Cascade model-based clinical cure prediction method for chronic hepatitis B

Abstract

The invention provides a clinical cure prediction method for chronic hepatitis B based on a cascade model, and belongs to the technical field of clinical prediction.The clinical cure prediction method comprises the steps that clinical data of a patient is collected, screened and preprocessed to obtain an effective sample, and the sample contains basic information, an ALT initial value and HBsAg quantitative indexes of multiple follow-up visit periods such as 0 week and 12 weeks; a first auxiliary vector is constructed by combining historical demands with historical body information, and a second auxiliary vector is constructed by capturing effect change indexes; establishing a global and periodic local mapping relation; training a cascade model containing a classification layer, a prediction layer and an auxiliary layer based on an ensemble learning algorithm; inputting a current sample and demand of a patient, and outputting the HBsAg negative conversion probability, the optimal healing cycle and the expected standard reaching probability of each cycle in the future. The method solves the problems that an existing model is low in prediction precision, poor in dynamic adaptability and single in dimension, and provides a reliable basis for personalized treatment.

Description

A cascade model-based method for predicting the clinical cure of chronic hepatitis B. Technical Field

[0001] This invention relates to the field of clinical prediction technology, and in particular to a method for predicting the clinical cure of chronic hepatitis B based on a cascade model. Background Technology

[0002] Chronic hepatitis B (CHB) is a global public health problem caused by hepatitis B virus (HBV) infection, affecting approximately 300 million people worldwide. Most patients require long-term drug treatment to achieve the goal of clinical cure, i.e., persistent negative HBsAg levels after treatment cessation. Currently, drugs such as PEG-IFNα-2b are commonly used in clinical practice; however, these drugs are expensive and prone to adverse reactions. Furthermore, patients may require 2-3 years or even longer of treatment to achieve clinical cure, and only a portion of patients achieve this outcome. Therefore, accurately predicting the probability of clinical cure and the potential cure period in the early stages of treatment is crucial for optimizing treatment strategies and reducing the economic burden and risk of adverse reactions for patients.

[0003] Existing methods for predicting clinical cure in chronic hepatitis B mainly suffer from three core problems: First, most prediction models are built using a single algorithm, which can easily lead to a low positive predictive value (PPV) when faced with an imbalanced dataset between seroconverted and non-seroconverted samples, resulting in insufficient reliability of the prediction results. Second, most models are statically designed, only able to predict the cure outcome at a specific time point based on input data from a fixed treatment cycle, and unable to adjust the prediction results in real time using dynamically updated clinical data during the patient's treatment process (such as changes in HBsAg levels in each cycle), making it difficult to adapt to the dynamic nature of the treatment process. Third, the prediction dimension is singular, only outputting the probability of cure in a fixed future cycle, unable to provide the seroconversion probability distribution for multiple subsequent treatment cycles or the range of cycles with the highest probability of cure, making it difficult to meet the comprehensive needs of clinical decision-making for treatment process planning.

[0004] Although some studies have attempted to use machine learning algorithms to build predictive models, such as logistic regression and single ensemble learning models, the aforementioned core problems have not yet been solved.

[0005] Therefore, this invention proposes a method for predicting the clinical cure of chronic hepatitis B based on a cascade model. Summary of the Invention

[0006] This invention provides a method for predicting the clinical cure of chronic hepatitis B based on a cascade model, in order to solve the aforementioned technical problems.

[0007] This invention provides a method for predicting the clinical cure of chronic hepatitis B based on a cascade model, comprising: Step 1: collecting clinical data of chronic hepatitis B patients to screen initial samples, and standardizing and preprocessing the initial samples to obtain effective samples; Step 2: decomposing the historical needs of chronic hepatitis B patients and combining them with the historical physical information of chronic hepatitis B patients when outputting their historical needs, mining all inferior mutation indicators based on the theoretical conservative treatment effect vector and the theoretical optimal treatment effect vector, and combining them to obtain a first auxiliary vector; simultaneously, capturing the effect change indicators of chronic hepatitis B patients in different follow-up periods and combining them to obtain a second auxiliary vector; Step 3: comprehensively establishing the first auxiliary vector. Step 4: Establish a global mapping relationship between the second auxiliary vector and the corresponding effective sample, and simultaneously establish periodic local mapping relationships between the first auxiliary vector, the second auxiliary vector and the corresponding effective sample; Step 5: Train and construct a cascaded model based on the ensemble learning algorithm and combining effective samples, global mapping relationships, periodic local mapping relationships, and historical needs, wherein the cascaded model includes: a classification layer, a prediction layer and an auxiliary layer; Step 6: Obtain the current sample and current needs of the patient's treatment and input them into the trained cascaded model, outputting the probability of the patient's hepatitis B surface antigen turning negative for multiple future treatment cycles, the treatment cycle with the highest probability of the patient achieving clinical cure, and the patient's expected probability of achieving the target based on each treatment cycle.

[0008] Preferably, the combination to obtain the first auxiliary vector includes: decomposing the historical needs of patients with chronic hepatitis B into several sub-needs and the purpose and constraint of each sub-need; matching candidate cure conditions related to each sub-need through a purpose-constraint-candidate checklist; determining the merits and demerits of each candidate cure condition based on the first distribution of all sub-needs and combined with the historical physical information of the patients with chronic hepatitis B; determining the conservative position point of the corresponding candidate cure condition; obtaining the theoretical conservative treatment effect vector that satisfies the historical needs; and obtaining the theoretical optimal treatment effect vector based on the merits and demerits mining results; screening inferior mutation indicators based on the theoretical conservative treatment effect vector and the theoretical optimal treatment effect vector; and combining all inferior mutation indicators to obtain the first auxiliary vector.

[0009] Preferably, the process of mining the merits of each candidate cure condition and determining the conservative position point for that candidate cure condition includes: obtaining historical cure effect sequences matching the candidate cure conditions from a historical database, traversing the cure itemset based on each historical cure effect, constructing an initial thread for the same cure item, and determining the thread type of the initial thread; determining the mining precision of the corresponding cure item based on the thread type, total thread cycle length, number of thread cycles, and average single cycle length of each initial thread; if the mining precision is less than a preset precision, using the average single cycle length as the cluster length for all initial threads... Cluster analysis is performed on a single cycle length, and the cluster with the largest total length is selected as the analysis target. The merits of the analysis target are evaluated, and the average effect with the highest probability density under the analysis target is combined to obtain the reference conservative point for the corresponding cure item. If the mining accuracy is not less than the preset accuracy, triples are constructed for each single cycle in the initial thread, wherein the triples include: single cycle length, single cycle effect, and corresponding cure participation. The merits of each triple are evaluated and a fitting merit curve is constructed. It is determined whether there are discrete points that are not located on the fitting merit curve itself. If so, Obtain the number of points N3 in the triplet of the goodness-of-fit curve itself and the goodness-of-fit distribution function F3. Determine the upper limit line ys and the lower limit line yx, wherein the vertical coordinate of the upper limit line ys is less than the maximum vertical value of the goodness-of-fit curve and greater than the minimum vertical value of the goodness-of-fit curve, the vertical coordinate of the lower limit line yx is greater than the minimum vertical value of the goodness-of-fit curve and less than the maximum vertical value of the goodness-of-fit curve, and the vertical coordinate of the upper limit line ys is greater than the vertical coordinate of the lower limit line yx. Count the first number N1 of discrete points located above the goodness-of-fit curve and the number of points located on the upper limit line. The system calculates the second number N2 of discrete points above ys and the first position distribution function F1 based on the first number N1, and the second position distribution function based on the second number N2; simultaneously, it calculates the third number N3 of discrete points located below the fitting good / bad curve, the fourth number N4 of discrete points located below the lower limit line yx, and the third position distribution function F3 based on the third number N3, and the fourth position distribution function F4 based on the fourth number N4; based on the statistical results, it determines the reference conservative point for the corresponding cure item; based on the reference conservative point for each cure item, it determines the conservative position point for the corresponding candidate cure condition.

[0010] Preferably, the thread type includes: full-thread type and half-thread type.

[0011] Preferably, based on statistical results, determining the reference conservative point for the corresponding cure item includes: normalizing the single-cycle effect corresponding to the discrete point to obtain the normalized single-cycle effect. Based on F1, F2, F3, and F4, and combined with Constructing the fusion distribution function ,in, Let be the weighting coefficients of the distribution function at the j-th location, and ; Let j be the distribution function at the j-th location; The weight coefficients of the distribution function at the j-th location; Let be the normalized average of the normalized single-period effect of the discrete points involved in the j-th location distribution function. Corresponding to F1, F2, F3, and F4 respectively; calculate the conservative constraint coefficients. ,in, Let be the single-cycle length corresponding to the i-th discrete point; The mean of all single-cycle lengths of the initial thread; n is the total number of discrete points; This is the maximum value of the single-cycle length of the initial thread; The minimum single-cycle length of the initial thread is determined; the fusion distribution function is determined. mode And calculate the initial value of the reference conservative point based on the conservative constraint coefficient. .

[0012] Preferably, the accuracy of the mining for the corresponding cure item is determined, including: ; ; ; ;in, The mining precision for the corresponding cure item; This is to initially measure the intermediate accuracy. The normalized value of the total thread cycle length of the initial thread; N is the number of thread cycles; Normalized average length of a single period; Thread cycle density; The average length concentration; For thread structure factors; Based on Determine the effective stability of the curative effect; The normalized mean of the therapeutic effect; The normalized standard deviation of the curative effect; m0 is the number of historical curative effects; Let m0 be the number of historical healing effects that follow a normal distribution, and T represents the value corresponding to the thread type.

[0013] Preferably, the process of training and building the cascaded model further includes: based on the global feature distribution law of the global mapping relationship and the periodic feature fluctuation law of the periodic local mapping relationship, performing feature space adaptation transformation on the initial clinical features of the validation samples to obtain adapted validation features that match the mapping logic of the cascaded model.

