less, while ensuring the controllability of overall screening efficiency.

[0028] In the above technical solution, in step S6, after the domain test is completed, the final potential risk value of the domain is mapped to the risk percentile in the empirical distribution of the calibration samples in the domain.

[0029] After domain-specific CAT (Comparative Assessment), the system does not directly compare theta values ​​across domains. Instead, it maps θ_s to the risk percentile p_s = F_s(θ_s) in the empirical distribution F_s of the calibration samples for that domain. This percentile represents the relative risk position of the subject within the calibration samples of the same domain and applicable population. This step addresses the comparability of measurement results across different domains. Different domains have different numbers of items, response levels, and calibration sample distributions. Theta values ​​for each domain are defined within their respective item banks and calibration samples, and their scales are not uniform. Therefore, theta values ​​from different domains cannot be directly added, averaged, or ranked. By converting theta to the risk percentile in the domain calibration samples, it is equivalent to unifying the measurement results on different scales into the same metric—the relative risk position within the school population of the same domain—thus providing a unified metrological basis for cross-domain comparisons and comprehensive risk calculations. The empirical distribution F_s is determined by the cumulative distribution function of the theta estimates of all samples in the calibration sample set C_s for that domain, or by smoothing it using methods such as kernel density estimation.

[0030] In the above technical solution, in step S7, when weighting and summing the risk percentiles in the set of applicable fields, the fields marked as inapplicable do not generate risk percentiles, nor are they included in the numerator and denominator of the comprehensive risk calculation.

[0031] This step addresses the denominator contamination problem in multi-domain risk fusion. Let A be the set of domains applicable to the current subject, A = {s : domains s are applicable to the current subject and have completed CAT}, and w_s be the domain weights. Then, the comprehensive risk percentile P_composite = round( sum_{s in A} w_s p_s / sum_{s in A} w_s ), and the highest risk domain s_max = argmax_{s in A} p_s. Domains marked as inapplicable are not estimated θ_s, not generated p_s, not included in the numerator and denominator of P_composite, and are not considered low-risk domains for ranking.

[0032] In traditional multidimensional composite scoring, scores from all dimensions are typically summed or averaged, even if some dimensions are not applicable to a particular individual. In the application scenario of this scheme, including inapplicable areas with zero or low scores in the composite leads to two technical biases: first, the denominator contains invalid information, resulting in decreased reliability of the composite indicator; second, inapplicable areas are assumed to be low-risk, when in reality, these areas pose no risk to the subject. By using the set of applicable areas A as the denominator for composite risk calculation, a conditional definition of the composite indicator is technically achieved. Composite risk reflects only the subject's relative position in areas where behavioral addiction risk may actually occur. θ_s is used for internal model estimation, p_s describes the child's relative position within the school population in the same area, P_composite summarizes the overall behavioral addiction risk, and s_max and area ranking indicate the focus of subsequent interventions.

[0033] In this system, the IRT parameters and CAT operation are not two independent modules, but rather form a continuous computational chain through probability, information content, posterior estimation, and percentile transformation. In the offline phase, a_i and b_ik first determine the class probability P_i (Y=k|θ) of an item at any theta position. The first derivative of the class probability further determines the item information function I_i(θ). In the online phase, the system calculates the posterior distribution based on the set of answered questions H and the corresponding class probabilities P_j (y_j|θ). The posterior mean yields the current risk estimate θ_hat, and the posterior standard deviation yields SE. The system then substitutes θ_hat into the I_i(θ_hat) of unanswered items, selecting the item with the highest discriminative power near the current risk level. Each answer synchronously updates H, the posterior distribution, θ_hat, SE, and the information content ranking for the next question. After the domain ends, θ_hat is not directly used as a cross-domain comparison value. Instead, it is entered into the domain experience distribution F_s(θ) to obtain p_s. p_s is then linked with the domain applicability state, domain weight w_s, and applicable domain set A to generate the comprehensive risk percentile P_composite, the highest risk domain s_max, and the domain risk ranking.

