Machine learning-based method and system for predicting vaccine adjuvant effect

Abstract

The application relates to the technical field of medical health information processing, and discloses a vaccine adjuvant effect prediction method and system based on machine learning. Medical observation data of vaccination events are acquired, and vaccination event level standardized observation results are constructed by taking single vaccination events as objects. Reaction trigger tracing is performed based on an event chain after vaccination, reaction trigger type intervention tracks and reaction role recognition results are generated, label pollution reconstruction is performed for immunization detection endpoints, intervention type label pollution results and double adjuvant effect representations constrained by label pollution are formed, and through double label machine learning training constrained by label credibility, prediction results for distinguishing non-intervention adjuvant effects and actual vaccination available adjuvant effects and population adaptability results are output. The application can reduce the influence of post-vaccination intervention and reaction pseudo-correlation on adjuvant effect prediction.

Description

Technical Field

[0001] This application relates to the field of medical and health information processing technology, and in particular to a method and system for predicting the effects of vaccine adjuvants based on machine learning. Background Technology

[0002] Vaccine adjuvants are used to enhance antigen-induced immune responses. In candidate adjuvant screening, dose design, population suitability evaluation, and real-world follow-up analysis, it is usually necessary to combine vaccination records, basic health information of subjects, post-vaccination response records, immune test results, and follow-up data to determine the adjuvant effect. As the scale of vaccine research and development data expands, machine learning methods are increasingly used to predict antibody titers, neutralizing antibody levels, cytokine changes, T-cell response intensity, or overall immune gain results after vaccination, in order to help determine the applicability of different adjuvants in different subject groups. In practical applications, children may receive antipyretic treatment due to fever after vaccination, and the elderly or people with chronic diseases may have their blood collection time adjusted due to discomfort after vaccination. In multi-center clinical studies, there may also be records of supplementary testing, enhanced follow-up, or withdrawal due to differences in treatment procedures. All of the above data belong to medical and health data formed after vaccination, and their changes will directly affect the training sample quality and output reliability of subsequent adjuvant effect prediction models.

[0003] Existing methods for predicting vaccine adjuvant effects often use post-vaccination immune test results as model labels and input post-vaccination reactions such as fever, redness, pain, and fatigue as common features. For post-vaccination events such as rescue medication, delayed blood collection, supplementary testing, and withdrawal from the group, sample removal, missing value imputation, outlier correction, covariate input, or stratified analysis are commonly used. These methods are suitable for improving data integrity and reducing the impact of some outlier records on model training. However, in scenarios where post-vaccination reactions further trigger rescue medication, blood collection adjustments, or supplementary testing, it is still difficult to fully characterize the continuous transmission relationship between post-vaccination reactions and subsequent interventions. Some post-vaccination reactions are induced by adjuvants. The immune response can trigger further interventions such as medication, blood collection adjustments, or supplementary testing. These interventions can alter the values, collection times, or observability of the immune endpoint, causing a deviation between the adjuvant effect label used in the model and the original induction effect of the adjuvant. Furthermore, there is no stable correlation between local reactions, systemic reactions, and protective immune outcomes under different adjuvants, different populations, and different time windows. The model risks learning to use reaction intensity as a basis for adjuvant effectiveness. Therefore, existing methods still have room for improvement in handling the linkage between post-vaccination intervention trajectories, reaction signal roles, and the credibility of immune endpoint labels. Consequently, it is necessary to improve the reliability of adjuvant effect prediction results in complex vaccination scenarios. Summary of the Invention

[0004] This application proposes a machine learning-based method and system for predicting the effects of vaccine adjuvants, in order to address the problems mentioned in the background art.

[0005] To achieve the above objectives, this application adopts the following technical solution: a machine learning-based method for predicting the efficacy of vaccine adjuvants, comprising: S1. Acquire medical observation data of vaccination events. Using a time-series merging method anchored to vaccination events, limit the post-vaccination event chain formed by post-vaccination response and post-vaccination intervention to the same single vaccination event and the same relative time axis of the immune detection endpoint, and generate standardized observation results at the vaccination event level. S2 is designed to trace response triggers in the post-vaccination event chain. It performs transmission and analysis on standardized observation results at the vaccination event level, enabling a traceable trigger relationship between post-vaccination response and post-vaccination intervention. It also marks the data role of post-vaccination response in the trigger relationship, generating response-triggered intervention trajectories and response role identification results. S3, through a label contamination reconstruction method oriented towards the immune detection endpoint, analyzes the process by which the response-triggered intervention trajectory and response role identification results act on the immune detection endpoint, transforming the immune detection endpoint affected by the intervention into a dual adjuvant effect characterization constrained by label contamination; S4 uses a dual-label machine learning training method constrained by label credibility to predict the adjuvant effect representation constrained by label contamination. This enables the model to output adjuvant effect prediction results that distinguish between the effect of uninterrupted adjuvant and the effect of adjuvant that is actually available after vaccination, and to form corresponding population suitability results.

[0006] This invention also provides a machine learning-based vaccine adjuvant efficacy prediction system, comprising: The standardized observation construction module acquires medical observation data of vaccination events. Using a time-series merging method anchored to vaccination events, it limits the post-vaccination event chain formed by post-vaccination response and post-vaccination intervention to the same single vaccination event and the same relative time axis of the immune detection endpoint, generating standardized observation results at the vaccination event level. The reaction tracing and identification module performs reaction trigger tracing for the post-vaccination event chain, performs transmission and analysis on standardized observation results at the vaccination event level, establishes a traceable trigger relationship between post-vaccination reactions and post-vaccination interventions, marks the data role of post-vaccination reactions in the trigger relationship, and generates reaction-triggered intervention trajectories and reaction role identification results. The label contamination reconstruction module analyzes the process by which the response-triggered intervention trajectory and response role identification results affect the immune detection endpoint through a label contamination reconstruction approach oriented towards the immune detection endpoint, thereby transforming the immune detection endpoint affected by the intervention into a dual adjuvant effect characterization constrained by label contamination. The prediction output module uses a dual-label machine learning training method constrained by label credibility to predict the adjuvant effect representation constrained by label contamination. This enables the model to output adjuvant effect prediction results that distinguish between the effect of the uninterrupted adjuvant and the effect of the adjuvant that can be used in actual vaccination, and to form corresponding population suitability results.

[0007] The beneficial effects of this invention are as follows: 1. This invention, through a time-series merging method anchored to vaccination events, limits post-vaccination reactions, post-vaccination interventions, and immune detection endpoints to the same single vaccination event and the same relative time axis, forming standardized observation results at the vaccination event level. This solves the problem of different doses, different vaccine adjuvant combinations, different blood collection cycles, and different follow-up records easily mixing into the same training sample, and ensures that subsequent response trigger tracing and label contamination reconstruction have unified object boundaries, time boundaries, and field availability boundaries, thereby improving the comparability and traceability of adjuvant effect prediction samples from the data organization level.

[0008] 2. This invention, through response trigger tracing and response role identification, analyzes the temporal transmission relationship between post-vaccination response and post-vaccination intervention as a response-triggered intervention trajectory, and marks post-vaccination response as an intermediate result with a clear data role. This solves the problem that post-vaccination response is directly learned by machine learning models as an adjuvant effect feature, enabling subsequent processing to distinguish between immune concomitant reactions, intervention-triggered reactions, irrelevant background reactions, and reactions with undetermined roles. This provides a more accurate constraint on whether the endpoint of immune detection is affected by post-vaccination intervention.

[0009] 3. This invention, through a label contamination reconstruction method oriented towards the immune detection endpoint, transforms numerical rewriting, collection time rewriting, and observability rewriting into interventional label contamination results. Furthermore, it constructs a dual adjuvant effect characterization constrained by label contamination, allowing the non-interventional adjuvant effect characterization and the actual available adjuvant effect characterization to enter the dual-label machine learning training process constrained by label credibility. This solves the problem of prediction bias caused by directly using the intervened immune detection endpoint as a single training label, enabling the model to output adjuvant effect prediction results that distinguish between the non-interventional adjuvant effect and the actual available adjuvant effect, and generates population suitability results based on label credibility and population stratification. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort: Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system framework diagram of the present invention. Detailed Implementation

[0011] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0012] Example 1

[0013] like Figure 1 As shown, this invention discloses a method for predicting the efficacy of vaccine adjuvants based on machine learning, including the following specific steps: In one implementation, S1 acquires medical observation data of vaccination events and uses a time-series merging method anchored to vaccination events to limit the post-vaccination event chain formed by post-vaccination reactions and post-vaccination interventions to the same single vaccination event and the same relative time axis of the immune detection endpoint, generating standardized observation results at the vaccination event level. This step is used to establish unified sample object boundaries, time boundaries, and field availability boundaries before entering reaction trigger tracing and label contamination reconstruction, reducing the situation where different doses, different vaccine adjuvant combinations, different blood collection cycles, and different follow-up records are mixed into the same training sample.

[0014] The medical observation data for vaccination events include vaccination identification information for determining a single vaccination event, post-vaccination event chain records, immune endpoint records, and basic health information of the subjects. Vaccination identification information is used to determine the subject boundaries of the vaccination event. Post-vaccination event chain records are used to record the temporal relationship between post-vaccination response and post-vaccination intervention. Immune endpoint records are used to record baseline testing before vaccination, post-vaccination immune testing, and the corresponding testing time. Basic health information of the subjects is used for subsequent construction of similar reference events and population adaptation processing. Post-vaccination event chain records consist of post-vaccination response records and post-vaccination intervention records. Immune endpoint records include the planned immunization testing time window, actual immunization testing time, and immune endpoint observation values.

[0015] When anchoring vaccination events, vaccination identification information is first individually aggregated according to subject identification, so that the vaccination records, post-vaccination event chain records, and immune test endpoint records corresponding to the same subject are included in the same candidate aggregation range. Then, the aggregated vaccination identification information is divided into candidate vaccination actions according to the number of doses and the vaccination time. The number of doses comes from at least one of the vaccination registration system, clinical trial protocol, and case report form. The vaccination time comes from at least one of the vaccination registration system, vaccination medical order execution record, and clinical trial vaccination record. The vaccination completion time in the vaccination execution record is used as the vaccination time sequence anchor point, and consistency verification is performed with the vaccination time from other sources.

