A method of disease prognosis review assessment

Abstract

This invention relates to the field of computer technology and proposes a method for disease prediction review assessment, comprising: standardizing the scale results and biological evidence results of the subject to be assessed to obtain scale representation and biological evidence representation; determining the correction direction and correction magnitude of the scale representation to the biological evidence representation; correcting the biological evidence representation based on the correction direction and correction magnitude to obtain a corrected prediction result; determining the correction consistency result by combining the scale representation, biological evidence representation, correction direction, and correction magnitude; obtaining a comprehensive confidence level by weighting the correction consistency result and confidence index based on preset weighting coefficients; and obtaining the assessment result based on the comprehensive confidence level. This method can effectively output the confidence level of a single sample and identify and prompt problems in the dynamic adjustment mechanism of scale participation in prediction, thereby achieving effective allocation and utilization of clinical diagnostic or review resources.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and more specifically, to a method for the reconciliation assessment of disease prediction. Background Technology

[0002] Currently, disease screening relies heavily on self-report scales for initial screening, but this approach has significant drawbacks: scales depend on subjective responses from subjects, which are susceptible to interference from various factors such as immediate emotions, physical condition, and environment, resulting in large fluctuations in test results and insufficient reliability of single scores. Existing disease prediction schemes that integrate scales and biological evidence use relatively fixed data fusion methods, simply superimposing the two types of data for risk calculation without distinguishing the respective sources of error in subjective scales and objective biological data. Furthermore, they cannot simultaneously identify issues such as fluctuations in scale threshold scores, biological sample noise, and model probability bias, and lack effective discrimination logic when contradictions arise between pieces of evidence.

[0003] Meanwhile, the lack of a unified and standardized review and evaluation mechanism for disease prediction results makes it impossible to identify and promptly send out questionable samples for review, resulting in a significant waste of clinical manpower and testing resources, and biased accuracy of prediction results.

[0004] Therefore, the verification mechanism for disease prediction results needs further research. Summary of the Invention

[0005] This invention aims to address, to a certain extent, the technical problems existing in the prior art. It proposes a method, device, equipment, medium, and product for assessing the verifiability of disease prediction, adding reliability markers and verification criteria to disease screening results. This effectively outputs the credibility of individual samples and identifies and alerts problems in the dynamic adjustment mechanism of scale-based prediction. It prioritizes limited interview, retesting, and clinical review resources for individuals with unstable results or inconsistent evidence, preserving the advantages of low-cost screening and the objectivity of biological evidence while further achieving effective allocation and utilization of clinical diagnostic or review resources and reducing the interference of factors such as data sources and model parameters on prediction or diagnostic results.

[0006] In a first aspect, the present invention provides a method for disease prediction verification assessment. According to an embodiment of the present invention, the method includes: standardizing the scale results and biological evidence results of the subject to be assessed to obtain scale representation and biological evidence representation; determining the correction direction and correction magnitude of the scale representation on the biological evidence representation based on the relative magnitude of the scale representation and the biological evidence representation and the distance between the biological evidence representation and the biological evidence threshold, respectively; performing correction processing on the biological evidence representation based on the correction direction and the correction magnitude to obtain a correction prediction result; inputting the scale representation, the biological evidence representation, the correction direction, and the correction magnitude into a consistency verification model to determine the correction consistency result; obtaining a comprehensive confidence level by weighting the correction consistency result and the confidence index based on a preset weighting coefficient; and obtaining an assessment result based on the comprehensive confidence level.

[0007] According to embodiments of the present invention, the above-described disease prediction verification assessment method may further have at least one of the following additional technical features: According to an embodiment of the present invention, the correction magnitude is determined by the following steps: determining the clarity of the biological evidence representation based on the absolute distance between the biological evidence result and the biological evidence threshold; determining the scale baseline gating strength based on the scale representation and the biological evidence representation via a machine learning network; performing mathematical operations based on the clarity of the biological evidence representation and a preset clarity attenuation coefficient to obtain a biological evidence clarity attenuation constraint term; and performing mathematical operations based on a preset correction upper limit, the scale baseline gating strength, and the biological evidence clarity attenuation constraint term to obtain the correction magnitude.

[0008] According to an embodiment of the present invention, the biological evidence result includes a disease risk prediction result based on metagenomic data, which is obtained by: constructing N feature matrices from the metagenomic data of the object to be evaluated, wherein the N feature matrices correspond to microbial features at different biological levels; performing feature extraction on the N feature matrices and cross-type fusion on the feature representations extracted from the feature matrices to obtain a fused representation; and performing disease risk prediction based on the fused representation to obtain a prediction result.

[0009] According to an embodiment of the present invention, the consistency verification model is trained in the following manner: a training dataset is obtained, the training dataset including: training scale results and corresponding training biological evidence results; the training dataset is input into a deep learning model, and the deep learning model is weakly supervised to obtain the consistency verification model by using the fit between the judgment trends of the training scale results and the corresponding training biological evidence results relative to their respective judgment thresholds, and / or the fit between the training correction prediction results and the true labels of the training samples as constraint targets.

[0010] According to an embodiment of the present invention, the confidence index includes at least one of the following: result baseline consistency index, correction deviation index, boundary proximity index, sample stability index, probability calibration index, and scale quality index.

[0011] According to an embodiment of the present invention, the baseline consistency index is used to characterize the degree of fit between the scale results and the biological evidence results relative to their respective judgment thresholds.

[0012] According to an embodiment of the present invention, the correction deviation index is used to characterize the degree of dependence of the correction prediction result on the scale characterization.

[0013] According to an embodiment of the present invention, the boundary proximity index is used to characterize the sensitivity of the conclusion reversal when the judgment interval in which the corrected prediction result is located is located.

[0014] According to an embodiment of the present invention, the sample stability index is used to characterize the volatility risk of a single sample prediction result.

[0015] According to an embodiment of the present invention, the probability calibration index is used to characterize the fitting deviation between the corrected prediction result and the true sample distribution.

[0016] According to an embodiment of the present invention, the scale quality index is used to characterize the data completeness of the scale results.

[0017] According to an embodiment of the present invention, obtaining the evaluation result based on the comprehensive confidence level further includes: determining a confidence level based on the comprehensive confidence level; determining the review reason based on the confidence index used in the comprehensive confidence level; determining a review recommendation based on the confidence level and the review reason; and outputting at least one of the confidence level, the review reason, and the review recommendation as the evaluation result.