[0014] Preferably, the process of training and constructing the cascaded model further includes: inputting the adaptation verification features into the trained cascaded model, simultaneously acquiring the intermediate classification results and corresponding classification weights output by the classification layer, the prediction results and prediction confidence output by the prediction layer, and the auxiliary features output by the auxiliary layer; comparing the prediction results with the original results to obtain a probability difference set; simultaneously, fusing the original cure features of the verification sample, the classification weights of the classification layer, and the auxiliary features of the auxiliary layer, and performing dot product attention calculation on the fused features to obtain fused reference features; capturing the feature extraction process logs of the classification layer, prediction layer, and auxiliary layer for the adaptation verification features to obtain periodic process sub-features of each layer; fusing the periodic process sub-features based on the functional weights of each layer in the cascaded model to obtain fused periodic process features; and comparing the fused periodic process features with the fused reference features to obtain a hierarchical feature difference set; and based on the global difference weights of the global mapping relationship and the periodic difference weights of the periodic local mapping relationship, calculating the difference probability elements and hierarchical feature difference sets under the same verification sample. Weighted encoding is performed on the comparison relationship to obtain a weighted comparison map. The weighted difference ladder corresponding to the difference probability element is determined by combining the distribution gradient of the global difference weights and the fluctuation range of the periodic difference weights in the weighted comparison map. Based on the weighted difference ladder of the same difference probability element across all validation samples, the difference ladder is weighted by reliability based on the prediction confidence output of the prediction layer, resulting in a weighted difference ladder sequence. The weighted sum of the mean square error and entropy of the weighted difference ladder sequence is used as the first interactive loss function. Simultaneously, based on all feature difference values ​​under the same difference feature, the probability difference influence coefficient is adjusted by combining the classification result confidence of the classification layer and the auxiliary feature contribution of the auxiliary layer. The cross-entropy of the feature difference value and the adjusted influence coefficient is used as the second interactive loss function. The global matching degree of the validation sample in the global mapping relationship and the periodic matching degree in the periodic local mapping relationship are calculated and used as the global loss adjustment factor and the periodic loss adjustment factor, respectively. Weights are allocated to the first and second interactive loss functions based on these adjustment factors to obtain the comprehensive loss function.

[0015] Compared with existing technologies, the beneficial effects of this application are as follows: By constructing a cascaded model containing a classification layer, a prediction layer, and an auxiliary layer, and combining patient clinical data, historical needs, and dynamic effect change indicators, early, dynamic, and multi-cycle prediction of clinical cure for chronic hepatitis B is achieved. This effectively solves the problems of existing single models having low positive predictive values, being unable to dynamically update prediction results, and only being able to predict cure status for a fixed period. It can output prediction results in the early stages of treatment and dynamically optimize prediction accuracy as the treatment cycle progresses. At the same time, it provides the probability of seroconversion and the most likely cure cycle for multiple future cycles, providing a comprehensive and reliable reference for personalized adjustment of clinical treatment plans and patient treatment decisions, and helping to reduce the economic burden on patients and the risk of adverse drug reactions.

[0016] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 is a flowchart of a clinical cure prediction method for chronic hepatitis B based on a cascade model, according to an embodiment of the present invention. Detailed Implementation

[0019] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0020] This invention provides a clinical cure prediction method for chronic hepatitis B based on a cascade model, as shown in Figure 1, comprising: Step 1: collecting clinical data of chronic hepatitis B patients to screen initial samples, and standardizing and preprocessing the initial samples to obtain effective samples; Step 2: decomposing the historical needs of chronic hepatitis B patients and combining the historical physical information of chronic hepatitis B patients when outputting historical needs, mining all inferior mutation indicators based on the theoretical conservative treatment effect vector and the theoretical optimal treatment effect vector, and combining them to obtain a first auxiliary vector; simultaneously, capturing the effect change indicators of chronic hepatitis B patients in different follow-up periods and combining them to obtain a second auxiliary vector; Step 3: comprehensively establishing the first auxiliary... Step 4: Establish a global mapping relationship between the first auxiliary vector, the second auxiliary vector, and the corresponding effective samples. Simultaneously, establish periodic local mapping relationships between the first auxiliary vector, the second auxiliary vector, and the corresponding effective samples. Step 5: Train and construct a cascaded model based on the ensemble learning algorithm, combined with effective samples, global mapping relationships, periodic local mapping relationships, and historical needs. The cascaded model includes a classification layer, a prediction layer, and an auxiliary layer. Step 6: Obtain the current samples and current needs of the patient's treatment and input them into the trained cascaded model. Output the probability of the patient's hepatitis B surface antigen turning negative for multiple future treatment cycles, the treatment cycle with the highest probability of clinical cure, and the patient's expected probability of achieving the target for each treatment cycle.

[0021] In this embodiment, clinical data is collected through channels such as hospital electronic medical record systems, laboratory testing data management platforms, and clinical research databases.

[0022] In this embodiment, the initial sample refers to the set of samples initially selected from the collected clinical data that contains the core indicators required for prediction. This includes basic patient information, initial ALT values, and quantitative indicators of hepatitis B surface antigen (HBsAg) at multiple follow-up periods (weeks 0, 12, 24, 36, and 48). This sample is the object of data preprocessing. For example, 1000 patient data points are selected from the Mount Everest Project database. Each data point includes gender, age at enrollment, initial ALT values, and quantitative HBsAg values ​​at weeks 0, 12, 24, 36, and 48. These 1000 data points constitute the initial sample.

[0023] In this embodiment, standardization uses the Z-score standardization method, and the formula is as follows: ,in, The original data, The sample mean. The standard deviation of the sample is used. Preprocessing includes missing value handling and outlier handling. For missing value handling, the KNNImputer method is used (K is set to 5, which is determined by cross-validation (K=3, 5, 7). When K=5, the accuracy of missing data imputation reaches 92%, which is higher than K=3 (88%) and K=7 (90%). Similar samples are selected for imputation by calculating the Euclidean distance between samples. In the outlier handling process, 1.5 times the interquartile range (IQR) is used as the outlier detection threshold of the box plot method. Outliers that are confirmed to be without entry errors are retained, and outliers with entry errors are corrected or removed.

[0024] In this embodiment, physical information corresponding to the historical demand time point is extracted from the patient's past medical records. The information dimensions include liver function indicators, HBV-related indicators, and disease assessment results. For example, when a patient makes a historical demand to achieve clinical cure within 2 years, the corresponding historical physical information is ALT value of 65 U / L, HBsAg value of 3200 IU / mL, no cirrhosis complications, and liver function grade A.

[0025] In this embodiment, the theoretical conservative treatment effect vector refers to the vector form of the minimum treatment effect that can meet the historical needs, obtained by combining the analysis of candidate cure conditions based on the patient's historical needs and historical physical information. The theoretical optimal treatment effect vector refers to the vector form of the treatment effect under the ideal state, obtained by combining the analysis of candidate cure conditions based on the patient's historical needs and historical physical information.

[0026] In this embodiment, the inferior mutation index refers to an index in which the actual treatment effect deteriorates significantly compared to the theoretical conservative treatment effect vector, or the gap between the actual treatment effect and the theoretical optimal treatment effect vector increases significantly during the treatment process.

[0027] In this embodiment, the first auxiliary vector refers to the vector formed by combining all the mined inferior mutation indicators according to a preset rule.

[0028] In this embodiment, the differences (e.g., the value of the indicator in the next cycle minus the value of the indicator in the previous cycle, with a decrease being negative and an increase being positive) and the rate of change (difference divided by the cycle interval, e.g., 12 weeks is one cycle interval, then the rate of change = difference / 12) of each treatment effect indicator between adjacent follow-up cycles are calculated. These differences and rates of change are used as effect variation indicators. For example, if a patient's HBsAg is 3000 IU / mL at week 0, 2100 IU / mL at week 12, and 1200 IU / mL at week 24, the corresponding effect variation indicators are: HBsAg decrease of 900 IU / mL from week 0 to week 12, with a decrease rate of 75 IU / mL / week; HBsAg decrease of 900 IU / mL from week 12 to week 24, with a decrease rate of 7... 5 IU / mL / week; initial ALT value was 70 U / L, 50 U / L at week 12, and 35 U / L at week 24. The corresponding effect change index was: ALT decreased by 20 U / L in week 0-12 and ALT decreased by 15 U / L in week 12-24. At this time, the second auxiliary vector was obtained by combining the HBsAg change and ALT change in the order: [900 IU / mL (decreased value in week 0-12), 75 IU / mL / week (decreased rate in week 0-12), 900 IU / mL (decreased value in week 12-24), 75 IU / mL / week (decreased rate in week 12-24), 20 U / L (decreased value in week 0-12), 15 U / L (decreased value in week 12-24)].

[0029] In this embodiment, the correlation strength and correlation pattern between each element in the first auxiliary vector and the second auxiliary vector and the cure status (turning from negative to positive, not turning from negative to negative) of the effective samples are analyzed to establish a correlation model, namely a global mapping relationship. For example, by analyzing 965 effective samples, it was found that the lower the value of the HBsAg inferior mutation index in the first auxiliary vector and the faster the HBsAg decrease rate in the second auxiliary vector, the higher the probability of the cure marker in the corresponding effective sample. This general correlation is the global mapping relationship.

[0030] In this embodiment, the effective samples are divided into different stages according to the follow-up period (e.g., 0-12 weeks, 12-24 weeks, 24-36 weeks, 36-48 weeks). For each stage, the correlation between the first auxiliary vector, the second auxiliary vector, and the treatment effect at the end of the stage is analyzed to establish a local correlation model for each stage, i.e., a periodic local mapping relationship. For example, in the 0-12 week treatment stage, the correlation weight of the ALT inferior mutation index in the first auxiliary vector is higher; in the 24-36 week treatment stage, the correlation weight of the HBsAg 12-24 week decline rate in the second auxiliary vector is higher. This staged correlation relationship is the periodic local mapping relationship.