[0016] When the differences in vaccination times recorded from different data sources are within the time error range allowed by the clinical trial protocol or data management specifications, the vaccination completion time in the vaccination execution record is used as the vaccination time sequence anchor point, and the time error range is preset by the clinical trial protocol or data management specifications. In scenarios where the vaccination time recording accuracy is at the minute level, the time error range is preferably 0 to 30 minutes. In scenarios where the vaccination time recording accuracy is at the date level, the time error range is preferably the same natural day. When there are vaccination records for multiple vaccine adjuvant combinations within the same natural day, a single vaccination event is still established according to the vaccine adjuvant combination. When the differences in vaccination times recorded from different data sources exceed the time error range, a vaccination time conflict marker is generated, and the post-vaccination event chain and the endpoint of immune detection are restricted from entering the binding range of unified processing of the vaccination relative time axis.

[0017] For each candidate vaccination action, a vaccine-adjuvant combination consistency check is performed using the vaccine identifier, adjuvant identifier, and vaccination batch. Specifically, the vaccine identifier, adjuvant identifier, and vaccination batch within the same candidate vaccination action are compared with the vaccination registration system, trial group records, and drug batch number records, respectively. When the vaccine identifier, adjuvant identifier, and vaccination batch can correspond to the same vaccination protocol, the candidate vaccination action is confirmed as the same single vaccination event. The output of the vaccine-adjuvant combination consistency check is the single vaccination event that passes the check, the candidate vaccination action that fails the check, and the corresponding vaccination boundary pending confirmation mark.

[0018] When the same subject has at least one of different vaccine identifiers or different adjuvant identifiers at the same vaccination time, a single vaccination event is established according to the vaccine adjuvant combination. In the scenarios of combined vaccination and sequential vaccination, the same day of vaccination is not used as the basis for merging, but the consistency verification result of the vaccine adjuvant combination is used as the basis for splitting. If either the vaccination batch or the vaccination time is missing, a vaccination boundary confirmation mark is generated. Vaccination events with a vaccination boundary confirmation mark are subject to binding scope restrictions. If the vaccination time is missing, the time of occurrence of post-vaccination reaction, the time of occurrence of post-vaccination intervention, and the time of immune detection are not converted into relative vaccination time. If the vaccination batch is missing, the vaccination event is not directly included in the construction of similar reference events under the same vaccine adjuvant combination.

[0019] After identifying the same single vaccination event, the vaccination time of that single vaccination event is used as the vaccination timeline anchor. The occurrence times in the post-vaccination event chain record and the immunization endpoint record are uniformly converted into relative vaccination times. The relative vaccination time represents the time difference between the occurrence time of the corresponding record and the vaccination timeline anchor. The time unit is days or hours, and it is consistent within the same application document. The input for this time standardization process is the original occurrence time of the vaccination timeline anchor and the record to be converted. The processing action is to convert the absolute time of different data sources into the relative vaccination time under the same reference. The output is the relative occurrence time of post-vaccination reaction, post-vaccination intervention, and immunization endpoint. The converted relative vaccination time is used to subsequently determine whether the post-vaccination reaction occurs before the post-vaccination intervention, whether the post-vaccination intervention occurs before the immunization endpoint collection, and whether the actual immunization testing time deviates from the planned immunization testing time window.

[0020] For pre-vaccination baseline test records, a negative relative vaccination time is formed according to the time difference between the test time and the vaccination timeline anchor point. The pre-vaccination baseline test records are limited to the baseline source of changes in the immunization endpoint and are not included in the post-vaccination event chain. This treatment is used to avoid using the pre-vaccination baseline test records as the post-vaccination immunization endpoint and to provide a baseline basis for the homogenization and standardized scaling of subsequent changes in immunization endpoints.

[0021] Next, the post-vaccination event chain records and the immune test endpoint records will be bound at the object level to form standardized observation results at the vaccination event level. The object-level binding includes standardized treatment of response, standardized treatment of intervention, and standardized treatment of immune test endpoint. Within the same single vaccination event, the relative time correspondence between the standardized records of response, standardized records of intervention, and standardized records of immune test endpoint will be established.

[0022] For post-vaccination reaction records, they are converted into standardized reaction records according to reaction type, reaction occurrence time, reaction duration, and reaction grade. The post-vaccination reaction type is determined according to at least one of the clinical trial protocol, follow-up form, and electronic medical record fields, including local post-vaccination reaction and systemic post-vaccination reaction. The reaction occurrence time is converted into the relative time of vaccination with respect to the vaccination timeline anchor point. The reaction duration is determined according to the reaction start time and reaction end time. The reaction grade is determined according to at least one of the clinical trial protocol, follow-up form, electronic medical record grading field, and preset reaction grading table. The reaction grade preferably adopts a grading method of 1 to 4. When the follow-up form clearly records no reaction, the no-reaction state is recorded as a separate state. Missing reaction records are not directly converted into a no-reaction state. When the reaction grades from different sources are inconsistent, the grade field in the clinical trial protocol or case report form is used as the standard grade field, and the reaction grades from other sources are used for consistency verification.

[0023] When only raw numerical values ​​are available but the reaction grade field is missing, the raw values ​​are converted into reaction grades according to the clinical trial protocol, case report form, or preset reaction grading table. Fever reactions are converted using body temperature ranges, local redness and swelling reactions are converted using redness and swelling diameter ranges, and pain reactions are converted using the standard pain grade of the follow-up questionnaire. When a traceable conversion table is missing, a reaction grade pending confirmation mark is generated. This reaction record retains the reaction type and reaction occurrence time, but does not directly enter subsequent reaction intensity support verification. When the reaction end time is missing, a reaction end status pending confirmation mark is generated. This reaction record retains the reaction occurrence time and reaction grade, but does not directly enter subsequent judgments dependent on reaction duration. When post-vaccination reaction records are missing, a reaction record missing mark is generated. Vaccination events with reaction record missing marks are not directly used as non-reaction events in subsequent reaction role identification and label credibility processing.

[0024] For post-vaccination intervention records, they are converted into standardized intervention records according to intervention type, intervention occurrence time, intervention duration, and intervention execution reason. Intervention types include rescue medication, delayed blood collection, early blood collection, supplementary testing, withdrawal from the group, and enhanced follow-up. The intervention occurrence time is converted to the relative vaccination time. The intervention duration is determined according to the intervention start time and intervention end time. The intervention execution reason comes from at least one of the reason fields in the post-vaccination intervention record, case report form, and electronic medical record. The standardization processing of post-vaccination intervention records does not directly determine that the intervention has changed the immune test endpoint. The output of this processing is the standardized intervention record and the corresponding data integrity mark. When the intervention execution reason is missing, an intervention reason missing mark is generated; when the intervention occurrence time is missing, an intervention time missing mark is generated; when the intervention end time is missing, an intervention end status pending confirmation mark is generated. Records with intervention reason missing mark are retained as background intervention records but do not directly participate in intervention reason consistency verification. Records with intervention time missing mark are not included in the time tracing of response-triggered intervention trajectory. Records with intervention end status pending confirmation mark retain the intervention start time and intervention type but do not participate in the judgment related to intervention duration.

[0025] For records of immune testing endpoints, standardized records are generated based on the planned immunization testing time window, actual immunization testing time, and observed immune endpoint values. Immune testing endpoints include antibody titer, neutralizing antibody concentration, cytokine level, T cell response intensity, and the comprehensive immune gain result recorded in the medical observation data of the vaccination event. If the comprehensive immune gain result is not recorded in the medical observation data of the vaccination event, it is not calculated in S1. The planned immunization testing time window is derived from at least one of the clinical trial protocol, follow-up plan, and testing plan table. The actual immunization testing time is derived from at least one of the sample collection record, laboratory testing record, and test report generation record, with the actual sample collection time being preferred as the actual immunization testing time. If the sample collection time is missing but the testing time exists, a missing collection time marker is generated, and the testing time is used as an auxiliary time field in subsequent reliability processing.

[0026] When the actual immunization testing time falls within the planned immunization testing time window, a planned testing marker is generated; when the actual immunization testing time is earlier or later than the planned immunization testing time window, a blood collection time window deviation marker is generated; when the planned immunization testing time window is missing, a planned window missing marker is generated, but no planned testing marker or blood collection time window deviation marker is generated for that immunization endpoint; immunization endpoints with a planned window missing marker do not participate in the direct judgment of the collection time rewriting support results in subsequent label contamination reconstruction; when at least one of the following situations occurs—invalid test result, test failure, and result replacement by supplementary test—a test result validity marker and a supplementary test replacement marker are generated. This marker is used in subsequent S3 to determine the observability rewriting status of the immunization endpoint.

[0027] After completing the standardized records of response, intervention, and immune endpoint, the basic health information of the subjects corresponding to the same single vaccination event is bound to the various standardized records at the object level to generate standardized observation results at the vaccination event level. The basic health information of the subjects is used for the construction of subsequent similar reference events and population adaptation processing, including age group, underlying disease status, immunosuppressive medication status, previous vaccination history, and previous infection history. The stratification method of the subjects' basic health information comes from the research protocol, clinical data dictionary, or historical sample distribution. When the historical samples are insufficient to support stable stratification, the established stratification in the research protocol or clinical data dictionary is adopted, and the stratification method is not changed temporarily in this step.

[0028] Standardized observation results at the vaccination event level include single vaccination event identifiers, vaccination timeline anchors, vaccination relative timelines, standardized response records, standardized intervention records, standardized records of immune test endpoints, basic health information of subjects, and data integrity markers. Data integrity markers are used to indicate whether the post-vaccination event chain and immune test endpoints have the fields required for subsequent processing. Data integrity markers include missing response record markers, complete intervention record markers, complete immune test endpoint markers, vaccination boundary confirmation markers, vaccination time conflict markers, unconfirmed response grade markers, unconfirmed response end status markers, missing intervention reason markers, missing intervention time markers, unconfirmed intervention end status markers, missing collection time markers, missing planning window markers, blood collection time window deviation markers, test result validity markers, and supplementary test replacement markers. Data integrity markers are not directly used as adjuvant effect prediction labels but are used to limit the range of available fields in subsequent response trigger tracing and label contamination reconstruction.