[0034] After completing the assessment of all applicable domains, the system outputs not only the overall risk percentile, domain-specific risk percentiles, risk levels, and domain rankings, but also a complete result package corresponding to this assessment. This result package includes at least six types of information: parameterized question bank information for each domain, including at least the item number, stem, options, discrimination index a_i, difficulty or multi-level threshold b_ik, calibration sample size, and information content; a list of individual CAT actual answer items, recording items, response levels, original answers, and times in the order of presentation; item-by-item measurement information, recording at least theta before item selection, SE before item selection, information content of the item at theta at that time, theta after answering, SE after answering, and the cumulative information content of answered items; an individual measurement information function curve obtained by summing the I_i(θ) of the individual's actual answer items point by point; the exposure applicability judgment rules for each domain and the judgment results for this assessment, explaining whether each domain has the exposure prerequisite, the judgment question used, the subject's answer for this assessment, and the final applicability status; and a risk percentile reference table for each exposure domain, explaining the mapping relationship between domain theta and the corresponding calibration sample risk percentiles. The results can be displayed on the front-end page or exported in a machine-readable format to support scientific research review, individual interpretation, model re-evaluation, and version auditing.

[0035] An exposure-conditioned IRT-CAT system for early warning of behavioral addiction risks in pre-adolescent children includes: The exposure condition calibration module is used to determine whether each domain has exposure prerequisites based on the domain configuration. For domains with exposure prerequisites, samples that meet the exposure conditions and have a qualified response rate are used as calibration samples; for domains without exposure prerequisites, samples that meet all valid response conditions are used as calibration samples. The item parameter estimation module is used to construct surrogate latent variables for each item based on the responses of other items after excluding the current item, and to fit the model to obtain the item's discrimination parameter and threshold parameter, thereby generating a parameterized item bank. The CAT operation module is used to perform applicability judgment on subjects by domain during online assessment: for domains with exposure prerequisites, if subjects do not meet the exposure conditions, they are marked as unsuitable and skipped; for applicable domains, computerized adaptive testing is performed: the item with the most information is selected and presented based on the current potential risk value, the potential risk value and standard error are updated after receiving the answers, and a sequential double threshold stop judgment is performed, stopping when the preset accuracy condition or number of questions condition is met; The risk synthesis module maps the final potential risk value of each applicable field to the risk percentile in the corresponding empirical distribution of the calibration sample, and performs a weighted synthesis of the risk percentile of each applicable field with the set of applicable fields as the denominator, outputting the comprehensive risk percentile and the highest risk field.

[0036] The system described above, through the coordinated operation of four core modules, translates the Exposure Conditioned IRT-CAT method into a deployable technical architecture. The system also includes a data preprocessing module for acquiring multi-domain item response data, exposure variables, and item stem coding tables, outputting unified response codes, missing data markers, and a valid response matrix; an item-rest proxy module for generating a proxy for each item's latent variable from the domain item response matrix; an information function module for calculating category probabilities, item information functions, and domain test information curves based on item discrimination and thresholds; a percentile reference module for generating domain experience distributions and percentile lookup tables based on the theta distribution of domain calibration samples; and an audit traceability module for recording item version, algorithm version, response sequence, theta trajectory, and outputting traceable information such as a parameterized item library, actual response list, item-by-item theta / SE / information trajectory, individual information curves, exposure applicability records, and a risk percentile reference table, supporting re-scoring and auditing.

[0037] The overall system design follows the principle of separating "offline calibration - online assessment": the exposure-conditioning calibration module and the item parameter estimation module complete the parameterization construction of the item bank offline, and only need to be re-executed when the population database is updated or the item bank is revised; the CAT operation module and the risk integration module respond to the test takers' assessment requests in real time online. This separation design has two technical advantages: First, it decouples the computationally intensive item parameter estimation process from the real-time response, so that online assessment only needs to perform posterior estimation and information calculation, ensuring the system's response speed; second, the parameterized item bank, as an intermediate product, can be version-managed and reused, supporting different assessment front-ends to call the same set of item bank parameters.

[0038] In the aforementioned system, the exposure-conditional calibration module is specifically used to: construct a calibration sample set for the default applicable domain using all samples whose effective response rate is not lower than a preset threshold; and construct a calibration sample set only for domains with exposure prerequisites, using samples whose exposure variable values ​​are equal to preset exposure values ​​and whose effective response rate is not lower than a preset threshold. Samples that do not meet the exposure conditions are not included in the calibration sample set for their respective domains. This module achieves accurate definition of the domain calibration reference population at the system level, maintaining an independent calibration sample set for each domain, so that the entry parameters for different domains are defined on their respective correct reference populations.

