Method for generating an aspartic acid-arginine hearing loss anxiety-related score

Abstract

This invention discloses a method for generating an aspartic acid-arginine deafness anxiety association score, relating to the field of bioinformatics. The method involves an electronic device acquiring the absolute quantitative concentrations of aspartic acid and arginine, along with sample number, batch number, concentration unit, quantitative range, and quality control status output by a testing institution for isolated serum samples. The concentrations are converted to reference units fixed in the current model version, and field, range, quality control, and applicable conditions are verified. A dual metabolite input vector is formed with aspartic acid first and arginine second. A pre-trained ExtraTrees extremely random forest model is used to generate an anxiety tendency association score ranging from 0 to 1. During model construction, the dual metabolite combination is determined based on SHAP importance ranking and Top-N combination comparison. When the model version is bound to a threshold, the method outputs "above / below / review prompt," and saves the model version and processing records.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics processing and machine learning, specifically to a method for generating an aspartic arginine-associated hearing loss anxiety score. Background Technology

[0002] Hearing is a vital ability for individuals to perceive sound, communicate, and acquire environmental information. Hearing loss (HL) refers to varying degrees of hearing impairment when an individual's hearing threshold is above the normal range. Due to auditory communication barriers, difficulties in understanding and completing scales, accompanying cognitive or linguistic factors, and the availability of assessment resources, deaf individuals may face challenges in completing standardized psychological scales or undergoing continuous follow-up, and relevant anxiety tendency information may not be readily available. Therefore, a technological solution is needed that can utilize objective testing data to generate information processing results for professional review.

[0003] In recent years, metabolomics and mass spectrometry detection technologies have developed rapidly. The concentration of small molecule metabolites in isolated serum samples can provide data related to stress, neurotransmitter metabolism, energy metabolism, and inflammation regulation. Targeted quantitative techniques such as UPLC-MS / MS can obtain absolute quantitative concentrations of metabolites through standard curves, internal standard correction, and quality control samples. Raw detection data usually contain many candidate metabolites. Directly using multi-indicator overall modeling may increase the workload, input dimensionality, and data processing complexity, and allow redundant information to affect the model's cross-cohort performance.

[0004] Existing solutions for processing anxiety predisposition information based on biomarkers or machine learning may still suffer from issues such as a large number of input indicators, unclear combinations of key metabolites, lack of standardization of batch fields and model preprocessing rules, and a lack of binding between model output and model version and threshold version. For deaf individuals, a more comprehensive information processing solution is needed to clearly define the boundaries of data quality, model version, and result interpretation. This solution should address how to convert the quality-controlled absolute quantitative concentrations of serum aspartic acid and arginine to a unified reference unit and arrange them in a fixed order to form a low-dimensional input, followed by the generation of continuous association scores using a traceable ExtraTrees extremely random forest model. Summary of the Invention

[0005] (a) Technical problems to be solved To address the aforementioned issues, this invention fixes serum aspartic acid and serum arginine as the dual metabolite inputs for generating the anxiety tendency association score for deaf subjects. The input vector is formed according to the reference units, unit conversion rules, quality control rules, and feature order bound to the model version, and then input into the ExtraTrees extremely random forest model to generate association scores ranging from 0 to 1. The system simultaneously saves model parameters, threshold versions, and quality control status; when a threshold is bound to the model version, it generates prompts indicating above the threshold, below the threshold, or pending review. This reduces redundant metabolite inputs, strengthens the correspondence between the two key metabolites and the model version, and improves the ability to verify and trace results.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: The aspartic-arginine hearing loss anxiety-associated score generation method, executed by an electronic device, includes: The absolute quantitative concentration data of serum metabolites from the isolated serum samples of the deaf subjects were obtained from the testing institution. The units of serum aspartic acid concentration and serum arginine concentration were unified according to the reference units and unit conversion rules fixed in the current model version. The sample number, test batch, quantitative range, quality control status and applicable conditions of the model version were verified. When both concentrations were valid, the dual metabolite input vector was formed according to the characteristic order of aspartic acid first and arginine second fixed in the model version. The dual metabolite input vector is input into a pre-trained ExtraTrees extremely random forest scoring model to generate an anxiety tendency association score for deaf individuals with values ​​ranging from 0 to 1. Save sample number, test batch, input concentration, quality control status, feature order, model version, and generation time; when the current model version is bound to a scoring threshold, further generate prompts for above the threshold, below the threshold, or pending review; the associated score is not interpreted as the absolute incidence probability in the population, and the output result does not directly constitute a medical diagnostic conclusion.

[0007] Furthermore, the aspartic acid and arginine corresponding to the dual metabolite input vectors are determined as follows: During the model building phase, a candidate metabolite matrix including the concentrations of various candidate serum metabolites is obtained, and input quality control is performed on the candidate metabolite matrix; the ExtraTrees candidate model is trained based on the internal training set, and the Shapley additive interpretability importance score of each candidate serum metabolite is calculated. The top N candidate metabolite combinations are formed by ranking them from highest to lowest according to the importance score of Shapley additive interpretation, where N is 1 to 10. For each of the top N candidate metabolite combinations, the ExtraTrees candidate model was retrained on the internal training set, and the area under the receiver operating characteristic curve, area under the precision-recall curve, sensitivity, specificity, and balanced accuracy were calculated on the internal validation set and on data labeled as the external evaluation cohort. Aspartic acid and arginine were selected as the dual metabolite input combination according to the priority order of the area under the receiver operating characteristic (ROC) curve for external evaluation, the area under the precision-recall curve for external evaluation, the area under the ROC curve for internal validation, the area under the precision-recall curve for internal validation, and the number of input features. Among these, the larger value of the first four items was given priority, and if the first four items were the same, the item with fewer input features was given priority. When external evaluation indicators were involved in the combination determination, the external evaluation queue was not used as the final independent external validation queue.

[0008] Furthermore, after data quality verification, the discovery queue is divided into an internal training set and an internal validation set in a stratified random manner. The external evaluation queue has no sample overlap with the discovery queue and does not participate in feature preprocessing parameter fitting, ExtraTrees candidate model training, Shapley additive interpretation importance ranking, probability calibration function fitting, or scoring threshold selection. When external evaluation indicators are used in candidate combination comparison, the relevant results are recorded as external evaluation results and are not used as the final independent external validation results.

[0009] Furthermore, the concentration data is obtained by the testing institution using a validated ultra-high performance liquid chromatography-tandem mass spectrometry or an equivalent absolute quantification method; the equivalent absolute quantification method meets the acceptance criteria of the model version solidification in terms of linear range, accuracy, precision, recovery, minimum quantitation limit, maximum quantitation limit, and inter-method deviation; when aspartic acid or arginine is below the minimum quantitation limit, above the maximum quantitation limit, or fails the quality control status, the sample is marked as pending verification and prohibited from entering the formal correlation scoring generation process.