[0029] Through the aforementioned processing, S1 confines medical observation data scattered across the vaccination registration system, follow-up records, medication records, withdrawal records, and laboratory test records to the same single vaccination event, forming a unified relative timeline for vaccination. The technical effects of this step are: reducing the mixing of data from different doses, different vaccine adjuvant combinations, and different testing cycles into the same training sample; providing object boundaries, time boundaries, and field boundaries for S2 to determine whether there is a traceable triggering relationship between post-vaccination response and post-vaccination intervention, and for S3 to determine whether post-vaccination intervention affects the endpoint of immunization testing.

[0030] In one implementation, S2, based on the standardized observation results at the vaccination event level generated by S1, performs response trigger tracing for the post-vaccination event chain, performs transmission analysis on the standardized observation results at the vaccination event level, establishes a traceable trigger relationship between the post-vaccination response and the post-vaccination intervention, and marks the data role of the post-vaccination response in the trigger relationship, generating response trigger-type intervention trajectory and response role identification results. This step is used to convert the post-vaccination response from ordinary input features into intermediate results with trigger relationships and data roles before entering the label contamination reconstruction.

[0031] The input to S2 is the standardized observation results at the vaccination event level formed by S1. These standardized observation results include standardized response records, standardized intervention records, standardized records of immune endpoint detection, a relative vaccination timeline, and data integrity markers within the same single vaccination event. Standardized response records serve as input records for post-vaccination responses, standardized intervention records serve as input records for post-vaccination interventions, and standardized records of immune endpoint detection serve as sources of change in immune endpoint detection during response role identification. Data integrity markers are used to limit the available scope of corresponding records for pre- and post-timeline verification, consistency verification of intervention causes, support verification of response intensity, and comparison with similar reference events. Standardized response records lacking a response occurrence time are not included in pre- and post-timeline verification; standardized intervention records lacking an intervention occurrence time are not included in the timeline tracing of response-triggered intervention trajectories; standardized intervention records lacking a reason for intervention execution are not included in the consistency verification of intervention causes; if at least one of the response grade or response duration is missing, the standardized response record is not included in the response intensity support verification; and vaccination events lacking evidence of changes in immune endpoint detection are not included in comparison with similar reference events.

[0032] When performing response trigger tracing, the timing of post-vaccination response and post-vaccination intervention is first verified within the same single vaccination event. Specifically, the response occurrence time of the current response record is extracted from the standardized response record, and the intervention occurrence time of the current intervention record is extracted from the standardized intervention record. The chronological relationship between the two is compared within the relative time axis of vaccination. Record pairs whose response occurrence time is no later than the intervention occurrence time are selected as candidate response-intervention pairs. Record pairs whose response occurrence time is later than the intervention occurrence time are not selected as candidates for response trigger relationship. This process outputs a set of candidate response-intervention pairs for subsequent trigger observation window verification, intervention cause consistency verification, and response intensity support verification.

[0033] The candidate response intervention is subjected to a trigger observation window verification. The trigger observation window verification is used to determine whether the interval between the response occurrence time and the intervention occurrence time is within the observation window of the corresponding intervention type. The observation window is preset according to the intervention type, and the time unit is days. The basis for its formation includes at least one of the following: clinical trial protocol, follow-up rules, case report form filling rules, and historical vaccination sample verification results.

[0034] The observation window for rescue medication is set to 0 to 2 days after the reaction occurs, to match rescue treatments following fever, pain, and allergic-like reactions; the observation window for enhanced follow-up is set to 0 to 7 days after the reaction occurs, to match short-term reaction-related follow-up after vaccination; the observation window for delayed blood collection, early blood collection, supplementary testing, and withdrawal is set to 0 to 14 days after the reaction occurs, to match adjustments to the follow-up plan and changes in testing arrangements. If the study protocol specifies otherwise for the follow-up period, blood collection adjustment period, supplementary testing period, and withdrawal determination period after vaccination, the observation window specified in the study protocol shall be used; if historical samples are insufficient to support a stable estimate of the observation window, the preset observation window in the clinical trial protocol or data management specifications shall be used, triggering the observation window validation output time proximity judgment result. This result is used to indicate that the candidate response intervention pair has a temporal triggering possibility, and does not separately determine the response triggering relationship.

[0035] For candidate response interventions that pass the observation window verification, the consistency of the intervention cause is checked. The consistency of the intervention cause is used to determine whether the intervention cause falls within the intervention cause mapping range of the corresponding response type. The intervention cause mapping range comes from at least one of the following: clinical trial protocol, case report form field, electronic medical record cause field, follow-up rules, and historical vaccination sample statistics. Fever reaction corresponds to antipyretic treatment, local pain corresponds to analgesic treatment, allergic-like reaction corresponds to anti-allergy treatment, persistent discomfort corresponds to enhanced follow-up, and when post-vaccination discomfort affects blood collection arrangements, it corresponds to at least one of delayed blood collection and supplementary testing. When the intervention cause field is missing, the current candidate response intervention does not pass the consistency of the intervention cause, and the corresponding standardized intervention record is retained as a background intervention record. The consistency of the intervention cause outputs the cause consistency judgment result, which is used to limit erroneous binding caused only by proximity of time.

[0036] For candidate response interventions that pass the observation window verification and intervention cause consistency verification, a response intensity support verification is performed. The response intensity support verification is used to determine whether the grade and duration of the current post-vaccination response meet the support conditions for triggering the corresponding post-vaccination intervention. The response grade is converted to a 0 to 1 evaluation scale according to the grade field in the clinical trial protocol or case report form. During the conversion, the highest grade allowed for this response type in the study protocol is used as the upper limit of the scale. The response duration is obtained according to the response start time and response end time, and the 95th percentile of the longest response duration defined by the study protocol or the duration of the same type of response in historical vaccination samples is used as the upper limit of the scale. When the response duration exceeds the upper limit of the scale, the duration support result is treated as 1. When the upper limit of the scale is missing, the effective upper limit preset by the study protocol or data management specifications is used, and an upper limit replacement marker is generated.

[0037] The response grade support result and the response duration support result are combined into the response intensity support result according to preset weights. The preset weights are derived from at least one of the study protocol, historical vaccination sample validation results, and development set validation results. When historical samples are insufficient, a fixed weight configuration in the study protocol or data management specifications is adopted. In one preferred implementation, the sum of the weights of the response grade support result and the response duration support result is 1. When the study protocol does not limit the weights and historical samples are insufficient, the response grade support result and the response duration support result are given the same weights. The response intensity support condition is a judgment condition on an evaluation scale of 0 to 1, and the value is derived from at least one of the clinical trial protocol, follow-up rules, and historical vaccination sample validation results.

[0038] The judgment value corresponding to the response intensity support condition is a fixed value in the range of 0.3 to 0.7. When the post-vaccination response record is complete, the reason for the intervention is clear, and the correspondence between the response and the intervention in the historical samples is stable, the judgment value is 0.3 to 0.5. When the source of the response record is scattered, the proportion of missing fields of the intervention reason reaches the limit of the data management specification, or it is necessary to reduce the false match rate, the judgment value is 0.5 to 0.7. The judgment value is determined by the research protocol, data management specification or development set validation results before the model runs, and remains fixed during the model runs.

[0039] When a candidate response intervention pair completes the trigger observation window verification, intervention cause consistency verification, and response intensity support verification, the candidate response intervention pair is determined as a response triggering relationship. All response triggering relationships within the same single vaccination event form a response triggering intervention trajectory. The response triggering intervention trajectory includes at least the triggering response, the triggered intervention, the response occurrence time, the intervention occurrence time, the intervention type, the trigger support result, and the corresponding data integrity marker. In the case where the same post-vaccination response triggers multiple post-vaccination interventions, the multiple post-vaccination interventions are written into the same response triggering intervention trajectory in the order of the relative time axis of vaccination. In the case where multiple post-vaccination responses point to the same post-vaccination intervention, the trigger support result corresponding to each post-vaccination response is retained and used as the input for pre-endpoint intervention tracing in subsequent S3.

[0040] After obtaining the response-triggered intervention trajectory, the data role of the post-vaccination response in the current single vaccination event is further identified. For each post-vaccination response record in the current single vaccination event, a set of similar reference events is established. The set of similar reference events is used to provide a benchmark for the identification of response roles. The vaccination events in the set of similar reference events meet the following conditions: they have the same vaccine adjuvant combination, the same response type, and the same subject baseline health stratification as the current single vaccination event, and they have not formed a response-triggered intervention trajectory. The subject baseline health stratification is derived from the subject baseline health information in S1, and the stratification criteria include at least one of age group, underlying disease status, immunosuppressive medication status, previous vaccination history, and previous infection history.

[0041] When the number of historical samples is sufficient to support stratified comparisons, the number of vaccination events in the same reference event set should preferably be no less than 5. If the initial reference event set has fewer than 5 vaccination events, the reference event set should be re-established according to the baseline health stratification of the subjects at the next higher level as pre-defined in the study protocol. If the re-established reference event set still has fewer than 5 vaccination events, the median result comparison of changes in the immune endpoint will not be performed, and the current post-vaccination response will be marked as a role pending response. Five vaccination events is the minimum usable number required for median result comparison, derived from at least one of the study protocol, data management guidelines, and historical sample stability verification results. If the study protocol has other minimum sample size requirements, the number requirement specified in the study protocol shall be adopted.

