Ensemble learning mediated tumor drug resistance risk early warning method

Abstract

This invention discloses an ensemble learning-mediated method for early warning of tumor drug resistance risks, addressing the problems of delayed detection of tumor drug resistance, patient injury from frequent imaging examinations, and high false positive rates in blood-based early warning systems. This invention constructs a tumor drug resistance time-series blood pairing dataset through non-invasive, time-series blood sampling using indwelling intravenous catheters anchored to the drug administration cycle. A core causal feature set of drug resistance is selected using an ensemble learning-mediated dual-constraint framework of temporal causality and drug resistance mechanisms. A two-layer heterogeneous ensemble learning early warning model embedded with medical knowledge graph constraints is then built, outputting individualized drug resistance risk results and matching them with evidence-based treatment plans, while simultaneously performing incremental model iterations. This invention significantly advances the drug resistance prediction window, reduces patient examination damage, improves early warning accuracy, and can provide support for the development of drugs to reverse drug resistance.

Description

Technical Field

[0001] This application relates to the field of methods for early warning of tumor drug resistance risks, and in particular to an integrated learning-mediated method for early warning of tumor drug resistance risks. Background Technology

[0002] Malignant tumors are a major public health problem that seriously threatens human health. Targeted therapy, immunotherapy, and chemotherapy are the core treatment methods for patients with intermediate and advanced malignant tumors. However, acquired resistance to anti-tumor drugs is the primary factor leading to treatment failure and shortened patient survival. Clinical practice has proven that timely intervention in the early stages of drug resistance can significantly delay disease progression and improve treatment outcomes. Therefore, developing non-invasive, routinely applicable technologies that can predict the risk of drug resistance in tumors is a core clinical need in the field of precision oncology.

[0003] Currently, routine clinical methods for monitoring drug resistance primarily rely on periodic imaging examinations combined with single blood tumor marker tests. Imaging examinations can only identify the progression of existing morphological lesions, resulting in a significant lag in drug resistance detection. Furthermore, frequent examinations can exacerbate the burden on patients due to ionizing radiation and contrast agent damage. Single blood tests only provide static results, failing to capture the dynamic evolution of drug resistance and are susceptible to confounding factors such as patient inflammation, concurrent medications, and underlying diseases, leading to a high false-positive rate. Existing AI-based tumor drug resistance early warning technologies mostly employ conventional correlation modeling methods, failing to explore the causal relationship between detection features and drug resistance development. These models exhibit poor generalization and insufficient clinical interpretability, and do not achieve precise matching between sampling and drug administration cycles, thus failing to effectively capture weak biological signals in the early stages of drug resistance and failing to meet the practical needs of precise clinical early warning.

[0004] The aforementioned shortcomings of existing technologies not only lead to a general lag in the detection of tumor drug resistance in clinical practice, causing patients to miss the optimal intervention time, but also prevent the clinical data generated by current monitoring models from forming a standardized, causally annotated database for drug resistance research. This hinders the effective support for elucidating drug resistance mechanisms and developing drugs to reverse resistance, creating a data barrier between clinical diagnosis and drug development. Therefore, developing a tumor drug resistance risk early warning technology that combines non-invasive, routine monitoring with high accuracy and clinical interpretability, while also bridging the data gap between clinical practice and research, has become an urgent technical problem to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to overcome the core defects of existing tumor drug resistance monitoring technologies, such as delayed detection, large damage from invasive examinations, high false positive rate of early warning, and disconnect between clinical diagnosis and drug development data. This application provides an integrated learning-mediated early warning method for tumor drug resistance risk.

[0006] To achieve the above objectives, this application discloses an integrated learning-mediated early warning method for tumor drug resistance risk, including:

[0007] Patients with the target tumor type receiving anti-tumor drug therapy were included in the cohort. Peripheral venous blood samples were collected non-invasively from patients using an indwelling venous catheter that was synchronized with the timing of the anti-tumor drug administration cycle. Blood time-series detection indicators were obtained.

[0008] Collect patient antitumor drug regimens and drug resistance standard validation data to construct a time-series blood paired dataset of tumor drug resistance;

[0009] By integrating the learning-mediated module, an integrated mining framework with dual constraints of temporal causality and drug resistance mechanism is constructed. A causal inference algorithm adapted to the temporal characteristics of tumor drug resistance is integrated to correct the individual patient baseline and treatment-related confounding factors, analyze the temporal lag effect of feature changes and drug resistance occurrence, and screen out the core drug resistance causal feature set that has a direct causal driving effect on tumor drug resistance from the tumor drug resistance temporal blood paired dataset.

[0010] A two-layer heterogeneous ensemble learning model for drug resistance risk warning was constructed, which is embedded with prior constraints of a panoramic knowledge graph of tumor drug resistance. The lower layer of the two-layer heterogeneous ensemble learning model for drug resistance risk warning is a cluster of temporal heterogeneous base learning bases corresponding to the core drug resistance causal feature set, and the upper layer of the two-layer heterogeneous ensemble learning model for drug resistance risk warning is a dynamic weight fusion meta-learner constrained by evidence-based medicine.

[0011] The integrated learning-mediated module coordinates the learning of various temporal heterogeneous bases to complete the single-dimensional primary drug resistance risk prediction. The dynamic weight fusion meta-learner constrained by evidence-based medicine, combined with the prior constraints of the pan-macro knowledge graph of tumor drug resistance, dynamically adjusts the fusion weight of the output results of each temporal heterogeneous base learner according to the modal completeness of the input data, clinical treatment scenarios, and core drug resistance mechanisms of anti-tumor drugs, to obtain the integrated learning-mediated drug resistance risk warning model.

[0012] During the patient's anti-tumor treatment, real-time blood time-series detection indicators are collected through an indwelling intravenous needle according to a preset cycle that is synchronized with the drug administration cycle. These indicators are then input into an integrated learning-mediated drug resistance risk early warning model. The integrated learning-mediated drug resistance risk early warning model outputs the patient's individualized drug resistance risk level, core driving drug resistance characteristics, and predicted drug resistance event windows.

[0013] Based on the patient's individualized drug resistance risk level and core driver drug resistance characteristics, we matched evidence-based clinical treatment protocols that had a causal relationship with the core driver drug resistance characteristics.

[0014] When the individualized drug resistance classification is determined to be of high drug resistance risk, a targeted examination instruction is triggered that is causally linked to the core driving drug resistance characteristics in order to determine the occurrence of drug resistance in the patient.

[0015] The patient's periodic blood test data, drug resistance risk warning results, examination verification results, and treatment outcome data are fed back to the integrated learning-mediated drug resistance risk warning model through the integrated learning-mediated module, and incremental iterative optimization is carried out in conjunction with the panoramic knowledge graph of tumor drug resistance.

[0016] Preferably, peripheral venous blood samples are collected non-invasively from patients using an indwelling venous catheter that is synchronized with the timing of antitumor drug administration, including:

[0017] Based on the time synchronization and anchoring with the drug administration cycle of anti-tumor drugs, the time-series detection indicators of the collected blood were obtained to match the drug action cycle.

[0018] A basic blood collection cycle is set according to the patient's treatment stage. In the early stage of treatment, individualized indicator baselines are established through continuous blood collection, and a fixed blood collection cycle is set during the stable period of treatment to complete routine monitoring.

[0019] The blood collection cycle is dynamically adjusted based on the patient's anti-tumor drug administration cycle, changes in treatment stage, and real-time detection results of blood test indicators corresponding to the core drug resistance causal feature set.

[0020] A linkage triggering mechanism is constructed for blood collection operations, sample quality control, and model pre-analysis. When abnormal fluctuations occur in blood test indicators corresponding to the core drug resistance causal feature set, the blood collection cycle is encrypted, the whole process of sample quality control is carried out simultaneously, and the pre-analysis of the integrated learning-mediated drug resistance risk early warning model is initiated.

[0021] Based on the real-time changes in the patient's antitumor drug administration cycle, treatment stage, and core drug resistance causal characteristics, the blood collection cycle is dynamically adjusted to achieve full-cycle adaptation, including baseline establishment, routine testing, and encrypted tracking of abnormal signals, synchronized with the drug administration cycle sequence.

[0022] Preferably, a core set of causal features for drug resistance that has a direct causal driving effect on drug resistance is selected from the time-series blood paired dataset of tumor drug resistance, including:

[0023] Using blood time-series detection indicators from a tumor drug resistance time-series paired blood dataset as input features and drug resistance standard validation data as outcome labels, an integrated mining framework constrained by both time-series causality and drug resistance mechanisms is used to integrate a causal inference algorithm adapted to the time-series characteristics of tumor drug resistance. The causal relationship results related to tumor drug resistance output by the causal inference algorithm are consistently fused to generate a causal directed acyclic graph between input features and outcome labels, thus obtaining a preliminary causal feature set.