[0031] In this embodiment, the performance of three ensemble learning algorithms—XGBoost, LightGBM, and Random Forest—was compared on an imbalanced dataset (positive to non-positive samples = 1:3): XGBoost had a positive prediction value (PPV) of 78% and an AUC of 0.85; LightGBM had a PPV of 72% and an AUC of 0.81; and Random Forest had a PPV of 65% and an AUC of 0.78. Therefore, XGBoost was selected as the core algorithm for both the classification and prediction layers.

[0032] In this embodiment, the classification layer of the cascaded model uses an XGBoost binary classification model to determine the patient's cure probability; the prediction layer contains two XGBoost regression models to predict the future negative conversion probability of patients with high and low cure probabilities, respectively; the auxiliary layer uses a feature enhancement model to output auxiliary features to optimize the classification and prediction results. The classification layer, being the first layer of the cascaded model, is used to classify the input patient samples according to their cure potential, outputting a classification result of high or low cure probability. For example, after receiving the patient's valid sample data, a first auxiliary vector, and a second auxiliary vector, the classification layer calculates using the trained XGBoost binary classification model, outputting a classification result indicating whether the patient has a high cure probability (probability value 0.85) or a low cure probability (probability value 0.32). A classification threshold of 0.5 is set; a probability value higher than the threshold is considered a high cure probability, and a probability value lower than the threshold is considered a low cure probability. The prediction layer, the second layer of the cascaded model, quantitatively predicts the patient's cure rate over multiple treatment cycles based on the classification results of the classification layer, outputting the probability of negative conversion for each cycle. This includes prediction models for patients with high cure probabilities and those for patients with low cure probabilities. Two regression models are constructed using the XGBoost algorithm as the prediction layer, trained separately for samples classified as high and low cure probabilities. The input features are consistent with the classification layer, and the output is the probability of negative conversion for each follow-up cycle (24 weeks, 36 weeks, 48 ​​weeks, etc.). The prediction weights for each cycle are adjusted through periodic local mapping relationships to adapt the prediction results to the features of different treatment stages. For example, if the classification layer determines a patient has a high cure probability, the prediction layer calls the high cure probability prediction model, outputting a negative conversion probability of 0.35 at week 24, 0.68 at week 36, and 0.82 at week 48. The auxiliary layer, a supplementary layer of the cascaded model, extracts deep and dynamically changing features from the input data, outputting auxiliary features to enhance the classification accuracy of the classification layer and the prediction accuracy of the prediction layer.A neural network is used to construct an auxiliary layer. Dynamic change indicators (such as the rate of change in treatment effect and the amplitude of fluctuation) from the first and second auxiliary vectors are used as input. Through model training, deep features reflecting the trend of treatment effect and potential risks are extracted. For example, after receiving the patient's effect change indicator sequence, the auxiliary layer extracts deep features such as the stability of the HBsAg decline trend and the amplitude of ALT fluctuation. These features are then used as auxiliary features input to the classification and prediction layers to optimize the classification results and prediction probabilities. It should be noted that the classification layer (XGBoost binary classification) has a tree depth of 5 and a learning rate of... The regularization parameter λ = 0.1 is determined through grid search (tree depth 3-7, learning rate 0.05-0.2, λ0-0.2); the prediction layer (XGBoost regression): tree depth = 6, learning rate = 0.08, regularization parameter λ = 0.05, the optimization objective is to minimize the mean squared error between the predicted probability and the actual cure probability; the auxiliary layer (neural network): input layer dimension = 12, number of hidden layers = 2 (number of neurons 32 and 16 respectively), output layer dimension = 4, activation function is ReLU, optimizer is Adam, learning rate = 0.001.

[0033] In this embodiment, the current sample refers to the clinical data of the patient requiring clinical cure prediction at the current treatment stage, including current basic information, detection indicators of completed follow-up cycles, etc., and serves as the input data for the cascade model. For example, a patient currently at the 12-week treatment node, their current sample includes gender: female, age: 38 years, initial ALT value: 55 U / L, HBsAg at week 0: 2800 IU / mL, and HBsAg at week 12: 1900 IU / mL.

[0034] In this embodiment, the current demand refers to the demand related to the treatment effect raised by the patient at the current stage of treatment that needs to be predicted. For example, a patient's current demand is: to achieve clinical cure within 36 weeks, and to consider adjusting the treatment plan if the probability of cure is low.

[0035] In this embodiment, the expected target achievement probability based on each treatment cycle refers to the probability that the patient will achieve clinical cure-related auxiliary target indicators (such as ALT returning to normal, HBsAg dropping to a specific level, etc.) in each future treatment cycle. It is a supplementary prediction to the seroconversion probability and provides a more comprehensive reference for the evaluation of treatment effect.

[0036] In this embodiment, based on clinical data from 1000 patients with chronic hepatitis B (600 in the training set and 400 in the validation set), the performance of the present invention was compared with that of the prior art (single XGBoost model), as shown in Table 1: Table 1 Performance Comparison Table Note: Dynamic prediction accuracy refers to the accuracy of the prediction result adjusted based on the updated clinical data (such as 12-week data) during the treatment process; multi-cycle prediction error refers to the average error between the probability of negative conversion in each prediction cycle and the actual cure rate.

[0037] The beneficial effects of the above technical solution are as follows: By constructing a cascaded model containing a classification layer, a prediction layer, and an auxiliary layer, and combining patient clinical data, historical needs, and dynamic effect change indicators, early, dynamic, and multi-cycle prediction of clinical cure for chronic hepatitis B is achieved. This effectively solves the problems of existing single models having low positive predictive values, being unable to dynamically update prediction results, and only being able to predict cure status for a fixed period. It can output prediction results in the early stages of treatment and dynamically optimize prediction accuracy as the treatment cycle progresses. At the same time, it provides the probability of seroconversion and the most likely cure cycle for multiple future cycles, providing a comprehensive and reliable reference for personalized adjustment of clinical treatment plans and patient treatment decisions, and helping to reduce the economic burden on patients and the risk of adverse drug reactions.

[0038] This invention provides a clinical cure prediction method for chronic hepatitis B based on a cascade model, which combines data to obtain a first auxiliary vector. The method includes: decomposing the historical needs of chronic hepatitis B patients into several sub-needs and the purpose and constraints for each sub-need; matching candidate cure conditions related to each sub-need through a purpose-constraint-candidate checklist; determining the merits of each candidate cure condition based on a first distribution of all sub-needs and combined with the historical physical information of the chronic hepatitis B patients; determining the conservative position point for each candidate cure condition and obtaining a theoretical conservative treatment effect vector that satisfies the historical needs; simultaneously, obtaining a theoretical optimal treatment effect vector based on the merits mining results; screening inferior mutation indicators based on the theoretical conservative treatment effect vector and the theoretical optimal treatment effect vector; and combining all inferior mutation indicators to obtain the first auxiliary vector.

[0039] In this embodiment, sub-requirements refer to the detailed requirements with clear objectives obtained by breaking down the patient's historical requirements. Each sub-requirement corresponds to a specific treatment goal. For example, if the patient's historical requirements are: to achieve clinical cure within 2 years, with no serious adverse drug reactions and treatment costs controlled within 50,000 yuan per year, the sub-requirements are: Sub-requirement 1: HBsAg decreases to below 1000 IU / mL at 24 weeks; Sub-requirement 2: HBsAg becomes negative at 48 weeks; Sub-requirement 3: ALT is maintained within the normal range during treatment; Sub-requirement 4: Annual treatment cost ≤ 50,000 yuan. Specifically, this is based on the core indicators of clinical cure (HBsAg negative conversion, normal liver function, etc.) and the patient's personalized needs (cost, adverse reactions), and is broken down according to the dimension, target type, time node, and constraints.

[0040] In this embodiment, the purpose of the demand refers to the core treatment goal or core appeal corresponding to each sub-demand. The purpose of sub-demand 1 is to rapidly reduce HBsAg levels; the purpose of sub-demand 4 is to control the economic cost of treatment.

[0041] Purpose constraints refer to the limitations that must be followed in the process of achieving each sub-requirement, including time constraints, numerical constraints, and safety constraints. For example, the purpose constraints for sub-requirement 1 are: completion within 24 weeks and no serious liver damage during the HBsAg decline process; the purpose constraints for sub-requirement 4 are: annual cost ≤ 50,000 yuan and no use of expensive out-of-pocket drugs.

[0042] In this embodiment, the purpose-constraint-candidate lookup table refers to a pre-established mapping table containing the desired purpose, purpose constraints, and corresponding candidate cure conditions. Part of the content is shown in Table 2: Table 2 Purpose-Constraint-Candidate Lookup Table In this embodiment, the first distribution refers to the coverage distribution of each candidate cure condition among all sub-demands, that is, the distribution of the number of sub-demands that the candidate cure condition can satisfy and the degree of satisfaction. For example, the candidate cure condition: HBsAg decrease ≥30% at 12 weeks can satisfy sub-demand 1 (rapid reduction of HBsAg) and sub-demand 2 (negative conversion at 48 weeks), with satisfaction degrees of 0.8 and 0.6 respectively. Its first distribution is: [sub-demand 1: 0.8, sub-demand 2: 0.6, sub-demand 3: 0, sub-demand 4: 0].

[0043] In this embodiment, the conservative position point refers to the key numerical node that ensures the treatment effect is not lower than the theoretical conservative level under the premise of meeting historical requirements for each candidate cure condition. That is, the lowest effective threshold of the condition. For example, the conservative position point for the HBsAg reduction ratio at 12 weeks for the candidate cure condition is 30%. That is, a ≥30% reduction in HBsAg at 12 weeks is the lowest threshold to ensure that the theoretical conservative treatment effect can be achieved in the future.