[0042] Before comparing changes in immune endpoints, baseline correction, standardization, and endpoint orientation alignment are performed on the immune endpoints of the current single vaccination event and the set of similar reference events. This results in standardized and uniformly scaled immune endpoint changes. The inputs to this process are pre-vaccination baseline test records and post-vaccination immune test records. The purpose of this process is to bring results from different testing platforms, different types of immune endpoints, and different endpoint orientations into the same comparison scale. The output is the immune endpoint change results used for role identification. For immune endpoints where an increase in value indicates enhanced adjuvant effect, the original orientation is maintained. For immune endpoints where an increase in value indicates weakened effect, orientation alignment is performed to ensure that the processed immune endpoint change results satisfy the condition that an increase in value corresponds to enhanced adjuvant effect. Vaccination events lacking pre-vaccination baseline test records or testing platform calibration data are not included in the median result comparison of similar reference events, and an immune endpoint change confirmation marker is generated.

[0043] After generating the results of changes in immune test endpoints after standardization and uniform scaling, the median result of the changes in immune test endpoints in the same set of reference events is calculated. The results of changes in immune test endpoints corresponding to the current post-vaccination response are compared with the median result. The median result is used as a comparison benchmark to reduce the impact of extreme immune test values, a small number of supplementary test values, and abnormal fluctuations in individual subjects on response role identification. The comparison results are only used to identify the data role of the post-vaccination response in the current single vaccination event and are not used as conclusions on clinical protective efficacy and medical safety risks.

[0044] If the current post-vaccination response does not form a response triggering relationship, the change in the immune endpoint corresponding to the current post-vaccination response is not lower than the median change in the immune endpoint in the same set of reference events, and the response occurs within the immune response observation window set in the study protocol, the current post-vaccination response will be marked as an immune co-occurrence response. Immune co-occurrence responses are used to characterize the co-occurrence relationship between the post-vaccination response and the change in the immune endpoint, and do not indicate that the post-vaccination response itself has a clinical protective effect.

[0045] If the current post-vaccination response does not form a response triggering relationship, the change in the corresponding immune endpoint is lower than the median change in the immune endpoint in the same set of reference events, and the response occurs outside the immune response observation window set in the study protocol, the current post-vaccination response will be marked as an irrelevant background response. The immune response observation window set in the study protocol is used to limit the comparable time range between the post-vaccination response and the target immune endpoint. The immune response observation window is derived from at least one of the time distributions of post-vaccination response and the target immune endpoint in the study protocol, follow-up rules, and historical vaccination samples. Days 0 to 7 post-vaccination are used as the reactivity observation window, and days 7 to 28 post-vaccination are used as the early immune detection association observation window. If at least one of the study protocol and follow-up rules has a separate time window, the corresponding time window will be used. If the current post-vaccination response does not meet the formation conditions of a co-immune reaction, an intervention-triggered reaction, or an irrelevant background reaction, the current post-vaccination response will be marked as a role pending reaction. A role pending reaction is used to characterize that the current data conditions are insufficient to support the data role determination of the post-vaccination response. This marking will not be used as a positive basis for the correction of the reliability of the immune endpoint in subsequent S3.

[0046] After data role labeling is completed for all post-vaccination response records within the same single vaccination event, a response role identification result is generated. The response role identification result includes at least the response record, response role, immunization endpoint support result, intervention trigger result, and response intensity support result. The response role identification result and the response trigger intervention trajectory are jointly output to S3. The response trigger intervention trajectory is used by S3 to screen post-vaccination interventions that occur before the collection of immunization endpoints. The response role identification result is used by S3 to determine whether the post-vaccination response should be used as constraint information in the label contamination reconstruction. Through this processing, the post-vaccination response is not directly entered into the prediction model as a normal machine learning input feature, but is first limited to an intermediate result with a data role, thereby reducing the risk that the model will mislearn the response intensity as the basis for adjuvant effect.

[0047] In one implementation, S3, based on the standardized observation results of the vaccination event level generated by S1, the response-triggered intervention trajectory generated by S2, and the response role identification results, analyzes the process by which the response-triggered intervention trajectory and the response role identification results act on the immune detection endpoint through a label contamination reconstruction method oriented towards the immune detection endpoint. This transforms the immune detection endpoint affected by the intervention into a dual adjuvant effect representation constrained by label contamination. This step is used to identify the state of the immune detection endpoint affected by post-vaccination intervention, collection time offset, supplementary testing substitution, and withdrawal deletion before machine learning training, and converts this state into the basis for label credibility constraints in subsequent dual-label machine learning training.

[0048] The inputs to S3 include standardized observation results at the vaccination event level, response-triggered intervention trajectories, and response role identification results. Standardized observation results at the vaccination event level provide standardized records of immunization endpoints, planned immunization testing time windows, actual immunization testing times, immunization endpoint observations, and data integrity markers. Response-triggered intervention trajectories provide the triggering relationship between post-vaccination reactions and post-vaccination interventions, the timing and type of interventions. Response role identification results provide the data role of the post-vaccination reaction in the current single vaccination event. The aforementioned inputs are used together to determine whether the post-vaccination intervention can serve as a source of label contamination for the current immunization endpoint.

[0049] First, the immune endpoints to be processed are extracted from the standardized records of immune endpoints. Pre-vaccination baseline records, post-vaccination immune endpoint records, planned immunization testing time windows, actual immunization testing times, validity markers for test results, supplementary testing replacement markers, and endpoint withdrawal status are determined. When both pre-vaccination baseline records and post-vaccination immune endpoint records are valid, baseline correction, unified scaling of the testing platform, and endpoint orientation consistency processing are performed to form a unified and standardized result of immune endpoint changes. The input to this processing is the pre-vaccination baseline records and post-vaccination immune endpoint records. The purpose of this processing is to bring immune endpoints from different testing platforms and with different endpoint orientations into the same comparative scale. The output is the result of immune endpoint changes. If at least one of the following is missing: pre-vaccination baseline records, post-vaccination immune endpoint records, or testing platform calibration basis, the immune endpoint will not be included in the construction of the uninterventional adjuvant effect characterization, and an immune endpoint change confirmation marker will be generated.

[0050] Subsequently, using response-triggered intervention trajectories as the tracing target, post-vaccination interventions occurring before the collection of immunization test endpoints were screened from these trajectories to form response-triggered endpoint pre-interventions. During screening, the relative time axis of vaccination was used as a unified time reference, and the occurrence time of post-vaccination interventions was compared with the actual immunization test time. Post-vaccination interventions occurring earlier than the actual immunization test time and originating from response-triggered intervention trajectories were included in the response-triggered endpoint pre-interventions. Post-vaccination interventions occurring after the actual immunization test time were not considered as sources of label contamination for the current immunization test endpoint. Post-vaccination interventions that did not form a response-triggered relationship were retained as background intervention records and were not used as the primary basis for label contamination reconstruction. Response-triggered endpoint pre-interventions served as input for the subsequent three types of rewriting identification.

[0051] For post-vaccination interventions that occur before the collection of immunization endpoints but do not form a response triggering relationship, if the intervention type, timing, and type of the immunization endpoint match a pre-defined impact mapping relationship, or if they cause the actual immunization testing time to deviate from the planned immunization testing time window, the validity of test results to change, supplementary testing to replace, or premature withdrawal from the endpoint group, a background intervention credibility limit marker will be generated. The background intervention credibility limit marker is not a primary source of contamination for response-triggered labels and is not used to form interventions before response-triggered endpoints. However, it is used to limit the corresponding immunization endpoint from being classified as an uncontaminated credibility endpoint and serves as a traceability basis for predicting actual usable effects.

[0052] For the process of a response-triggered pre-endpoint intervention acting on an immune detection endpoint, endpoint numerical rewriting identification is performed. The preset impact mapping relationship includes at least the intervention type, immune endpoint type, the impact observation window between the intervention occurrence time and the actual immune detection time, and a standardized support level. The standardized support level is a value ranging from 0 to 1, preferably using 0, 0.25, 0.5, 0.75, and 1 as discrete levels, corresponding to no numerical rewriting support, low-level numerical rewriting support, medium-level numerical rewriting support, high-level numerical rewriting support, and upper limit numerical rewriting support, respectively. For the current response-triggered pre-endpoint intervention, the intervention type is read. The intervention occurrence time, immune endpoint type, and actual immune detection time are matched with the corresponding standardized support level in the preset influence mapping relationship to form a numerical rewriting support result. If the current intervention type does not match the preset influence mapping relationship, or if the interval between the intervention occurrence time and the actual immune detection time exceeds the corresponding influence observation window, no numerical rewriting support result is formed. The preset influence mapping relationship is derived from at least one of the following: clinical trial protocol, case report form, medication restriction rules, historical sample validation results, and pre-established data dictionary. If historical samples are insufficient, a fixed mapping relationship in the clinical trial protocol or data management specifications is used.

[0053] For response-triggered endpoint interventions acting on the immune test endpoint, endpoint collection time rewriting identification is performed. If the actual immune test time falls within the planned immune test time window, no collection time rewriting support result is generated. If the actual immune test time is earlier or later than the planned immune test time window, a collection time rewriting support result is generated based on the deviation of the actual immune test time from the boundary of the planned immune test time window. The deviation is converted into a standardized support result within the range of 0 to 1. When the deviation reaches the preset upper limit of deviation, the collection time rewriting support result is set to 1. The preset upper limit of deviation is in days and is derived from the permitted blood collection window, testing schedule, and follow-up window rules in the clinical trial protocol. If at least one of the following criteria is met: the study protocol, the data management guidelines, and the historical sample data, the upper limit of the pre-set deviation for the short-term immunization endpoint within 7 days post-vaccination is 1 to 3 days; the upper limit of the pre-set deviation for the intermediate-term immunization endpoint between 14 and 28 days post-vaccination is 3 to 7 days; and the upper limit of the pre-set deviation for the long-term follow-up immunization endpoint exceeding 28 days post-vaccination is 7 to 14 days. If historical samples are insufficient, the fixed upper limit in the study protocol or data management guidelines shall be adopted. If the study protocol specifies other provisions for the immunization time window, the upper limit of deviation specified in the study protocol shall be adopted. If the planned immunization time window is missing, no data collection time rewriting support result shall be directly generated, and a data integrity marker corresponding to the missing planned window shall be generated.