[0024] By adapting a causal inference algorithm to the temporal characteristics of tumor drug resistance, we correct for the interference of individual patient baseline differences and treatment-related confounding factors, quantify the average causal driving effect of each feature in the preliminary causal feature set on the drug resistance outcome, analyze the temporal lag effect window of feature changes leading to drug resistance, and eliminate false positive features without significant causal driving effects.

[0025] Based on the strength of causal driving effects, the significance of time-series lag effects, and the matching degree of biological mechanisms, a multi-dimensional causal consistency verification was completed, and a core set of causal features for drug resistance was obtained.

[0026] Preferably, the configuration of a temporal heterogeneous base learner cluster includes:

[0027] Based on the biological origin of the core drug resistance causal feature set and the corresponding driving mechanism type of tumor drug resistance, the core drug resistance causal feature set is divided into three categories: genomic clonal evolution time series data, immune microenvironment time series data, and systemic state biochemical time series data, and the specific tumor drug resistance biological mechanism corresponding to each type of data is clarified.

[0028] For genomic clonal evolution time series data, a time-based learner with embedded clonal evolution mechanism prior constraints is configured to learn drug resistance-related features that conform to the laws of tumor clonal evolution.

[0029] For time-series data of the immune microenvironment, a long short-term memory network base learner with embedded prior constraints of immune drug resistance mechanisms is configured to distinguish between drug resistance-driven immune signals and immune fluctuations caused by non-disease factors.

[0030] For whole-body biochemical time-series data, a gradient boosting tree-based learner with prior constraints embedded in the whole-body state-drug resistance association mechanism is configured to correct the interference of non-disease factors on drug resistance risk prediction through prior constraints.

[0031] Complete the mutual binding of each base learner with the corresponding time series data and the corresponding drug resistance biological mechanism, so that each base learner only processes the time series data of the corresponding category; each bound base learner independently completes deep learning of the time series features of the corresponding dimension, outputs the primary drug resistance risk prediction results of the corresponding dimension, and matches the corresponding drug resistance biological mechanism annotations to the prediction results to form primary prediction results with biological annotations.

[0032] Preferably, before training, the temporal heterogeneous base learner cluster undergoes unsupervised pre-training through drug-resistant clonal evolution-guided self-supervised contrastive learning, including:

[0033] An unlabeled pre-trained subset was split from the tumor drug resistance time-series blood pair dataset and aligned and standardized preprocessed to be synchronized with the drug administration cycle time series.

[0034] Based on the biological laws of tumor clonal evolution and drug resistance, we construct positive and negative sample pairs for comparative learning, and at the same time, we mine difficult sample pairs corresponding to the inflection point of clonal evolution.

[0035] Using the contrastive learning loss function as the core of optimization, and combining the drug resistance biology interpretability constraint loss, unsupervised pre-training is performed on learners of heterogeneous bases at different time intervals.

[0036] After pre-training is completed, the biological mechanism matching of the pre-trained features is verified, noisy features without clear biological significance of drug resistance are removed, and the weights of the underlying common features of each base learner are frozen.

[0037] Preferably, the fusion weights of the outputs of each temporally heterogeneous base learner are dynamically adjusted, including:

[0038] By integrating the learning-mediated module, the modal completeness of the input data, the number of patient treatment lines, the type of anti-tumor drugs, and the tumor stage are identified to determine the priority of the core features of the current early warning scenario.

[0039] Based on the prior medical constraints of the panoramic knowledge graph of tumor drug resistance, the initial fusion weights of the primary prediction results of each temporal heterogeneous base learner are calculated through the attention mechanism, and the minimum basic weight thresholds are set for the base learners corresponding to the core drug resistance biomarkers specified in clinical guidelines.

[0040] Based on the core drug resistance mechanism of anti-tumor drugs, the drug-specific adaptation of fusion weights is completed, the dominant weights are assigned to the base learners to which the corresponding core drug resistance features belong, and the evidence-based medicine evidence level of the fusion weight adjustment results is verified.

[0041] Based on the final fusion weights after adaptation, the initial prediction results of each temporal heterogeneous base learner are weighted and fused to output the final comprehensive prediction result of drug resistance risk.

[0042] Preferred evidence-based clinical treatment protocols that are causally associated with the core driver of drug resistance include:

[0043] Based on the drug resistance probability, core driving drug resistance characteristics, and drug resistance occurrence time window output by the converged ensemble learning-mediated drug resistance risk warning model, the individualized drug resistance risk level is divided into three levels: low risk, medium risk, and high risk according to evidence-based medicine standards.

[0044] For low-risk patients, maintain routine monitoring synchronized with the medication cycle and cancel imaging examinations;

[0045] For patients at medium risk, increase the frequency of blood sampling and monitoring, conduct basic compatibility tests that are radiation-free and non-invasive, and track the changing trends of indicators corresponding to the core drivers of drug resistance.

[0046] For high-risk patients, the clinical review of the early warning results is completed within a preset time limit. Based on the causal association matching mechanism between the core driving drug resistance characteristics and the examination items, the targeted examination items are customized for the patients, including the examination of the target points and lesion sites corresponding to the driving characteristics, in order to confirm the drug resistance status.

[0047] Preferably, incremental iterative optimization of linking the tumor drug resistance panoramic knowledge graph includes:

[0048] By integrating the learning-mediated module, the elastic weight consolidation algorithm is used to calculate the importance score of the core weights related to the common characteristics of drug resistance in the converged early warning model and apply protective restrictions.

[0049] We screened representative drug-resistant and sensitive samples covering multiple tumor types, drug types, and populations, and constructed a core sample memory bank with drug resistance biological annotations.

[0050] When new clinical data is input, the task-specific branch layer of each time-series heterogeneous base learner and the dynamic weight fusion meta-learner of evidence-based medicine constraints are adjusted, and the core samples in the memory bank are replayed simultaneously to complete joint training. At the same time, the tumor drug resistance panoramic knowledge graph is linked to complete the autonomous learning of new drug resistance mechanisms and the dynamic update of the core drug resistance causal feature set.

[0051] After each incremental training is completed, model performance is validated and clinical compliance is verified through an independent validation set, and version updates and full-process traceability records are maintained.

[0052] Preferably, the integration learning-mediated module feeds back to the early warning model, completing incremental iterative optimization of compliance in conjunction with the tumor drug resistance panoramic knowledge graph, including:

[0053] By integrating learning-mediated modules, novel blood biomarkers strongly associated with tumor drug resistance are discovered, and the corresponding drug resistance driving mechanisms and potential drug targets are analyzed.

[0054] High-throughput virtual screening and drug repositioning analysis of drugs for drug resistance reversal were completed based on the core drug resistance causal feature set.

[0055] By using biomarkers corresponding to the core drug resistance causal feature set, we can enrich the subjects in clinical trials of anti-tumor drugs and construct a prospective stratification tool for predicting drug resistance risk in clinical trial subjects.

[0056] Using the dynamic changes of blood biomarkers corresponding to the core drug resistance causal feature set as a surrogate efficacy endpoint in clinical trials of anti-tumor drugs, a drug resistance biological model is constructed based on the core characteristics of drug-resistant patients, providing real-world data support for the in vitro validation of drug resistance reversal drugs.

[0057] Preferably, it also includes algorithmic fairness optimization for performance balancing across the entire population mediated by ensemble learning, including:

[0058] Construct a set of two-dimensional sensitive attributes specifically for early warning of tumor drug resistance, and complete the causal irrelevance verification of all sensitive attributes;

[0059] A hierarchical fairness optimization architecture matching the two-layer heterogeneous model is constructed, consisting of a causal-fair joint constraint unit at the base learner level and a fairness attention fusion unit at the meta learner level.

[0060] During the training of the two-layer heterogeneous model, a causal-fair joint loss function corresponding to the feature dimension is constructed through a causal-fair joint constraint unit to eliminate spurious causal associations between sensitive attributes and drug resistance causal features, and to eliminate the feature dependence of the base learner on sensitive attributes.

[0061] The weight allocation mechanism of the meta-learner is optimized by using a fair attention fusion unit, and the lower limit of the basic weight of the corresponding base learner is increased for subgroups with insufficient sample size.

[0062] For drug-resistant small sample populations, a cross-scenario feature enhancement fairness adaptation branch is constructed to complete cross-subgroup feature transfer enhancement and generate a dedicated lightweight early warning branch.

[0063] Establish a continuous fairness verification and update mechanism, complete the performance balance verification of all subgroups before each incremental update, and optimize the fairness constraint parameters simultaneously during the incremental update process;

[0064] After the two-layer heterogeneous model is trained and updated, it is validated through a clinically oriented, multi-dimensional fairness verification system, enabling different subgroups of people to receive drug resistance warnings.