[0018] Secondly, the present invention provides a disease prediction verification assessment device. According to an embodiment of the present invention, the device includes a standardization module for standardizing the scale results and biological evidence results of the subject to be assessed to obtain scale representation and biological evidence representation; a correction coefficient module for determining the correction direction and correction magnitude of the scale representation on the biological evidence representation based on the relative magnitude of the scale representation and the biological evidence representation and the distance between the biological evidence representation and the biological evidence threshold, respectively; a correction module for correcting the biological evidence representation based on the correction direction and the correction magnitude to obtain a corrected prediction result; a verification module for inputting the scale representation, the biological evidence representation, the correction direction, and the correction magnitude into a consistency verification model to determine the correction consistency result; a confidence module for obtaining a comprehensive confidence level by weighting the consistency result and the acceptance index based on a preset weighting coefficient; and an assessment module for obtaining an assessment result based on the comprehensive confidence level.

[0019] Thirdly, the present invention provides an electronic device, according to an embodiment of the present invention, the device comprising: a processor and a memory; the aforementioned memory for storing a computer program; the aforementioned processor for executing the computer program to implement the disease prediction review assessment method as provided in the first aspect.

[0020] Fourthly, the present invention provides a computer-readable storage medium, according to an embodiment of the invention, wherein the medium stores computer instructions or programs that, when executed on a computer, cause the disease prediction review assessment method provided in the first aspect to be performed.

[0021] Fifthly, the present invention provides a computer program product, according to an embodiment of the invention, the product comprising computer instructions that, when some or all of the computer instructions are executed on a computer, cause the disease prediction review assessment provided in the first aspect to be performed.

[0022] In summary, the evaluation method provided by this invention, on the one hand, obtains the correction result through a dynamic correction mechanism based on the scale structure and biological evidence results. Under the premise of ensuring that biological evidence dominates and has objective prediction, the scale correction strength and direction are adaptively adjusted according to the clarity of single-sample evidence. This can not only make full use of scale information to make up for the uncertainty of biological evidence when the conclusion of biological evidence judgment is ambiguous or close to the judgment threshold, thus improving the prediction sensitivity, but also actively constrain the scale intervention range when the biological evidence has sufficient certainty, avoiding the interference of subjective scale information with the objective conclusion of biological evidence judgment. This effectively reduces the prediction bias and evidence conflict problems caused by multimodal fusion, and improves the interpretability, robustness and clinical credibility of the overall prediction results.

[0023] On the other hand, identifying and detecting multiple indicators, including correction consistency, sample stability, and scale quality, in the corrected prediction results enables fine-grained breakdown and tracing of prediction credibility, accurately distinguishing the sources of prediction bias, and overcoming the limitations of generalizing and judging based on a single confidence level conclusion. It can not only target abnormal samples and trigger a review reminder mechanism, but also quantitatively assess the credibility level of each prediction conclusion, improve the interpretability of the model output results, facilitate the rapid identification of low-confidence prediction cases in the clinic, reduce the risk of misjudgment and omission, and improve the rigor, controllability, and clinical applicability of the overall risk prediction system. Attached Figure Description

[0024] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 A flowchart illustrating the method for verifying disease prediction; Figure 2 A schematic block diagram of a disease prediction verification assessment device; Figure 3 A schematic block diagram of an electronic device; Figure 4 This is a schematic block diagram of an electronic device that includes a transceiver component. Detailed Implementation

[0025] The embodiments of the present invention are described in detail below. These embodiments are exemplary and are only used to explain the present invention, and should not be construed as limiting the invention. Where specific techniques or conditions are not specified in the embodiments, they are performed according to the techniques or conditions described in the literature in the art or according to the product instructions. Reagents or instruments used, unless otherwise specified, are all commercially available conventional products.

[0026] It should be noted that in the embodiments of the present invention, the terms "comprising" or "including" are open-ended expressions, that is, they include the contents specified in the present invention, but do not exclude other aspects.

[0027] In embodiments of the present invention, the terms “optionally,” “optionally,” or “optionally” generally refer to events or conditions described subsequently that may but may not occur, and the description includes both cases in which the event or condition occurs and cases in which the event or condition does not occur.

[0028] The technical problem to be solved, the inventive concept, and the system architecture of the embodiments of the present invention will be described below: Currently, for disease prediction, especially for mental illnesses like postpartum depression that are difficult to diagnose, existing postpartum depression risk prediction schemes suffer from two core defects. First, the prediction data sources are often singular or the feature fusion methods are crude. Relying solely on psychological scales can lead to subjective bias and delayed warnings, or relying solely on biological data without sufficient clinical subjective information. Second, the fusion method of simply combining biological and scale features fails to distinguish the primary and secondary logic of these two types of heterogeneous evidence. This can result in subjective information from the scales masking objective biological judgments, significant fluctuations in prediction results when evidence conflicts, and difficulty in adaptively matching the clarity of evidence to different samples. Furthermore, the model output only provides a single risk probability value, lacking a corresponding credibility assessment system. It cannot distinguish the specific causes of unreliable prediction results, making it difficult to pinpoint the source as one or more of the following: incomplete scale completion, contradictions between biological and scale evidence, abnormal sample feature distribution, or judgment results approaching the risk threshold. Clinicians cannot intuitively assess the reference value of the prediction conclusions, leading to a high risk of misjudgment and missed diagnoses in large-scale screening scenarios, and insufficient interpretability of the results.

[0029] To address the shortcomings of existing technologies, on the one hand, a dynamic correction mechanism is established, prioritizing biological evidence and supplementing it with scale information. This mechanism precisely defines and calculates the correction direction and magnitude, dynamically attenuating the scale correction strength based on the clarity of microbial evidence. When the biological evidence conclusion is clear, the scale intervention weight is autonomously reduced, and scale information is only fully introduced to supplement correction when the biological evidence is ambiguous or close to the judgment threshold. This balances the primary and secondary relationship between objective biological data and subjective scale data, mitigating prediction bias caused by conflicts in multimodal evidence fusion. On the other hand, a multi-indicator credibility detection system covering dimensions such as correction consistency, sample stability, and scale quality is constructed. This system conducts fine-grained quantitative evaluation of the corrected prediction results, accurately tracing the causes of insufficient credibility, enabling graded identification and review prompts for low-reliability samples, and balancing prediction accuracy with clinical interpretability.

[0030] This effectively eliminates the shortcomings of existing technologies, such as the imbalance between primary and secondary evidence from multiple sources and the lack of credibility traceability in prediction results, thereby achieving orderly dynamic fusion of multi-source heterogeneous data and quantifiable confidence assessment of prediction results.

[0031] The technical solution of this application will be described in detail below: In one aspect of the invention, a method for disease prediction verification assessment is proposed, referring to... Figure 1 As shown, the method may include: S100: Standardize the scale results and biological evidence results of the subjects to be assessed to obtain scale representations and biological evidence representations.

[0032] This step can eliminate interference caused by differences in the dimensions and numerical distributions of heterogeneous data from scale results and biological evidence results, unify the scale of data expression, avoid model weight bias caused by large differences in numerical ranges of different types of data, and at the same time, rectify invalid fluctuation information in the original data, simplify redundant content, and ensure smooth and efficient subsequent information processing.