[0054] For the process of intervention before the endpoint of a reaction-triggered endpoint acting on the endpoint of an immune test, the observability rewriting identification of the endpoint is performed. When the test result is valid and the retest is only used for laboratory verification, no observability rewriting support result is formed. When the retest replaces the invalid or missing first test result, an intermediate observability rewriting support result is formed, with a preferred value of 0.5 to 0.7. When the test result is invalid, the immune test endpoint is missing, or the immune test endpoint is unobservable due to withdrawal before the endpoint, an upper limit observability rewriting support result is formed, with a value of 1. The values ​​of the intermediate observability rewriting support result and the upper limit observability rewriting support result are derived from at least one of the following: data management specifications, laboratory verification rules, historical sample validation results, and a pre-established data dictionary. When historical samples are insufficient, a fixed value in the data management specifications is used. Withdrawal records are only used as the observability rewriting source for the immune test endpoint when the withdrawal time is earlier than the end time of the current planned immune test time window and results in the missing corresponding immune test endpoint.

[0055] When multiple pre-interventions for response-triggered endpoints correspond to the same immunization endpoint, candidate support results are generated for each post-inoculation intervention in a corresponding dimension. When multiple candidate support results exist for the same dimension, the highest candidate support result in that dimension is taken as the final support result for that dimension. This process is used to avoid the overlapping of multiple post-inoculation interventions, which would cause the standardized support results to exceed the range of 0 to 1. It also ensures that the degree of intervention label contamination reflects the highest support status of the current immunization endpoint affected by post-inoculation interventions. For dimensions that have no basis for rewriting after complete field verification, the support result for that dimension is set to 0. For dimensions that cannot form effective support results due to missing key fields, a corresponding data integrity label is generated, and the immunization endpoint is restricted from entering the uncontaminated reliable endpoint category when grading endpoint reliability.

[0056] After obtaining the numerical rewriting support results, the data acquisition time rewriting support results, and the observable rewriting support results, the degree of intervention label contamination is calculated according to the following formula: ; in, This indicates the degree of intervention label contamination at the immune test endpoint corresponding to the current single vaccination event. It is the output of this formula and serves as the basis for subsequent endpoint confidence grading, label confidence result formation, and S4 training sample weight determination. This indicates the numerical rewriting support result, which corresponds to the result obtained from the aforementioned endpoint numerical rewriting identification, and is used to characterize the degree of influence of response-triggered intervention on the endpoint observation value of immune detection. This indicates the support result for data collection time rewriting, corresponding to the result obtained from the aforementioned endpoint data collection time rewriting identification, used to characterize the degree of deviation of the actual immunization detection time from the planned immunization detection time window; This indicates the observability rewriting support result, corresponding to the results obtained from the aforementioned endpoint observability rewriting identification. It is used to characterize the degree to which the observability of an immune test endpoint is altered due to invalid test results, retest substitution, or endpoint shifting. , and All results are converted to a standardized value between 0 and 1 before being entered into the formula. When there are multiple candidate support results for the same dimension, the final support result for that dimension is obtained according to the aforementioned rule of the highest candidate support result. The value range is from 0 to 1.

[0057] This formula employs a union-based fusion logic, without presupposing that the numerical rewriting support results, collection time rewriting support results, and observability rewriting support results are independent of each other. The endpoint numerical rewriting, endpoint collection time rewriting, and endpoint observability rewriting characterize the impact of post-vaccination intervention on machine learning labels from three aspects: the numerical value of the immune detection endpoint, collection time, and observability, respectively. When all three support results are 0, the current immune detection endpoint has not formed interventional label contamination. When any support result is greater than 0, the degree of interventional label contamination increases with the corresponding support result. When any support result is 1, the current immune detection endpoint reaches the label contamination upper limit in the corresponding dimension. This calculation brings different types of post-vaccination interventions into a unified label contamination constraint scale and converts salvage medication, blood sampling deviation, supplementary testing substitution, and withdrawal censoring into label contamination measurement results that can be used for endpoint credibility grading and S4 training contribution control. This solves the problem of adjuvant effect prediction bias caused by directly using the intervened immune detection endpoint as a machine learning training label.

[0058] After establishing the level of intervention label contamination, the endpoints of the immunoassay tests are graded according to the level of intervention label contamination. The endpoint confidence grading includes uncontaminated reliable endpoints, correctable contamination endpoints, and endpoints that cannot be directly labeled. Both the low and high thresholds for contamination identification are within the range of 0 to 1, with the low threshold being lower than the high threshold. The low threshold is preferably 0.2 to 0.4, and the high threshold is preferably 0.6 to 0.8. When data records are complete, the intervention cause is clear, and historical sample verification shows that the label contamination effect is stable, the low and high thresholds for contamination identification are taken from the lower limit of their respective ranges. When data sources are scattered, the proportion of missing intervention causes reaches the data management specification limit, or the proportion of supplementary testing reaches the data management specification limit, the low and high thresholds for contamination identification are taken from the lower limit of their respective ranges. The upper limit of the range should be used; if the research protocol, development set validation results, or data management specifications stipulate otherwise, the corresponding specified value shall be adopted. When the degree of intervention label contamination is lower than the low threshold for contamination identification and there is no data integrity marker that restricts access to the uncontaminated credible endpoint, the corresponding immune detection endpoint shall be classified as an uncontaminated credible endpoint. When the degree of intervention label contamination reaches the low threshold for contamination identification but is lower than the high threshold for contamination identification and the detection results are valid, the immune detection endpoint shall be classified as a correctable contamination endpoint. When the degree of intervention label contamination reaches the high threshold for contamination identification, or the detection results are invalid, the supplementary test replacement is not traceable, or the immune detection endpoint is missing due to the endpoint being dropped from the previous group, the immune detection endpoint shall be classified as an endpoint that cannot be directly labeled. When historical samples are insufficient to stably determine the threshold, the threshold preset by the research protocol or data management specifications shall be adopted and kept fixed during the model operation phase.

[0059] After obtaining the degree of interventional label contamination, a label confidence result is further formed. The label confidence result is obtained by inversely converting the degree of interventional label contamination, and the value ranges from 0 to 1. In a preferred implementation, the label confidence result is 1 minus the degree of interventional label contamination. The closer the value corresponding to the degree of interventional label contamination is to 1, the closer the value corresponding to the label confidence result is to 0. The label confidence result is used for training sample weights and prediction confidence processing in S4, and does not directly change the value of the actual usable effect prediction label.

[0060] For uncontaminated credible endpoints, the changes in the corresponding immune detection endpoint after homogenization and uniform scaling are used as the direct source of the characterization of the non-interventional adjuvant effect. For correctable contaminated endpoints, a set of similar non-interventional reference endpoints is established. The set of similar non-interventional reference endpoints includes vaccination events with the same vaccine adjuvant combination as the current single vaccination event, the same or similar baseline health stratification of subjects, and whose corresponding immune detection endpoints are effective and classified as uncontaminated credible endpoints. The set of similar non-interventional reference endpoints is used to provide a reference correction benchmark. The number of vaccination events in the set is preferably no less than 5. If there are fewer than 5 vaccination events, the reference range is expanded according to the baseline health stratification of subjects at the next higher level as preset in the study protocol. If there are still fewer than 5 vaccination events after expansion, no correction source for the characterization of the non-interventional adjuvant effect is generated for the correctable contaminated endpoint, and the corresponding endpoint is marked as an endpoint that cannot be directly labeled.

[0061] When the set of similar uninterrupted reference endpoints meets the minimum quantity requirement, the median change in the set of similar uninterrupted reference endpoints is calculated, and reference correction is performed based on the degree of intervention label contamination. During reference correction, the degree of intervention label contamination is used as the contribution ratio of the median change in the set of similar uninterrupted reference endpoints, and the result obtained by subtracting the degree of intervention label contamination from 1 is used as the contribution ratio of the change in the immune detection endpoint after homogenization and uniform scaling corresponding to the current correctable contamination endpoint. This forms the correction source for the characterization of the uninterrupted adjuvant effect. This reference correction process is used to reduce the impact of post-vaccination intervention on the characterization of the uninterrupted adjuvant effect, while avoiding the loss of training samples caused by directly discarding correctable contamination endpoints.

[0062] The actual usable adjuvant effect characterization is used to characterize the usability of the target adjuvant under real-world vaccination conditions after being affected by post-vaccination response, rescue medication, blood sampling deviation, supplementary testing substitution, and withdrawal deletion. This characterization retains the results of changes in immune detection endpoints after homogenization and uniform scaling, the degree of intervention label contamination, endpoint confidence level, response role identification results, and response-triggered intervention trajectories. When it is necessary to generate actual usable effect prediction labels for use by S4, labels are extracted only from the actual usable adjuvant effect characterization with effective immune detection endpoint changes. The actual usable label confidence result does not directly change the value of the actual usable effect prediction label, but serves as the basis for subsequent actual usability training weights and prediction confidence processing. Endpoints cannot be directly labeled, and no uninterventional effect prediction labels are generated. Only the structured actual usable adjuvant effect characterization is retained.

[0063] Finally, S3 outputs a dual adjuvant effect representation constrained by label contamination. This dual adjuvant effect representation includes the non-intervention adjuvant effect representation, the actual available adjuvant effect representation, the interventional label contamination result, the non-intervention label confidence result, and the actual available label confidence result. The non-intervention adjuvant effect representation is used for constructing the non-intervention effect prediction label in S4; the actual available adjuvant effect representation is used for constructing the actual available effect prediction label in S4; the interventional label contamination result, the non-intervention label confidence result, and the actual available label confidence result are used to control the contribution of training samples and prediction confidence in S4. Through this processing, the immune detection endpoint is not directly used as the machine learning training label, but is first processed by label contamination reconstruction and endpoint confidence grading, thereby reducing the impact of post-vaccination intervention, collection time offset, supplementary testing substitution, and withdrawal deletion on the adjuvant effect prediction results.