[0039] In the aforementioned system, the sequential dual-threshold stop judgment executed by the CAT operation module has the following priority: the first judgment condition is that the number of answered questions is not less than the preset minimum number of questions and the standard error is not greater than the preset accuracy threshold. The second judgment condition is only applied if this condition is not met, i.e., the number of answered questions reaches half of the total number of questions in the question bank for that domain, rounded up. If neither condition is met, the next item is selected based on the maximum information content. This module achieves a dynamic balance between accuracy and workload at the system operation level.

[0040] In the aforementioned system, within the risk integration module, areas marked as inapplicable are not estimated for potential risk values, do not generate risk percentiles, and are not included in the numerator and denominator of the integrated risk weighting calculation. This module implements the conditional definition of integrated risk at the system level, and the "inapplicable" state is propagated throughout the entire system chain: from the inapplicable marking in the CAT running module, to the non-generation of risk percentiles, and then to the exclusion of the numerator and denominator in the integrated calculation.

[0041] In the aforementioned system, the question parameter estimation module is specifically used for: for each item in the domain, deleting the current item from the domain response matrix, constructing an item-rest surrogate score using the mean of valid responses from the remaining items and standardizing it to obtain the surrogate latent variable; constructing a binary variable for each ordered response threshold of the item, fitting a cumulative logistic model, and obtaining the discrimination parameter and multi-level threshold parameters from the model coefficients; during the fitting process, at least one robust processing is also performed: ridge regularization of the slope term, deleting threshold fittings with insufficient sample size, limiting the threshold to a preset range, monotonically sorting the thresholds, or generating a backoff threshold based on the empirical cumulative proportion when the threshold fitting fails. This module achieves a stable conversion from school census data to a parameterized question bank at the system level, avoids homogeneity bias through item-rest design, and addresses various non-ideal situations that may occur in real data through multiple robust processing mechanisms.

[0042] This system can be extended to other areas of children's behavioral health with exposure prerequisites or applicable conditions, such as sleep problems, attention problems, emotional problems, insufficient physical activity, or risks associated with social media use. In other embodiments, question parameter estimation can be replaced by marginal maximum likelihood, Bayesian estimation, generalized partial scoring models, or multidimensional IRT models; online estimation can employ EAP, MAP, MLE, or particle approximation; and question selection strategies can be based on maximum information content, superimposed with exposure rate control, content balancing, or domain coverage constraints. Any implementation that still includes exposure conditional calibration, domain applicability determination, domain percentile mapping, and comprehensive risk calculation for applicable domains falls under the scope of this method and system.

[0043] Compared with existing technologies, the advantages of this invention are as follows: In the offline stage, only actual exposed samples are used to calibrate the parameters; in the online stage, non-exposed individuals are directly marked as inapplicable and skipped. This process ensures that the scale is established on the correct reference population during both calibration and assessment, fundamentally solving the structural misjudgment problem of non-exposed individuals in behavioral addiction screening. First, theta of each domain is mapped to the risk percentile in the school population of that domain; then, weighted summation is performed using only the set of applicable domains of the subjects as the denominator. Inapplicable domains are not estimated, involved, or contaminated. This solves the technical problem of incomparability across domains and ensures the reasonable interpretability of the comprehensive indicators. Prioritizing accuracy and providing a safety net for the number of questions, this allows children with different risk levels to receive differentiated assessment lengths, balancing measurement quality and implementation efficiency in large-scale school screening scenarios. Integrating multiple robust processing methods such as ridge regularization, threshold truncation, monotonic sorting, and backoff, this ensures that even non-ideal census data can stably produce a parameterized question bank that can be used by CAT. Attached Figure Description

[0044] Figure 1 This is the core method flowchart of the present invention; Figure 2 This is a flowchart of the IRT question bank calibration and parameterization construction process of the present invention; Figure 3 This is the exposure condition calibration logic diagram of the present invention; Figure 4 This is a detailed flowchart of the computerized adaptive test execution of the present invention. Detailed Implementation

[0045] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. Example 1

[0046] like Figure 1-4 As shown, the basic database in this embodiment is derived from pre-adolescent school-age children collected in school settings. It is a population-based database, rather than a clinical-based database primarily based on outpatient cases. This database is used to form a reference for the distribution of behavioral addiction risk among school children and to establish an IRT parameter library suitable for early warning.