[0044] In this embodiment, for example, a patient's theoretical conservative treatment effect vector is: [ALT returns to normal (≤40U / L), HBsAg decreases to below 1000IU / mL at week 24, decreases to below 100IU / mL at week 48, no serious adverse reactions], where each element represents the minimum requirement for a treatment effect indicator. Specifically, based on the patient's historical needs and objectives (such as treatment cycle and minimum effect threshold), combined with the patient's historical physical information and corresponding disease baseline, and referring to the basic requirements for the treatment effect of chronic hepatitis B in clinical treatment guidelines, the minimum target value for each treatment effect indicator is determined, and these minimum target values ​​are arranged in a preset order to form the theoretical conservative treatment effect vector.

[0045] For example, the theoretical optimal treatment outcome vector for the above patients is: [ALT reduced to 25 U / L, HBsAg reduced to below 1500 IU / mL in week 12, reduced to below 500 IU / mL in week 24, seroconversion in week 36, and no adverse reactions]. Specifically, it is determined by referring to the best treatment outcome data of patients with similar conditions in clinical studies, the ideal treatment goals recommended in clinical treatment guidelines, and combining the optimal improvement potential of the patient's historical physical information. The ideal target value for each treatment outcome indicator is then determined and arranged in the same order as the theoretical conservative treatment outcome vector to form the theoretical optimal treatment outcome vector.

[0046] In this embodiment, the numerical differences between the actual treatment effect indicators and the corresponding indicators in the theoretical conservative treatment effect vector and the theoretical optimal treatment effect vector are compared, and a difference threshold is set (e.g., HBsAg difference ≥ 500 IU / mL, adverse reaction grade exceeding the allowable range by 1 level or more). When the difference between the actual indicator and the theoretical conservative indicator exceeds the threshold, or the difference with the theoretical optimal indicator shows a continuously increasing trend, the indicator is determined to be a bad mutation indicator. For example, in a patient's theoretical conservative treatment effect vector, HBsAg ≤ 1000 IU / mL at week 24, but the actual HBsAg at week 24 is 2800 IU / mL, which is significantly higher than the theoretical conservative value. The HBsAg level at week 24 is the bad mutation indicator. In another patient, there are no serious adverse reactions in the theoretical optimal treatment effect vector, but grade 3 myelosuppression occurs during actual treatment. The adverse reaction grade is the bad mutation indicator.

[0047] In this embodiment, the identified inferior mutation indicators are classified (such as HBV-related indicators, liver function indicators, adverse reaction indicators, etc.), and sorted according to the classification order. The specific value or level of each indicator is quantified (such as adverse reaction level 1 being quantified as 1, level 2 being quantified as 2, etc.), and arranged in sequence to form the first auxiliary vector. For example, the inferior mutation indicators of a patient include: HBsAg level at week 24 | adverse reaction level, ALT re-rise magnitude. The first auxiliary vector is obtained by combining the HBsAg-related indicators in the order of HBsAg-related indicators-liver function indicators-adverse reaction indicators: [2800 IU / mL (actual value at week 24), level 3 (adverse reaction level), 30 U / L (ALT re-rise magnitude)].

[0048] The beneficial effects of the above technical solution are as follows: by refining the construction process of the first auxiliary vector, the accurate mining of the correlation features between the patient's historical needs and treatment effects is realized. Through steps such as sub-need decomposition, candidate condition matching, and effect vector comparison, it is ensured that the first auxiliary vector can comprehensively and accurately reflect the key deterioration factors affecting the cure effect, providing high-quality feature input for the subsequent establishment of mapping relationships and cascade model training, and improving the basic reliability of model prediction.

[0049] This invention provides a method for predicting the clinical cure of chronic hepatitis B based on a cascade model. The method involves evaluating the merits of each candidate cure condition and determining a conservative position for that condition. The method includes: retrieving historical cure effect sequences matching the candidate cure conditions from a historical database; traversing the cure itemset based on each historical cure effect; constructing an initial thread for each cure item and determining the thread type of the initial thread; determining the mining accuracy of the corresponding cure item based on the thread type, total thread cycle length, number of thread cycles, and average length per cycle for each initial thread; and if the mining accuracy is less than a preset accuracy, using the average length per cycle... The average length is the cluster length. Cluster analysis is performed on all single-cycle lengths of the initial thread, and the cluster with the largest total length is selected as the analysis target. The merits of the analysis target are evaluated, and the average effect with the highest probability density under the analysis target is combined to obtain the reference conservative point for the corresponding cure item. If the mining accuracy is not less than the preset accuracy, triples are constructed for each single cycle in the initial thread, wherein the triples include: single-cycle length, single-cycle effect, and corresponding cure participation. The merits of each triple are evaluated and a fitting merit curve is constructed, and it is determined whether there are any triples that are not located in the fitting merit curve. If discrete points exist on the curve itself, obtain the number N3 of triplet points of the goodness-of-fit curve and the goodness-of-fit distribution function F3, and determine the upper limit line ys and the lower limit line yx. The vertical coordinate of the upper limit line ys is less than the maximum vertical value of the goodness-of-fit curve and greater than the minimum vertical value of the goodness-of-fit curve; the vertical coordinate of the lower limit line yx is greater than the minimum vertical value of the goodness-of-fit curve and less than the maximum vertical value of the goodness-of-fit curve; and the vertical coordinate of the upper limit line ys is greater than the vertical coordinate of the lower limit line yx. Count the first number N1 of discrete points located above the goodness-of-fit curve, and the number of discrete points located above the goodness-of-fit curve. The system calculates the second number N2 of discrete points above the upper limit line ys and the first position distribution function F1 based on the first number N1, and the second position distribution function based on the second number N2; simultaneously, it calculates the third number N3 of discrete points below the fitting good / bad curve, the fourth number N4 of discrete points below the lower limit line yx and the third position distribution function F3 based on the third number N3, and the fourth position distribution function F4 based on the fourth number N4; based on the statistical results, it determines the reference conservative point for the corresponding cure item; based on the reference conservative point for each cure item, it determines the conservative position point for the corresponding candidate cure condition.

[0050] Preferably, the thread type includes: full-thread type and half-thread type. It should be noted that the full-thread type refers to the historical healing effect sequence of the healing item completely covering all preset follow-up periods of 0 weeks, 12 weeks, 24 weeks, 36 weeks and 48 weeks, and the effect data of each period is not missing. The half-thread type refers to the historical healing effect sequence of the healing item covering less than 3 preset follow-up periods, or the effect data of 1 or more periods is missing.

[0051] In this embodiment, the historical cure effect sequence refers to the set of past patient cure effect data obtained from the historical database, which are arranged in chronological order and match the current candidate cure condition. For example, if the candidate cure condition is: HBsAg decrease ≥30% at 12 weeks, 500 patients who meet this condition are selected from the historical database, and their HBsAg values, ALT values, and other cure effect indicators are extracted in chronological order of treatment time at 12 weeks, 24 weeks, 36 weeks, and 48 weeks, forming 500 historical cure effect sequences. Each sequence is [changes in HBsAg at 12 weeks, HBsAg at 24 weeks, HBsAg at 36 weeks, HBsAg at 48 weeks, and ALT].

[0052] In this context, the "cure itemset" refers to the set of key indicators related to the cure outcome contained in each historical cure effect sequence, that is, the indicator name and value corresponding to each element in the sequence. For example, a historical cure effect sequence is: [12 weeks HBsAg=2100IU / mL, 24 weeks HBsAg=1000IU / mL, 36 weeks HBsAg=300IU / mL, 48 weeks HBsAg=0IU / mL, ALT remains normal], its cure itemset is {12 weeks HBsAg (2100IU / mL), 24 weeks HBsAg (1000IU / mL), 36 weeks HBsAg (300IU / mL), 48 weeks HBsAg (0IU / mL), ALT status (normal)}.

[0053] In this embodiment, the initial thread refers to the dataset formed by concatenating data of the same cure item in different historical cure effect sequences in sequence order, which is used to analyze the change pattern of the cure item. For example, if the data of the cure item "12-week HBsAg decrease rate" in the three historical cure effect sequences are 35%, 42%, and 38%, then its initial thread is [35%, 42%, 38%]; if the data of the cure item "24-week HBsAg value" in the three sequences are 1800 IU / mL, 1500 IU / mL, and 1600 IU / mL, then the initial thread is [1800 IU / mL, 1500 IU / mL, 1600 IU / mL].

[0054] In this embodiment, the total thread cycle length is the sum of the lengths of all single cycles in the initial thread. The length of a single cycle refers to the time interval of the preset follow-up period (e.g., 12 weeks for 0-12 weeks, 12 weeks for 12-24 weeks). If the follow-up period corresponding to the initial thread is 0-12 weeks, 12-24 weeks, and 24-36 weeks (3 single cycles), then the total thread cycle length = 12 weeks + 12 weeks + 12 weeks = 36 weeks.

[0055] In this embodiment, the number of thread cycles is the number of preset follow-up cycle segments actually included in the initial thread, that is, the number of single cycles covered by the initial thread (regardless of whether the data is missing, only the number of covered cycle segments is counted). For example, if the initial thread covers segment 1, segment 2, and segment 3, the number of thread cycles is 3 regardless of whether the data is missing.

[0056] In this embodiment, the average length of a single cycle is the ratio of the total length of the thread cycles to the number of thread cycles, which is used to reflect the average time span of each covered cycle segment in the initial thread.

[0057] In this embodiment, the mining accuracy refers to the reliability of the mining results when mining the merits of candidate cure conditions. It is represented by a value between 0 and 1, with the value closer to 1 indicating a more reliable mining result. For example, if the mining accuracy of a candidate cure condition is calculated to be 0.85, it indicates that the reliability of its merits mining results is relatively high; if the calculated result is 0.62, it indicates that the reliability is average.