[0064] In one implementation, S4, based on the label contamination-constrained dual adjuvant effect representation output by S3, uses a label credibility-constrained dual-label machine learning training method to predict the label contamination-constrained dual adjuvant effect representation. This enables the model to output adjuvant effect prediction results that distinguish between the effect of the adjuvant without intervention and the effect of the adjuvant that can be used in actual vaccination, and forms corresponding population suitability results. This step is used to avoid the machine learning model using the immune detection endpoint affected by post-vaccination intervention as a single training label, so that the model can learn the effect representation of the target adjuvant under the condition of no intervention and the usable effect representation under the actual vaccination environment.

[0065] The inputs to S4 include the non-intervention adjuvant effect characterization, the actual available adjuvant effect characterization, the intervention label contamination result, the non-intervention label confidence result, and the actual available label confidence result output by S3. Specifically, the non-intervention adjuvant effect characterization is used to construct the non-intervention effect prediction label, the actual available adjuvant effect characterization is used to construct the actual available effect prediction label, the intervention label contamination result is used to constrain the difference between the predicted non-intervention effect value and the predicted actual available effect value, the non-intervention label confidence result is used to determine the training sample weights for the non-intervention effect prediction task, and the actual available label confidence result is used to determine the training sample weights for the actual available effect prediction task. The non-intervention label confidence result is generated by S3 based on the degree of intervention label contamination, and the actual available label confidence result is generated based on the validity of the test results, the traceability of the actual immunization testing time, and the completeness of the immunization endpoint observation. S4 does not regenerate the non-intervention label confidence result or the actual available label confidence result.

[0066] Training sample objects are determined by a single vaccination event, an immune detection time window, and an immune endpoint type. For each training sample object, the corresponding dual adjuvant effect representation constrained by label contamination is read, and it is determined whether the training sample object has a valid uninterrupted adjuvant effect representation and a valid actual vaccination usable adjuvant effect representation. Training sample objects with uncontaminated credible endpoints or correctable contamination endpoints and having formed uninterrupted adjuvant effect representations are included in the supervised training sample set for the uninterrupted effect prediction task. Training sample objects with valid immune detection endpoint change results after homogenization and uniform scaling are included in the supervised training sample set for the actual usable effect prediction task. Training sample objects corresponding to endpoints that cannot be directly labeled and have not formed uninterrupted adjuvant effect representations are not included in the supervised training sample set for the uninterrupted effect prediction task, but are included in the supervised training sample set for the actual usable effect prediction task if they have valid actual vaccination usable adjuvant effect representations. This process outputs the training sample set and its corresponding label entry status, which is used to limit the calculation range of the subsequent model training loss.

[0067] The predictive input feature vector of the machine learning model is constructed. The predictive input feature vector includes the subject's basic health information, vaccine adjuvant combination identifier, immune detection time window, immune endpoint type, response role identification result, response-triggered intervention trajectory, and intervention label contamination result. The subject's basic health information is used to characterize the subject's basic health stratification. The vaccine adjuvant combination identifier is used to limit the vaccination target corresponding to the target adjuvant. The immune detection time window and immune endpoint type are used to limit the prediction target. The response role identification result is used to characterize the data role of the post-vaccination response. The response-triggered intervention trajectory is used to characterize the triggering relationship between the post-vaccination response and the post-vaccination intervention. The intervention label contamination result is used to characterize the degree to which the immune detection endpoint is affected by the post-vaccination intervention when used as a training label. This processing allows the post-vaccination response and post-vaccination intervention to participate in the predictive learning with the intermediate results formed by S2 and S3.

[0068] For training labels, the uninterventional adjuvant effect representation in the dual adjuvant effect representation constrained by label contamination is used as the uninterventional effect prediction label, and the change in the immune detection endpoint after homogenization and uniform scaling in the dual adjuvant effect representation constrained by label contamination is used as the actual usable effect prediction label. Both the uninterventional effect prediction label and the actual usable effect prediction label are adjuvant effect representations after homogenization and uniform scaling, and are unified to an evaluation scale of 0 to 1. An increase in value indicates an enhanced adjuvant effect at the corresponding immune detection endpoint. The original antibody titer, neutralizing antibody concentration, cytokine concentration, and T cell response intensity are not included in the model training loss calculation before homogenization and uniform scaling are completed.

[0069] The unintervention label confidence results output by S3 are converted into unintervention training weights, and the actual usable label confidence results output by S3 are converted into actual usable training weights. The values ​​of both unintervention training weights and actual usable training weights range from 0 to 1. Values ​​closer to 1 indicate higher confidence of the corresponding training label, increasing its contribution to the training loss of the corresponding model; values ​​closer to 0 indicate lower confidence of the corresponding training label, decreasing its contribution to the training loss of the corresponding model. Unintervention training weights corresponding to uncontaminated confidence endpoints are preferentially close to 1. Unintervention training weights corresponding to contaminated endpoints that can be corrected are determined based on the unintervention label confidence results. Endpoints that cannot be directly labeled do not form unintervention effect prediction labels and are not included in the supervised loss calculation of the unintervention effect prediction task. When uninterventional endpoints still have effective representation of the actual usable adjuvant effect of vaccination, they are only included in the supervised loss calculation of the actual usable effect prediction task. Their actual usable training weights are determined by the validity of the detection results, the traceability of the actual immunization detection time, and the completeness of the immunization endpoint observation. When post-vaccination intervention itself is part of the actual usable effect of vaccination, the actual usable training weights are not reduced separately.

[0070] When constructing the machine learning model, the model input is the predicted input feature vector, and the model output includes the predicted value of the uninterventional effect and the predicted value of the actual usable effect. The machine learning model adopts a supervised learning model, which includes at least one of gradient boosting tree, random forest, support vector regression, neural network and graph model. The model type is used to provide the prediction function and is not a distinguishing feature of this invention. The difference in this step is that the model training labels, training sample weights and difference constraints are all derived from the standardized observation results of vaccination events, response-triggered intervention trajectories, response role recognition results and intervention label contamination results formed in S1 to S3.

[0071] To simultaneously learn the effects of the uninterventional adjuvant and the effects of the adjuvant actually available for vaccination, and to constrain the difference between the two to maintain a correlation with the degree of interventional label contamination, the following two-label task loss function constrained by label credibility is adopted: ; in, This represents the model training loss, which is the output of this formula. It is used to train the machine learning model and constrain it to learn both the effect of the uninterrupted adjuvant and the effect of the adjuvant that is actually available after vaccination. This represents the training sample index, which refers to a training sample object determined by a single vaccination event, the immune detection time window, and the immune endpoint type. This represents the set of training samples with labels predicting the effect of no intervention. The set is derived from training sample objects that have already formed a characterization of the effect of no intervention adjuvant. This represents the set of training samples with predictive labels that have practically usable effects. The set is derived from training sample objects that have effective immune detection endpoint change results after homogenization and uniform scaling. This represents the set of training samples that simultaneously possess both the label for predicting the effect without intervention and the label for predicting the actual usable effect. The source of this set is the training sample objects that simultaneously enter both the task of predicting the effect without intervention and the task of predicting the actual usable effect. Indicates the training sample index The corresponding uninterrupted training weights are obtained by converting the uninterrupted label confidence results output by S3, and the values ​​range from 0 to 1. Indicates the training sample index The corresponding actual usable training weights are obtained by converting the actual usable label confidence results output by S3, and the value range is from 0 to 1. Indicates the training sample index The corresponding predictive label for the effect without intervention is formed by characterizing the effect of the adjuvant without intervention; Indicates the training sample index The corresponding predicted value of the uninterrupted effect is the output of the machine learning model; Indicates the training sample index The corresponding actual usable effect prediction label is formed from the changes in the effective immune detection endpoint in the characterization of the actual usable adjuvant effect of vaccination; Indicates the training sample index The corresponding predicted actual usable effect is the output of the machine learning model; Indicates the training sample index The corresponding level of intervention label contamination is obtained from the intervention label contamination results output by S3, with a value ranging from 0 to 1; this is used in the difference constraint term. As a constraint weight, it is used to reduce the contribution of the difference constraint when the confidence of either the uninterrupted effect prediction label or the actual usable effect prediction label is low, so as to avoid the low confidence label forming an excessive constraint on the model output difference. This represents the loss weight for the actual usable effect prediction task, used to adjust the contribution of the actual usable effect prediction task to the model training loss. The loss weights representing the dual-label difference constraint are used to adjust the contribution of the upper bound constraint on the difference between the predicted value of the uninterrupted effect and the actual usable effect in the model training loss. This indicates that a non-negative truncation is applied to the difference within the parentheses, so that no difference constraint penalty is generated if the absolute amount of the difference between the predicted value of the uninterventional effect and the actual usable effect prediction does not exceed the level of intervention label contamination.

[0072] In this formula, the summation range of the first term is the training sample set with the uninterventional effect prediction label, the summation range of the second term is the training sample set with the actual usable effect prediction label, and the summation range of the third term is the training sample set with both the uninterventional effect prediction label and the actual usable effect prediction label. When any training sample set is empty, the corresponding loss term is 0, and the remaining loss terms continue to participate in the model training loss calculation. The uninterventional effect prediction label, the actual usable effect prediction label, the uninterventional effect prediction value, and the actual usable effect prediction value are all within the 0 to 1 evaluation scale after homogenization and unified scaling. The degree of intervention label contamination is also a standardized result within the 0 to 1 range. Therefore, the absolute amount of difference in the third term is on the same evaluation scale as the degree of intervention label contamination. The training sample weights, the loss weights of the actual usable effect prediction task, and the loss weights of the dual-label difference constraint are all non-negative numbers.