[0010] Furthermore, the main model version uses the original numerical scale. After converting the input concentration to the reference unit fixed in the model version, only field name verification, feature order verification, and quality control status verification are performed. Missing value imputation, logarithmic transformation, z-score standardization, or rank-Gaussian transformation are not performed. If any core input is missing, exceeds the quantitative range, or fails the quality control status, a "pending review" prompt is output, and no formal correlation score is generated. When only the field name or record format changes and the input value remains unchanged, interface consistency verification is re-executed. When the relationship between input values ​​changes due to unit mapping, standardization, logarithmic transformation, outlier replacement, or changes in the detection platform, a new model version is generated, and the model is retrained, the scoring threshold is determined, and the model performance is evaluated.

[0011] Furthermore, the ExtraTrees extremely random forest scoring generation model includes multiple extremely random decision trees. Each extremely random decision tree randomly selects candidate features and candidate segmentation thresholds, outputs a single tree score for the positive category of anxiety tendency, and integrates the single tree scores of multiple trees to obtain the anxiety tendency association score of the deaf subject.

[0012] Furthermore, the model version file should at least store the number of trees, maximum depth, minimum number of split samples, minimum number of leaf node samples, number of candidate features for node splits, classification criteria, class weights, whether bootstrap sampling is enabled, class encoding rules, positive class index, random seed for data partitioning, random seed for model training, feature order, training data version, software environment version, model file hash value, threshold version, and model version number; association scores that have not been independently probability-calibrated are used as 0 to 1 model output scores and are not interpreted as the absolute probability of disease incidence in the population.

[0013] Furthermore, when the current model version is bound to a scoring threshold, the associated score is compared with the scoring threshold, and a prompt indicating above or below the threshold is generated; when the quality control status, quantitative range, or applicable conditions fail, a prompt indicating that review is pending is generated; the scoring threshold is determined based on the internal validation set and is fixed with the model version. During the deployment phase, the scoring threshold is not redefined based on a single test sample or an external evaluation queue; when the threshold is adjusted, a new threshold version is generated while the original threshold record is retained.

[0014] An aspartic acid-arginine-based deafness anxiety-related scoring system includes: The data acquisition module is used to acquire the serum aspartic acid concentration, serum arginine concentration, sample number, test batch, concentration unit, lower limit of quantitation, upper limit of quantitation, and quality control status of the isolated serum samples of the deaf subjects to be tested by the testing institution. The quality control verification module is used to unify units according to the reference units and unit conversion rules fixed in the current model version, and to verify the field integrity, quantitative range and quality control status of the concentration data. The vector generation module is used to generate a dual metabolite input vector according to the characteristic order of aspartic acid first and arginine last, which is fixed in the current model version, when both core concentrations are valid. The applicability verification module is used to verify the sample type, applicable labels for the test object, detection platform information, and the applicability scope of the current model version. The ExtraTrees Extremely Random Forest score generation module is used to call the pre-trained ExtraTrees Extremely Random Forest model to generate anxiety tendency association scores ranging from 0 to 1 based on the dual metabolite input vector. The results output module is used to output the anxiety tendency association score, and when the model version is bound to a scoring threshold, it outputs a prompt indicating above the threshold, a prompt indicating below the threshold, or a prompt indicating that it needs to be reviewed. The model version management module is used to save reference units, unit conversion rules, model parameters, feature order, preprocessing parameters, threshold version, training data version, software environment version, model file hash value, and model version number. The traceability record module is used to save the sample number, detection batch, input concentration, quality control status, feature order, model version, threshold version, correlation score, prompt result, and generation time for each correlation score generation process.

[0015] An electronic device includes a processor, a memory, and a communication interface. The memory stores a computer program, reference units, unit conversion rules, feature order, feature preprocessing parameters, and an ExtraTrees extremely random forest model file. When the processor executes the computer program, it implements the association score generation method.

[0016] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the associated score generation method.

[0017] (III) Beneficial Effects This invention provides a method for generating an aspartic acid-arginine deafness anxiety association score, which has the following beneficial effects: By fixing the absolute quantitative concentrations of serum aspartic acid and serum arginine output by the testing agency as dual metabolite inputs, and verifying the sample number, testing batch, concentration unit, quantitative range, and quality control status, it is possible to avoid low-quality or mismatched data from entering the model, making the model input and its source verifiable and traceable.

[0018] By performing SHAP importance ranking and Top-N forward selection on ExtraTrees, aspartic acid and arginine are solidified into a low-dimensional input combination, which can reduce the input of redundant metabolites, the detection burden and model complexity, and highlight the role of the two key metabolites in the generation of anxiety tendency risk in deaf subjects.

[0019] By binding fixed feature order, preprocessing rules, model parameters, and threshold versions, the ExtraTrees model can also obtain continuous scores of the same or similar values ​​within the range of 0 to 1. When the threshold is validated and bound to the model version, it will also generate prompts for scores above or below the threshold, or prompts for manual review. The output records are not used as medical diagnostic conclusions alone but can be used for model version verification, batch tracking, and audit trails. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall architecture of the anxiety tendency association scoring system for deaf individuals according to the present invention; Figure 2 This is a schematic diagram of the method for generating an anxiety tendency association score for deaf individuals according to the present invention; Figure 3 This is a schematic diagram of the concentration data acquisition and quality control verification process of the testing agency in this invention; Figure 4 This is a schematic diagram of the screening and external evaluation process for dual metabolite candidate combinations in this invention; Figure 5 This is a schematic diagram illustrating the process of generating association scores and optional binary classification prompts using the ExtraTrees model in this invention; Figure 6 This is a schematic diagram illustrating model evaluation, threshold determination, and data usage boundaries in this invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figures 1-6 This invention provides a method for generating an aspartic acid-arginine deafness anxiety association score, including the following: The method processes absolute quantitative concentration data of serum metabolites output by testing institutions for isolated serum samples. The core inputs are serum aspartate concentration and serum arginine concentration. This method does not include blood collection procedures performed on humans, nor does it focus on the specific experimental procedures within testing institutions. The model output is an anxiety tendency correlation score, which is not interpreted as an absolute probability of disease in the population and does not directly constitute a medical diagnostic conclusion.

[0023] The anxiety tendency referred to in this scheme is a supervisory label defined by the GAD-7 labeling rules pre-fixed during the model construction phase. The current model version is applicable to deaf subjects with a complaint of hearing loss, medical records, or follow-up records, who meet the inclusion and exclusion criteria of the model version. The model version records at least the age range, the basis for determining hearing loss, the side of hearing loss, the course of the disease, the severity, the type of audiology, the status of hearing aid or cochlear implant use, the use of psychiatric medications, and the exclusion criteria for major metabolic diseases; if the subject lacks necessary applicable information or exceeds this range, only a prompt indicating that the applicable range needs to be reviewed will be output. If extended to other populations, the corresponding sample criteria, model version, preprocessing parameters, scoring thresholds, and evaluation results should be re-established. Aspartic acid and arginine are not two arbitrarily selected metabolites, but rather a fixed combination of two metabolites determined based on the candidate metabolite matrix, calculated in the ExtraTrees candidate model using SHAP importance scores, and after performing Top-N forward selection. The external queue metrics in this embodiment were previously used to compare the external performance of candidate combinations. Therefore, the relevant data are uniformly referred to as "external evaluation results" and are not presented as the final independent external validation results that have not been affected by model selection. Before formal deployment, it is advisable to conduct independent validation with a new queue that has not participated in the screening.