[0058] In this embodiment, based on historical data from 1,000 patients with chronic hepatitis B, it was verified that when the preset precision was 0.7, the accuracy rate of the reference conservative point was 85%, and the false positive rate was controlled below 12%, which met the reliability requirements of clinical prediction.

[0059] In this embodiment, cluster analysis refers to a data analysis method that groups all single-cycle lengths of the initial thread using the average length of a single cycle as the cluster length. This is used to select representative analysis targets, with the K-means clustering algorithm being preferred. For example, if the single-cycle lengths of the initial thread are 12 weeks, 12 weeks, 24 weeks, 24 weeks, 24 weeks, and 36 weeks, and the average length of a single cycle is 20 weeks, K-means clustering (K=3) yields cluster result 1: [12 weeks, 12 weeks] (total length 24 weeks), cluster result 2: [24 weeks, 24 weeks, 24 weeks] (total length 72 weeks), and cluster result 3: [36 weeks] (total length 36 weeks). Cluster result 2, with the largest total length, is selected as the analysis target.

[0060] In this embodiment, the average effect with the highest probability density refers to the value with the highest probability density corresponding to the average of the effects of each single period in the analysis target (clustering results), that is, the most representative average effect value. For example, if the single-period effects (HBsAg reduction rate) of the analysis target (clustering result 2) are 32%, 35%, and 33%, respectively, and the average effect is 33.3%, its probability density is 0.8 (the highest among all possible average effect values), then this value is the average effect with the highest probability density.

[0061] In this embodiment, the reference conservative point refers to a preliminary numerical node that reflects the level of conservatism of the cure item, obtained based on the analysis objective or statistical results. It is the basis for determining the conservative position of candidate cure conditions. For example, the reference conservative point obtained by cluster analysis is: HBsAg decrease ≥32% at 12 weeks; the reference conservative point obtained by the triplet-fit curve method is: HBsAg ≤1400 IU / mL at 24 weeks.

[0062] In this embodiment, a triplet refers to a three-dimensional data combination of single-cycle length, single-cycle effect, and healing participation for each single cycle in the initial thread. For example, if the single-cycle length of an initial thread is 12 weeks, the single-cycle effect is a 35% decrease in HBsAg, and the healing participation (the contribution of this cycle effect to the final cure) is 0.3, then the corresponding triplet is (12 weeks, 35%, 0.3); the triplet for another single cycle is (24 weeks, 48%, 0.45).

[0063] In this embodiment, the fitted superiority / inferiority curve refers to a curve reflecting the relationship between periodic characteristics and superiority / inferiority, constructed using a curve fitting algorithm based on the superiority / inferiority scores of all triples, preferably a polynomial fitting algorithm. For example, fitting the superiority / inferiority scores (0-1 points) of 10 triples to a single period length yields the fitted superiority / inferiority curve equation as follows: Where x is the period length and y is the degree of quality, the curve shows a trend of first decreasing and then increasing.

[0064] In this embodiment, a discrete point refers to a triplet data point that is not located on the fitted goodness-basis curve, i.e., a point where the deviation between its actual goodness-basis score and the predicted value of the fitted curve exceeds a preset threshold (preferably 0.1). For example, if the actual goodness-basis score of a triplet is 0.65, and the goodness-basis score predicted by the fitted curve is 0.48, the deviation is 0.17 > 0.1, then this triplet is a discrete point; another triplet has an actual score of 0.52, a predicted score of 0.45, and a deviation of 0.07 < 0.1, and is not a discrete point.

[0065] In this embodiment, the upper limit line ys and the lower limit line yx refer to two parallel lines set based on the numerical range of the fitted goodness-of-failure curves. They are used to divide the position intervals of discrete points. ys is located in the upper part of the curve, and yx is located in the lower part of the curve. For example, if the maximum vertical value (maximum value of goodness-of-failure degree) of the fitted goodness-of-failure curve is 0.85 and the minimum vertical value is 0.32, the vertical coordinate of ys is set to 0.7 (<0.85 and >0.32), and the vertical coordinate of yx is set to 0.45 (>0.32 and <0.85), and 0.7 > 0.45. Specifically, based on the maximum and minimum vertical values ​​of the fitted goodness-of-failure curves, the values ​​are set according to the rule ys = (max_y + min_y) × 0.75 and yx = (max_y + min_y) × 0.5, and fine-tuned according to the distribution of discrete points.

[0066] The positional distribution function refers to a function describing the distribution pattern of discrete points based on the number and location of discrete points in different positional intervals (above the curve, above ys, below the curve, below yx). A normal distribution function is preferred for fitting. For example, the positional distribution function F1 for discrete points located above the curve (first quantity N1=8) is N (mean 18 weeks, standard deviation 3.2), indicating that the period length of these discrete points is mainly concentrated around 18 weeks, with a standard deviation of 3.2. The distribution function F4 for discrete points located below yx (fourth quantity N4=5) is N (mean 30 weeks, standard deviation 4.1).

[0067] The beneficial effects of the above technical solution are as follows: by designing a method for determining conservative position points based on different scenarios (high and low mining accuracy), constructing an initial thread by combining historical cure effect sequences, and accurately mining reference conservative points using cluster analysis or triplet-fitting curve-statistical methods, the reliability and rationality of the conservative position points of candidate cure conditions are ensured, providing key support for the construction of theoretical conservative treatment effect vectors, further improving the effectiveness of the first auxiliary vector, and indirectly optimizing the prediction accuracy of the model.

[0068] This invention provides a method for predicting the clinical cure of chronic hepatitis B based on a cascade model. Based on statistical results, it determines a reference conservative point for the corresponding cure item, including: normalizing the single-cycle effect corresponding to the discrete point to obtain a normalized single-cycle effect. Based on F1, F2, F3, and F4, and combined with Constructing the fusion distribution function ,in, Let be the weighting coefficients of the distribution function at the j-th location, and ; Let j be the distribution function at the j-th location; The weight coefficients of the distribution function at the j-th location; Let be the average value of the normalized single-period effect of the discrete points involved in the j-th location distribution function. Corresponding to F1, F2, F3, and F4 respectively; calculate the conservative constraint coefficients. ,in, Let be the single-cycle length corresponding to the i-th discrete point; The mean of all single-cycle lengths of the initial thread; n is the total number of discrete points; This is the maximum value of the single-cycle length of the initial thread; The minimum single-cycle length of the initial thread is determined; the fusion distribution function is determined. mode And calculate the initial value of the reference conservative point based on the conservative constraint coefficient. .

[0069] In this embodiment, the normalized single-cycle effect refers to the value obtained after standardizing the single-cycle effect corresponding to the discrete point, eliminating the influence of dimensions, and facilitating subsequent calculations. The Min-Max normalization method is adopted.

[0070] In this embodiment, the fusion distribution function refers to the distribution function that combines the four location distribution functions (F1-F4) and the normalized single-cycle effect to comprehensively reflect the location and effect characteristics of all discrete points, and is used to extract the most representative location parameter (mode).

[0071] For example, if the weights of F1 (w1=0.2), F2 (w2=0.2), F3 (w3=0.3), and F4 (w4=0.25), the fusion distribution function is: =(0.25F1+0.2F2+0.3F3+0.25F4) / (0.25+0.2+0.3+0.25)=0.25F1+0.2F2+0.3F3+0.25F4. Its distribution curve is a weighted superposition of the four distribution functions, where the weights are determined by the weight coefficients of the distribution functions at each position. The average value of the normalized single-period effect is used to calculate the weighting coefficient wj of each distribution function, which is then substituted into the fusion distribution function formula. It should be noted that the importance score of each location distribution function to the reference conservative point is determined by calculating the feature importance score of each location distribution function to the reference conservative point through the random forest algorithm. Among them, F3 (the location distribution function of discrete points below the fit goodness-of-fit curve) has the highest importance, so it is given the highest weight, and the specific values ​​are as follows: =0.25、 =0.2、 =0.3、 =0.25.

[0072] In this embodiment, the mode The probability density function (PDC) refers to the value with the highest probability of occurrence in the fusion distribution function, i.e., the x-axis (single period length) corresponding to the peak of the distribution curve. For example, if the distribution curve of the fusion distribution function Ffusion peaks at the 24th cycle, with a probability density of 0.92 (the highest among all x-axis values), then... =24 weeks.

[0073] The beneficial effects of the above technical solution are as follows: through normalization processing, construction of fusion distribution function, calculation of conservative constraint coefficient, etc., the quantitative calculation of reference conservative point is realized. It fully integrates the position distribution, effect characteristics and period length dispersion of discrete points, ensuring the accuracy and conservatism of reference conservative point, providing a scientific basis for the final determination of the conservative position point of candidate cure condition, and further improving the reliability of theoretical conservative treatment effect vector.

[0074] This invention provides a method for predicting the clinical cure rate of chronic hepatitis B based on a cascade model, determining the mining accuracy of corresponding cure items, including: ; ; ; ;in, The mining precision for the corresponding cure item; This is to initially measure the intermediate accuracy. The normalized value of the total thread cycle length of the initial thread; N is the number of thread cycles; Normalized average length of a single period; Thread cycle density; The average length concentration; For thread structure factors; Based on Determine the effective stability of the curative effect; The normalized mean of the therapeutic effect; The normalized standard deviation of the curative effect; m0 is the number of historical curative effects; Let m0 be the number of historical healing effects that follow a normal distribution, and T represents the value corresponding to the thread type.

[0075] In this embodiment, T=1 when it is a full-threaded type and T=0.5 when it is a half-threaded type.

[0076] The beneficial effects of the above technical solution are: by comprehensively considering multi-dimensional parameters (thread structure, stability of healing participation, periodic characteristics, etc.), the accuracy of the mining is accurately evaluated, providing an objective basis for the reasonable selection of subsequent conservative point determination methods, avoiding method selection deviations caused by single factors, and ensuring the scientificity and reliability of the conservative location point mining process.