[0047] This embodiment includes four domains: food addiction, internet addiction, short video addiction, and gaming addiction. The food addiction question bank contains 35 items with a response level of 5, belonging to the default applicable domain, and is calibrated using all 8314 valid school samples. The internet addiction question bank contains 20 items with a response level of 5, belonging to the default applicable domain, and is calibrated using all 8314 valid school samples. The short video addiction question bank contains 17 items with a response level of 5, belonging to the domain with exposure preconditions; its behavioral exposure variable is whether or not children watch short videos, and only 4319 children with actual short video exposure were used as the calibration sample. The gaming addiction question bank contains 9 items with a response level of 2, belonging to the domain with exposure preconditions; its behavioral exposure variable is whether or not children play online games, and only 3258 children with actual gaming exposure were used as the calibration sample.

[0048] A unified sequential double-threshold stopping rule is adopted for each domain: the minimum precision judgment number of questions is 3, the standard error threshold is 0.4, and the number of questions that must be stopped in a domain is half of the total number of questions in the domain's question bank, rounded up. The mandatory stopping limit for the food addiction domain is 18 questions, for the internet addiction domain it is 10 questions, for the short video addiction domain it is 9 questions, and for the gaming addiction domain it is 5 questions.

[0049] Examples of median results for the item discrimination parameters are as follows: The median discrimination for the food addiction domain is 1.36, with a peak domain test information of 18.48; the median discrimination for the internet addiction domain is 1.49, with a peak domain test information of 13.30; the median discrimination for the short video addiction domain is 1.80, with a peak domain test information of 16.76; and the median discrimination for the gaming addiction domain is 1.48, with a peak domain test information of 4.49. These values ​​represent one embodiment of the current prototype data, used to illustrate the executable parameterized item bank construction process of this invention. They do not constitute a limitation on the scope of protection. In formal implementation, the parameters can be recalibrated based on new samples, revised items, or different regional populations. Example 2

[0050] This embodiment describes the specific execution process of exposure-conditional calibration. The system first establishes a domain configuration for each behavioral domain. The domain configuration includes at least a domain identifier, a list of questions, a response level range, whether an exposure prerequisite exists, the name of the exposure variable, the exposure value, the minimum effective response ratio, and CAT stopping parameters. The CAT stopping parameters include at least a minimum precision judgment number of 3 questions, a standard error threshold of 0.4, and a domain-forced stopping number calculated by rounding up half the total number of questions in the domain's question bank.

[0051] For default application domains such as food addiction and internet addiction, which lack exposure prerequisites, the system constructs a calibration matrix from all samples that meet the valid response criteria. Let D_s be the original sample set of domain s, τ_s be the minimum valid response ratio, and C_s be the calibration sample set of all samples that meet the requirement that the valid response ratio is not less than τ_s, i.e., C_s = {r in D_s :valid_ratio(r,s)>= τ_s}.

[0052] For domains with exposure prerequisites, such as short video addiction and gaming addiction, the system first filters actual exposure samples based on exposure variables, and then performs valid response judgment and question parameter estimation on the exposure samples. Let E_s be the exposure variable for this domain, v_s be the exposure value, and the calibration sample set C_s be the samples that simultaneously satisfy the exposure variable value equal to the preset exposure value and the valid response ratio not less than τ_s, i.e., C_s = {r in D_s : E_s(r)=v_s and valid_ratio(r,s)>= τ_s}.

[0053] Samples that do not meet the exposure criteria are not included in the calibration sample set for the corresponding domain. This process ensures that children who have not participated in the short videos or games are not mistakenly input into the model as low-risk samples during the calibration phase, thereby reducing parameter bias caused by differences in exposure structure. Example 3

[0054] This embodiment describes a multi-level item parameter estimation process based on item-rest surrogate latent variables. For item i in domain s, the system first removes the current item i from the domain response matrix and constructs the item-rest surrogate score using the mean of valid responses from the remaining items. Let R_s be the set of items in domain s, Y_rj be the response score of sample r on item j, and the surrogate score M_ri be the mean of valid responses from all items except item i: M_ri = mean{Y_rj : j in R_s, j != i, Y_rj valid}. Then, the mean and standard deviation of M_i are calculated in the calibration sample set C_s of domain s, and the surrogate score is standardized to obtain the surrogate latent variable θ_ri(-i) = [M_ri - mean(M_i)] / sd(M_i).