[0024] During the main model construction, the candidate input consists of 163 serum metabolites that can be matched by name in both the discovery and external evaluation queues. The 205 cases in the discovery queue are randomly divided into an internal training set and an internal validation set in a 7:3 stratified ratio, with a random seed of 2024. The main model can employ ExtraTrees extremely random forest, with a random seed of 20260515 for model training. The main model preprocessing strategy is raw; after converting the input concentrations to the reference units fixed in the model version, only field name validation, feature order validation, and quality control status validation are performed, without performing missing value imputation, log2 transformation, z-score standardization, or rank-Gauss transformation.

[0025] In this embodiment, serum aspartic acid concentration is denoted as C_Asp, and serum arginine concentration is denoted as C_Arg. Both are concentration values ​​output by the testing institution using a validated absolute quantitative method and having passed quality control. The unique concentration unit recorded in both the training data file and the model version file is defined as the model reference unit.

[0026] In the main model, the input feature value F_Asp of the aspartic acid model is equal to C_Asp after being converted to the model reference unit and passed quality control verification, and the input feature value F_Arg of the arginine model is equal to C_Arg after being converted to the model reference unit and passed quality control verification; if any core concentration is missing or invalid, no imputation is performed and no formal correlation score is generated.

[0027] The dual metabolite input vector is denoted as: In the formula: the anxiety tendency association score is denoted as S, which takes a value of 0 to 1, representing the degree of model association between the test sample and the positive anxiety tendency category under the current model version.

[0028] Considering that the calibration results shown in Table 7 still have biases, S is used as the model output score in this embodiment, and cannot be directly interpreted as the absolute incidence rate in the population, nor can it be directly used as a medical diagnostic conclusion.

[0029] The SHAP importance score is the mean of the absolute SHAP values ​​of candidate metabolites in the training samples, used to characterize their average contribution to the model output. Forward selection involves including candidate metabolites one by one according to their SHAP importance ranking, and retraining the ExtraTrees candidate model for each Top-N combination. Then, the selection strategy of combining internal validation and external evaluation metrics is compared. When external evaluation metrics are involved in the combination comparison, the relevant cohort is no longer called the final independent external validation cohort.

[0030] The overall process is as follows: S100: Obtain the absolute quantitative concentration data of serum aspartic acid and arginine output by the testing agency after quality control, as well as its sample number, test batch, concentration unit, quantitative range and quality control status. S200 verifies the completeness of fields, concentration units, quantitative ranges, quality control status, and model version applicability; if any core concentration is missing or invalid, it outputs a "required for review" prompt and terminates the formal inference. S300, the dual metabolite input vector is formed according to the fixed order of aspartic acid first and arginine last in the model version; S400, in the model building phase, calculates the SHAP importance score of candidate metabolites based on the ExtraTrees candidate model, and performs Top-N forward selection according to the SHAP ranking to determine the combination of aspartic acid and arginine dual metabolites. S500, using dual metabolite input vectors as input and anxiety tendency labels as supervision signals, trains and solidifies the ExtraTrees extremely random forest rating generation model; During the deployment phase, the S600 inputs the dual metabolite input vectors of the test samples into the pre-trained model and outputs an anxiety tendency association score for deaf subjects. S700 saves sample number, test batch, input concentration, quality control status, feature order, model version, output correlation score and optional threshold prompts to form a traceable record; S800, Based on the above method, construct an anxiety tendency association score generation system for deaf subjects and a computer-readable storage medium.

[0031] This embodiment does not limit the company name of the testing institution, the brand of the commercial reagent kit, or the complete testing SOP. The testing institution should use a validated UPLC-MS / MS or equivalent absolute quantitation method to deliver to the system the concentration values, concentration units, lower limit of quantitation (LVC), upper limit of quantitation (UPVC), sample number, batch number, and quality control status of serum aspartic acid and arginine. The equivalent method should meet the acceptance criteria of the model version in terms of linearity, accuracy, precision, recovery, limit of quantitation, and inter-method bias.

[0032] The model building phase can also accept other candidate metabolite concentrations for SHAP ranking and Top-N combination comparison; the formal inputs for the deployment phase are fixed to the concentrations of aspartic acid and arginine.

[0033] The sample type for the main embodiment of this scheme is limited to in vitro serum samples, and fecal samples, urine samples or other sample types are not used as the main embodiment. If it is extended to other sample types, the corresponding detection method, feature preprocessing parameters, feature screening results and evaluation data should be re-established.

[0034] The system performs model applicability verification, confirming at least the following: the sample type is in vitro serum; the test subjects belong to the deaf population covered by the current model version; the aspartic acid and arginine fields are complete; the input concentration can be converted to the model reference units; the batch quality control has passed; and the testing platform meets the applicability conditions. If any condition is not met, the official scoring model is not invoked; only a "review pending" message is output.

[0035] The inclusion criteria for deaf subjects are recorded according to the rules pre-defined during the model construction phase, including at least the age range, the basis for determining hearing loss, the side of hearing loss, the duration of the condition, the severity, the type of audiology, the GAD-7 scale, in vitro serum test information, the status of hearing aid or cochlear implant use, and the main exclusion criteria. Audiological information is managed as sample background information or applicable scope labels and is not a required metabolite feature input to the main model.

[0036] The system will only proceed to the formal correlation score generation process if both core metabolites are within the quantitative range confirmed by the testing institution, the quality control status is passed, and the fields are complete; otherwise, a "pending review" message will be output. Sample processing volumes, centrifugation conditions, derivatization conditions, column types, elution gradients, mass spectrometry parameters, and calibration formulas used internally by the testing institution are not the focus of this application.

[0037] The system records at least the sample number, batch number, aspartic acid concentration, arginine concentration, concentration unit, lower limit of quantitation (LQ), upper limit of quantitation (UPQ), quality control status, detection method version, and data processing version for each sample. The feature preprocessing module only receives the concentration data when all fields are complete and both core metabolites have passed quality control.

[0038] The absolute quantitative concentrations of serum aspartic acid and serum arginine were subjected to feature preprocessing that was tied to the model version to obtain the input feature values ​​of the aspartic acid model and the arginine model; different model versions should not use the same preprocessing parameters.

[0039] The main model version preferentially uses the raw numerical scale, converting the input concentration to the model reference unit only during the unit conversion stage. Subsequently, it performs field name validation, feature order validation, and quality control status validation, without performing missing value imputation, log2 transformation, z-score standardization, or rank-Gauss transformation. If any core input is missing, below the minimum quantitation limit, above the maximum quantitation limit, or fails quality control, a "pending review" message is output, and no formal correlation score is generated.