[0065] Compared with existing technologies, this invention provides an integrated learning-mediated early warning method for tumor drug resistance risk, which has the following beneficial effects:

[0066] 1. By using a non-invasive, time-sequential blood collection system with an indwelling venous catheter that is synchronized with the dosing cycle of anti-tumor drugs, this system replaces traditional repeated venous punctures and fixed-cycle imaging examinations. This significantly reduces the pain and radiation-related physical damage to patients, as well as contrast agent-related injuries. At the same time, it advances the prediction window for anti-tumor drug resistance from the stage of imaging-based diagnosis of progression to the early molecular biological signal stage, thus achieving drug resistance early warning. Combined with a risk-based, graded, and precise treatment plan, this system avoids the waste of medical resources and the physical burden on patients caused by indiscriminate comprehensive examinations. It also provides patients with the best intervention opportunity, which can significantly prolong the progression-free survival of patients and comprehensively improve the management quality and patient treatment experience throughout the entire anti-tumor treatment cycle.

[0067] 2. Through an integrated mining framework constrained by both temporal causality and drug resistance mechanisms, interference from confounding factors such as inflammation, concomitant medication, and underlying diseases is accurately eliminated, and the causal feature set that truly drives drug resistance is obtained. Combined with a two-layer heterogeneous integrated learning model that embeds prior constraints of a panoramic knowledge graph of tumor drug resistance, individualized adaptation to different types of anti-tumor drugs and different tumor scenarios is achieved. At the same time, the hierarchical fairness optimization mechanism solves the model performance bias problem for rare tumor types and small sample populations, and achieves stable and accurate early warning for the entire population.

[0068] 3. By establishing a comprehensive database of tumor drug resistance hematological characteristics with causal annotations, the barriers between clinical diagnosis and treatment and anti-tumor drug development are broken down. This database provides full-dimensional real-world data support for the discovery of novel drug resistance biomarkers, the analysis of drug resistance driving mechanisms, and the discovery of drug targets for drug resistance reversal. At the same time, it can achieve precise enrichment of subjects in anti-tumor drug clinical trials through core drug resistance causal characteristics, and shorten the clinical trial cycle by using dynamic changes in hematological biomarkers as alternative efficacy endpoints. This significantly reduces the R&D costs and failure risks of anti-tumor drugs, and enables the dual empowerment of precision diagnosis and treatment of tumors and the development of drug resistance reversal drugs. Attached Figure Description

[0069] Figure 1 This is a flowchart illustrating the steps of the integrated learning-mediated early warning method for tumor drug resistance risk in this application embodiment.

[0070] Figure 2 This is a flowchart illustrating the steps of the integrated learning-mediated tumor drug resistance risk warning method in this application, which involves non-invasively collecting peripheral venous blood samples from patients using an indwelling venous catheter that is synchronously anchored to the timing of the anti-tumor drug administration cycle.

[0071] Figure 3 This is a flowchart illustrating the steps of the temporal heterogeneous base learner cluster in the integrated learning-mediated tumor drug resistance risk warning method of this application, which completes unsupervised pre-training through drug resistance clonal evolution-guided self-supervised comparative learning before training.

[0072] Figure 4 This is a flowchart illustrating the steps of dynamically adjusting the fusion weights of the output results of each temporal heterogeneous base learner in the integrated learning-mediated tumor drug resistance risk warning method according to the embodiments of this application. Detailed Implementation

[0073] The following is in conjunction with the appendix Figure 1-4 This application will be described in further detail.

[0074] This application discloses an integrated learning-mediated method for early warning of tumor drug resistance risk. (Refer to...) Figure 1 Integrated learning-mediated early warning methods for tumor drug resistance risk include:

[0075] S1. A cohort of patients receiving anti-tumor drug treatment for the target tumor type was included. Peripheral venous blood samples were non-invasively collected from the patients using an indwelling venous catheter that was synchronized with the timing of the anti-tumor drug administration cycle. Blood timing detection indicators were then obtained.

[0076] S2. Collect patient antitumor drug regimens and drug resistance standard verification data to construct a time-series blood paired dataset of tumor drug resistance.

[0077] S3. Through the integrated learning-mediated module, an integrated mining framework with dual constraints of temporal causality and drug resistance mechanism is constructed. A causal inference algorithm adapted to the temporal characteristics of tumor drug resistance is integrated to correct the individual patient baseline and treatment-related confounding factors. The temporal lag effect of feature changes and drug resistance occurrence is analyzed. The core drug resistance causal feature set with a direct causal driving effect on tumor drug resistance is obtained from the tumor drug resistance temporal blood paired dataset.

[0078] S4. Construct a two-layer heterogeneous integrated learning drug resistance risk warning model with prior constraints embedded in the panoramic knowledge graph of tumor drug resistance. The lower layer of the two-layer heterogeneous integrated learning drug resistance risk warning model is a temporal heterogeneous base learning base cluster corresponding to the core drug resistance causal feature set, and the upper layer of the two-layer heterogeneous integrated learning drug resistance risk warning model is a dynamic weight fusion meta-learner with evidence-based medicine constraints.

[0079] S5. By coordinating the various temporal heterogeneous base learners through the integrated learning-mediated module, a single-dimensional primary drug resistance risk prediction is completed. Through the dynamic weight fusion meta-learner constrained by evidence-based medicine, combined with the prior constraints of the pan-macro knowledge graph of tumor drug resistance, the fusion weight of the output results of each temporal heterogeneous base learner is dynamically adjusted according to the modal completeness of the input data, clinical treatment scenarios, and core drug resistance mechanisms of anti-tumor drugs, thus obtaining the integrated learning-mediated drug resistance risk early warning model.

[0080] S6. During the anti-tumor treatment of patients, real-time blood time-series detection indicators are collected through indwelling intravenous needles according to a preset cycle that is synchronized with the drug administration cycle. The data are input into the integrated learning-mediated drug resistance risk early warning model. The integrated learning-mediated drug resistance risk early warning model outputs the patient's individualized drug resistance risk level, core driving drug resistance characteristics, and drug resistance event window prediction results.

[0081] S7. Based on the patient's individualized drug resistance risk level and core driver drug resistance characteristics, match evidence-based clinical treatment protocols that have a causal relationship with the core driver drug resistance characteristics.

[0082] S8. When the individualized drug resistance classification level is determined to be of high drug resistance risk, a targeted examination instruction that is causally linked to the core driving drug resistance characteristics is triggered to determine the occurrence of drug resistance in the patient.

[0083] S9. The patient's periodic blood test data, drug resistance risk warning results, examination verification results, and treatment outcome data are fed back to the integrated learning-mediated drug resistance risk warning model through the integrated learning-mediated module, and incremental iterative optimization is carried out in conjunction with the tumor drug resistance panoramic knowledge graph.

[0084] Reference Figure 2 Specifically, the steps for non-invasively collecting peripheral venous blood samples from patients using an indwelling venous catheter that is synchronized with the timing of anti-tumor drug administration include:

[0085] S11. Based on the synchronization and anchoring with the drug administration cycle of anti-tumor drugs, the time-series detection indicators of the collected blood are accurately matched with the drug action cycle.

[0086] S12. Set a basic blood collection cycle according to the patient's treatment stage. In the early stage of the patient's treatment, establish an individualized baseline of indicators through continuous blood collection. In the stable period of the patient's treatment, set a fixed blood collection cycle to complete routine monitoring.

[0087] S13. Adjust the blood collection cycle dynamically based on the patient's anti-tumor drug administration cycle, changes in treatment stage, and real-time detection results of blood test indicators corresponding to the core drug resistance causal feature set;

[0088] S14. Construct a linkage triggering mechanism for blood collection operations, sample quality control, and model pre-analysis. When blood test indicators corresponding to the core drug resistance causal feature set show abnormal fluctuations, the blood collection cycle is encrypted, the whole process of sample quality control is carried out simultaneously, and the pre-analysis of the integrated learning-mediated drug resistance risk warning model is initiated.

[0089] S15. Based on the real-time changes in the patient's anti-tumor drug administration cycle, treatment stage, and core drug resistance causal characteristics, dynamically adjust the blood collection cycle to achieve full-cycle adaptation of baseline establishment, routine detection, and encrypted tracking of abnormal signals synchronized with the drug administration cycle sequence.

[0090] The key to this approach is to precisely match blood sampling timelines with the drug's dosing cycle, using synchronization with the drug's action cycle as a benchmark. This involves closely linking blood sampling timelines to the drug's pharmacokinetics, pharmacodynamics, and the development of drug resistance mechanisms. This ensures that the collected blood samples accurately reflect the drug's state of action and potential resistance signals within the patient's body. Specifically, blood can be collected within a fixed time window when the drug reaches its steady-state blood concentration, or at the trough concentration point before each dose, to accurately capture changes in biological response during the drug's action cycle.