[0033] In some embodiments, the applicable diseases include, but are not limited to, mental and behavioral disorders, neurological diseases, endocrine disorders, nutritional and metabolic diseases, musculoskeletal disorders, and connective tissue diseases. Preferably, the applicable diseases are mental and behavioral disorders, including but not limited to depression, postpartum depression, bipolar disorder, narcolepsy, insomnia, and Alzheimer's disease. More preferably, the applicable diseases include depression and postpartum depression.

[0034] In some embodiments, the sources of scale results include, but are not limited to, the Edinburgh Postnatal Depression Scale, the Self-Rating Depression Scale, the Hamilton Depression Rating Scale, the Self-Rating Anxiety Scale, the Pittsburgh Sleep Quality Index, and the Montreal Cognitive Assessment Scale. Preferably, the sources of scale results include the Edinburgh Postnatal Depression Scale and the Self-Rating Depression Scale. In a preferred embodiment, if a scale item is missing, it can be included in the statistics according to a preset default value during the training phase, and the mean of the training set can be used to fill the missing item during the deployment and inference phase. When recording the missing status at the deployment end, a missing mask can also be retained for subsequent confidence level correction and verification prompts.

[0035] In some embodiments, the standardization processing of scale results and biological evidence results includes, but is not limited to, Z-score standardization, extreme value normalization, norm normalization, logarithmic normalization, etc.

[0036] According to embodiments of the present invention, biological evidence results include disease risk prediction results based on metagenomic data. In some embodiments, disease risk prediction results based on metagenomic data are obtained as follows: N feature matrices are constructed from the metagenomic data of the object to be evaluated, each corresponding to microbial features at different biological levels; feature extraction is performed on the N feature matrices, and cross-type fusion is performed on the extracted feature representations to obtain a fused representation; based on the fused representation, disease risk prediction is performed to obtain the prediction result. Biological evidence results obtained by constructing feature matrices at multiple biological levels and fusing them across levels can integrate microbial information at different scales, suppress sequencing noise and feature bias, making the output of biological evidence risk probability more stable and reliable. While improving the rationality of the quantified results of evidence clarity, it also provides a more robust judgment basis for subsequent dynamic matching of correction magnitudes, enhancing the stability and interpretability of the overall confidence assessment system.

[0037] S200: Based on the relative magnitude of the scale representation and the biological evidence representation, and the distance between the biological evidence representation and the biological evidence threshold, determine the correction direction and correction magnitude of the scale representation to the biological evidence representation.

[0038] In some embodiments, the correction direction is obtained by independent gating network computation and is used to determine the correction bias trend of scale representation to biological evidence representation. For example, referring to equation (1), the standardized scale features and biological evidence representation are used as inputs to calculate the output, with a value range constraint of -1 to 1. A non-negative value indicates that the scale representation tends to increase the biological evidence representation, and a negative value indicates that the scale representation tends to decrease the biological evidence representation, where [h m [,q] represents the vector concatenation operation, h m For biological evidence representation, q represents the scale representation, and W represents the biological evidence representation. d b d represents the weight matrix and bias term obtained from training the network in the correction direction, respectively, which are used to perform a linear transformation on the concatenated joint features. tanh represents the hyperbolic tangent activation function with output values ​​between -1 and 1.

[0039] (1).

[0040] According to an embodiment of the present invention, the correction magnitude is determined through the following steps: determining the basic gating strength of the scale based on the scale representation and the biological evidence representation via a machine learning network; determining the clarity of the biological evidence representation based on the absolute distance between the biological evidence result and the biological evidence threshold; performing mathematical operations based on the clarity of the biological evidence representation and a preset clarity attenuation coefficient to obtain a biological evidence clarity attenuation constraint term; and performing mathematical operations based on a preset correction upper limit, the basic gating strength of the scale, and the biological evidence clarity attenuation constraint term to obtain the correction magnitude. By adaptively constraining the correction magnitude through the aforementioned steps, the biological evidence and scale information are balanced, avoiding excessive interference of subjective data with the prediction results.

[0041] In some embodiments, the scale's base gating strength is obtained by independent gating network computation. For example, referring to equation (2), the standardized scale features and biological evidence representations are used as inputs to learn the output, with a value range constraint of 0 to 1, to characterize the scale's base activation potential for correction, where [h m [,q] represents the vector concatenation operation, h m For biological evidence representation, q represents the scale representation, and W represents the biological evidence representation. g b g represents the weight matrix and bias term obtained from training the gated network, respectively, used to perform a linear transformation on the concatenated joint features, and represents the Sigmoid activation function with output values ​​between 0 and 1.

[0042] (2).

[0043] In some embodiments, the clarity of a biological evidence representation is obtained by the absolute distance between the standardized biological evidence representation and the mapped biological evidence threshold. The greater the clarity of the biological evidence representation, the further the biological evidence result deviates from the biological evidence threshold, and the clearer the judgment structure of the biological evidence result itself. In some embodiments, the biological evidence threshold can be obtained by training an independent biological evidence prediction model. For example, the original log-probability of the risk of the biological evidence result is first mapped to the basic risk probability by referring to equation (3), then the maximum computable distance from the risk probability to the judgment threshold is unified by referring to equation (4) to eliminate the numerical scale difference caused by the threshold offset, and finally the clarity is calculated by referring to equation (5) and the calculation result is constrained to the interval between 0 and 1, where z m r represents the original log-odds ratio of biological evidence outcomes. m τ represents the base risk probability mapped to the interval between 0 and 1. m D represents the threshold of biological evidence for mapping. m Represents the limited basic risk probability r m With threshold τ m The maximum possible difference between them, clip(•,0,1) is the cutoff function, which constrains the result to the interval between 0 and 1, p m This indicates the clarity of the representation of biological evidence.

[0044] (3); (4); (5).

[0045] In some embodiments, the clarity of the bio-evidence representation can be obtained by the absolute distance between the standardized bio-evidence representation and a preset bio-evidence threshold. For example, the preset bio-evidence threshold is 0.5.

[0046] In some embodiments, the preset sharpness attenuation coefficient is constrained to the range of 0 to 1, and the correction magnitude is determined with reference to equation (6), where g ma x represents the preset correction upper limit, α p This represents the preset sharpness attenuation coefficient, and m represents the correction amplitude.

[0047] (6).

[0048] S300: Based on the correction direction and the correction magnitude, perform correction processing on the biological evidence characterization to obtain the correction prediction result.

[0049] In some embodiments, referring to equation (7), the original log odds of risk are corrected based on the correction magnitude and correction direction to further obtain the corrected prediction result, where z represents the corrected log odds of risk after correction, and γ s This indicates a correction scale parameter. For example, γ s It can be learned through the training process, or determined within a preset range using a validation set, such as γ. s Version 3.0 is acceptable.

[0050] (7).