[0040] After completing quality control verification and unit unification, construct the dual metabolite input vector according to the feature order fixed during the training phase: The feature order remains aspartic acid first, followed by arginine, throughout the training, evaluation, and deployment phases. Before each inference iteration, the field name, reference unit, unit conversion factor, quantitative range, quality control status, feature order, and preprocessing parameter version are verified; if any of these are inconsistent, formal association scoring generation is halted and a "review pending" message is output.

[0041] Feature preprocessing parameters are saved as part of the model version file, and at least the model reference unit, unit conversion factor, number of significant decimal places, field order, missing value rejection rules, outlier handling rules, quality control judgment rules, and applicable detection platform information are recorded. The main model version is marked as raw and does not undergo missing value imputation, log2 transformation, z-score normalization, or rank-Gauss transformation.

[0042] The units should be uniformly calculated as C_ref=C_in×k_u, where C_in is the concentration output by the detection agency, k_u is the unit conversion factor fixed in the current model version, and C_ref is the reference unit concentration entering the model. The specific symbol, conversion factor, and number of significant decimal places of the model reference unit are saved with the model version and should not be changed temporarily during deployment.

[0043] If only the field name or record format changes while the input values ​​remain the same, the interface consistency verification should be re-executed. If new unit mapping, standardization, distribution mapping, outlier replacement, or other preprocessing methods that change the relationship of input values ​​are subsequently introduced, a new model version should be created, the model should be retrained, the scoring threshold should be determined, and the evaluation results should be recorded.

[0044] A one-to-one correspondence is established between the preprocessing parameter file and the ExtraTrees extremely random forest model file. When the system calls the model, it simultaneously verifies the model version number, preprocessing parameter version number, feature order version number, and threshold version number; if any version number is inconsistent, the output of the formal association score is stopped, and a model configuration review prompt is generated.

[0045] During the model building phase, a candidate metabolite matrix is ​​established. Candidate metabolites are derived from common serum metabolites that can be matched by metabolite name in both the discovery queue and the external evaluation queue. In one specific embodiment, there are 163 candidate input features; candidate metabolites include at least aspartic acid, arginine, glutamine, glutamic acid, lactic acid, butyric acid, palmitoylcarnitine, phenylalanine, propionic acid, and phenylalanine, and may also include extended candidate metabolites such as alanine and phenylpyruvic acid.

[0046] Before candidate metabolites enter SHAP ranking, they undergo input quality control. Input quality control includes at least missing rate checks, checks for the proportion below the minimum limit of quantitation, checks for the relative standard deviation of quality control samples, batch drift checks, abnormal peak verification, and unit consistency checks. Metabolites that fail input quality control are not included in ExtraTrees extremely random forest training, nor are they included in Top-N forward selection.

[0047] In one approach, a candidate metabolite is removed from the candidate matrix when the missing proportion of a candidate metabolite in the training samples exceeds a preset proportion threshold, or when the relative standard deviation of a candidate metabolite in the mixed biomass control samples exceeds a preset proportion threshold. If the concentration of a candidate metabolite in an individual sample is abnormal but the quality control samples in the same batch are normal, it is recorded as a sample to be reviewed, rather than directly changing the preprocessing parameters of all samples.

[0048] The above input quality control rules ensure that subsequent SHAP contribution ranking and Top-N forward selection are based on a stable and verifiable metabolite concentration matrix, thus preventing low-quality peaks, batch drift, or unit mixing from causing unreproducible combined screening results. The dataset is preferably hierarchically structured to form internal training and internal validation sets, keeping the ratio of positive to negative anxiety-prone samples relatively constant.

[0049] The data was randomly seeded with seed 2024, and the cohort was divided in a 7:3 ratio, with 143 cases in the internal training set and 62 cases in the internal validation set. The internal validation set included 31 positive and 31 negative samples with anxiety tendencies. The preferred independent random seed for model training was 20260515. The internal validation set was not used for model training or preprocessing parameter fitting to evaluate internal performance and determine scoring thresholds.

[0050] The external evaluation queue and the discovery queue do not have sample overlap and do not participate in preprocessing parameter fitting, ExtraTrees model training, SHAP importance ranking, probability calibration function fitting, or scoring threshold selection; however, since its performance metrics are used for Top-N candidate combination comparison, this embodiment positions it as an external evaluation queue rather than a final independent external validation queue.

[0051] In one implementation, the dataset consisted of 205 individuals with hearing loss, including 102 with negative anxiety tendencies and 103 with positive anxiety tendencies; the external evaluation cohort comprised 103 individuals, including 63 with negative anxiety tendencies and 40 with positive anxiety tendencies. Inclusion criteria included a complaint of hearing loss or related medical records, completion of hearing loss-related information collection, GAD-7 scale assessment, and in vitro serum sample testing. Exclusion criteria included malignant tumors, diabetes, metabolic diseases such as thyroid dysfunction, or other major physical illnesses. The model version record should also specify the applicable boundaries, including actual age range, audiological type, severity, and medication use.

[0052] Before model building, if the queue contains samples with inconsistent types, missing core metabolites, substandard quality control, abnormal testing batches, or does not meet the pre-defined inclusion / exclusion criteria, they will be removed according to the above rules. These rules are solidified and written into the model version record before model building to avoid selecting samples based on the final model output direction. An ExtraTrees extremely random forest is trained on the internal training set, and then the SHAP value of the model output for each candidate metabolite on each training sample is calculated. The SHAP importance score corresponding to the j-th candidate metabolite is... Calculate using the following formula: Where n represents the number of samples in the internal training set. ij represents the SHAP value of the j-th candidate metabolite in the i-th training sample, and Ij represents the SHAP importance score of the j-th candidate metabolite; the candidate metabolites are sorted from high to low according to Ij.

[0053] Table 1. SHAP Importance Ranking Results (Scores are rounded to 6 decimal places) Then, a Top-N forward selection was performed. A Top-1 combination was formed from the top-ranked metabolite, and this was expanded sequentially to Top-N combinations. For each combination, the ExtraTrees candidate model was retrained on the internal training set, and internal validation and external evaluation metrics were calculated separately. Combination selection was prioritized according to the following order: external evaluation AUC, external evaluation PR-AUC, internal validation AUC, internal validation PR-AUC, and the number of input features. The combination with the larger values ​​for the first four criteria was given priority; if the first four criteria were the same, the combination with fewer input features was selected. Since the external evaluation metrics were involved in combination determination, the data in Table 2 were not used to demonstrate the independent external generalization performance unaffected by model selection.

[0054] Table 2 Evaluation results of Top-N forward selection Table 2 shows that the external evaluation AUC of the Top-2 combination was 0.810 and the external evaluation PR-AUC was 0.751, both the highest among the Top-1 to Top-10 combinations compared. Therefore, aspartic acid and arginine were determined as the dual metabolite input combination according to the pre-fixed comparison order. This conclusion is based on the candidate combination selection results of the existing external evaluation cohort and still needs to be independently validated with a new cohort that has not participated in any screening.

[0055] The external evaluation queue has no sample overlap with the discovery dataset and does not participate in preprocessing parameter fitting, model training, SHAP ranking, probability calibration, or threshold selection; its metrics only participate in the external performance comparison of candidate combinations and are marked as external evaluation results in the model version record.