[0091] Setting differentiated baseline blood collection cycles based on patients' treatment stages is crucial for balancing comprehensive data acquisition with patient adherence. In the early stages of treatment, patients' physiological states and drug responses fluctuate significantly; continuous blood collection allows for intensive data acquisition, enabling the precise construction of individualized baseline indicators and providing accurate reference standards for subsequent drug resistance risk assessment. During the stable treatment phase, patients' conditions are relatively stable, and the risk of drug resistance is low; therefore, longer, fixed blood collection cycles can be set, ensuring data continuity while reducing the operational burden on patients.

[0092] The core principle of dynamically adjusting blood collection cycles based on the patient's clinical status is to ensure a high degree of flexibility and adaptability in the blood collection strategy. When a patient's anti-tumor dosing regimen is adjusted, such as by changing the dosage, switching drugs, discontinuing or resuming medication, or when the treatment phase changes (e.g., from induction therapy to maintenance therapy), or when disease progression occurs, or when blood test indicators corresponding to the core drug resistance causal characteristic set show abnormal fluctuations, the blood collection cycle is dynamically adjusted accordingly to re-anchor the drug action cycle and promptly capture potential drug resistance signals.

[0093] The core of constructing a linked triggering mechanism for blood collection, quality control, and model pre-analysis is to establish a rapid response and quality assurance system for potential drug resistance risks. When the indicators corresponding to the core drug resistance causal feature set show abnormal fluctuations, this mechanism is immediately activated. On the one hand, it intensifies the blood collection cycle to obtain more intensive dynamic data; on the other hand, it simultaneously completes full-process quality control of samples to ensure the authenticity of abnormal signals and avoid misjudgments due to sample quality issues. At the same time, it initiates the pre-analysis of the early warning model to complete the preliminary assessment of potential risks using existing data, thus gaining valuable time for clinical decision-making.

[0094] This ensures that the acquired blood time-series detection indicators can accurately reflect the dynamic process of drug action and drug resistance, providing high-quality and timely input data for subsequent early warning models.

[0095] A specific example is as follows: For a lung cancer patient receiving targeted therapy in a 21-day cycle, blood was collected every 7 days during the first cycle of treatment, and three consecutive blood samples were collected to establish an individualized baseline for the patient. The blood collection time was fixed within 24 hours before each administration, precisely anchored to the administration cycle. After the patient entered the stable period, the blood collection cycle was adjusted to once per administration cycle. When the test results of the patient in the 5th treatment cycle showed a significant increase in the abundance of the EGFRT790M mutation, which is a core causal feature of drug resistance, the linkage triggering mechanism was immediately activated, and the blood collection cycle was increased to once every 7 days. Simultaneously, the entire process of sample quality control and model pre-analysis were completed, realizing the real-time capture of early drug resistance signals.

[0096] Furthermore, this embodiment further optimizes the selection of a core drug resistance causal feature set that has a direct causal driving effect on tumor drug resistance from the tumor drug resistance time-series blood paired dataset, including:

[0097] S31. Using blood time-series detection indicators from the tumor drug resistance time-series paired blood dataset as input features and drug resistance gold standard verification data as outcome labels, an integrated mining framework with dual constraints of time-series causality and drug resistance mechanism is used to integrate a causal inference algorithm adapted to the time-series characteristics of tumor drug resistance. The causal relationship results related to tumor drug resistance output by the causal inference algorithm are consistently fused to generate a causal directed acyclic graph between input features and outcome labels, thus obtaining a preliminary causal feature set.

[0098] S32. By adapting the causal inference algorithm to the time-series characteristics of tumor drug resistance, the algorithm corrects for the interference of individual patient baseline differences and treatment-related confounding factors, quantifies the average causal driving effect of each feature in the preliminary causal feature set on the drug resistance outcome, analyzes the time-series lag effect window of feature changes leading to drug resistance, and eliminates false positive features without significant causal driving effects.

[0099] S33. Based on the strength of causal driving effect, significance of time lag effect, and matching degree of biological mechanism, a multi-dimensional causal consistency verification is completed, and the core drug resistance causal feature set is obtained by screening.

[0100] The core of generating a preliminary causal feature set through multi-algorithm integration and consensus fusion is to address the issues of single causal inference algorithms being susceptible to data noise and lacking robustness. Specifically, multiple causal inference algorithms adapted to the temporal characteristics of tumor drug resistance can be run in parallel, including time-series causal algorithms based on Granger causality tests, PC algorithms based on structural causal models, and FCI algorithms. Consistency fusion is achieved by using majority voting or weighted consensus on the causal relationship results output by each algorithm. Only features identified as causally related to drug resistance outcomes in all algorithms are retained. A causal directed acyclic graph between input features and outcome labels is generated, eliminating non-correlated features with reverse causality or indirect correlation, resulting in a preliminary causal feature set, thus reducing the proportion of false positive features from the source.

[0101] The core of correcting for confounding factors and quantifying causal driving effects is to ensure that the association between the selected features and the drug resistance outcome is a genuine causal relationship, rather than a spurious association caused by confounding factors. Specifically, causal inference methods such as inverse probability weighting, propensity score matching, and dual robust estimation can be used to correct for interference from individual baseline and treatment-related confounding factors such as patient age, gender, underlying diseases, concomitant medications, and infection events. Subsequently, the average causal driving effect of each feature on the drug resistance outcome in the preliminary causal feature set is quantified. Simultaneously, methods such as time-series cross-correlation analysis and dynamic time warping are used to analyze the time lag window between feature changes and the occurrence of drug resistance, clarifying the time value of the feature as an early warning signal. Finally, false positive features with no statistically significant causal driving effect are eliminated.

[0102] Multidimensional causal consistency verification, at its core, involves finalizing the selection of features across three dimensions: statistical significance, temporal dimension, and biological rationality. This ensures that the core set of causal features for drug resistance not only possesses statistical significance but also clear clinical warning value and biological significance. Specifically, this involves assessing the contribution of features to drug resistance outcomes based on the strength of their causal driving effect, evaluating their early warning value based on the significance of time-lag effect windows, and verifying the biological mechanism matching degree based on a comprehensive knowledge graph of tumor drug resistance. This ensures that the features align with known biological mechanisms of tumor drug resistance. Ultimately, features that simultaneously meet the requirements of all three dimensions are selected, forming the final core set of causal features for drug resistance.

[0103] This implementation method effectively solves the problems of high false positive rate and lack of biological interpretability of features caused by traditional correlation analysis through a multi-stage and multi-dimensional refined screening strategy. A specific example is as follows: For a time-series blood-paired dataset of tumor drug resistance in colorectal cancer chemotherapy patients, blood time-series detection indicators are used as input features, and disease progression results defined by the RECIST standard are used as outcome labels. Three time-series causal inference algorithms are run in parallel, and the intersection of the results yields a preliminary causal feature set containing 28 features. After correcting for confounding factors such as patient age, liver and kidney function, and concomitant medications using an inverse probability weighting algorithm, the average causal driving effect of each feature is quantified. Analysis reveals that changes in 12 features lead the occurrence of drug resistance by 2-4 weeks, and the remaining features without significant causal effects are removed. Finally, the biological mechanism matching degree is verified through a tumor drug resistance panoramic knowledge graph, confirming that 8 features are highly matched with the core mechanisms of colorectal cancer chemotherapy resistance, such as DNA damage repair, epithelial-mesenchymal transition, or immune escape. This ultimately forms a core drug resistance causal feature set containing 8 features, providing high-quality input for subsequent model construction.

[0104] Furthermore, configure a temporal heterogeneous base learner cluster, including:

[0105] S41. Based on the biological origin of the core drug resistance causal feature set and the corresponding driving mechanism type of tumor drug resistance, the core drug resistance causal feature set is divided into three categories: genomic clonal evolution time series data, immune microenvironment time series data, and systemic state biochemical time series data, and the specific tumor drug resistance biological mechanism corresponding to each type of data is clarified.

[0106] S42. For genomic clonal evolution time series data, configure a time-based learner that embeds prior constraints of clonal evolution mechanism to learn drug resistance-related features that conform to the laws of tumor clonal evolution.

[0107] S43. For time-series data of the immune microenvironment, configure a long short-term memory network base learner with embedded prior constraints of immune resistance mechanisms to distinguish between immune signals driven by resistance and immune fluctuations caused by non-disease factors.

[0108] S44. For whole-body biochemical time-series data, configure a gradient boosting tree-based learner with prior constraints embedded in the whole-body state-drug resistance association mechanism, and correct the interference of non-disease factors on drug resistance risk prediction through prior constraints.

[0109] S45. Complete the mutual binding of each base learner with the corresponding time series data and the corresponding drug resistance biological mechanism, so that each base learner only processes the time series data of the corresponding category; each bound base learner independently completes the deep learning of the corresponding dimension time series features, outputs the primary drug resistance risk prediction result of the corresponding dimension, and matches the corresponding drug resistance biological mechanism annotation to the prediction result to form the primary prediction result with biological annotation.