[0055] For a multi-level entry with m+1 ordered response levels, the system constructs a binary variable Z_rik = 1 (Y_ri>= k), k=1,...,m, for each threshold k, and fits a cumulative logistic model: logit P(Z_rik=1 | θ_ri(-i)) = α_ik + β_ik θ_ri(-i). The difficulty parameter b_ik = -α_ik / β_ik corresponding to the threshold, and the discriminative parameter a_i = median{β_ik : β_ik satisfies the validity condition}, are obtained from the model coefficients. The cumulative probability is P_i(Y_i>= k | θ) = 1 / [1 + exp{-a_i(θ - b_ik)}].

[0056] For binary categories in the field of gaming addiction, the above model degenerates into a two-parameter logistic model with a single threshold.

[0057] During the fitting process, the system performs several robust processing steps: ridge regularization on the slope term, deletion of threshold fittings with insufficient sample numbers, limiting the threshold to a preset theta range, monotonically sorting the thresholds, or generating a backoff threshold based on empirical cumulative proportions when threshold fitting fails. For example, the logistic model uses IRLS iteration and sets a ridge penalty on the slope term. If the number of samples in both classes on both sides of a certain threshold is not less than a preset lower limit and the slope is greater than the lower limit, the threshold fitting result is retained; otherwise, empirical proportion backoff is initiated.

[0058] The system calculates the probability of each response category based on the item discrimination parameter and the threshold parameter. First, the cumulative probability S_k(θ) = 1 / [1 + exp{-a_i(θ - b_ik)}] is calculated. Then, the category probability is obtained from the difference between adjacent cumulative probabilities: P_i(Y=0|θ) = 1 - S_1(θ), P_i(Y=k|θ) = S_k(θ) - S_{k+1}(θ), k=1,...,m-1, P_i(Y=m|θ) = S_m(θ). The item information function is I_i(θ) = sum_k [P'_ik(θ)]^2 / P_ik(θ).

[0059] The system calculates the information function and domain test information curve for each item on a preset theta grid, and writes the item stem, option label, original code, discrimination index, threshold, calibration sample size, missing rate, information peak, domain applicability rules, CAT stopping parameters, risk threshold, percentile lookup table, and version information into a machine-readable item base file. The CAT stopping parameters include minimum_items=3, se_threshold=0.4, and maximum_items=ceil(N_s / 2). Example 4

[0060] This embodiment describes the execution process of a sequential dual-threshold stop rule. During online assessment, the system executes CAT sequentially by domain. For domains with exposure prerequisites, it first reads existing exposure variables or poses applicability questions to the subjects. If the subjects do not meet the exposure conditions for that domain, the domain is marked as inapplicable and skipped, moving on to the next domain. For applicable domains, the system initializes the potential risk value to 0 and the standard error to 1, or initializes it based on prior records.

[0061] For the applicable domain, the system calculates the posterior distribution on the theta grid. Let the prior π(θ) be a standard normal distribution or an empirical prior determined by calibration samples. For the set of answered items H, the system accumulates the class probability of each answered response in log space to obtain the posterior weight, and uses EAP as the latent risk value estimate, with the posterior standard deviation as the standard error: logposterior_g = log π(θ_g) + sum_{j in H} log P_j(y_j | θ_g), w_g = exp(logposterior_g - max_h log posterior_h) / sum_h exp(log posterior_h - max_h logposterior_h), θ_hat = sum_g θ_g w_g, SE = [sum_g (θ_g - θ_hat)^2 w_g]^(1 / 2).

[0062] After each response, the system calculates the information content of all unanswered items at the current θ_hat. Let U be the set of unanswered items. Select the item with the largest information content i* = argmax_{i in U} I_i(θ_hat) and present it to the subject. After receiving the response, update H, θ_hat and SE.