[0051] In some instances, after obtaining the corrected log-odds, further scaling processes are applied, including but not limited to temperature scaling, log-odds affine transformation, extremum truncation preprocessing, dataset statistical normalization, learnable temperature parameters, and Tanh interval compression. This scaling process enables the adjustment of the slope of the corrected log-odds mapping, suppresses Sigmoid probability saturation, calibrates prediction confidence, and adjusts the steepness of classification decisions.

[0052] S400: Input the scale characterization, biological evidence characterization, correction direction and correction magnitude into the consistency verification model to determine the correction consistency results.

[0053] In some embodiments, the biological evidence representation, scale representation, correction magnitude, and correction direction are concatenated into vectors, and the concatenated features are input into a nonlinear mapping network for computation to obtain a correction consistency result. In some embodiments, the nonlinear mapping network includes at least one of a multilayer perceptron, logistic regression, gradient boosting tree, one-dimensional convolutional network, and self-attention network. For example, referring to equation (8), the computation is completed through the nonlinear mapping network of the multilayer perceptron, where H α Describing a multilayer perceptron, s α This indicates the result of the calibration consistency; the meanings of the other parameters are the same as those described above.

[0054] (8).

[0055] According to an embodiment of the present invention, the consistency verification model is trained in the following manner: a training dataset is obtained, which includes training scale results and corresponding training biological evidence results; the training dataset is input into a deep learning model, and the deep learning model is weakly supervised to obtain the consistency verification model by using the fit between the judgment trends of the training scale results and the corresponding training biological evidence results relative to their respective judgment thresholds, and / or the fit between the training correction prediction results and the true labels of the training samples as constraint targets.

[0056] Preferably, the consistency verification model is obtained by using the degree of fit between the training scale results and the corresponding training biological evidence results relative to their respective judgment thresholds, and the degree of fit between the training correction prediction results and the true labels of the training samples as constraint objectives, to perform weakly supervised training on the deep learning model. Using dual constraint objectives for weakly supervised training of the consistency verification model can eliminate the dependence of traditional supervised learning on a large number of accurately paired labeled samples, allowing model convergence to be achieved with incomplete labeled data, thus reducing labeling costs. During training, score fitting constraints are used to avoid score drift distortion in weakly supervised training. Simultaneously, cross-modal bias constraints are used to explore the intrinsic matching relationship between biological evidence representation and scale representation, effectively quantifying the degree of agreement between the two types of heterogeneous data. This improves the model's generalization performance on small samples, outputs robust verification results, and effectively enhances the adaptability and stability of the evaluation system in weakly labeled scenarios.

[0057] For example, referring to equation (9), the indicator function is used to determine whether the trends of the scale representation and the biological evidence representation relative to their respective judgment thresholds are in the same direction. A value of 1 is taken when the trends are consistent, and a value of 0 is taken when there is a conflict. Referring to equation (10), the judgment category obtained by the training threshold division of the corrected prediction result is compared with the sample label. A value of 1 is taken when the judgments are consistent, and a value of 0 is taken when the judgments are inconsistent. Based on the foregoing, a joint loss function is constructed, which includes at least one of the following: binary cross-entropy loss, mean squared error loss, mean absolute error loss, hinge loss, and focusing loss. For example, referring to equation (11), the joint loss is composed of the consistency branch binary cross-entropy loss and the correction matching branch binary cross-entropy loss multiplied by their respective weights and then summed. Where y cons This is the consistency branch binary cross-entropy loss, representing the cross-entropy loss indicating whether the trends of scale representation and biological evidence representation relative to their respective judgment thresholds are in the same direction. corr This is the corrected matching branch binary cross-entropy loss, which represents the cross-entropy loss that compares the corrected prediction result with the sample label based on the training threshold. Ⅱ(•) represents the indicator function; it outputs 1 if the condition within the parentheses is true, and 0 if it is false. α τ represents the scale's characteristics. α τ represents the scale threshold and the corrected prediction result. train Let y represent the classification threshold used in training, y represent the true binary label of the sample, BCE represent the binary cross-entropy loss function, and s represent the classification threshold used in training. α Indicates the result of the correction consistency, λ cons Representation, λ corr Let φ represent the weight hyperparameters of the two loss branches, respectively. auth This represents the total loss. The scale threshold τ is also included. α It can be learned from a predictive model trained on scale data, or it can be obtained by mapping a preset threshold.

[0058] (9); (10); (11).

[0059] S500: The overall confidence level is obtained by weighting the calibration consistency results and confidence index based on preset weighting coefficients.

[0060] According to embodiments of the present invention, the confidence index includes at least one of the following: baseline consistency index, correction deviation index, boundary proximity index, sample stability index, probability calibration index, and scale quality index. Setting multi-dimensional independent confidence indices can effectively overcome the shortcomings of a single evaluation dimension, accurately identify different types of outcome uncertainty and data defects, improve the overall assessment's resistance to interference, and enhance the reliability, traceability, and interpretability of overall prediction results, thus adapting to the application requirements of clinical disease risk assessment.

[0061] According to embodiments of the present invention, the baseline consistency index is used to characterize the degree of fit between the scale results and the biological evidence results relative to their respective decision thresholds. In some embodiments, referring to equation (12), the difference between the scale representation and the scale threshold, and the difference between the biological evidence representation and the biological evidence threshold are calculated. After multiplying the two differences, the evidence consistency index C is constructed through an indicator function. strat In some embodiments, as shown in Equation (13), the consistency index is represented by a continuous consistency component. This component is close to 1 when the scale representation and the biological evidence representation are on the same side of their respective thresholds and far from the thresholds, and close to 0 when they are significantly opposite. An intermediate value is given when any representation value is close to the corresponding threshold to avoid the 0 / 1 jump in the output when the representation value is close to the corresponding threshold being too sensitive. Where C strat-soft λ represents the continuous consistency index. s The slope parameter is greater than 0. ε is a preset minimum positive constant used for numerical calculation error tolerance to prevent abnormal situations such as the denominator being zero or the logarithmic independent variable being equal to zero during the calculation process, thereby improving the numerical solution stability of the evidence continuity consistency index.

[0062] (12); (13).

[0063] According to embodiments of the present invention, the correction deviation index characterizes the degree of dependence of the correction prediction result on the scale representation. In some embodiments, referring to equation (14), the absolute difference between the correction prediction risk value and the biological evidence representation is calculated. This difference is divided by a preset scale adjustment parameter and the negative number is taken. Then, it is substituted into the natural index to obtain the correction deviation index. The index value is between 0 and 1. The smaller the value, the greater the correction change of the scale representation on the biological evidence representation, and the stronger the dependence of the correction prediction result on the scale representation. Wherein C gap η represents the correction deviation index, η represents the preset scale adjustment parameter, and exp(•) represents the natural exponential function.

[0064] (14).