[0077] This invention provides a method for predicting the clinical cure of chronic hepatitis B based on a cascade model. In the process of training and constructing the cascade model, the method further includes: based on the global feature distribution law of the global mapping relationship and the periodic feature fluctuation law of the periodic local mapping relationship, performing feature space adaptation transformation on the initial clinical features of the validation sample to obtain the adapted validation features that match the mapping logic of the cascade model.

[0078] Preferably, the process of training and constructing the cascaded model further includes: inputting the adaptation verification features into the trained cascaded model, simultaneously acquiring the intermediate classification results and corresponding classification weights output by the classification layer, the prediction results and prediction confidence output by the prediction layer, and the auxiliary features output by the auxiliary layer; comparing the prediction results with the original results to obtain a probability difference set; simultaneously, fusing the original cure features of the verification sample, the classification weights of the classification layer, and the auxiliary features of the auxiliary layer, and performing dot product attention calculation on the fused features to obtain fused reference features; capturing the feature extraction process logs of the classification layer, prediction layer, and auxiliary layer for the adaptation verification features to obtain periodic process sub-features of each layer; fusing the periodic process sub-features based on the functional weights of each layer in the cascaded model to obtain fused periodic process features; and comparing the fused periodic process features with the fused reference features to obtain a hierarchical feature difference set; and based on the global difference weights of the global mapping relationship and the periodic difference weights of the periodic local mapping relationship, calculating the difference probability elements and hierarchical feature difference sets under the same verification sample. Weighted encoding is performed on the comparison relationship to obtain a weighted comparison map. The weighted difference ladder corresponding to the difference probability element is determined by combining the distribution gradient of the global difference weights and the fluctuation range of the periodic difference weights in the weighted comparison map. Based on the weighted difference ladder of the same difference probability element across all validation samples, the difference ladder is weighted by reliability based on the prediction confidence output of the prediction layer, resulting in a weighted difference ladder sequence. The weighted sum of the mean square error and entropy of the weighted difference ladder sequence is used as the first interactive loss function. Simultaneously, based on all feature difference values ​​under the same difference feature, the probability difference influence coefficient is adjusted by combining the classification result confidence of the classification layer and the auxiliary feature contribution of the auxiliary layer. The cross-entropy of the feature difference value and the adjusted influence coefficient is used as the second interactive loss function. The global matching degree of the validation sample in the global mapping relationship and the periodic matching degree in the periodic local mapping relationship are calculated and used as the global loss adjustment factor and the periodic loss adjustment factor, respectively. Weights are allocated to the first and second interactive loss functions based on these adjustment factors to obtain the comprehensive loss function.

[0079] In this embodiment, ,in, , L1 is the first interaction loss function, and L2 is the second interaction loss function. ,and , .

[0080] In this embodiment, the global feature distribution pattern refers to the distribution pattern of each feature of the first auxiliary vector and the second auxiliary vector in all valid samples within the global mapping relationship, including the feature's mean, standard deviation, and distribution type (normal distribution, skewed distribution, etc.). For example, in the global mapping relationship, the mean of the 12-week HBsAg inferior mutation value feature of the first auxiliary vector is 2200 IU / mL, and the standard deviation is 500 IU / mL, exhibiting a normal distribution; the mean of the 0-12-week HBsAg decline rate feature of the second auxiliary vector is 85 IU / mL / week, and the standard deviation is 30 IU / mL / week, exhibiting a right-skewed distribution.

[0081] The periodic characteristic fluctuation pattern refers to the fluctuation pattern of each characteristic of the first auxiliary vector and the second auxiliary vector in different treatment stages (0-12 weeks, 12-24 weeks, etc.) in a periodic local mapping relationship, including the fluctuation amplitude and fluctuation trend (increasing, decreasing, stable). For example, in the 0-12 week stage, the fluctuation amplitude of the HBsAg inferior mutation value at week 12 is 300-800 IU / mL, showing a decreasing trend; in the 24-36 week stage, the fluctuation amplitude of this characteristic is 100-300 IU / mL, showing a stable trend; the HBsAg decrease rate in the 0-12 week stage fluctuates between 40-120 IU / mL / week, showing an increasing trend.

[0082] Initial clinical features refer to the original clinical features in the validation sample that have not undergone feature processing, including patient basic information, initial ALT value, quantitative HBsAg indicators for each cycle, etc., which are consistent with the initial sample indicators.

[0083] Feature space adaptation transformation refers to the feature transformation processing of initial clinical features based on the global feature distribution law and the periodic feature fluctuation law, so that the transformed features match the mapping logic (global and local mapping relationship) of the cascade model, and optimizes methods such as feature standardization, feature scaling, and feature reconstruction. For example, according to the global feature distribution law, the global mean of HBsAg value at 12 weeks is 2000 IU / mL and the standard deviation is 400 IU / mL. After standardizing the HBsAg value of 1900 IU / mL at 12 weeks of validation sample, the adaptation feature value z = (1900-2000) / 400 = -0.25 is obtained; according to the periodic feature fluctuation law, the fluctuation range of HBsAg decline rate in the 0-12 week stage is 40-120 IU / mL / week. After scaling the decline rate of 80 IU / mL / week of validation sample, the adaptation feature value = (80-40) / (120-40) = 0.5 is obtained.

[0084] In this embodiment, the adaptation verification features refer to the feature data obtained after feature space adaptation transformation for cascaded model verification. Their dimensions are consistent with the initial clinical features or optimized according to the mapping logic, and they can be directly input into the model for verification. For example, the adapted validation features of the above validation samples after adaptation and transformation are: gender code: 1 (female), age standardized value: 0.32, ALT standardized value: 0.15, 0-week HBsAg standardized value: 0.45, 12-week HBsAg standardized value: -0.25, 24-week HBsAg standardized value: -0.85, 0-12 week HBsAg decline rate scaling value: 0.5, 12-24 week HBsAg decline rate scaling value: 0.62. The beneficial effects of the above technical solution are: through feature space adaptation and transformation, the clinical features of the validation samples are accurately matched with the global and local mapping logic of the cascade model, eliminating model validation bias caused by differences in feature distribution, improving the accuracy of model validation, and thus ensuring the reliability of the cascade model training process. This allows the trained model to better adapt to the feature patterns of actual clinical data, improving the final prediction effect.

[0085] In this embodiment, the probability difference set refers to the set of differences between the prediction results output by the prediction layer and the original results (probabilities corresponding to the actual cure status) of the verification samples after the adaptive verification features are input into the trained cascaded model. Each difference element is the difference between the predicted probability and the original probability. For example, if the original results of a verification sample are: 0.3 for turning negative at 24 weeks, 0.6 for turning negative at 36 weeks, and 0.8 for turning negative at 48 weeks, and the model prediction results are: 0.25 for turning negative at 24 weeks, 0.65 for turning negative at 36 weeks, and 0.78 for turning negative at 48 weeks, then the probability difference set is [-0.05, 0.05, -0.02].

[0086] Fusion reference features refer to the reference features calculated using dot product attention after fusing the original cure features, classification weights of the classification layer, and auxiliary features of the auxiliary layer of the validation sample. These features are used for comparative analysis and comprehensively reflect the core features of the sample and the processing results of each layer of the model. For example, the original cure features of the validation sample are [2700 IU / mL (0-week HBsAg), 1900 IU / mL (12-week HBsAg), 1300 IU / mL (24-week HBsAg)], the classification weights of the classification layer are [0.35 (HBsAg decline rate weight), 0.25 (ALT normalization weight), 0.2 (no cirrhosis weight), 0.2 (age weight)], and the auxiliary features of the auxiliary layer are [0.72 (HBsAg decline stability), 0.68 (ALT fluctuation amplitude)]. The three features are concatenated as [2700, 1900, 1300, 0.35, 0.25, 0.2, 0.2, 0.72, 0.68]. By analyzing the three... Each element in the concatenated feature vector is normalized, and the dot product attention is calculated using the ScaledDot-ProductAttention algorithm (attention weights are optimized through model training, such as [0.15, 0.15, 0.15, 0.1, 0.1, 0.08, 0.08, 0.12, 0.07]). This yields the final fused reference feature value. It should be noted that this value is obtained using the Transformer model's self-attention mechanism, with training data consisting of fused features (original cure features + classification weights + auxiliary features) from 800 valid samples. The training objective is to minimize the mean squared error (MSE ≤ 0.01) between the fused reference feature and the actual cure feature. The initial range of the attention weights after training is [0.05-0.2], dynamically optimized with model iteration.

[0087] The feature extraction process log refers to the structured data recorded by each layer of the model when processing adaptation and validation features, including feature transformation steps, parameter calls, and intermediate output results, used to trace the feature processing logic. For example, the feature extraction process log of the classification layer includes the following: input feature dimension: 9-dimensional → through the first fully connected layer (weight matrix W1: 9×16) → activation function ReLU → output 16-dimensional features → through the second fully connected layer (weight matrix W2: 16×2) → output classification result (high cure probability: 0.82) and classification weights; the log of the prediction layer includes the input classification result and auxiliary vector → through the XGBoost regressor (tree depth: 5, learning rate: 0.1) → output prediction probabilities for each period.

[0088] Periodic process sub-features refer to the core feature processing results related to a specific treatment cycle extracted from the feature extraction process log, reflecting the feature extraction focus of each layer in different cycles. For example, the periodic process sub-features of the classification layer are divided into 0-12 weeks, 12-24 weeks, and 24-36 weeks, respectively [0.82 (classification confidence in 0-12 weeks), 0.75 (classification weight update value in 12-24 weeks), 0.68 (feature contribution in 24-36 weeks)]; the periodic process sub-features of the prediction layer are [0.32 (prediction confidence in 24 weeks), 0.65 (prediction confidence in 36 weeks), 0.78 (prediction confidence in 48 weeks)]; the periodic process sub-features of the auxiliary layer are [0.72 (auxiliary feature contribution in 0-12 weeks), 0.65 (auxiliary feature contribution in 12-24 weeks), 0.58 (auxiliary feature contribution in 24-36 weeks)].