[0073] The design logic of this formula is as follows: the first term is used to constrain the model to learn the effect of the adjuvant without intervention; the second term is used to constrain the model to learn the actual usable adjuvant effect after vaccination; and the third term is used to constrain the absolute amount of the difference between the predicted value of the effect without intervention and the predicted value of the actual usable effect, so that the absolute amount of the difference does not exceed the allowable difference range characterized by the degree of intervention label contamination. The third term uses the absolute amount of the difference and forms an upper bound constraint through non-negative truncation. It does not presuppose that post-vaccination intervention will necessarily lead to the actual usable effect being lower than the effect of the adjuvant without intervention, nor does it force that a high degree of intervention label contamination necessarily corresponds to a large prediction difference. This processing can adapt to real vaccination scenarios where the direction of the impact of salvage medication, delayed blood collection, supplementary testing, and withdrawal and censoring on the actual usable effect is inconsistent, and supports S4 output of adjuvant effect prediction results that distinguish between the effect of the adjuvant without intervention and the actual usable adjuvant effect after vaccination.

[0074] The loss weights for both the actual usable effect prediction task and the dual-label difference constraint are unitless weights, determined by the research protocol, development set validation results, cross-validation results, or data management specifications, and all values ​​are not less than 0. In one preferred implementation, the loss weight for the actual usable effect prediction task is between 0.5 and 2; the loss weight for the dual-label difference constraint is between 0.1 and 1. In scenarios where the prediction of the actual usable adjuvant effect after vaccination is the main output target, the loss weight for the actual usable effect prediction task is taken at the upper limit of the corresponding range. In scenarios where the source of label contamination is complex and the correspondence between the dual-label difference and the degree of intervention label contamination is unstable, the loss weight for the dual-label difference constraint is taken at the lower limit of the corresponding range. When historical samples are insufficient to stably determine the loss weights, a fixed weight configuration in the research protocol or data management specifications is adopted and kept fixed during the model training phase. This treatment is used to avoid temporarily changing the loss weights during the training phase due to insufficient sample size, ensuring that the training contribution control has a fixed basis.

[0075] After the model training is completed, for the target vaccination event samples, the same process from S1 to S3 is used to generate standardized observation results at the vaccination event level, response-triggered intervention trajectory, response role identification results, intervention label contamination results, dual adjuvant effect characterization constrained by label contamination, non-intervention label credibility results, and actual usable label credibility results. A prediction input feature vector is constructed and input into the trained machine learning model to output the non-intervention effect prediction value and the actual usable effect prediction value, forming the adjuvant effect prediction result. The adjuvant effect prediction result includes the prediction result that distinguishes between the non-intervention adjuvant effect and the actual usable adjuvant effect.

[0076] Furthermore, prediction confidence results are generated based on the label confidence results and model uncertainty results for the corresponding prediction task. The label confidence results for the corresponding prediction task include the confidence results of the non-intervention label and the confidence results of the actually usable label. For the predicted value of the non-intervention effect, the corresponding prediction confidence is determined by the non-intervention label confidence result and the model stability result. For the predicted value of the actually usable effect, the corresponding prediction confidence is determined by the confidence results of the actually usable label and the model stability result. The model uncertainty result represents the stability of the machine learning model output. The model uncertainty result is obtained by any one of the cross-validation error, the model ensemble prediction variance, and the calibration set residual normalization result, and is uniformly converted into an uncertainty standardization result in the range of 0 to 1. The model stability result is taken as 1 minus the uncertainty standardization result. The prediction confidence result takes a value range of 0 to 1. The higher the value, the higher the confidence of the adjuvant effect prediction result in both label confidence and model output stability. The prediction confidence result is used to illustrate the data confidence of the machine learning prediction results and is not used as a medical safety conclusion.

[0077] For target population stratification, the population is divided based on the subjects' basic health information. The predicted values ​​of the non-intervention effect and the actual usable effect of each target vaccination event sample within the target population stratum are weighted and summarized according to the prediction confidence results. This generates the non-intervention effect group result and the actual usable effect group result for the target population stratum. The input for the weighted summarization is the adjuvant effect prediction result and the prediction confidence result within the same target population stratum. The processing action is to use the prediction confidence result as the sample weight and perform a weighted average on the predicted values ​​of the non-intervention effect and the actual usable effect, respectively. The output is the group prediction result corresponding to the target population stratum. The target population stratification is derived from the subjects' basic health information in S1. The stratification criteria include at least one of the following: age group, underlying disease status, immunosuppressive drug status, previous vaccination history, and previous infection history. When there are no target vaccination event samples with a prediction confidence result greater than 0 within the target population stratum, no group prediction result is output for that target population stratum, and a stratification sample insufficiency marker is generated.

[0078] Subsequently, population fit results are generated based on the results of the non-intervention effect group, the results of the actual usable effect group, the effect judgment threshold, and the fit difference threshold. The effect judgment threshold is used to determine whether the non-intervention adjuvant effect in the target population stratification meets the target effect requirements set in the study protocol. The fit difference threshold is used to determine whether the absolute amount of difference between the results of the actual usable effect group and the results of the non-intervention effect group reaches a level that requires notification. Both the effect judgment threshold and the fit difference threshold are thresholds on an evaluation scale of 0 to 1, derived from at least one of the historical validation set, the study protocol, and the data management specifications, and remain fixed during the model operation phase. The effect judgment threshold is set to 0.5 to 0.8, and the fit difference threshold is set to 0.1 to 0.3. When historical samples are insufficient, the fixed threshold configuration in the study protocol or the data management specifications is adopted.

[0079] When the results of the non-intervention group reach the effect judgment threshold, and the absolute amount of the difference between the actual usable effect group results and the non-intervention group results does not reach the fit difference threshold, the fit is satisfied result is output; when the results of the non-intervention group reach the effect judgment threshold, and the absolute amount of the difference between the actual usable effect group results and the non-intervention group results reaches the fit difference threshold, the fit is pending confirmation result is output; when the results of the non-intervention group do not reach the effect judgment threshold, the fit is insufficient result is output. The population fit result is used to characterize the applicability of the target adjuvant in different baseline health states, different post-vaccination response sensitivities, and different post-vaccination intervention trigger risk populations.

[0080] Finally, S4 outputs the adjuvant effect prediction results, prediction confidence results, non-intervention label confidence results, actual usability label confidence results, and population suitability results. The adjuvant effect prediction results include the predicted values ​​of the non-intervention effect and the actual usability effect. The prediction confidence results are used to indicate the data confidence of the adjuvant effect prediction results. The non-intervention label confidence results are used to indicate the confidence level of the non-intervention effect prediction label after being affected by post-vaccination intervention and endpoint confidence grading. The actual usability label confidence results are used to indicate the confidence level of the actual usability effect prediction label in terms of the validity of test results, the traceability of actual immunization testing time, and the completeness of immunization endpoint observation. The population suitability results are used to characterize the applicability of the target adjuvant in the target population stratification. This output is used to provide data processing results for vaccine adjuvant effect prediction and population stratification evaluation, and is not directly used as a clinical diagnostic conclusion or medical safety conclusion.

[0081] Example 2

[0082] like Figure 2 As shown, this invention discloses a vaccine adjuvant effect prediction system based on machine learning, comprising: The standardized observation construction module acquires medical observation data of vaccination events. Using a time-series merging method anchored to vaccination events, it limits the post-vaccination event chain formed by post-vaccination response and post-vaccination intervention to the same single vaccination event and the same relative time axis of the immune detection endpoint, generating standardized observation results at the vaccination event level. The reaction tracing and identification module performs reaction trigger tracing for the post-vaccination event chain, performs transmission and analysis on standardized observation results at the vaccination event level, establishes a traceable trigger relationship between post-vaccination reactions and post-vaccination interventions, marks the data role of post-vaccination reactions in the trigger relationship, and generates reaction-triggered intervention trajectories and reaction role identification results. The label contamination reconstruction module analyzes the process by which the response-triggered intervention trajectory and response role identification results affect the immune detection endpoint through a label contamination reconstruction approach oriented towards the immune detection endpoint, thereby transforming the immune detection endpoint affected by the intervention into a dual adjuvant effect characterization constrained by label contamination. The prediction output module uses a dual-label machine learning training method constrained by label credibility to predict the adjuvant effect representation constrained by label contamination. This enables the model to output adjuvant effect prediction results that distinguish between the effect of the uninterrupted adjuvant and the effect of the adjuvant that can be used in actual vaccination, and to form corresponding population suitability results.

[0083] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A machine learning-based method for predicting the effectiveness of vaccine adjuvants, characterized in that, include: Acquire medical observation data of vaccination events, and use a time-series merging method anchored by vaccination events to limit the post-vaccination event chain formed by post-vaccination response and post-vaccination intervention to the same single vaccination event and the same relative time axis of the immune detection endpoint, thereby generating standardized observation results at the vaccination event level. The system performs response trigger tracing for post-vaccination event chains, performs transmission analysis on standardized observation results at the vaccination event level, establishes a traceable trigger relationship between post-vaccination response and post-vaccination intervention, marks the data role of post-vaccination response in the trigger relationship, and generates response trigger-type intervention trajectories and response role identification results. By reconstructing the label contamination around the immune detection endpoint, the process by which the response-triggered intervention trajectory and response role identification results affect the immune detection endpoint is analyzed, so that the immune detection endpoint affected by the intervention is transformed into a dual adjuvant effect characterization constrained by label contamination. By using a dual-label machine learning training method constrained by label credibility, the model predicts the adjuvant effect representation constrained by label contamination. This enables the model to output adjuvant effect predictions that distinguish between the effect of uninterrupted adjuvants and the effect of adjuvants that are actually available after vaccination, and to generate corresponding population suitability results.

2. The method for predicting vaccine adjuvant efficacy based on machine learning according to claim 1, characterized in that, The medical observation data for vaccination events include vaccination identification information used to determine individual vaccination events, post-vaccination event chain records, immune test endpoint records, and basic health information of the subjects; The timing merging method for anchoring vaccination events includes: first, individualizing vaccination identification information according to subject identification; then, dividing the collected vaccination identification information into candidate vaccination actions according to the number of doses and vaccination time; for each candidate vaccination action, performing vaccine-adjuvant combination consistency verification using vaccine identification, adjuvant identification, and vaccination batch, and confirming candidate vaccination actions that pass the vaccine-adjuvant combination consistency verification as the same single vaccination event; when the same subject has at least one of different vaccine identification or different adjuvant identification at the same vaccination time, establishing single vaccination events according to the vaccine-adjuvant combination; when either the vaccination batch or the vaccination time is missing, generating a vaccination boundary confirmation mark, and restricting the binding range of the post-vaccination event chain and the immune detection endpoint to be uniformly processed according to the vaccination relative time axis based on the vaccination boundary confirmation mark.