[0065] According to an embodiment of the present invention, the boundary proximity index is used to characterize the sensitivity of the judgment interval in which the corrected prediction result is located to a conclusion reversal. In some embodiments, referring to equation (15), the absolute distance between the corrected prediction result and a preset judgment threshold is calculated, the distance is divided by a preset scale parameter, the negative is taken, and a natural exponentiation is performed. The boundary proximity index is obtained by subtracting the exponentiation result from the constant 1, where C bd The threshold value represents the boundary proximity index, and δ represents the preset scale parameter. This index ranges from 0 to 1. A smaller index value indicates that the corrected prediction result is closer to the critical boundary, and the probability of the judgment conclusion shifting due to small numerical perturbations is higher. In some embodiments, the preset judgment threshold is a judgment threshold learned through training or a mapping result of a biological evidence threshold.

[0066] (15).

[0067] According to embodiments of the present invention, a sample stability index is used to characterize the volatility risk of a single sample prediction result. In some embodiments, as shown in equation (16), where C unc This represents the sample stability index, where β1, β2, and β3 represent the preset non-negative weight parameters, u e The model calculates the volatility index, u α H( represents the fluctuation index of input data) ) represents the predicted information entropy, u eHigh values ​​usually indicate insufficient similar samples in the training set or unstable judgment between models. High uα usually indicates large fluctuations in input quality or individual state. The uncertainty components of the model calculation fluctuation index, input data fluctuation index and the prediction information entropy corresponding to the correction prediction result are obtained respectively. The three types of uncertainty components are weighted and summed by preset non-negative weight hyperparameters to obtain the weighted uncertainty sum. The negative of the sum is substituted into the natural exponential function for mapping operation to finally obtain the sample stability index. The sample stability index ranges from 0 to 1. The smaller the index value, the worse the stability of the prediction conclusion. In some embodiments, as shown in Equation (17), multiple sets of prediction values ​​under the same set of input are obtained through multiple forward propagation. The variance of the prediction value is used to characterize the prediction fluctuation caused by the randomness of the internal parameters of the model. Where b=1,...,B represents the total number of sampling times of multiple forward propagation of the model, (b) represents the single correction prediction result obtained by the b-th random forward propagation, and Var(•) represents the variance operation. In some embodiments, as shown in Equation (18), the input data fluctuation index is calculated based on the degree of insufficient sequencing depth, the proportion of missing features and the degree of scale abnormality in biological evidence. Where q depth Indicates the degree of insufficient sequencing depth, q miss Indicates the proportion of missing features, q scale The scale indicates the degree of abnormality, and c1, c2, and c3 represent the non-negative weights of the three calculation parameters, respectively. In some embodiments, the predictive information entropy H is obtained by referring to equation (19). A high value usually indicates that the closer the correction prediction result is to the decision threshold, the more hesitant the model is in its decision-making, and the higher the ambiguity of the decision conclusion.

[0068] (16); (17); (18); (19).

[0069] According to embodiments of the present invention, a probabilistic calibration index is used to characterize the fitting deviation between the corrected prediction result and the true sample distribution. In some embodiments, the probabilistic calibration result is determined based on an expected calibration error, which includes at least one of a local expected calibration error, a global expected calibration error, or a smoothed expected calibration error. Exemplarily, referring to equation (20), the probabilistic calibration index is determined based on a local expected calibration error, wherein the local expected calibration error E... cal This represents the average difference between the "model output probability" and the "actual occurrence proportion of the validation set" within the current probability interval. In some embodiments, the validation set samples are assigned to K probability bins B according to the calibrated probability results. kIn some embodiments, equal-width binning or equal-frequency binning can be used, and a minimum sample size can be set for each probability bin. Then, the local expected calibration error is obtained by referring to equation (21) for the probability bin containing the current object to be evaluated. For small sample validation data, equal-frequency binning is preferred. Wherein, B k Let y represent the set of all samples that fall into the k-th bin. i E represents the true binary classification label of the i-th sample. cal (k) represents the local desired calibration error, C cal This indicates the probability calibration result.

[0070] (20); (twenty one).

[0071] According to embodiments of the present invention, a scale quality index is used to characterize the data completeness of scale results. In some embodiments, the scale quality index is calculated using formula (22), where C qual Indicates the quality index of the scale, m j This indicates that the j-th item in the scale is missing, out of bounds, or cannot be parsed, where J is the number of scale dimensions. For example, during the training phase, preset default values ​​can be used to maintain the integrity of the scale data matrix, and during the deployment and inference phase, it can be filled with the mean of the training set. When recording the missing mask at the deployment end, this indicator can be further used to correct the overall confidence.

[0072] (twenty two).

[0073] In some embodiments, the comprehensive confidence score is calculated from the calibration consistency results and the aforementioned six confidence indices. In some embodiments, the comprehensive confidence score is further calculated from the independent confidence header output results in the prediction network backbone of the biological evidence results. For example, the comprehensive confidence score is calculated as shown in Equation (23), wherein the preset weighting coefficients can be obtained by optimization on the validation set with constraints of verification accuracy, calibration error, or actual business objectives, or the preset weighting coefficients can be obtained by presetting based on fixed rules and then confirming them on the validation set. w1, w2, w3, w4, w5, w6, w7, and w8 are the preset weighting coefficients of each calculation parameter, and the sum of all preset weighting coefficients is 1, wherein w1~w7 are non-zero preset coefficients, and w8 is equal to 0 when the confidence header output results are not enabled.

[0074] (twenty three).

[0075] In some embodiments, the overall confidence level is further adaptively adjusted. For example, as shown in Equation (24), when the deviation between the corrected prediction result and the biological evidence result is too large and exceeds the preset deviation threshold, it indicates that the scale correction magnitude is abnormally aggressive and the correction logic deviates from the original biological evidence, and the overall confidence level is reduced. For example, as shown in Equation (25), when the fluctuation index calculated by the model or the fluctuation index of the input data exceeds the preset index threshold, it indicates that the stability of the modeling operation or the input data source is too low. For example, as shown in Equation (26), when the correction magnitude is greater than the preset magnitude threshold and the correction consistency result is lower than the preset correction consistency threshold, it indicates that the correction effect of the object to be evaluated is too poor, and the overall confidence level is reduced. Wherein, κ1, κ2, and κ3 are preset reduction coefficients with values ​​from 0 to 1, Δ1 is the preset deviation threshold, Δ2 is the preset index threshold for the fluctuation index calculated by the model, Δ3 is the preset index threshold for the fluctuation index of the input data, m0 is the preset magnitude threshold, and s0 is the preset correction consistency threshold.

[0076] (twenty four); (25); (26).

[0077] S600: The evaluation results are obtained based on the overall confidence level.

[0078] According to an embodiment of the present invention, obtaining the evaluation result based on the comprehensive confidence level further includes: determining the confidence level based on the comprehensive confidence level; determining the review reason based on the confidence index used in the comprehensive confidence level; determining the review recommendation based on the confidence level and the review reason; and outputting at least one of the confidence level, review reason and review recommendation as the evaluation result.