[0110] The core causal feature set for drug resistance is classified based on its biological origin and drug resistance mechanism. The key is to structurally decompose the selected causal features to provide a basis for the subsequent configuration of the base learner. Specifically, bioinformatics annotation tools can be used to perform GO and KEGG pathway enrichment analysis on each feature. Combined with a comprehensive knowledge graph of tumor drug resistance, the biological origin and corresponding drug resistance driving mechanism of each feature are clarified, ultimately classifying them into three categories: The first category is genomic clonal evolution time-series data, corresponding to the core driving mechanisms of targeted therapy and chemotherapy resistance, including ctDNA resistance mutation abundance and tumor clonal evolution-related indicators; the second category is immune microenvironment time-series data, corresponding to the core driving mechanisms of immunotherapy resistance, including the proportion of immune cell subsets, expression of immune checkpoint molecules, and levels of inflammatory factors; the third category is systemic biochemical time-series data, corresponding to the patient's systemic treatment response status and chemotherapy resistance-related mechanisms, including complete blood count, liver and kidney function, and metabolic-related biochemical indicators.

[0111] The core of configuring dedicated base learners and embedding corresponding prior constraints for different types of data is to adapt each base learner to the temporal characteristics and biological laws of the corresponding data, thereby improving the accuracy and interpretability of feature learning. For genomic clonal evolution time-series data, which has long-term and non-linear clonal evolution characteristics, a temporal Transformer base learner is configured, and prior constraints on clonal evolution mechanisms are added to the loss function to force the model to prioritize learning drug resistance features that conform to the selection and expansion laws of tumor clones. For immune microenvironment time-series data, which has medium-term and continuous dynamic change characteristics, a Long Short-Term Memory (LSTM) network base learner is configured, and prior constraints on immune drug resistance mechanisms are embedded in the model to filter immune fluctuations caused by non-disease factors such as infection and autoimmune responses through gating mechanisms. For systemic biochemical time-series data, which has short-term and discrete temporal characteristics with multiple influencing factors, a gradient boosting tree base learner is configured, and prior constraints on systemic state-drug resistance association mechanisms are embedded in feature engineering and model training to correct for interference from non-disease factors such as malnutrition and liver and kidney dysfunction on drug resistance prediction.

[0112] The core of achieving dedicated binding and independent prediction for base learners lies in the specialized processing of drug resistance signals across different dimensions, while simultaneously imbuing the prediction results with biological interpretability. In practice, a data splitting mechanism ensures that each base learner receives only time-series data of its corresponding category, enabling independent feature deep learning and primary drug resistance risk prediction. Furthermore, each primary prediction result is matched with a corresponding biological mechanism annotation for drug resistance; for example, the prediction result of the genomic base learner is annotated as clonal evolution driving drug resistance, while the prediction result of the immune base learner is annotated as immune escape driving drug resistance. This ultimately forms biologically annotated primary prediction results, providing high-quality, highly interpretable input for the fusion of upper-level meta-learners.

[0113] This implementation effectively avoids mutual interference between different biological signals and improves the model's ability to identify different drug resistance mechanisms through a specialized and modular configuration of base learner clusters. Specific examples are as follows: For the EGFR T790M mutation abundance time-series data in the core drug resistance causal feature set, it is classified as genomic clonal evolution time-series data and bound to a time-series Transformer base learner with embedded clonal evolution prior constraints. After independently completing feature learning, this base learner outputs a primary drug resistance risk prediction result and annotates it as EGFR pathway clonal evolution driving targeted drug resistance. For the PD-L1 expression level time-series data, it is classified as immune microenvironment time-series data and bound to a long short-term memory network base learner with embedded immune drug resistance prior constraints. This outputs a primary prediction result with annotations indicating immune escape driving immunotherapy resistance, ultimately providing multi-dimensional and interpretable primary prediction inputs for the upper-level meta-learner.

[0114] Furthermore, refer to Figure 3Before training, the temporal heterogeneous base learner cluster undergoes unsupervised pre-training through drug-resistant clonal evolution-guided self-supervised contrastive learning, including:

[0115] S411. Split an unlabeled pre-trained subset from the tumor drug resistance time series blood pair dataset and complete the alignment and standardization preprocessing synchronized with the drug administration cycle time series.

[0116] S412. Construct contrastive learning positive and negative sample pairs based on the biological laws of tumor clonal evolution and drug resistance, and at the same time mine difficult sample pairs corresponding to the inflection point of clonal evolution.

[0117] S413. Using the contrastive learning loss function as the core of optimization, and combining the drug resistance biology interpretability constraint loss, unsupervised pre-training is performed on the learners of each temporal heterogeneous base.

[0118] S414. After pre-training, the biological mechanism matching of the pre-trained features is verified, noisy features without clear biological significance of drug resistance are removed, and the weights of the common features of each base learner are frozen.

[0119] The core of constructing positive and negative sample pairs and difficult sample pairs based on biological principles is to integrate knowledge of tumor drug resistance into self-supervised learning, guiding the model to learn effective features related to drug resistance mechanisms, rather than general temporal features. Specifically, positive sample pairs are formed by using homologous temporal data from adjacent dosing cycles of the same patient, ensuring that the sample pairs have a consistent clonal evolutionary background and drug resistance development trend; negative sample pairs are formed by using similar temporal data from different patients with opposite drug resistance phenotypes, ensuring that the sample pairs have differentiated drug resistance biological characteristics; simultaneously, an inflection point detection algorithm is used to identify key time nodes in tumor clonal evolution, mining difficult sample pairs before and after the clonal evolution inflection point, strengthening the model's ability to capture weak early signals of drug resistance, and addressing the problem of insufficient attention to key clinical signals in traditional self-supervised learning.

[0120] The core of unsupervised pre-training optimized by dual loss functions is balancing the model's feature discrimination ability with its biological interpretability. Specifically, the InfoNCE contrastive learning loss function is used as the core optimization tool, guiding the model to maximize the feature consistency of positive sample pairs and minimize the feature consistency of negative sample pairs, thus learning the inherent patterns in time-series data. Simultaneously, a drug resistance biological interpretability constraint loss is added to the loss function. Through a comprehensive knowledge graph of tumor drug resistance, this penalty is applied to features unrelated to known drug resistance pathways and mechanisms, ensuring that the features learned by the model have clear biological meaning and avoiding the learning of meaningless, noisy features.

[0121] Pre-training feature validation and weight freezing are crucial for consolidating the results of pre-training and laying a solid foundation for subsequent supervised training. Specifically, after pre-training, gene enrichment analysis and pathway association analysis are used to validate the biological mechanism matching of the pre-training features extracted by each base learner, eliminating noisy features that are not significantly related to drug resistance mechanisms. Subsequently, the weights for extracting the common underlying features of each base learner are frozen, retaining only the trainable permissions of the top-level feature mapping layer. This avoids feature degradation and overfitting due to insufficient labeled data in subsequent supervised training, while also accelerating the model's convergence speed.

[0122] Furthermore, refer to Figure 4 The fusion weights of the outputs of each temporally heterogeneous base learner are dynamically adjusted, including:

[0123] S51. By using the integrated learning-mediated module to identify the modal completeness of the input data, the number of patient treatment lines, the type of anti-tumor drugs, and the tumor stage, the core feature priority of the current early warning scenario is determined.

[0124] S52. Based on the prior medical constraints of the panoramic knowledge graph of tumor drug resistance, the initial fusion weights of the primary prediction results of each temporal heterogeneous base learner are calculated through the attention mechanism, and the minimum basic weight threshold is set for the base learners corresponding to the core drug resistance markers specified in the clinical guidelines.

[0125] S53. Based on the core drug resistance mechanism of anti-tumor drugs, complete the drug-specific adaptation of fusion weights, assign dominant weights to the base learners to which the corresponding core drug resistance features belong, and at the same time complete the evidence-based medicine evidence level verification of the fusion weight adjustment results.

[0126] S54. Based on the final fusion weights after adaptation, the initial prediction results of each temporal heterogeneous base learner are weighted and fused to output the final comprehensive prediction result of drug resistance risk.

[0127] The key to identifying clinical scenarios and prioritizing core features lies in adapting weights to the individualized clinical context of each patient, rather than using fixed weight allocation rules. Specifically, an ensemble learning-mediated module automatically identifies the modal completeness of the input data, the number of treatment lines for the patient, the type of anti-tumor drug, and the tumor stage. Combined with a comprehensive knowledge graph of tumor drug resistance, it determines the most critical feature dimensions for drug resistance warning in the current scenario and their corresponding base learners, clarifying the priority of core features and providing clinical basis for subsequent weight calculations.