[0056] Comparative Example 1, representing the Top-1 single metabolite combination, showed an external evaluation AUC of 0.747, an external evaluation PR-AUC of 0.629, and an external evaluation equilibrium accuracy of 0.649 when only aspartic acid was used. After adding arginine to form the Top-2 combination, these metrics improved to 0.810, 0.751, and 0.672, respectively. This comparison illustrates the performance of the two-metabolite combination relative to the single-metabolite combination in the existing evaluation cohort and is not equivalent to the final independent validation conclusion.

[0057] Comparative Example 2 consists of Top-3 to Top-10 multi-metabolite combinations. Although some combinations have high internal validation metrics, their external evaluation AUC and PR-AUC do not exceed those of Top-2. Within the data range of this example, further improvements were made to glutamine, glutamate, lactate, butyrate, palmitoylcarnitine, or other metabolites, but these failed to achieve higher external evaluation AUC and PR-AUC. This was achieved by increasing detection costs, input dimensions, and model complexity.

[0058] Comparative Example 3 uses alternative models such as Random Forest or LightGBM, which can also output model scores under the same dual metabolite input conditions. In this embodiment, the main model uniformly uses ExtraTrees Extremely Random Forest. SHAP ranking, Top-N candidate models, and final association score generation all use the same algorithm framework to avoid mixing model names and training procedures. Training labels are formed using the GAD-7 scale; a GAD-7 score greater than or equal to 5 is considered positive for anxiety tendency, and less than 5 is considered negative for anxiety tendency. This label is only used for model training and evaluation and does not constitute a medical diagnostic conclusion during the deployment phase.

[0059] In an additional implementation, a GAD-7 score greater than or equal to 10 can be used as a label for moderate to high anxiety risk, for sensitivity analysis or model retraining in specific application scenarios. Regardless of the label threshold used, the label formation rules, training sample versions, inclusion and exclusion criteria, preprocessing parameters, model parameters, and evaluation results should be saved in the model version record; different label thresholds correspond to different model versions and should not be mixed within the same model version. The ExtraTrees extremely random forest rating generation model is trained using the dual metabolite input vectors of each training sample as model input and the anxiety tendency label as a supervision signal.

[0060] The ExtraTrees extremely random forest scoring generation model consists of multiple extremely random decision trees; each tree randomly determines candidate features and candidate splitting thresholds during node splitting, and outputs a 0 to 1 model score for the test sample belonging to the anxiety tendency group by integrating multiple base learners.

[0061] In one implementation, the range of candidate hyperparameters for the ExtraTrees model includes: The number of trees ranges from 100 to 1000, the maximum depth from 2 to 10, the minimum number of leaf node samples from 2 to 5 (not exceeding 10% of the training set samples), the minimum number of split samples from 2 to 20, and the number of candidate features for node splits is 1 or 2. The classification criterion uses Gini impurity or information gain, and the class weights use a balanced mode or are set according to the distribution of classes in the training set. The candidate range is only used to illustrate the optional search space and does not replace the actual fixed parameters of the final model.

[0062] The final model parameters for generating the results in Tables 4 to 7 are: n_estimators=800, max_depth=3, min_samples_split=2, min_samples_leaf=5, max_features=sqrt, criterion=gini, class_weight=balanced, bootstrap=False, random_state=20260515. For the main model containing only two input features, max_features=sqrt corresponds to randomly considering one candidate feature for each node. Running parallelism n_jobs does not change the model logic and cannot be considered as the technical boundary of the model.

[0063] ExtraTrees is an ensemble tree model, not a neural network model trained iteratively by epoch, nor is it a model fitted through boosting iterations. Therefore, this main model does not have early stopping rules. Model complexity is controlled by fixing the number of trees, tree depth, minimum number of samples per leaf node, number of candidate features for node splitting, whether bootstrap sampling and random seeds are enabled.

[0064] Suppose that ExtraTrees extremely random forest consists of M extremely random decision trees, and the m-th extremely random decision tree outputs a single-tree score s_m(X) for the positive anxiety tendency category given the input vector X. Then the model output association score S is the average of the scores from each single tree: In this model, the single-tree score is determined by the proportion of positive training samples with anxiety tendency that the test sample falls into the leaf node, and the model output adopts the standard predict_proba calculation rule implemented in the current ExtraTrees; the positive category code and the read category index are fixed in the model version file. Laplace smoothing is not introduced separately in this embodiment.

[0065] It should be noted that the current master model version does not perform additional probability calibration on the model output. Table 7 shows that the calibration intercept, calibration slope, and Hosmer-Lemeshow test still indicate calibration bias. Therefore, the output S is used as a model association score from 0 to 1, rather than being directly interpreted as an absolute incidence probability. If subsequent Platt scaling, isotonic regression, or other validated calibration methods are used, a new model version should be formed based on the training cohort or an independent calibration cohort, and its calibration performance should be evaluated separately.

[0066] After model training is completed, save the model structure, actual hyperparameters, random seed for data partitioning, random seed for model training, class encoding rules, positive class index, feature names and order, preprocessing parameters, missing value rejection rules, quality control rules, output threshold, training data version, label formation rules, software environment version, model file hash value, model version number, and timestamp; unenabled probability calibration functions are not written to the current model version.

[0067] In one embodiment, the main model uses aspartic acid and arginine as inputs, and selects 163 common serum metabolites for SHAP ranking and Top-N combination comparison to determine the Top-2. Preprocessing is done in raw form, with a random seed of 2024 for data partitioning and a random seed of 20260515 for model training. Simultaneously, the binary classification hint threshold is determined by the Youden index of the internal validation set, and the model file retains its full accuracy, denoted as approximately 0.520721. This threshold is used for internal validation and external evaluation, and is not refitted by the external evaluation queue.

[0068] Table 3. Main Model Methodology Parameters Table 4. Discriminative performance and statistical tests of the main model (external cohorts represent external evaluation results). Note: ROC-AUC represents the area under the receiver operating characteristic curve, PR-AUC represents the area under the precision-recall curve, AUC represents the area under the curve, and CI represents the confidence interval.

[0069] Table 5. Binary classification prompting performance under fixed scoring thresholds. Note: PPV represents positive predictive value, NPV represents negative predictive value, and CI represents confidence interval.

[0070] Table 6. Confusion matrix at a fixed scoring threshold of approximately 0.520721 Note: TP (True Positive) indicates a true positive, FP (False Positive) indicates a false positive, TN (True Negative) indicates a true negative, and FN (False Negative) indicates a false negative.

[0071] At a scoring threshold of approximately 0.520721, in the external evaluation cohort, 22 out of 40 samples with a positive anxiety tendency received a hint above the threshold, and 18 received a hint below the threshold; among the 63 samples with a negative anxiety tendency, 50 received a hint below the threshold, and 13 received a hint above the threshold. This yielded a sensitivity of 0.550, a specificity of 0.794, a PPV of 0.629, an NPV of 0.735, and a balanced accuracy of approximately 0.672. The confusion matrix is ​​consistent with the indicators listed in Table 5 and is used to explain the threshold hint performance without altering the technical characteristics of the model's continuously correlated scoring output.