[0063] Let n_s be the number of questions answered in the current domain, SE_s be the posterior standard deviation of this round of EAP, and N_s be the total number of questions in the question bank for this domain. The mandatory stopping limit is M_s = ceil(N_s / 2). After each answer and EAP update, the system executes the stopping judgment according to the following priority: First priority: If n_s≥3 and SE_s≤0.4, then stop testing in this area if the accuracy target is met, and record the reason for stopping as "accuracy target met".

[0064] Second priority: Only when the aforementioned accuracy stopping condition is not satisfied, if n_s≥M_s, regardless of whether SE_s has dropped below 0.4, the test in this domain is forcibly stopped at the upper limit of the number of questions, and the stop reason is recorded as reaching the upper limit of the number of questions. At this time, if SE_s is still greater than 0.4, the system retains the final potential risk value and standard error, and marks that the estimation uncertainty is high, so as to prompt for subsequent retest or supplementary assessment.

[0065] Continuing condition: If n_s<3, or n_s≥3 but SE_s>0.4, and n_s<M_s, the system calculates the information amount I_i(θ_hat) of all unanswered items at the current θ_hat, selects the item with the highest information amount as the next question, and enters the next iteration. The above judgment order cannot be exchanged. Embodiment 5

[0066] This embodiment is a specific execution process. After the domain CAT is completed, the system does not directly compare the potential risk values of each domain, but maps θ_s to the risk percentile p_s = F_s(θ_s) in the empirical distribution F_s of the calibration samples of this domain. This percentile represents the relative risk position of the subject in the calibration samples of the same domain and the same applicable population.

[0067] Let A be the set of domains applicable to the current subject, A = {s: domain s is applicable to the current subject and CAT has been completed}, and the domain weight w_s can be set to 1 by default. The composite risk percentile P_composite = round( sum_{s in A}w_s p_s / sum_{s in A} w_s ), and the highest risk domain s_max = argmax_{s in A} p_s. Domains marked as not applicable do not estimate θ_s, do not generate p_s, do not enter the numerator and denominator of P_composite, and are not involved in the ranking as low-risk domains. Embodiment 6

[0068] In this embodiment, the system generates a traceable result package after the assessment is completed. After the system completes the assessment of all applicable domains, in addition to outputting the composite risk percentile, domain-specific risk percentiles, risk levels and domain rankings, it also generates a complete result package in one-to-one correspondence with this assessment, which includes the following six types of information.

[0069] The first category contains parameterized question bank information for each domain, including at least the item number, question stem, options, discrimination parameters, multi-level threshold parameters, calibration sample size, and information content. The second category is a list of individual CAT (Computer-Aided Instruction) responses, recorded in presentation order, including the item, response level, original answer, and response time. The third category is item-by-item measurement information, recording at least the potential risk value before item selection, the standard error before item selection, the information content of the item at the current potential risk value, the potential risk value after answering, the standard error after answering, and the cumulative information content of answered items. The fourth category is the individual measurement information function curve, obtained by summing the information functions of the actual responses point by point. The fifth category is the exposure applicability judgment rules for each domain and the judgment results for this instance, explaining whether each domain has the exposure prerequisite, the judgment question used, the subject's response for this instance, and the final applicability status. The sixth category is a risk percentile reference table for each exposure domain, illustrating the mapping relationship between the domain's potential risk value and the corresponding calibration sample risk percentile.

[0070] The results can be displayed on the front-end page or exported as machine-readable JSON, printed reports, or server-side structured records to support scientific research review, individual interpretation, model re-evaluation, and version auditing. Example 7

[0071] This embodiment illustrates the system's anomaly handling and extended implementation methods. If a subject withdraws midway, the system saves the current answer state and supports resuming the test at the breakpoint. If the answer time is abnormally short, or the answer pattern exhibits obvious linear or random answer characteristics, the system marks the answer as low confidence and prompts for a retest. If the test is forcibly stopped due to reaching half of the total number of questions in the domain question bank (rounded up) and the standard error is still higher than 0.4, the system outputs an estimation uncertainty prompt and suggests a subsequent retest or supplementary assessment, but the online CAT will not continue to exceed half of the domain question bank limit.