[0128] The core principle of calculating initial fusion weights based on prior constraints and setting minimum weight thresholds is to ensure that weight allocation is grounded in evidence-based medicine, avoiding weight allocations that are not consistent with clinical logic due to purely data-driven approaches. Specifically, based on prior medical constraints from a comprehensive knowledge graph of tumor drug resistance, an attention mechanism is used to automatically calculate the initial fusion weights of the primary prediction results of each base learner. Simultaneously, for base learners corresponding to core drug resistance biomarkers explicitly recommended by clinical guidelines, an insurmountable minimum basic weight threshold is set to ensure that these clinically validated core biomarkers consistently play a fundamental role in early warning systems, preventing the underestimation of critical information due to data distribution issues.

[0129] Drug-specific weight adaptation and evidence-based validation aim to ensure that weight allocation accurately matches the core resistance mechanism of the currently used drug, thereby improving the targeting of early warnings. Specifically, based on the core resistance mechanism of the anti-tumor drug used by the patient, the initial fusion weights are drug-specifically adapted, and dominant weights are assigned to base learners directly related to the drug's core resistance mechanism.

[0130] The core of the weighted fusion and result output is to integrate multi-dimensional prediction results based on the final fusion weights that have undergone multiple verifications, generating a final comprehensive early warning result. In specific implementation, based on the adapted final fusion weights, the initial prediction results of each base learner are weighted and fused to generate the final comprehensive prediction result of drug resistance risk. At the same time, the evidence-based basis and logical explanation of the weight allocation are output simultaneously, providing clinicians with interpretable early warning results and improving clinical acceptance.

[0131] Furthermore, evidence-based clinical treatment protocols that are causally associated with core driver drug resistance characteristics include:

[0132] S71. Based on the drug resistance probability, core driving drug resistance characteristics and drug resistance occurrence time window output by the converged ensemble learning-mediated drug resistance risk warning model, the individualized drug resistance risk level is divided into three levels: low risk, medium risk and high risk according to evidence-based medicine standards.

[0133] S72. For low-risk patients, maintain routine monitoring synchronized with the dosing cycle and cancel imaging examinations;

[0134] S73. For medium-risk patients, increase the frequency of blood sampling and monitoring, conduct basic compatibility tests without radiation or damage, and track the changing trends of indicators corresponding to the core driving drug resistance characteristics.

[0135] S74. For high-risk patients, complete the clinical review of the early warning results within the preset time limit. Based on the causal association matching mechanism between the core driving drug resistance characteristics and the examination items, customize targeted examination items for patients, including examining the target points and lesion sites corresponding to the driving characteristics to confirm the drug resistance status.

[0136] The core of the standardized risk grading based on multi-dimensional information is to provide clear and unified evidence-based basis for subsequent stratified treatment, avoiding subjectivity and arbitrariness in grading. Specifically, it integrates information from three dimensions—the probability of drug resistance, the causal effect strength of the core driving drug resistance feature, and the time window for drug resistance development—output by the comprehensive model, and classifies the risk into three levels according to evidence-based medicine standards: low risk is a drug resistance probability of <30% in the next two treatment cycles, with no clear driving drug resistance feature; medium risk is a drug resistance probability of 30%-70% in the next two treatment cycles, with a clear core driving drug resistance feature; high risk is a drug resistance probability of >70% in the next two treatment cycles, with the core driving drug resistance feature continuing to progress, and a predicted drug resistance development time window of <2 months, ensuring a high degree of alignment between the grading standards and the clinical risk level.

[0137] The core of the stratified treatment strategy for different risk levels is to achieve precise and differentiated clinical management, avoiding overtreatment while ensuring timely intervention for high-risk patients. For low-risk patients, routine blood monitoring is maintained in sync with the medication cycle, and unnecessary imaging examinations are waived, ensuring safety while reducing the physical burden and economic costs on patients. For medium-risk patients, the frequency of blood sampling and monitoring is increased, and only basic, non-radiation and non-invasive examinations such as ultrasound are scheduled to closely track the changing trends of indicators corresponding to the core drivers of drug resistance, achieving continuous monitoring of early signals and avoiding invasive and radiation-related examinations. For high-risk patients, clinical review of early warning results is completed within a preset timeframe of 24-72 hours, and the most proactive and precise intervention process is initiated.

[0138] The core of the causal-based targeted testing matching mechanism is to directly link testing items with core driver drug resistance characteristics, avoiding patient harm and waste of medical resources caused by indiscriminate comprehensive testing. In practice, based on a comprehensive knowledge graph of tumor drug resistance, causal-based matching rules between core driver drug resistance characteristics and testing items are pre-constructed to clarify the optimal testing methods corresponding to different driver characteristics. For example, when the core driver characteristic is EGFRT790M mutation, high-sensitivity liquid biopsy gene testing is matched; when the core driver characteristic is immune microenvironment dysregulation, immune-related biopsy and ultrasound examinations are matched; when the core driver characteristic is local lesion progression, precise imaging examinations of the corresponding sites are matched. Testing is only performed on the target points and lesion sites corresponding to the driver characteristics, ensuring accurate confirmation of drug resistance while minimizing patient harm caused by testing.

[0139] Furthermore, incremental iterative optimization of the interconnected knowledge graph of tumor drug resistance will be carried out, including:

[0140] S91. Through the integrated learning-mediated module, the elastic weight consolidation algorithm is used to calculate the importance score of the core weights related to the common characteristics of drug resistance in the converged early warning model and apply protective restrictions.

[0141] S92. Screen representative drug-resistant and sensitive samples covering multiple tumor types, drug types, and populations, and construct a core sample memory bank with drug resistance biological annotations;

[0142] S93. When new clinical data is input, the task-specific branch layer of each time-series heterogeneous base learner and the dynamic weight fusion meta-learner of evidence-based medicine constraints are adjusted. The core samples in the memory bank are played back simultaneously to complete joint training. At the same time, the tumor drug resistance panoramic knowledge graph is linked to complete the autonomous learning of the new drug resistance mechanism and the dynamic update of the core drug resistance causal feature set.

[0143] S94. After each incremental training is completed, the model performance is verified and clinical compliance is checked through an independent validation set, and version updates and full-process traceability records are retained.

[0144] The core weight protection mechanism based on the elastic weight consolidation algorithm aims to address the catastrophic forgetting problem during incremental updates, ensuring that the model retains its ability to identify common drug resistance features while learning new knowledge. Specifically, the elastic weight consolidation algorithm calculates the importance score of core weights related to common drug resistance features in the converged early warning model. A higher score indicates that the weight is more important to the drug resistance prediction task already mastered by the model. Subsequently, protective restrictions are imposed on highly important weights. During incremental training, a regularization penalty is applied to their update magnitude; the higher the weight's importance, the smaller the allowed update magnitude, thereby protecting the common knowledge already learned by the model and preventing catastrophic forgetting.

[0145] The core sample memory with biological annotations is constructed primarily to provide representative samples to support incremental training and further solidify the model's generalization ability. Specifically, from the tumor drug resistance time-series blood paired dataset, stratified sampling is used to select representative drug-resistant and sensitive samples covering multiple tumor types, drug types, populations, and drug resistance mechanisms. Each sample is matched with a corresponding drug resistance biological mechanism annotation to construct the core sample memory. The samples in this memory comprehensively reflect the various drug resistance scenarios already mastered by the model, providing a foundation for sample replay during incremental training.

[0146] The core of incremental training and knowledge graph-linked updates is to achieve synchronous iteration of model parameters and the knowledge system, allowing the model to continuously adapt to new drug resistance mechanisms. Specifically, when new clinical data is input, only the task-specific branch layers and meta-learners of each temporal heterogeneous base learner are fine-tuned, while the protected core weights are not significantly updated to avoid disrupting already learned common knowledge. Simultaneously, core samples in the memory bank are replayed to complete joint training with the new data, balancing the learning of new knowledge with the consolidation of old knowledge. During incremental training, the tumor drug resistance panoramic knowledge graph is linked to enable autonomous learning of new drug resistance mechanisms, synchronously updating the core drug resistance causal feature set. This allows the model to not only learn new data but also integrate new domain knowledge, continuously improving its ability to identify novel drug resistance patterns.

[0147] Performance validation, compliance verification, and version management are fundamentally about ensuring that incrementally updated models consistently meet clinical application requirements while maintaining full traceability. Specifically, after each incremental training iteration, core performance metrics such as accuracy, sensitivity, and specificity are validated using a validation set completely independent of the training data to ensure no significant performance degradation. Simultaneously, clinical compliance verification is performed to ensure that the model's output and weight allocation consistently conform to clinical guidelines and standards. Finally, model version updates are implemented, with a fully traceable record of training data, parameter changes, performance validation results, and compliance verification results, complying with regulatory requirements for medical device software.

[0148] Furthermore, by integrating the learning-mediated module and feeding it back to the early warning model, incremental iterative optimization of compliance is completed in conjunction with the tumor drug resistance panoramic knowledge graph, including:

[0149] X1. By using an integrated learning-mediated module, we can discover novel blood biomarkers that are strongly associated with the occurrence of tumor drug resistance and analyze the corresponding resistance driving mechanisms and potential drug targets.