3. The method for predicting vaccine adjuvant efficacy based on machine learning according to claim 2, characterized in that, Standardized observation results at the vaccination event level are formed by object-level binding of post-vaccination event chain records and immunization endpoint records; Object-level binding includes: converting post-vaccination reaction records into standardized reaction records according to reaction type, reaction occurrence time, reaction duration, and reaction level; converting post-vaccination intervention records into standardized intervention records according to intervention type, intervention occurrence time, intervention duration, and intervention execution reason; converting immunization endpoint records into standardized immunization endpoint records according to planned immunization testing time window, actual immunization testing time, and immunization endpoint observation value; and establishing a relative time correspondence between standardized reaction records, standardized intervention records, and standardized immunization endpoint records within the same single vaccination event. When at least one of the post-vaccination event chain and immunization endpoint lacks a necessary field for subsequent processing, a data integrity flag is generated. The data integrity flag is used to limit the range of available fields in reaction trigger tracing and label contamination reconstruction.

4. The method for predicting vaccine adjuvant efficacy based on machine learning according to claim 1, characterized in that, The response trigger tracing method includes: within the standardized observation results at the vaccination event level, verifying the time sequence of post-vaccination response and post-vaccination intervention, and designating record pairs whose response time is no later than the intervention time as candidate response-intervention pairs; subsequently, performing trigger observation window verification, intervention cause consistency verification, and response intensity support verification on the candidate response-intervention pairs. The trigger observation window verification is used to determine whether the interval between the response time and the intervention time is within the observation window of the corresponding intervention type. The intervention cause consistency verification is used to determine whether the intervention execution cause falls within the intervention cause mapping range of the corresponding response type. The response intensity support verification is used to determine whether the support results of the response level and response duration after uniform scaling meet the trigger support conditions. The trigger support conditions are derived from at least one of the clinical trial protocol, follow-up rules, and historical vaccination sample verification results. Candidate response-intervention pairs that have completed the trigger observation window verification, intervention cause consistency verification, and response intensity support verification are identified as response trigger relationships, and response trigger type intervention trajectories are formed from the response trigger relationships.

5. The method for predicting vaccine adjuvant efficacy based on machine learning according to claim 4, characterized in that, The reaction role recognition results are generated by comparing similar reference events. The comparison method for similar reference events includes: selecting vaccination events with the same vaccine adjuvant combination, the same response type, and the same baseline health stratification of the subjects as the current single vaccination event, and which have not formed a response-triggered intervention trajectory, as the set of similar reference events; comparing the change in immune detection endpoints after homogenization and uniform scaling of the current post-vaccination response with the median change in immune detection endpoints in the set of similar reference events, and assigning data roles to the current post-vaccination response based on the response triggering relationship; and selecting post-vaccination responses that do not form a response triggering relationship and whose changes in immune detection endpoints are not lower than those in the set of similar reference events. When the median change in the detection endpoint and the reaction occurs within the immune response observation window set in the study protocol, it is marked as an immune concomitant reaction; when the current post-vaccination reaction forms a reaction triggering relationship, it is marked as an intervention-triggered reaction; when the current post-vaccination reaction does not form a reaction triggering relationship, the change in the immune detection endpoint is lower than the median change in the immune detection endpoint in the same set of reference events, and the reaction occurs outside the immune response observation window set in the study protocol, it is marked as an irrelevant background reaction; when the current post-vaccination reaction does not meet the formation conditions of immune concomitant reaction, intervention-triggered reaction, and irrelevant background reaction, it is marked as a role pending reaction.

6. The method for predicting vaccine adjuvant efficacy based on machine learning according to claim 1, characterized in that, The label contamination reconstruction method includes: using response-triggered intervention trajectories as the tracing object, screening post-vaccination interventions that occurred before the collection of immune test endpoints from the response-triggered intervention trajectories to form response-triggered endpoint pre-interventions; performing three types of rewriting identification on the process of response-triggered endpoint pre-interventions acting on immune test endpoints, wherein, endpoint numerical rewriting identification generates numerical rewriting support results according to the preset influence mapping relationship between intervention type, intervention occurrence time, immune endpoint type and immune test time window; endpoint collection time rewriting identification generates collection time rewriting support results according to the deviation relationship between the actual immune test time and the planned immune test time window; endpoint observability rewriting identification generates observability rewriting support results according to the validity of test results, supplementary testing substitution status and pre-endpoint withdrawal status; the preset influence mapping relationship comes from at least one of the following: clinical trial protocol, case report form, medication restriction rules, historical sample validation results and pre-established data dictionary.

7. The method for predicting vaccine adjuvant efficacy based on machine learning according to claim 6, characterized in that, The characterization of dual adjuvant effects constrained by label contamination was generated through endpoint confidence grading and reference correction mechanisms; The endpoint confidence grading and reference correction mechanism includes: uniformly converting the numerical rewriting support results, the collection time rewriting support results, and the observability rewriting support results into standardized support results in the range of 0 to 1, and obtaining the degree of intervention label contamination based on the three standardized support results; classifying the immune detection endpoints into uncontaminated reliable endpoints, correctable contamination endpoints, and endpoints that cannot be directly labeled according to the degree of intervention label contamination. The changes in immune detection endpoints after homogenization and uniform scaling of uncontaminated credible endpoints serve as a direct source for characterizing the effect of non-interventional adjuvants. A set of similar non-interventional reference endpoints is established for correctable contamination endpoints, and reference correction is performed between the changes in the homogenized and uniformly scaled immune detection endpoints corresponding to the current correctable contamination endpoints and the median changes in the set of similar non-interventional reference endpoints, according to the degree of contamination of the intervention label, forming a corrected source for characterizing the effect of non-interventional adjuvants. The generation of prediction labels for non-interventional effects is stopped when endpoints cannot be directly labeled, and the structured characterization of the actual adjuvant effect available for vaccination is retained.

8. The method for predicting vaccine adjuvant efficacy based on machine learning according to claim 1, characterized in that, The dual-label machine learning training method includes: determining the training sample objects based on a single vaccination event, the immune detection time window, and the type of immune endpoint; using the uninterventional adjuvant effect representation in the dual adjuvant effect representation constrained by label contamination as the uninterventional effect prediction label, and using the immune detection endpoint change results after homogenization and uniform scaling in the dual adjuvant effect representation constrained by label contamination as the actual usable effect prediction label; obtaining the uninterventional label credibility result based on the degree of interventional label contamination, and obtaining the actual usable label credibility result based on the validity of the detection results, the traceability of the actual immune detection time, and the completeness of the immune endpoint observation; using the uninterventional label credibility result and the actual usable label credibility result as the training sample weights for the uninterventional effect prediction task and the actual usable effect prediction task, respectively; both the uninterventional effect prediction label and the actual usable effect prediction label are adjuvant effect representations after homogenization and uniform scaling, and the original antibody titer, neutralizing antibody concentration, cytokine concentration, and T cell response intensity are not included in the model training loss calculation before homogenization and uniform scaling are completed; training sample objects that cannot be directly labeled with endpoints are not included in the supervised training sample set for the uninterventional effect prediction task.

9. The method for predicting vaccine adjuvant efficacy based on machine learning according to claim 8, characterized in that, The label credibility constraints include: during the machine learning model training process, adjusting the contribution of the unintervention effect prediction label to the model loss through the unintervention label credibility result, adjusting the contribution of the actual usable effect prediction label to the model loss through the actual usable label credibility result, and constraining the absolute amount of difference between the unintervention effect prediction value and the actual usable effect prediction value through the degree of intervention label contamination, so that the machine learning model outputs adjuvant effect prediction results that distinguish between the unintervention adjuvant effect and the actual usable adjuvant effect; subsequently, the adjuvant effect prediction results are weighted according to the baseline health stratification of the subjects to form the population suitability results; wherein, the prediction credibility is jointly determined by the label credibility result of the corresponding prediction task and the model uncertainty result. The label credibility result of the corresponding prediction task includes the unintervention label credibility result and the actual usable label credibility result, and the model uncertainty result is obtained through any one of the cross-validation error, the model ensemble prediction variance, and the calibration set residual normalization result.

10. A machine learning-based vaccine adjuvant efficacy prediction system, employing the machine learning-based vaccine adjuvant efficacy prediction method as described in any one of claims 1-9, characterized in that, include: The standardized observation construction module acquires medical observation data of vaccination events. Using a time-series merging method anchored to vaccination events, it limits the post-vaccination event chain formed by post-vaccination response and post-vaccination intervention to the same single vaccination event and the same relative time axis of the immune detection endpoint, generating standardized observation results at the vaccination event level. The reaction tracing and identification module performs reaction trigger tracing for the post-vaccination event chain, performs transmission and analysis on standardized observation results at the vaccination event level, establishes a traceable trigger relationship between post-vaccination reactions and post-vaccination interventions, marks the data role of post-vaccination reactions in the trigger relationship, and generates reaction-triggered intervention trajectories and reaction role identification results. The label contamination reconstruction module analyzes the process by which the response-triggered intervention trajectory and response role identification results affect the immune detection endpoint through a label contamination reconstruction approach oriented towards the immune detection endpoint, thereby transforming the immune detection endpoint affected by the intervention into a dual adjuvant effect characterization constrained by label contamination. The prediction output module uses a dual-label machine learning training method constrained by label credibility to predict the adjuvant effect representation constrained by label contamination. This enables the model to output adjuvant effect prediction results that distinguish between the effect of the uninterrupted adjuvant and the effect of the adjuvant that can be used in actual vaccination, and to form corresponding population suitability results.

Images