[0072] The AUC, PR-AUC, confidence intervals, and P-values ​​in Table 4 are based on a sample-by-sample verification of the true labels, predicted scores, and a fixed statistical script. When using bootstrap resampling to calculate the 95% confidence interval, the model version record simultaneously saves the number of bootstrap samplings, the random seed, the stratification method, the confidence interval algorithm, and the statistical script version. Each sampling only recalculates the evaluation metrics; it does not retrain the model, reselect features, or refit preprocessing parameters or scoring thresholds. The sample-by-sample true labels, predicted scores, and verification results are saved as supplementary files to the model version record.

[0073] The consistency between the model association score and the actual positive rate can be described using the Brier score, calibration intercept, calibration slope, Hosmer-Lemeshow test, and calibration curve. Currently, the calibration function has not been fitted on an external evaluation cohort; Table 7 is only used to illustrate the calibration performance of the existing model association score and as a basis for subsequent model version updates.

[0074] Table 7. Model Association Scoring Calibration Analysis Results Note: HL is the abbreviation for Hosmer-Lemeshow test; the chi-square value is the χ² statistic; and the degrees of freedom are abbreviated as df.

[0075] Table 7 shows the calibration slopes for internal validation and external evaluation as 7.132 and 4.564, respectively, and the Hosmer-Lemeshow test p-values ​​as 0.002 and 0.004, respectively, both indicating a significant calibration bias between the model's correlation score and the actual positive rate. Therefore, the current version output is uniformly referred to as the anxiety tendency correlation score and is not directly interpreted as the absolute probability of disease. If a probability calibration function is introduced in the future, it should be fitted based on the training cohort or an independent calibration cohort to form a new model version. During deployment, the system first verifies the sample number, testing batch, quantitative range, reference unit, unit conversion rules, and quality control status.

[0076] After quality control is passed, the preprocessing parameters bound to the current model version are invoked to form a dual metabolite input vector, with aspartic acid first and arginine last. This vector is then input into the pre-trained ExtraTrees extremely random forest scoring model to obtain the associated score S. The system directly outputs S. When the model version is bound to a preset scoring threshold T, if S is greater than or equal to T, a prompt indicating that the score is above the threshold or requiring manual review is output; if S is less than T, a prompt indicating that the score is below the threshold is output. This embodiment does not set up multi-level anxiety stratification such as mild, moderate, and severe anxiety.

[0077] In one specific embodiment, the threshold used to generate above- or below-threshold prompts is approximately 0.520721, and the model file retains its full accuracy. This threshold is determined by an internal validation set based on the Youden index and is fixed with the model version; this threshold must not be refitted using an external evaluation queue or a single test sample, and a new threshold version record should be generated when the threshold is adjusted.

[0078] The system output may include: sample number, test batch, aspartic acid concentration, arginine concentration, dual metabolite input vector, model version number, correlation score, optional threshold prompts, quality control status, generation time, and review prompts. Furthermore, the output interface or file should clearly state that the results are based on information processing of in vitro serum metabolite data and do not replace the comprehensive judgment of professionals.

[0079] The following example is only used to illustrate interface processing, scoring threshold judgment, and traceability logic. It is not intended to prove model performance, nor does it represent real samples in the research cohort. The two concentrations of sample A are converted into input vector X_A after unit standardization, quantitative range verification, and quality control validation. Assuming that the model actually calculates S_A=0.630000 and the scoring threshold T=0.520721, the system outputs a prompt indicating that the threshold is exceeded and a prompt for manual review, and records the input data, model version, threshold version, and generation time.

[0080] The two concentrations of sample B are validated using the same rules to form the input vector X_B. Assuming the model actually calculates S_B = 0.410000 and the scoring threshold T = 0.520721, the system outputs a message indicating that the score is below the threshold. This message only indicates that the model score is below the current version's threshold and does not equate to excluding anxiety-related states.

[0081] If the arginine concentration of sample C is lower than the minimum quantitation limit given by the testing agency, or if its quality control status fails, the system will not perform missing value imputation, will not call the formal scoring generation model, will not output the correlation score, will only generate a pending review prompt, and will save the reason why it did not enter the inference process.

[0082] To ensure the model inference process is verifiable, a traceable record is saved each time an association score is generated. This record includes at least the sample number, sample batch, testing institution number, testing time, testing method version, quality control status, aspartic acid concentration, arginine concentration, model reference unit, unit conversion factor, preprocessing parameter version, input feature order, model algorithm type, model version number, threshold version, association score, optional threshold hints, generation time, and system task number.

[0083] The model version management module saves training data versions, label formation rules, inclusion and exclusion criteria, candidate metabolite sets, SHAP ranking results, Top-N combination comparison results, final feature combinations, actual hyperparameters of ExtraTrees, preprocessing parameters, quality control rules, output thresholds, internal validation results, external evaluation results, calibration analysis results, and deployment time. After validation of a newly added independent cohort that did not participate in the screening process, the result should be recorded separately from the existing external evaluation results.

[0084] When new samples are used for model retraining, the system does not directly overwrite the existing model, but generates candidate model versions. The candidate model versions re-execute candidate metabolite input quality control, preprocessing parameter fitting, SHAP ranking, Top-N forward selection, ExtraTrees extremely random forest training, internal validation, and external evaluation; only after meeting the preset performance and quality control requirements can they be launched as official model versions.

[0085] When there are substantial changes to the detection platform, sample preparation process, or target metabolite quantification method, the system re-verifies the model reference units, unit conversion factors, standard curve range, lower limit of quantitation (LQ), upper limit of quantitation (LQ), quality control sample pass rate, and comparability with the original detection method for aspartic acid and arginine. If the bridging evaluation cannot prove equivalence within the deviation range of the model version, the model is retrained, the scoring threshold is determined, and the model performance is evaluated.

[0086] Corresponding to the methods described above, this solution also provides an aspartic acid-arginine deafness anxiety association score generation system, specifically including: a data acquisition module, a quality control and verification module, a vector generation module, an applicability condition verification module, an ExtraTrees score generation module, a result output module, a model version management module, and a traceability record module. The model building environment can also include a feature combination management module to save candidate metabolite sets, SHAP importance ranking, and Top-N combination comparison results; however, this module is not a necessary module for the deployment end to complete single-sample score inference.

[0087] The data acquisition module is used to acquire the serum aspartic acid concentration, serum arginine concentration, sample number, test batch, concentration unit, minimum limit of quantitation, maximum limit of quantitation, and quality control status output by the testing institution.

[0088] The quality control verification module is used to read the sample number, test batch, concentration unit, quantitative range, quality control status and test method version provided by the testing institution, complete the unit unification according to the reference unit and conversion factor fixed in the model version, and determine whether the aspartic acid and arginine data can be entered into the model; the system does not require the testing institution to re-execute the peak identification, peak integration or calibration curve conversion.