[0150] X2. High-throughput virtual screening and drug relocation analysis of drugs for drug resistance reversal based on core drug resistance causal feature sets;

[0151] X3. By using biomarkers corresponding to the core drug resistance causal feature set, enrich the subjects in the clinical trials of anti-tumor drugs and construct a prospective stratification tool for predicting drug resistance risk in clinical trial subjects;

[0152] X4. Using the dynamic changes of blood biomarkers corresponding to the core drug resistance causal feature set as a surrogate efficacy endpoint in clinical trials of anti-tumor drugs, a drug resistance biological model is constructed based on the core characteristics of drug-resistant patients, providing real-world data support for the in vitro validation of drug resistance reversal drugs.

[0153] The core of novel biomarker mining and drug target discovery is to transform clinical data into source clues for drug development. Specifically, based on a comprehensive database of tumor drug resistance blood characteristics with causal annotations, a causal mining framework integrating learning-mediated modules is used to discover novel blood biomarkers strongly correlated with the occurrence of tumor drug resistance. This framework analyzes their driving roles and molecular mechanisms in drug resistance development, thereby identifying potential targets for drug intervention and providing innovative directions for the development of new drugs to reverse drug resistance.

[0154] Drug virtual screening and repositioning based on core features aims to accelerate the discovery of candidate drugs by directly utilizing causal characteristics of drug resistance. Specifically, based on the drug resistance pathways and targets corresponding to the core causal feature set, high-throughput virtual screening of drug resistance reversal drugs is achieved through computational methods such as molecular docking, pharmacophore matching, and molecular dynamics simulation. This screens candidate drugs that can target and block the core drug resistance mechanism from a large-scale compound library. Simultaneously, combined with a drug action mechanism database, drug repositioning analysis is performed to screen marketed drugs that can act on the core drug resistance target, significantly shortening the development cycle and reducing development costs of drug resistance reversal.

[0155] The core of clinical trial participant enrichment and risk stratification is to optimize clinical trial design and improve the success rate of clinical trials by utilizing key drug resistance biomarkers. Specifically, during the recruitment phase of anti-tumor drug clinical trials, precise participant enrichment is achieved using biomarkers corresponding to the core drug resistance causal characteristic set. This screens patients sensitive to the investigational drug with low resistance risk for enrollment, while excluding patients with high resistance risk, thereby improving the efficacy detection rate of the investigational drug. Simultaneously, a prospective stratification tool for predicting drug resistance risk in clinical trial participants is constructed based on these core biomarkers. This allows for stratified management of participants during the trial, optimizing trial protocol design and reducing the risk of clinical trial failure.

[0156] The core of constructing surrogate efficacy endpoints and supporting them with in vitro models is to transform real-world clinical data into a comprehensive support system for drug development. Specifically, this involves using the dynamic changes in blood biomarkers corresponding to core drug resistance causal characteristic sets as surrogate efficacy endpoints in anti-tumor drug clinical trials. These endpoints replace traditional endpoints such as overall survival and progression-free survival, which require long-term follow-up, significantly shortening the observation period for clinical trials. Simultaneously, based on the core characteristics of drug-resistant patients in the database, corresponding drug-resistant cell models, organoid models, and humanized animal models are constructed to realistically simulate the drug resistance mechanisms of patients. This provides real-world data support for the in vitro screening and efficacy validation of drug resistance reversal drugs, improving the conversion rate between preclinical research and clinical trials.

[0157] In practical applications, based on a comprehensive database of chemotherapy resistance in colorectal cancer, a novel blood biomarker strongly correlated with oxaliplatin resistance was discovered through an ensemble learning-mediated module. The DNA damage repair resistance mechanism mediated by this biomarker was analyzed, and potential drug targets were identified. High-throughput virtual screening was performed based on this target, identifying three candidate compounds for resistance reversal. Simultaneously, drug repositioning analysis revealed that an already marketed cardiovascular drug could target this target. In subsequent clinical trials of oxaliplatin resistance reversal drugs, this biomarker was used for subject enrichment, screening patients at high risk of oxaliplatin resistance for enrollment. The dynamic changes of this biomarker were used as a surrogate efficacy endpoint, significantly shortening the trial period. Based on the core characteristics of resistant patients, an organoid model of oxaliplatin-resistant colorectal cancer was constructed, providing real-world model support for the in vitro validation of candidate drugs and realizing the entire chain of transformation from clinical data to drug development.

[0158] Furthermore, it also includes an algorithmic fairness optimization process mediated by ensemble learning to achieve performance balancing across the entire population, including:

[0159] A1. Construct a set of two-dimensional sensitive attributes specifically for early warning of tumor drug resistance, and complete the causal irrelevance verification of all sensitive attributes;

[0160] A2. Construct a hierarchical fairness optimization architecture that matches the two-layer heterogeneous model, which is divided into a causal-fair joint constraint unit at the base learner level and a fairness attention fusion unit at the meta learner level.

[0161] A3. During the training of the two-layer heterogeneous model, a causal-fair joint loss function corresponding to the feature dimension is constructed through the causal-fair joint constraint unit to eliminate the false causal association between sensitive attributes and drug resistance causal features, and to eliminate the feature dependence of the base learner on sensitive attributes.

[0162] A4. Optimize the weight allocation mechanism of the meta-learner through the fairness attention fusion unit, and increase the lower limit of the basic weight of the corresponding base learner for subgroups with insufficient sample size.

[0163] A5. Construct a cross-scenario feature enhancement fairness adaptation branch for drug-resistant small sample populations, complete cross-subgroup feature transfer enhancement, and generate a dedicated lightweight early warning branch;

[0164] A6. Construct a continuous fairness verification and update mechanism, complete the performance balance verification of all subgroups before each incremental update, and simultaneously optimize the fairness constraint parameters during the incremental update process;

[0165] A7. After the training and updating of the two-layer heterogeneous model is completed, the model is validated through a clinically oriented multi-dimensional fairness verification system, so that different subgroups of people can receive drug resistance warnings.

[0166] The core of the dual-dimensional sensitive attribute set construction and causal irrelevance verification is to clarify the goal of fairness optimization and ensure that the optimization process does not affect the model's learning of real drug resistance driving features. Specifically, a dual-dimensional sensitive attribute set dedicated to tumor drug resistance early warning is constructed: the first category consists of patient individual attributes that are not directly causally related to tumor drug resistance, including demographic attributes such as age, gender, race, and region; the second category consists of scenario attributes related to clinical data distribution, including tumor rarity, number of treatment lines, and completeness of blood time-series data. Subsequently, causal irrelevance verification of all sensitive attributes is performed using causal inference methods to confirm that there is no direct causal driving effect between these attributes and the occurrence of tumor drug resistance, avoiding misclassification of real drug resistance prognostic factors as sensitive attributes and ensuring that the direction of fairness optimization is correct.

[0167] The core of building a hierarchical fairness optimization architecture is to integrate fairness optimization into each layer of the two-layer heterogeneous model, achieving end-to-end fairness control from feature learning to result fusion, rather than just performing post-processing corrections at the model output. Specifically, a hierarchical fairness optimization architecture perfectly matched to the two-layer heterogeneous model is constructed, consisting of two core units: a causal-fairness joint constraint unit at the base learner level, responsible for eliminating algorithmic bias at the source of feature learning; and a fairness attention fusion unit at the meta-learner level, responsible for balancing the predictive performance of different subgroups during the result fusion stage, achieving end-to-end fairness optimization.

[0168] The core of causal-fair joint constraints and fairness attention optimization is to eliminate algorithmic bias without sacrificing the model's core warning performance. Specifically, during model training, a causal-fair joint constraint unit constructs a causal-fair joint loss function for each base learner based on its corresponding feature dimension. This loss function, based on the model's predictive performance loss and combined with a fairness regularization term, eliminates spurious causal associations between sensitive attributes and drug resistance causal features, thus eliminating the base learner's feature dependence on sensitive attributes and ensuring the model predicts only based on genuine drug resistance causal features. Simultaneously, a fairness attention fusion unit optimizes the weight allocation mechanism of the meta-learners. For subgroups with insufficient sample sizes, such as rare tumors and elderly patients, the lower bound of the base weights corresponding to the core drug resistance features is increased to prevent the prediction results of these subgroups from being ignored due to insufficient sample size, thus balancing the predictive performance of different subgroups.