[0079] In some embodiments, the confidence levels include high confidence, medium confidence, and low confidence. For example, a composite confidence level of 0.75 or higher is considered high confidence, a composite confidence level of 0.5 or higher but lower than 0.75 is considered medium confidence, and a composite confidence level lower than 0.5 is considered low confidence. In some embodiments, the aforementioned high, medium, and low confidence levels correspond to review recommendations for routine follow-up, review recommendation, and priority review, respectively.

[0080] In some embodiments, as shown in Equation (27), a review score can be calculated based on the comprehensive confidence level, the probability change before and after correction, the model calculation fluctuation index, the input data fluctuation index, and the correction consistency result. Review suggestions can be obtained through the belonging interval of the review score, where ω1, ω2, ω3, ω4, and ω5 represent the calculation weights of the aforementioned parameters.

[0081] (27).

[0082] In some embodiments, the reason for review is determined based on anomalies in the confidence index used for the overall confidence level. For example, a low correction consistency result but a high correction magnitude indicates that the corrected prediction result is significantly influenced by the scale, but this influence is not adequately supported by biological evidence. The corresponding review reason could be output as "The scale correction is large, but the evidence consistency is insufficient; it is recommended to prioritize reviewing scale completion and interview records." For example, a low boundary proximity index indicates that the corrected prediction result is close to the decision threshold. This does not equate to a model error, but rather indicates that the binary classification label is sensitive to threshold settings. The corresponding review reason could be output as "The risk probability is close to the threshold; it is recommended to make a judgment after follow-up or a second assessment."

[0083] In some embodiments, the reason for review can also be determined by the abnormal indicators in the aforementioned review score calculation method. For example, when the model calculates a high fluctuation index, it usually indicates that the sample is not similar enough to common samples in the training set, or that there is a combination relationship between multiple sets of biological evidence features that the model has not yet fully learned. The corresponding review reason can be output as "Insufficient similar training samples; review combined with clinical interviews or retest data is recommended." For example, when the input data fluctuation index is high, taking microbial metagenomic data as an example, even with fixed model parameters, the sample itself still has significant observational noise. Common reasons include low sequencing depth, a high proportion of missing abundance matrices, missing or abnormal scale scores, and short-term fluctuations in individual status. The corresponding review reason can be output as "High noise in the input data; verification of sequencing quality or completeness of scale completion is recommended."

[0084] In another aspect of the invention, a disease prediction verification assessment device is provided, with reference to Figure 2 As shown, the device includes: a standardization module 1000, used to standardize the scale results and biological evidence results of the object to be evaluated to obtain scale representation and biological evidence representation; a correction coefficient module 2000, used to determine the correction direction and correction magnitude of the scale representation to the biological evidence representation based on the relative magnitude of the scale representation and the biological evidence representation and the distance between the biological evidence representation and the biological evidence threshold, respectively; a correction module 3000, used to correct the biological evidence representation based on the correction direction and the correction magnitude to obtain a correction prediction result; a verification module 4000, used to input the scale representation, biological evidence representation, correction direction and correction magnitude into the consistency verification model to determine the correction consistency result; a confidence module 5000, used to obtain a comprehensive confidence level by weighting the consistency result and the acceptance index based on a preset weighting coefficient; and an evaluation module 6000, used to obtain an evaluation result based on the comprehensive confidence level.

[0085] It should be understood that the device embodiments and method embodiments can correspond to each other, and similar descriptions can be referred to the method embodiments. To avoid repetition, they will not be repeated here. Specifically, this device can execute the above-described corresponding disease prediction review assessment method embodiments individually or as a whole, and the foregoing and other operations and / or functions of each unit in the device are respectively to implement the corresponding processes in the above method, which will not be repeated here for the sake of brevity.

[0086] In another aspect of the invention, an electronic device is provided, with reference to Figure 3 As shown. The electronic device includes: a memory 7000 and a processor 8000. The memory 7000 is used to store computer programs; the processor 8000 is used to execute the aforementioned disease prediction review assessment method. Figure 3 The dashed line indicates that the memory 7000 and processor 8000 can transmit data and instructions through dual links, specifically including a physical connection link and a virtual information transmission link.

[0087] For example, processor 8000 can be used to execute the steps in the above method according to the instructions in the computer program.

[0088] Processor 8000 may include, but is not limited to: General-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0089] The 7000 memory includes, but is not limited to: Volatile memory and / or non-volatile memory. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static RAM (SRAM), Dynamic RAM (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDR SDRAM), Enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), and Direct Rambus RAM (DR RAM).

[0090] A computer program can be divided into one or more modules, which are stored in memory 7000 and executed by processor 8000 to perform the method provided by the present invention. The one or more modules can be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in an electronic device.

[0091] refer to Figure 4 As shown, the electronic device may further include: Transceiver 9000, which can be connected to processor 8000 or memory 7000.

[0092] The processor 8000 can control the transceiver 9000 to communicate with other devices; specifically, it can send information or data to other devices or receive information or data sent by other devices. The transceiver 9000 may include a transmitter and a receiver. The transceiver 9000 may further include antennas, and the number of antennas may be one or more.

[0093] It should be understood that the various components in an electronic device are connected through a bus system, which includes not only a data bus, but also a power bus, a control bus, and a status signal bus.

[0094] In another aspect, the present invention provides a computer-readable storage medium storing computer instructions or programs that, when executed on a computer, cause the aforementioned disease prediction review assessment method to be performed.

[0095] In another aspect of the invention, a computer program product is provided, comprising computer instructions that, when some or all of the computer instructions are run on a computer, cause the aforementioned disease prediction review assessment method to be executed.

[0096] The embodiments described below are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0097] Example 1 Based on the data sources of scale results and microbial metagenomic evidence (hereinafter referred to as microbial results) from a sample of 201 parturients, a fixed random seed and stratified cross-validation were used to evaluate model training and validation. Specifically, the basic random seed was 42, and the three ensemble random seeds were 42, 59, and 76, respectively. Five-fold stratified cross-validation was performed under each random seed, resulting in 15 fold models. In each fold, the training fold was used to determine model parameters, normalization parameters, gating parameters, probability calibration parameters, and risk assessment thresholds; the corresponding validation fold samples did not participate in the training and parameter determination of that fold model. The results of each validation fold were summarized as the cross-validation prediction results, used to evaluate the internal generalization performance and configuration contribution of this method within this cohort. In this embodiment, the risk assessment threshold of the optimized model trained based on the scale results and microbial results was 0.30. The stratification thresholds for the overall confidence and review scores were determined by the risk-coverage relationship on the validation set and the review resource capacity. The confidence results included high confidence, medium confidence, and low confidence, and the review recommendations included routine follow-up, recommended review, and priority review.