[0089] The vector generation module invokes the raw strategy bound to the current model version, performing unit unification, field name validation, feature order validation, and quality control status validation on the two absolute quantitative concentrations. Missing value imputation, log2 transformation, z-score normalization, or rank-Gauss transformation are not performed. The module then forms the dual metabolite input vectors in a fixed order. The feature combination management module in the model building environment stores the candidate metabolite set, the ExtraTrees model SHAP importance ranking results, the Top-N combination comparison results, and the aspartic acid and arginine dual metabolite input combinations.

[0090] Furthermore, during system deployment, the input features are locked as aspartic acid and arginine. The ExtraTrees scoring generation module calls a pre-trained model to infer the dual metabolite input vector consisting of two concentrations in a fixed order, and outputs an anxiety tendency association score for the deaf subject.

[0091] The results output module outputs an anxiety tendency association score and, when the threshold version is valid, outputs prompts indicating above the threshold, below the threshold, manual review, or quality control anomalies. It does not output mild, moderate, or severe anxiety stratification, nor does it directly output a diagnosis of anxiety disorder. The model version management module saves the model's actual parameters, reference units, unit conversion rules, preprocessing parameters, quality control rules, threshold version, training data version, software environment version, model file hash value, internal validation results, external evaluation results, confusion matrix, and calibration analysis results. The traceability record module saves the sample number, testing batch, input concentration, preprocessing results, model version, output results, and generation time for each inference. The applicability verification module reads the sample type, age range, hearing loss determination criteria, hearing loss duration, hearing loss side, hearing loss severity, audiological type, hearing aid or cochlear implant usage status, main medication information, or model version applicability label. When the sample being tested exceeds the applicability range of the current model version, it only outputs a "model applicability range pending review" prompt and does not output a formal association score.

[0092] The system may also include a field consistency verification module, used to verify whether the field names, model reference units, unit conversion factors, detection method versions, preprocessing parameter versions, feature order versions, and threshold versions of the input data are consistent. If the field names, model reference units, unit conversion factors, detection method versions, preprocessing parameter versions, feature order versions, or threshold versions are inconsistent, the ExtraTrees scoring generation module will be stopped to avoid output errors due to configuration mismatches.

[0093] This solution also provides an electronic device, including a processor, a memory, and a communication interface. The memory stores a computer program, feature preprocessing parameters, feature order, ExtraTrees extremely random forest model file, quality control rules, and optional output thresholds; when the processor executes the computer program, it implements the above-mentioned method for generating anxiety tendency association scores for deaf individuals.

[0094] This solution also provides a computer-readable storage medium. When the computer program stored in the medium is executed by a processor, it performs the following steps: obtaining serum aspartate concentration and serum arginine concentration; verifying data quality and applicable conditions; generating dual metabolite input vectors in a fixed order; calling a pre-trained ExtraTrees model to output association scores; generating prompts or review prompts for values ​​above or below the threshold when the model version is bound to a threshold; and saving traceable records.

[0095] The computer-readable storage medium may be a hard disk, optical disk, solid-state drive, flash memory, read-only memory, random access memory, server storage space, or cloud storage space. The computer program may be deployed on local testing equipment, laboratory information management systems, hospital internal servers, scientific research data management platforms, or cloud computing platforms.

[0096] In an alternative implementation, a separate LightGBM control model or candidate model version can be created for the same dual metabolite input. This model should use an independent model version number, training parameters, preprocessing parameters, scoring threshold, and evaluation results, and should not be mixed with the current ExtraTrees main model version. Unless the new model is trained and validated and a formal version is formed, this alternative implementation does not change the technical positioning of the main embodiment, which uses the ExtraTrees extremely random forest model as the scoring generation model.

[0097] In another alternative implementation, a random forest, XGBoost, logistic regression, support vector machine, or neural network control model can be built for the same dual metabolite inputs. All of these models should use preprocessing parameters, feature ordering, class coding rules, and evaluation results tied to their respective versions; if intended for official deployment, a new model version should be created and validated accordingly. Besides aspartic acid and arginine, metabolites such as glutamine, glutamic acid, lactic acid, butyric acid, palmitoylcarnitine, propionic acid, alanine, or phenylpyruvic acid can be used for mechanism analysis and control experiments, but the current main model's official inputs are fixed aspartic acid and arginine.

[0098] Regarding the expansion of applicable subjects, this embodiment applies to deaf subjects whose exclusion criteria are covered by the current model version; if it is expanded to other populations, the sample queue, applicable scope label, preprocessing parameters, model parameters, scoring thresholds and evaluation results should be re-established and managed using the new model version.

[0099] In the evaluation of candidate models, if LightGBM, Random Forest, or other models perform close to the main model on a certain internal validation set, the final model should still be determined by comprehensively considering external evaluation, feature quantity, model score stability, version reproducibility, and deployment complexity. Unless a new validation report and model version are generated, the statement that the main embodiment of this application uses the ExtraTrees extremely random forest model as the formal scoring generation model remains unchanged.

[0100] In the model construction results of this scheme, aspartic acid and arginine ranked first and second in importance among candidate metabolites SHAP, respectively. In the dataset used, the serum arginine level was generally higher in samples with a positive anxiety tendency, while aspartic acid showed a decreasing trend. Arginine is related to nitric oxide metabolism, and aspartic acid is an excitatory amino acid. The above information can provide biological explanations for the combination of dual metabolites. This explanation does not constitute a definitive conclusion on the mechanism of anxiety, is not included in the essential technical features of the claims, and is not intended to directly make a medical diagnosis.

[0101] It should be noted that the above metabolic pathway explanations are only used to illustrate that aspartic acid and arginine, as model inputs, have biologically interpretable clues and do not constitute a definitive limitation on the mechanism of anxiety tendency. The technical effectiveness of this approach comes from the standardization of units, quality control, feature screening, ExtraTrees extremely random forest scoring generation, and result traceability of absolute quantitative data of isolated serum metabolites, and does not rely on the complete proof of specific molecular mechanisms.

[0102] In summary, this implementation method uses the quality-controlled absolute quantitative concentrations of serum aspartic acid and arginine as fixed dual metabolite inputs. It employs model version-bound reference units, unit conversion rules, raw preprocessing, fixed feature order, ExtraTrees model SHAP ranking and Top-N combination comparison, correlation score generation, and process logging to highlight the two key metabolites and their correspondence with model parameters. The external cohort data listed in Tables 4 to 7 represent external evaluation results after participating in candidate combination comparisons; independent validation using a new cohort that has not participated in any screening is recommended before formal deployment.

[0103] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0104] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0105] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0106] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0107] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for generating an aspartic acid-arginine-based deafness anxiety association score, characterized in that, Performed by electronic devices, including: Obtain absolute quantitative concentration data of serum metabolites from the isolated serum samples of the deaf subjects to be tested by the testing institution. The data shall include at least the serum aspartate concentration, serum arginine concentration, concentration unit, sample number, test batch, lower limit of quantitation, upper limit of quantitation, and quality control status. Based on the reference units and unit conversion rules fixed in the current model version, the units of the serum aspartic acid concentration and serum arginine concentration are unified, and the field integrity, quantitative range, quality control status and model version applicability are verified; when both concentrations are valid, the dual metabolite input vector is formed according to the characteristic order of aspartic acid first and arginine last, which is fixed in the current model version. The dual metabolite input vector is input into the pre-trained ExtraTrees extremely random forest scoring generation model to obtain an anxiety tendency association score with a value range of 0 to 1. The anxiety tendency association score is the model output score under the current model version, and is not interpreted as the absolute probability of disease in the population, nor is it directly output as a medical diagnosis conclusion.