[0089] In this embodiment, the functional weight refers to the weight set according to the importance of the core functions of the classification layer, prediction layer, and auxiliary layer in the cascaded model, and is used to fuse the periodic process sub-features of each layer. The classification layer is responsible for screening cure potential and is the foundation of the model, with a functional weight of 0.4; the prediction layer is responsible for core probability prediction and is the core of the model, with a functional weight of 0.5; the auxiliary layer is responsible for feature optimization and is a supplement to the model, with a functional weight of 0.1, and the sum of the weights is 1. For example, the weighted sub-features of the periodic process in the classification layer [0.82, 0.75, 0.68] are [0.82×0.4=0.328, 0.75×0.4=0.3, 0.68×0.4=0.272]; the weighted sub-features of the prediction layer [0.32, 0.65, 0.78] are [0.32×0.5=0.16, 0.65×0.5=0.325, 0.78×0.5=0.39]; and the weighted sub-features of the auxiliary layer [0.72, 0.65, 0.58] are [0.72×0.1=0.072, 0.65×0.1=0.065, 0.58×0.1=0.058].

[0090] In this embodiment, the fusion periodic process feature refers to the summation of the weighted periodic process sub-features of each layer for the same treatment cycle, resulting in a comprehensive feature that reflects the processing effect of the cycle model. For example, combining the above weighted sub-features, the fusion periodic process feature for 0-12 weeks (corresponding to 24-week prediction) = 0.328 (classification layer) + 0.16 (prediction layer) + 0.072 (auxiliary layer) = 0.56; 12-24 weeks (corresponding to 36-week prediction) = 0.3 + 0.325 + 0.065 = 0.69; 24-36 weeks (corresponding to 48-week prediction) = 0.272 + 0.39 + 0.058 = 0.72, and the final fusion periodic process feature is [0.56, 0.69, 0.72].

[0091] In this embodiment, the hierarchical feature difference set refers to the set of differences between the fused periodic process features and the fused reference features. First, the fused periodic process features need to be standardized to the same dimension as the fused reference features before calculating the difference value for each period. For example, if the fused reference feature value is H1, and the fused periodic process features are [0.56, 0.69, 0.72] and standardized using Min-Max, the standardized value is calculated. Then, the hierarchical feature difference set = standardized fused periodic process features - fused reference features.

[0092] In this embodiment, the global difference weight refers to the weight determined based on the feature importance in the global mapping relationship, reflecting the degree of influence of global feature differences on the model loss. It is calculated using the feature importance score of the random forest algorithm, and the preferred value range is 0.4-0.6. For example, in the global mapping relationship, the total feature importance score of the first auxiliary vector is 0.55, and that of the second auxiliary vector is 0.45. Setting the global difference weight to 0.5 represents that the influence of global feature differences on the loss accounts for 50%.

[0093] In this embodiment, the cycle difference weight refers to the weight determined based on the feature importance of the local mapping relationship of each treatment cycle, reflecting the degree of influence of the feature differences of different cycles on the loss. The sum of the weights of each cycle is 1. For example, in the local mapping relationship of 0-12 weeks (corresponding to 24-week prediction), the feature importance of the HBsAg decline rate is 0.3, and the cycle difference weight is set to 0.3; in 12-24 weeks (corresponding to 36-week prediction), the feature importance of the HBsAg absolute value is 0.4, and the weight is set to 0.4; in 24-36 weeks (corresponding to 48-week prediction), the feature importance of ALT stability is 0.3, and the weight is set to 0.3. The final cycle difference weight is [0.3, 0.4, 0.3].

[0094] In this embodiment, the weighted comparison chart refers to the structured comparison data (which can be presented in matrix form) formed by weighting the probability difference set and the hierarchical feature difference set according to global difference weight and periodic difference weight, and is used to intuitively reflect the correlation strength between the two. For example, the probability difference set is [-0.05, 0.05, -0.02], the hierarchical feature difference set is [-85.226, -3.976, 14.774], the global difference weight is 0.5, and the periodic difference weight is [0.3, 0.4, 0.3]. The comparison matrix obtained after weighted encoding is shown in Table 3: Table 3 Comparison Matrix Obtained After Weighted Encoding In this embodiment, the distribution gradient of the global difference weights refers to the range of values ​​of the global difference weights during multiple model training iterations, and is used to determine the stratification threshold of the difference probability elements. For example, if the global difference weights after 10 training iterations are [0.48, 0.52, 0.49, 0.51, 0.47, 0.53, 0.46, 0.54, 0.45, 0.55], the distribution gradient is 0.45-0.55, and the median gradient value of 0.5 is taken as the core threshold reference.

[0095] In this embodiment, the values ​​of the difference weights in each cycle during multiple training sessions are recorded, and the fluctuation range of the weights in each cycle (maximum value - minimum value) is calculated. The overall fluctuation range is then obtained by summing the values. For example, the fluctuation range of the difference weights in the 0-12 week cycle is [0.28-0.32], the 12-24 week cycle is [0.38-0.42], and the 24-36 week cycle is [0.29-0.31]. The overall fluctuation range is concentrated in the 0.28-0.42 range.

[0096] In this embodiment, based on the weight distribution gradient and fluctuation range, a stratification threshold is set using statistical methods (such as the median method). The weighted probability difference elements are divided into 2-3 steps according to the threshold, forming a weighted difference ladder. For example, based on the global difference weight distribution gradient of 0.45-0.55 and the periodic difference weight fluctuation range of 0.28-0.42, the stratification threshold is set to 0.02 (absolute value). The weighted probability difference elements are [-0.025, 0.025, -0.01]. Elements with an absolute value ≥ 0.02 (-0.025, 0.025) are divided into the first step, and elements with an absolute value < 0.02 (-0.01) are divided into the second step. The final weighted difference ladder is [first step: -0.025, 0.025; second step: -0.01].

[0097] In this embodiment, the predict_proba function of the XGBoost regressor outputs both the predicted probability and the confidence level. The confidence level is calculated by the degree of fit of the prediction error during model training. For example, the confidence level of a 0.25 predicted negative conversion probability at 24 weeks for a certain validation sample is 0.8, the confidence level of 0.65 at 36 weeks is 0.85, and the confidence level of 0.78 at 48 weeks is 0.9.

[0098] Reliability weighting refers to multiplying each weighted difference ladder element by the prediction confidence level for the corresponding prediction period.

[0099] A weighted difference ladder sequence refers to the reliability-weighted values ​​of the same difference type (such as 24-week prediction difference and 36-week prediction difference) across all validation samples, arranged in sample order. For example, for the 24-week prediction difference, the reliability-weighted values ​​of the three validation samples are -0.02, -0.022, and -0.018, respectively. Arranging them in sample order, the weighted difference ladder sequence is [-0.02, -0.022, -0.018].

[0100] In this embodiment, the first interactive loss function L1 refers to a loss function based on the weighted sum of the mean squared error (MSE) and entropy of the weighted difference ladder sequence. , where α is the weight value of 0.6, adjusted through cross-validation.

[0101] In this embodiment, the predict_proba function of the XGBoost binary classifier is used to obtain the probability value corresponding to the classification result. This probability value is the confidence level of the classification result. For example, if the classification layer determines that a patient has a high probability of being cured, the confidence level of the output classification result is 0.85; if another patient is determined to have a low probability of being cured, the confidence level is 0.78.

[0102] The contribution of auxiliary features refers to the degree of influence of each auxiliary feature output by the auxiliary layer on the final prediction result. The value ranges from 0 to 1 and is calculated by the SHAP value. For example, the SHAP value of the HBsAg decline stability feature output by the auxiliary layer is 0.75, and the SHAP value of the ALT fluctuation amplitude feature is 0.68. That is, the contribution of the auxiliary features of the two are 0.75 and 0.68, respectively.

[0103] Influence coefficient of probability difference ,in, The weight is set to 0.5 to balance the confidence of the two classes. Conf_class is the confidence of the classification result, and Contrib_aux is the mean contribution of the auxiliary features.

[0104] The second interactive loss function, L2, refers to a loss function that uses the cross-entropy of the standardized feature difference values ​​and the adjusted probability difference coefficients as its core. ,in, B1 represents the standardized value of the i-th feature difference. This represents the probability difference influence coefficient for the corresponding period.

[0105] The global matching degree refers to the degree of fit between the adaptive verification features of the verification sample and the feature distribution pattern in the global mapping relationship. It is calculated using cosine similarity. For example, the feature center vector of the global mapping relationship (the mean vector of the adaptive features of all valid samples) is [2500, 1800, 1200, 0.3, 0.25, 0.2, 0.2, 0.7, 0.65], and the adaptive verification feature vector of a certain verification sample is [2700, 1900, 1300, 0.35, 0.25, 0.2, 0.2, 0.72, 0.68]. The cosine similarity between the two is 0.98, that is, the global matching degree is 0.98.

[0106] The cycle matching degree refers to the degree of fit between the adaptation verification features of the validation sample and the fluctuation pattern of the corresponding local mapping relationship features in each treatment cycle. It is the mean of the matching degree for each cycle, with a value ranging from 0 to 1. For example, the feature fluctuation center vector of the local mapping relationship in weeks 0-12 is [2600, 1850, 0.32, 0.24, 0.71], and the adaptation feature vector of the validation sample in this cycle is [2700, 1900, 0.35, 0.25, 0.72], with a matching degree (cosine similarity) of 0.97; the matching degree in weeks 12-24 is 0.96; and the matching degree in weeks 24-36 is 0.95. The average of the three is taken as the cycle matching degree.

[0107] In this embodiment, when the value of the comprehensive loss function is less than 0.01 for five consecutive iterations and the prediction accuracy on the validation set is ≥85%, the model training is stopped to avoid overfitting.