[0098] The prediction results of the evaluation model are shown in Table 1. Under the same data and a 5-fold × 3 random seed validation, the basic model trained using only microbiological results achieved an accuracy of 0.7512, an F1 score of 0.7596, an area under the curve (AUC) of 0.7994, and a Brier score of 0.1001. After further incorporating scale data, scale dynamic correction, temperature scaling probability calibration, and a comprehensive confidence level splitting mechanism based on the microbiological results data, the optimized model based on the risk assessment threshold improved its accuracy to 0.9204, its F1 score to 0.9208, its AUC to 0.9652, and its Brier score to 0.0696. These comparisons demonstrate that scale-controlled dynamic correction can effectively improve the accuracy of prediction results.

[0099] The reliability assessment results of the optimized model were analyzed. In this optimized model, the accuracy rate was 0.9700 for the routine follow-up group, the error rate was 0.0333 for the recommended review group, and the error rate was 0.2683 for the priority review group. The review cohort, after merging the output of the recommended review and priority review assessment results, covered 81.25% of the false prediction samples. This indicates that after the overall accuracy was improved, the comprehensive confidence level and review score can effectively provide risk warnings for false prediction samples, thereby avoiding the misallocation or waste of clinical diagnostic or review resources.

[0100] Table 1

[0101] To verify the effectiveness of the main configuration in the method of this invention, ablation experiments were conducted. The experimental results are shown in Table 2. Under the same 5-fold × 3 random seed validation, the accuracy of the complete mechanism model based on the risk judgment threshold was 0.9204. After removing the scale dynamic correction, the accuracy decreased to 0.7512, the F1 score decreased to 0.7596, and the area under the curve decreased to 0.7994, indicating that the scale information, after dynamic gating, makes a significant contribution to the overall prediction accuracy. After removing the confidence assessment constraint, the accuracy was 0.8856, a relatively small decrease, but the coverage of erroneous prediction samples in the review cohort, including suggested review and priority review, decreased significantly to 0.5217, indicating that the confidence assessment mechanism makes a significant contribution to the identification ability of individuals requiring review.

[0102] Table 2

[0103] Furthermore, based on the prediction results of 5-fold × 3 random seed cross-validation of the same 201 samples, a quantitative analysis was conducted on the relationship between consistency score and prediction error. The prediction error indicator variable was set to 1 for prediction error and 0 for prediction correctness. The results showed that the correlation coefficient between consistency score and prediction correctness was 0.2802, and the correlation coefficient between consistency score and the prediction error indicator variable was -0.2802. After dividing the sample into low, middle, and high quantile groups according to the tertiles of the consistency score, the error rates of the three groups were 0.1940, 0.0448, and 0.0000, respectively. A linear probability model was then fitted with the prediction error indicator variable as the dependent variable and the consistency score as the independent variable. The slope of the fit was -0.7178, and the coefficient of determination was 0.0785. These results indicate that as the consistency score decreases, the overall risk of prediction error increases; however, this relationship is not strictly linear. Therefore, the consistency score is more suitable as an indicator for enriching and reviewing error risk, and should be combined with boundary proximity indicators, sample stability indicators, etc., for comprehensive judgment.

[0104] Under the same data caliber, the correlation coefficient between the overall confidence level and prediction accuracy was 0.2670, and the correlation coefficient with the prediction error indicator variable was -0.2670. After dividing the sample into low quantile, middle quantile, and high quantile groups according to the ternary quantile of the overall confidence level, the error rates of the three groups were 0.1642, 0.0746, and 0.0000, respectively. This result indicates that the overall confidence level is not simply equivalent to the classification probability, but can concentrate more mispredicted samples in the low confidence or verification intervals, thus providing a basis for the allocation of verification resources.

[0105] Further comparisons were made between scale-based dynamic gating correction and fixed-weight correction. The final prediction result of the fixed-weight correction method was calculated using the formula: Final Prediction Result = (1-w) × Microbial Risk + w × Scale Risk. The scale risk was obtained by averaging the scores of the Edinburgh Postnatal Depression Scale and the SDS Self-Rating Depression Scale, truncated to the 0-1 interval. For missing or annotated scale values, resolvable values ​​were extracted first; those still unresolvable were filled with the mean of the training cohort. When the fixed weight w = 0.4, the accuracy was 0.8259, the F1 score was 0.8498, and the Brier score was 0.1464; when the fixed weight w = 0.6, the accuracy was 0.8756, the F1 score was 0.8879, and the Brier score was 0.1568. The scale-based dynamic gating correction resulted in an accuracy of 0.9204, an F1 score of 0.9208, and a Brier score of 0.0696. Compared to fixed-weight correction, scale dynamic gating correction can adaptively adjust the scale participation amplitude according to the clarity of biological evidence, while improving threshold determination results and probability error.

[0106] Example 2 Based on the assessment model obtained in Example 1, the subjects who underwent a follow-up examination 6 weeks postpartum were reassessed. After completing the Edinburgh Postpartum Depression Scale and the Self-Rating Depression Scale, the raw scores received by the system were 15 and 58, respectively. The two scales were normalized to the training set mean and standard deviation before being fed into the scale encoder, with the scale indicating a high risk. The risk probability output by the microbial evidence branch was 0.78, with only minor adjustments to the scale gating. The corrected prediction result after scale involvement was 0.82, and the judgment threshold was 0.30. The risk probability of this sample was much higher than the threshold, and both the scale result and the microbial evidence branch result were in the high-risk range. The input data fluctuation was low, and the corrected consistency score was 0.72. The output result of this example was: high confidence, routine follow-up, with the review reason being "both the scale score and the microbial risk indicate high risk, the risk probability is far from the threshold, and the current result is highly stable." In this scenario, the scale result and microbial evidence jointly support the high-risk indication, and the review record can be retained according to the routine follow-up procedure.

[0107] Example 3 Based on the assessment model obtained in Example 1, the subjects followed up 8 weeks postpartum were reassessed. The subject scored 17 on the Edinburgh Postpartum Depression Scale and 61 on the Self-Rating Depression Scale, indicating a high risk. The microbial evidence branch output was 0.18, below the judgment threshold of 0.30, placing it in the low-risk range. After the scale was incorporated, the corrected prediction result rose to 0.42, the correction size was 0.31, and the corrected consistency score was 0.34. Because the scale results were inconsistent with the microbial evidence branch results, and the final probability is highly sensitive to the scale input, the output of this example was: low confidence, priority review. The reason for review was: "The scale indicates high risk, but the microbial evidence indicates low risk; the scale correction is significant, it is recommended to review the scale completion environment, understanding of items, and interview records." In this scenario, the model did not reject the scale score but instead listed the subject as a priority review target, applicable to situations where the subject was influenced by accompanying persons during completion, experienced recent sudden life events, or where interview information still needs supplementary confirmation.