2. The method for generating correlation scores according to claim 1, characterized in that, The aspartic acid and arginine were determined in the following manner: During the model building phase, a candidate metabolite matrix including the concentrations of various candidate serum metabolites is obtained, and input quality control is performed on the candidate metabolite matrix; the ExtraTrees candidate model is trained based on the internal training set, and the Shapley additive interpretability importance score of each candidate serum metabolite is calculated. The top N candidate metabolite combinations are formed according to their importance scores from high to low, where N is 1 to 10; for each of the top N candidate metabolite combinations, the ExtraTrees candidate model is retrained on the internal training set, and the area under the receiver operating characteristic curve, the area under the precision-recall curve, sensitivity, specificity, and balanced accuracy are calculated on the internal validation set and on data labeled as the external evaluation cohort. Aspartic acid and arginine were selected as the dual metabolite input combination based on the priority order of the area under the receiver operating characteristic (ROC) curve for external evaluation, the area under the precision-recall curve for external evaluation, the area under the ROC curve for internal validation, the area under the precision-recall curve for internal validation, and the number of input features.

3. The method for generating correlation scores according to claim 2, characterized in that, After data quality verification, the discovery queue is divided into an internal training set and an internal validation set in a stratified random manner. The random seed for data partitioning, the random seed for model training, the sample partitioning rules, the label formation rules, and the input quality control rules are written into the model version record. The external evaluation queue has no sample overlap with the discovery queue and does not participate in feature preprocessing parameter fitting, ExtraTrees candidate model training, Shapley additive interpretation importance ranking, probability calibration function fitting, or scoring threshold selection.

4. The association score generation method according to claim 1, characterized in that, The concentration data is obtained by the testing institution using validated ultra-high performance liquid chromatography-tandem mass spectrometry or equivalent absolute quantification methods, and includes at least the sample number, testing batch, analyte name, concentration value, concentration unit, minimum limit of quantitation (LLQ), maximum limit of quantitation (UPQ), and quality control status. When aspartic acid or arginine is below the LQ, above the UPQ, or fails the quality control, the corresponding sample is marked as pending verification and is prohibited from entering the formal correlation scoring generation process.

5. The method for generating correlation scores according to claim 1, characterized in that, The feature preprocessing parameters bound to the current model version include reference units, unit conversion factors, field order, missing value rejection rules, outlier handling rules, quality control judgment rules, and applicable detection platform information. The main model version uses the original numerical scale and only performs unit unification, field name verification, feature order verification, and quality control status verification. It does not perform missing value imputation, logarithmic transformation, z-score normalization, or rank-Gaussian transformation. When any core input of aspartic acid or arginine is missing, below the minimum limit of quantitation, above the maximum limit of quantitation, or fails the quality control status, a "pending review" prompt is output, and no formal correlation score is generated. When only the field name or record format changes and the input values ​​remain the same, the interface consistency verification is re-executed; when the relationship between input values ​​changes due to unit mapping, standardization, logarithmic transformation, outlier replacement, or changes in the detection platform, a new model version is generated, and the model is retrained, the scoring threshold is determined, and the model performance is evaluated.

6. The method for generating correlation scores according to claim 1, characterized in that, The ExtraTrees extremely random forest scoring generation model includes multiple extremely random decision trees; each extremely random decision tree randomly selects candidate features and candidate splitting thresholds during node splitting, outputs the single-tree score of the dual metabolite input vector belonging to the positive anxiety tendency category, and integrates the single-tree scores of multiple extremely random decision trees to obtain the anxiety tendency association score.

7. The method for generating correlation scores according to claim 6, characterized in that, The model version file should at least save the number of trees, maximum depth, minimum number of split samples, minimum number of leaf node samples, number of candidate features for node splits, classification criteria, class weights, whether bootstrap sampling is enabled, class encoding rules, positive class index, random seed for data partitioning, random seed for model training, feature order, training data version, software environment version, model file hash value, threshold version, model version number, and the actual fixed parameter values ​​for the corresponding model version. When a probabilistic calibration function is introduced, a new model version is fitted and generated based on the training queue or an independent calibration queue.

8. The method for generating correlation scores according to claim 1, characterized in that, When the current model version is bound to a scoring threshold, the anxiety tendency association score is compared with the scoring threshold, and a prompt above or below the threshold is generated; when the quality control status, quantitative range, or model version applicability conditions are not met, a prompt for review is generated; the scoring threshold is determined based on the internal validation set and is fixed with the model version. During the deployment phase, the scoring threshold is not re-determined based on a single test sample or external evaluation queue. When the threshold is adjusted, a new threshold version is generated and the original threshold record is retained. The prompts do not directly constitute a medical diagnostic conclusion.

9. An aspartic acid-arginine-based deafness anxiety association scoring system, characterized in that, include: The data acquisition module is used to acquire the absolute quantitative concentrations of serum aspartic acid and arginine, as well as the sample number, batch number, concentration unit, limit of quantitation (LOQ), limit of quantitation (LOQ), and quality control status of the isolated serum samples from the deaf subjects to be tested, output by the testing institution. The quality control verification module is used to unify units according to the reference units and unit conversion rules fixed in the current model version, and to verify the field integrity, quantitative range and quality control status of the concentration data. The vector generation module is used to generate a dual metabolite input vector according to the characteristic order of aspartic acid first and arginine last, which is fixed in the current model version, when both core concentrations are valid. The applicability verification module is used to verify the sample type, applicable labels for the test object, detection platform information, and the applicability scope of the current model version. The ExtraTrees Extremely Random Forest score generation module is used to call the pre-trained ExtraTrees Extremely Random Forest model to generate anxiety tendency association scores ranging from 0 to 1 based on the dual metabolite input vector. The results output module is used to output the anxiety tendency association score, and when the current model version is bound to a score threshold, it outputs a prompt indicating above the threshold, a prompt indicating below the threshold, or a prompt indicating that it needs to be reviewed. The model version management module is used to save reference units, unit conversion rules, model parameters, feature order, preprocessing parameters, threshold version, training data version, software environment version, model file hash value, and model version number. The traceability record module is used to save the sample number, detection batch, input concentration, quality control status, feature order, model version, threshold version, correlation score, prompt result, and generation time for each correlation score generation process.

10. An electronic device, characterized in that, The system includes a processor and a memory, wherein the memory stores a computer program, reference units, unit conversion rules, feature order, feature preprocessing parameters, and an ExtraTrees extremely random forest model file, and the computer program, when executed by the processor, implements the association score generation method according to any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the association score generation method according to any one of claims 1 to 8.

Images