[0108] The beneficial effects of the above technical solution are as follows: By constructing a dual-interaction loss function and introducing global and periodic matching degree adjustment factors, a scientific and comprehensive integrated loss function is formed. This function accurately captures the difference between the model's predicted results and the actual cure status, while also fully considering the fit between the feature extraction process and the reference features. The global loss adjustment factor and the periodic loss adjustment factor can dynamically adapt to the feature matching situation of different samples, enabling the model training to specifically optimize errors under different scenarios and effectively reduce the training bias caused by a single loss function. The cascaded model optimized by this integrated loss function significantly improves prediction accuracy and generalization ability.

[0109] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for predicting the clinical cure of chronic hepatitis B based on a cascade model, characterized in that, include: Step 1: Collect clinical data from patients with chronic hepatitis B to screen initial samples, and standardize and preprocess the initial samples to obtain effective samples; Step 2: Deconstruct the historical needs of patients with chronic hepatitis B and combine them with their historical physical information when outputting their historical needs to mine all inferior mutation indicators based on the theoretical conservative treatment effect vector and the theoretical optimal treatment effect vector, and combine them to obtain the first auxiliary vector. At the same time, capture the effect change indicators of patients with chronic hepatitis B in different follow-up periods and combine them to obtain the second auxiliary vector; Step 3: Comprehensively establish the first auxiliary vector, the second auxiliary vector, and the corresponding effective samples to form a complete... Step 4: Based on the ensemble learning algorithm and combining the effective samples, global mapping relationship and periodic local mapping relationship, and historical needs, train and construct a cascade model, wherein the cascade model includes: classification layer, prediction layer and auxiliary layer; Step 5: Obtain the current sample and current needs of the patient's treatment and input them into the trained cascade model, output the probability of the patient's hepatitis B surface antigen turning negative in multiple future treatment cycles, the treatment cycle with the highest probability of the patient obtaining clinical cure, and the patient's expected probability of achieving the target based on each treatment cycle.

2. The method for predicting the clinical cure of chronic hepatitis B based on a cascade model according to claim 1, characterized in that, The first auxiliary vector is obtained by combining the following steps: breaking down the historical needs of patients with chronic hepatitis B into several sub-needs and the purpose and constraints for each sub-need; matching candidate cure conditions related to each sub-need through a purpose-constraint-candidate checklist; determining the merits and demerits of each candidate cure condition based on the first distribution of all sub-needs and combined with the historical physical information of the patients with chronic hepatitis B; determining the conservative position point of the corresponding candidate cure condition; obtaining the theoretical conservative treatment effect vector that satisfies the historical needs; and obtaining the theoretical optimal treatment effect vector based on the merits and demerits mining results. Finally, filtering inferior mutation indicators based on the theoretical conservative treatment effect vector and the theoretical optimal treatment effect vector, and combining all inferior mutation indicators to obtain the first auxiliary vector.

3. The method for predicting the clinical cure of chronic hepatitis B based on a cascade model according to claim 2, characterized in that, For each candidate cure condition, a merit-based analysis is performed to determine a conservative position point for that condition. This includes: retrieving historical cure effect sequences matching the candidate cure conditions from a historical database, traversing the cure item set based on each historical cure effect, constructing an initial thread for the same cure item, and determining the thread type of the initial thread; determining the mining precision for the corresponding cure item based on the thread type, total thread cycle length, number of thread cycles, and average single cycle length of each initial thread; if the mining precision is less than a preset precision, using the average single cycle length as the cluster length, clustering all single cycles of the initial thread... Cluster analysis is performed on the period length, and the cluster with the largest total length is selected as the analysis target. The merits of the analysis target are evaluated, and the average effect with the highest probability density under the analysis target is combined to obtain the reference conservative point for the corresponding cure item. If the mining accuracy is not less than the preset accuracy, triples are constructed for each single period in the initial thread, wherein the triples include: single period length, single period effect, and corresponding cure participation effect. The merits of each triple are evaluated and a fitting merit curve is constructed. It is determined whether there are discrete points that are not located on the fitting merit curve itself. If so, the merits are obtained. The number of points N3 in the triplet of the goodness-of-fit curve itself and the goodness-of-fit distribution function F3 are used to determine the upper limit line ys and the lower limit line yx. The vertical coordinate of the upper limit line ys is less than the maximum vertical value and greater than the minimum vertical value of the goodness-of-fit curve, and the vertical coordinate of the lower limit line yx is greater than the minimum vertical value and less than the maximum vertical value of the goodness-of-fit curve. The vertical coordinate of the upper limit line ys is greater than the vertical coordinate of the lower limit line yx. A first number N1 of discrete points located above the goodness-of-fit curve and the number of points located on the upper limit line ys are counted. The system calculates the second number N2 of discrete points above position s and the first position distribution function F1 based on the first number N1, and the second position distribution function based on the second number N2; simultaneously, it calculates the third number N3 of discrete points located below the fitting good / bad curve, the fourth number N4 of discrete points located below the lower limit line yx, and the third position distribution function F3 based on the third number N3, and the fourth position distribution function F4 based on the fourth number N4; based on the statistical results, it determines the reference conservative point for the corresponding cure item; based on the reference conservative point for each cure item, it determines the conservative position point for the corresponding candidate cure condition.

4. The method for predicting the clinical cure of chronic hepatitis B based on a cascade model according to claim 3, characterized in that, The thread types include: full-threaded type and half-threaded type.

5. The method for predicting the clinical cure of chronic hepatitis B based on a cascade model according to claim 3, characterized in that, Based on the statistical results, a reference conservative point is determined for the corresponding cure item, including: normalizing the single-cycle effect corresponding to the discrete point to obtain the normalized single-cycle effect. Based on F1, F2, F3, and F4, and combined with Constructing the fusion distribution function ,in, Let be the weighting coefficients of the distribution function at the j-th location, and ； Let j be the distribution function at the j-th location; The weight coefficients of the distribution function at the j-th location; Let be the normalized average of the normalized single-period effect of the discrete points involved in the j-th location distribution function. Corresponding to F1, F2, F3, and F4 respectively; calculate the conservative constraint coefficients. ,in, Let be the single-cycle length corresponding to the i-th discrete point; The mean of all single-cycle lengths of the initial thread; n is the total number of discrete points; This is the maximum value of the single-cycle length of the initial thread; The minimum single-cycle length of the initial thread is determined; the fusion distribution function is determined. mode And calculate the initial value of the reference conservative point based on the conservative constraint coefficient. 。 6. The method for predicting the clinical cure of chronic hepatitis B based on a cascade model according to claim 3, characterized in that, Determine the mining precision for the corresponding cure item, including: ； ； ； ;in, The mining precision for the corresponding cure item; This is to initially measure the intermediate accuracy. The normalized value of the total thread cycle length of the initial thread; N is the number of thread cycles; Normalized average length of a single period; Thread cycle density; The average length concentration; For thread structure factors; For based on Determine the effective stability of the curative effect; The normalized mean of the therapeutic effect; The normalized standard deviation of the curative effect; m0 is the number of historical curative effects. Let m0 be the number of historical healing effects that follow a normal distribution, and T represents the value corresponding to the thread type.

7. The method for predicting the clinical cure of chronic hepatitis B based on a cascade model according to claim 1, characterized in that, The process of training and building the cascaded model also includes: based on the global feature distribution law of the global mapping relationship and the periodic feature fluctuation law of the periodic local mapping relationship, performing feature space adaptation transformation on the initial clinical features of the validation samples to obtain the adapted validation features that match the mapping logic of the cascaded model.

8. The method for predicting the clinical cure of chronic hepatitis B based on a cascade model according to claim 7, characterized in that, The process of training and building the cascaded model also includes: inputting the adaptation verification features into the trained cascaded model, simultaneously acquiring the intermediate classification results and corresponding classification weights output by the classification layer, the prediction results and prediction confidence output by the prediction layer, and the auxiliary features output by the auxiliary layer; comparing the prediction results with the original results to obtain a probability difference set; simultaneously, fusing the original healing features of the verification sample, the classification weights of the classification layer, and the auxiliary features of the auxiliary layer, and performing dot product attention calculation on the fused features to obtain fused reference features; capturing the feature extraction process logs of the classification layer, prediction layer, and auxiliary layer for the adaptation verification features to obtain periodic process sub-features of each layer; fusing the periodic process sub-features based on the functional weights of each layer in the cascaded model to obtain fused periodic process features; and comparing the fused periodic process features with the fused reference features to obtain a hierarchical feature difference set; and comparing the global difference weights of the global mapping relationship and the periodic difference weights of the periodic local mapping relationship with the difference probability elements and hierarchical feature difference sets under the same verification sample. The relationships are weighted and encoded to obtain a weighted comparison map. The weighted difference ladder corresponding to the difference probability element is determined by combining the distribution gradient of the global difference weights and the fluctuation range of the periodic difference weights in the weighted comparison map. Based on the weighted difference ladder of the same difference probability element across all validation samples, the difference ladder is weighted by reliability based on the prediction confidence output of the prediction layer, resulting in a weighted difference ladder sequence. The weighted sum of the mean square error and entropy of the weighted difference ladder sequence is used as the first interaction loss function. Simultaneously, based on all feature difference values ​​under the same difference feature, the probability difference influence coefficient is adjusted by combining the classification result confidence of the classification layer and the auxiliary feature contribution of the auxiliary layer. The cross-entropy of the feature difference value and the adjusted influence coefficient is used as the second interaction loss function. The global matching degree of the validation sample in the global mapping relationship and the periodic matching degree in the periodic local mapping relationship are calculated and used as the global loss adjustment factor and the periodic loss adjustment factor, respectively. Weights are then assigned to the first and second interaction loss functions based on these adjustment factors to obtain the comprehensive loss function.

Images