[0108] Example 4 Based on the assessment model obtained in Example 1, the subjects who underwent a follow-up outpatient visit 5 weeks postpartum were reassessed. The Edinburgh Postpartum Depression Scale score was 6, and the self-rating depression scale score was 39, indicating a low risk. The microbiological evidence branch output was 0.74, higher than the judgment threshold of 0.30, and far from the threshold. Since the microbiological evidence was relatively clear, the scale gating module only allowed minor adjustments to the scale, resulting in a combined risk probability of 0.68 after scale involvement. The system calculated a low degree of consistency between the scale and the microbiological evidence, but the corrected prediction result was far from the threshold, with a corrected consistency score of 0.41. The output result of this example was: low confidence, recommendation for review. The reason for review was "the scale score is low but the microbiological risk is high; the scale result is inconsistent with the biological evidence; it is recommended to review in conjunction with sleep status, breastfeeding stress, family support, and the results of the next follow-up visit." In this scenario, the model did not determine that the subject concealed or misreported information; it only indicated that the consistency between the current scale result and the combined risk probability was insufficient, requiring further confirmation through follow-up and interviews.

[0109] Example 5 Based on the assessment model obtained in Example 1, the subjects followed up 10 weeks postpartum were reassessed. The subjects scored 10 on the Edinburgh Postpartum Depression Scale and 46 on the Self-Rating Depression Scale. The output of the microbial evidence branch was 0.29, the corrected prediction result after scale involvement was 0.31, and the judgment threshold was 0.30. This sample's corrected prediction result only exceeded the threshold by 0.01, indicating that the result is close to the high / low risk judgment threshold, and the model result fluctuation level is moderate. In this example, even though the scale result is basically consistent with the microbial evidence branch result, the output of this example is: moderate confidence, review recommended. The reason for review is that "the risk probability is close to the threshold; a small fluctuation in input or change in probability calibration may change the high / low risk indication; it is recommended to combine the next follow-up results for judgment." In this scenario, the main reason for recommending review is that the risk result is too close to the threshold, rather than the inconsistency between the scale and the microbial evidence.

[0110] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0111] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for disease prediction and verification assessment, characterized in that, include: Standardize the scale results and biological evidence results of the subjects to be evaluated to obtain scale representations and biological evidence representations; Based on the relative magnitude of the scale representation and the biological evidence representation and the distance between the biological evidence representation and the biological evidence threshold, the correction direction and correction magnitude of the scale representation on the biological evidence representation are determined respectively. Based on the correction direction and the correction magnitude, the biological evidence characterization is corrected to obtain the correction prediction result; The scale representation, the biological evidence representation, the correction direction, and the correction magnitude are input into the consistency verification model to determine the correction consistency result. The overall confidence level is obtained by weighting the correction consistency results and the confidence index based on preset weighting coefficients. The evaluation results are obtained based on the comprehensive confidence level.

2. The method according to claim 1, characterized in that, The correction range is determined through the following steps: The basic gating strength of the scale is determined based on the scale representation and the biological evidence representation via a machine learning network. The clarity of the biological evidence representation is determined based on the absolute distance between the biological evidence result and the biological evidence threshold. Based on the clarity of the biological evidence representation and the preset clarity attenuation coefficient, mathematical operations are performed to obtain the biological evidence clarity attenuation constraint term. Based on the preset correction upper limit, the scale's basic gating strength, and the biological evidence clarity attenuation constraint, mathematical operations are performed to obtain the correction magnitude.

3. The method according to claim 1, characterized in that, The biological evidence results include disease risk prediction results based on metagenomic data, which are obtained through the following methods: N feature matrices are constructed from the metagenomic data of the object to be evaluated, and the N feature matrices correspond to microbial features at different biological levels. Feature extraction is performed on the N feature matrices respectively, and the feature representations extracted from the features are fused across types to obtain a fused representation; Based on the fusion representation, disease risk is predicted, and the prediction results are obtained.

4. The method according to claim 1, characterized in that, The consistency verification model is trained in the following manner: Obtain a training dataset, which includes: training scale results and corresponding training biological evidence results; The training dataset is input into the deep learning model. The deep learning model is weakly supervised and trained using the fit between the training scale results and the corresponding training biological evidence results relative to their respective judgment thresholds, and / or the fit between the training correction prediction results and the true labels of the training samples as constraint targets, to obtain the consistency verification model.

5. The method according to claim 1, characterized in that, The confidence index includes at least one of the following: baseline consistency index, correction deviation index, boundary proximity index, sample stability index, probability calibration index, and scale quality index. The baseline consistency index is used to characterize the degree of fit between the scale results and the biological evidence results relative to their respective judgment thresholds. The correction deviation index is used to characterize the degree of dependence of the correction prediction result on the scale characterization; The boundary proximity index is used to characterize the sensitivity of the conclusion reversal when the judgment interval in which the corrected prediction result is located is located. The sample stability index is used to characterize the volatility risk of single-sample prediction results. The probability calibration index is used to characterize the fitting deviation between the corrected prediction result and the true sample distribution; The scale quality indicators are used to characterize the data completeness of the scale results.

6. The method according to claim 5, characterized in that, The process of obtaining the evaluation result based on the comprehensive confidence level further includes: The confidence level is determined based on the overall confidence level; The reasons for review are determined based on the confidence index adopted by the comprehensive confidence level. Based on the confidence level and the reasons for review, a review recommendation is determined; Output at least one of the confidence level, the reason for review, and the review recommendation as the evaluation result.

7. A disease prediction and verification assessment device, characterized in that, include: The standardization module is used to standardize the scale results and biological evidence results of the subjects to be evaluated, so as to obtain scale representations and biological evidence representations. The correction coefficient module is used to determine the correction direction and correction magnitude of the scale representation on the biological evidence representation based on the relative magnitude of the scale representation and the biological evidence representation and the distance between the biological evidence representation and the biological evidence threshold, respectively. A correction module is used to perform correction processing on the biological evidence characterization based on the correction direction and the correction magnitude to obtain a correction prediction result; The verification module is used to input the scale representation, the biological evidence representation, the correction direction, and the correction magnitude into the consistency verification model to determine the correction consistency result; The confidence module is used to calculate the overall confidence level by weighting the consistency results and the acceptance index based on preset weighting coefficients. The evaluation module is used to obtain the evaluation result based on the comprehensive confidence level.

8. An electronic device, characterized in that, include: Memory and processor; The memory is used to store computer programs; The processor is configured to execute the computer program to implement the disease prediction review assessment method as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions or programs that, when executed on a computer, cause the disease prediction review assessment method as described in any one of claims 1 to 7 to be performed.

10. A computer program product, characterized in that, The computer program product includes computer instructions that, when some or all of the computer instructions are run on a computer, cause the disease prediction review assessment method as described in any one of claims 1 to 7 to be executed.

Images