[0072] In other embodiments, the topic parameter estimation can be replaced by marginal maximum likelihood, Bayesian estimation, generalized partial scoring model, or multidimensional IRT model; online estimation can employ EAP, MAP, MLE, or particle approximation; the topic selection strategy can be based on maximum information content, superimposed with exposure rate control, content balancing, or domain coverage constraints. Any implementation that still includes exposure conditional calibration, domain applicability determination, domain percentile mapping, and comprehensive risk calculation for applicable domains falls under the category of variations of the inventive concept.

[0073] The system can be deployed as a web page, mini-program, mobile device, local offline program, or server-side API, and can also be extended to areas such as sleep, mood, attention, insufficient physical activity, or other behavioral health risks in children.

[0074] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

[0075] Although this paper uses terms such as exposure condition calibration module, parameter estimation module, CAT operation module, and risk synthesis module frequently, the possibility of using other terms is not excluded. These terms are used merely for the convenience of describing and explaining the essence of this invention; interpreting them as any additional limitation would contradict the spirit of this invention.

Claims

1. An exposure-conditioned IRT-CAT method for early warning of behavioral addiction risks in pre-adolescent children, characterized in that, include: S1: Obtain multi-domain item response data and behavioral exposure variables corresponding to at least one domain; S2: Determine whether each domain has exposure prerequisites based on the domain configuration. For domains with exposure prerequisites, use the samples that meet the exposure conditions and have a high effective response rate as the calibration samples for that domain. For areas that do not have the prerequisite for exposure, all samples that meet the valid response criteria are used as calibration samples; S3: For the calibration samples in each field, for each item, construct surrogate latent variables based on the responses of other items after excluding the current item, and fit the model to obtain the discrimination parameter and threshold parameter of the item, and generate a parameterized item bank; S4: During online assessment, the applicability of each subject to the subject is judged on a domain-by-domain basis: for domains with exposure prerequisites, if the subject does not meet the exposure conditions for that domain, the domain is marked as unsuitable and skipped; For applicable fields, perform computerized adaptive testing; S5: Perform computerized adaptive testing, select the item with the most information based on the current potential risk value, update the potential risk value and standard error after receiving the answer, and perform sequential double threshold stop judgment. Stop the test in this area when the preset accuracy condition or number of questions is met. S6: After the domain testing is completed, the final potential risk value of the domain is mapped to the risk percentile in the empirical distribution of the calibration samples in that domain; S7: Using the set of applicable fields of the current subjects as the denominator, perform a weighted summation of the risk percentiles of each applicable field, and output the overall risk percentile and the highest risk field.

2. The exposure-conditioned IRT-CAT method for early warning of behavioral addiction risks in pre-adolescent children according to claim 1, characterized in that, In step S2, for areas with exposure prerequisites, samples that meet the exposure conditions and have a high effective response rate are used as calibration samples, specifically including: For the default applicable domain, the calibration sample set C_s is all samples that satisfy the requirement that the effective response ratio is not less than the preset threshold τ_s; For domains with exposure prerequisites, the calibration sample set C_s is a set of samples that simultaneously satisfy the condition that the exposure variable value is equal to the preset exposure value and the effective response ratio is not lower than the preset threshold τ_s. Samples that do not meet the exposure conditions will not be included in the calibration sample set for the corresponding field.

3. The exposure-conditioned IRT-CAT method for early warning of behavioral addiction risks in pre-adolescent children according to claim 1, characterized in that, In step S3, the proxy latent variable is obtained by calculating the mean of valid responses to other entries in the domain calibration sample excluding the current entry and then standardizing it; the fitting model is a fitted cumulative logistic model; at least one robust processing is also performed during the fitting process: ridge regularization of the slope term, deletion of threshold fitting with insufficient sample size, limiting the threshold to a preset range, monotonically sorting the threshold, or generating a backoff threshold based on the empirical cumulative ratio when the threshold fitting fails.

4. The exposure-conditioned IRT-CAT method for early warning of behavioral addiction risks in pre-adolescent children according to claim 1, characterized in that, In step S5, the sequential dual-threshold stop determination has the following priority: The first priority is to determine if the number of questions answered is not less than the minimum precision judgment number and the standard error is not greater than the preset precision threshold. If these conditions are met, the process stops when the precision target is met. If the aforementioned conditions are not met, the second priority check will be performed to determine whether the number of questions answered has reached half of the total number of questions in the question bank for that field and then rounded up. If the condition is met, the process will be forcibly stopped at the maximum number of questions. If neither of the two conditions is met, continue selecting a topic.

5. The exposure-conditioned IRT-CAT method for early warning of behavioral addiction risks in pre-adolescent children according to claim 1, characterized in that, In step S7, the domains marked as unsuitable do not generate risk percentiles, nor are they included in the numerator and denominator of the comprehensive risk calculation; the risk percentiles are determined by the empirical distribution of the domain calibration samples; the weighted aggregation adopts an equal-weighted average or a weighted average method.

6. An exposure-conditioned IRT-CAT system for early warning of behavioral addiction risks in pre-adolescent children, characterized in that, include: The exposure condition calibration module is used to determine whether each domain has an exposure prerequisite based on the domain configuration. For domains with exposure prerequisites, samples that meet the exposure conditions and have a qualified response rate are used as calibration samples. For areas that do not have the prerequisite for exposure, all samples that meet the valid response criteria are used as calibration samples; The item parameter estimation module is used to construct surrogate latent variables for each item based on the responses of other items after excluding the current item, and to fit the model to obtain the item's discrimination parameter and threshold parameter, thereby generating a parameterized item bank. The CAT operation module is used to perform applicability judgment on subjects domain by domain during online assessment: for domains with exposure prerequisites, if subjects do not meet the exposure conditions, they are marked as inapplicable and skipped; For applicable fields, computerized adaptive testing is performed: the item with the most information is selected and presented based on the current potential risk value, the potential risk value and standard error are updated after receiving the answer, and a sequential double threshold stop judgment is performed, stopping when the preset accuracy condition or number of questions is met; The risk synthesis module maps the final potential risk value of each applicable field to the risk percentile in the corresponding empirical distribution of the calibration sample, and performs a weighted synthesis of the risk percentile of each applicable field with the set of applicable fields as the denominator, outputting the comprehensive risk percentile and the highest risk field.

7. The exposure-conditioned IRT-CAT system for early warning of behavioral addiction risks in pre-adolescent children according to claim 6, characterized in that, The exposure conditioned calibration module is specifically used for: For the default applicable domain, a calibration sample set is constructed using all samples that meet the requirement that the proportion of valid responses is not less than a preset threshold; For domains with exposure prerequisites, a calibration sample set is constructed using samples that simultaneously satisfy the conditions that the exposure variable value is equal to the preset exposure value and the effective response ratio is not lower than the preset threshold. Samples that do not meet the exposure conditions will not be included in the calibration sample set for the corresponding field.

8. The exposure-conditioned IRT-CAT system for early warning of behavioral addiction risks in pre-adolescent children according to claim 6, characterized in that, In the sequential dual-threshold stop judgment executed by the CAT operation module, the first judgment condition is that the number of answered questions is not less than the minimum precision judgment number of questions and the standard error is not greater than the preset precision threshold. This condition has priority. Only when this condition is not met will the second judgment condition be entered, that is, the number of answered questions reaches half of the total number of questions in the question bank of this field and is rounded up. If neither of the two conditions is met, the next item will continue to be selected according to the maximum information content.

9. The exposure-conditioned IRT-CAT system for early warning of behavioral addiction risks in pre-adolescent children according to claim 6, characterized in that, In the risk integration module, areas marked as unsuitable are not estimated for potential risk values, are not generated for risk percentiles, and are not included in the numerator and denominator of the integrated risk weighting calculation.

10. The exposure-conditioned IRT-CAT system for early warning of behavioral addiction risks in pre-adolescent children according to claim 6, characterized in that, The parameter estimation module is specifically used for: For each entry in the domain, the current entry is removed from the domain response matrix, and the proxy score is constructed using the mean of the valid responses of the remaining entries and then standardized to obtain the proxy latent variable. For each ordered response threshold of an item, construct a binary variable, fit a cumulative logistic model, and obtain the discrimination parameter and multi-level threshold parameter from the model coefficients; During the fitting process, at least one robust processing is also performed: ridge regularization of the slope term, deletion of threshold fitting with insufficient sample size, limiting the threshold to a preset range, monotonically sorting the thresholds, or generating a backoff threshold based on the empirical cumulative ratio when the threshold fitting fails.

Images