[0169] The core of the small-sample population-specific adaptation branch and continuous fairness update mechanism is to address the model fairness issue in scenarios with extremely scarce data, while ensuring that the model maintains consistent performance throughout its lifecycle. Specifically, for small-sample populations with drug resistance, such as rare tumors and pediatric cancer patients, a cross-scenario feature enhancement fairness adaptation branch is constructed. Through transfer learning, shared causal features of drug resistance are transferred and enhanced across tumor types and populations, strengthening the model's ability to capture drug resistance features in small samples. Simultaneously, a dedicated lightweight warning branch is generated for small-sample populations to address overfitting issues with limited data. Furthermore, a continuous fairness verification and update mechanism is constructed. Before each incremental model update, performance balance verification is performed across all subgroups to identify subgroup performance biases introduced by new data. During incremental updates, fairness constraint parameters are simultaneously optimized to prevent incremental updates from amplifying algorithmic bias.

[0170] Clinically-oriented multidimensional fairness validation focuses on ensuring that model fairness optimization aligns with actual clinical needs, rather than merely satisfying algorithm-level fairness metrics. Specifically, after model training and updates, a clinically-oriented multidimensional fairness validation system is used for verification. In addition to traditional classification performance balance metrics, it further verifies the balance of drug resistance warning lead time, positive predictive value, and negative predictive value among different subgroups. This ensures that patients of different ages, genders, and tumor types receive stable, accurate, and consistent drug resistance warning services, completely eliminating algorithmic bias at the clinical application level and achieving fairness across the entire population.

[0171] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. An ensemble learning-mediated early warning method for tumor drug resistance risk, characterized in that, include: Patients with the target tumor type receiving anti-tumor drug therapy were included in the cohort. Peripheral venous blood samples were collected non-invasively from patients using an indwelling venous catheter that was synchronized with the timing of the anti-tumor drug administration cycle. Blood time-series detection indicators were obtained. Collect patient antitumor drug regimens and drug resistance standard validation data to construct a time-series blood paired dataset of tumor drug resistance. By integrating the learning-mediated module, an integrated mining framework with dual constraints of temporal causality and drug resistance mechanism is constructed. The integrated mining framework is adapted to the causal inference algorithm of tumor drug resistance temporal characteristics, corrects the individual patient baseline and treatment-related confounding factors, analyzes the temporal lag effect of feature changes and drug resistance occurrence, and screens out the core drug resistance causal feature set that has a direct causal driving effect on tumor drug resistance from the tumor drug resistance temporal blood paired dataset. A two-layer heterogeneous ensemble learning model for drug resistance risk warning was constructed, embedding prior constraints from a comprehensive knowledge graph of tumor drug resistance. The lower layer of the two-layer heterogeneous ensemble learning model for drug resistance risk warning consists of a cluster of temporal heterogeneous base learning bases corresponding to the core causal feature set of drug resistance. The upper layer of the two-layer heterogeneous ensemble learning model for drug resistance risk warning consists of a dynamic weight fusion meta-learner constrained by evidence-based medicine. The ensemble learning-mediated drug resistance risk warning model dynamically adjusts the fusion weights of the output results of each temporal heterogeneous base learner based on the modal completeness of the input data, the clinical treatment scenario, and the core drug resistance mechanism of anti-tumor drugs, through the dynamic weight fusion meta-learner constrained by evidence-based medicine and the prior constraints of the comprehensive knowledge graph of tumor drug resistance. During the patient's anti-tumor treatment, real-time blood time-series detection indicators are collected through an indwelling intravenous needle according to a preset cycle that is synchronized with the drug administration cycle. These indicators are then input into an integrated learning-mediated drug resistance risk early warning model. The integrated learning-mediated drug resistance risk early warning model outputs the patient's individualized drug resistance risk level, core driving drug resistance characteristics, and prediction results of the drug resistance event window. Based on the patient's individualized drug resistance risk level and core driver drug resistance characteristics, we matched evidence-based clinical treatment protocols that had a causal relationship with the core driver drug resistance characteristics. When the individualized drug resistance classification is determined to be of high drug resistance risk, a targeted examination instruction is triggered that is causally linked to the core driving drug resistance characteristics in order to determine the occurrence of drug resistance in the patient. The patient's periodic blood test data, drug resistance risk warning results, examination verification results, and treatment outcome data are fed back to the integrated learning-mediated drug resistance risk warning model through the integrated learning-mediated module, and incremental iterative optimization is carried out in conjunction with the panoramic knowledge graph of tumor drug resistance.

2. The integrated learning-mediated early warning method for tumor drug resistance risk according to claim 1, characterized in that, Peripheral venous blood samples were non-invasively collected from patients using an indwelling venous catheter that was synchronized with the dosing cycle of antitumor drugs. These samples included: Based on the time synchronization and anchoring with the drug administration cycle of anti-tumor drugs, the time-series detection indicators of the collected blood were obtained to match the drug action cycle. A basic blood collection cycle is set according to the patient's treatment stage. In the early stage of treatment, individualized indicator baselines are established through continuous blood collection, and a fixed blood collection cycle is set during the stable period of treatment to complete routine monitoring. The blood collection cycle is dynamically adjusted based on the patient's anti-tumor drug administration cycle, changes in treatment stage, and real-time detection results of blood test indicators corresponding to the core drug resistance causal feature set. A linkage triggering mechanism is constructed for blood collection operations, sample quality control, and model pre-analysis. When abnormal fluctuations occur in blood test indicators corresponding to the core drug resistance causal feature set, the blood collection cycle is encrypted, the whole process of sample quality control is carried out simultaneously, and the pre-analysis of the integrated learning-mediated drug resistance risk early warning model is initiated. Based on the real-time changes in the patient's antitumor drug administration cycle, treatment stage, and core drug resistance causal characteristics, the blood collection cycle is dynamically adjusted to achieve full-cycle adaptation, including baseline establishment, routine testing, and encrypted tracking of abnormal signals, all synchronized with the drug administration cycle.

3. The integrated learning-mediated early warning method for tumor drug resistance risk according to claim 1, characterized in that, From the time-series blood paired dataset of tumor drug resistance, a core set of causal features with direct causal driving effects on tumor drug resistance was screened, including: Using blood time-series detection indicators from the tumor drug resistance time-series paired blood dataset as input features and drug resistance standard verification data as outcome labels, the causal relationship results related to tumor drug resistance output by the causal inference algorithm are fused to generate a causal directed acyclic graph between the input features and the outcome labels, thus obtaining a preliminary causal feature set. By adapting a causal inference algorithm to the temporal characteristics of tumor drug resistance, we correct for the interference of individual patient baseline differences and treatment-related confounding factors, quantify the average causal driving effect of each feature in the preliminary causal feature set on the drug resistance outcome, analyze the temporal lag effect window of feature changes leading to drug resistance, and eliminate false positive features without significant causal driving effects. Based on the strength of causal driving effects, the significance of time-series lag effects, and the matching degree of biological mechanisms, a multi-dimensional causal consistency verification was completed, and a core set of causal features for drug resistance was obtained.

4. The integrated learning-mediated early warning method for tumor drug resistance risk according to claim 1, characterized in that, Configure a temporal heterogeneous base learner cluster, including: Based on the biological origin of the core drug resistance causal feature set and the corresponding driving mechanism type of tumor drug resistance, the core drug resistance causal feature set is divided into three categories: genomic clonal evolution time series data, immune microenvironment time series data, and systemic state biochemical time series data, and the specific tumor drug resistance biological mechanism corresponding to each type of data is clarified. For genomic clonal evolution time series data, a time-based learner with embedded clonal evolution mechanism prior constraints is configured to learn drug resistance-related features that conform to the laws of tumor clonal evolution. For time-series data of the immune microenvironment, a long short-term memory network base learner with embedded prior constraints of immune drug resistance mechanisms is configured to distinguish between drug resistance-driven immune signals and immune fluctuations caused by non-disease factors. For whole-body biochemical time-series data, a gradient boosting tree-based learner with prior constraints embedded in the whole-body state-drug resistance association mechanism is configured to correct the interference of non-disease factors on drug resistance risk prediction through prior constraints. Each base learner is bound to its corresponding time series data and corresponding drug resistance biological mechanism, so that each base learner processes only the time series data of its corresponding category. Each bound base learner independently performs deep learning of the time series features of its corresponding dimension, outputs the primary drug resistance risk prediction result of the corresponding dimension, and matches the prediction result with the corresponding drug resistance biological mechanism annotation to form the primary prediction result with biological annotation.

5. The integrated learning-mediated early warning method for tumor drug resistance risk according to claim 4, characterized in that, Before training, the temporal heterogeneous base learner cluster undergoes unsupervised pre-training through drug-resistant clonal evolution-guided self-supervised contrastive learning, including: An unlabeled pre-trained subset was split from the tumor drug resistance time-series blood pair dataset and aligned and standardized preprocessed to be synchronized with the drug administration cycle time series. Based on the biological laws of tumor clonal evolution and drug resistance, we construct positive and negative sample pairs for comparative learning, and at the same time, we mine difficult sample pairs corresponding to the inflection point of clonal evolution. Using the contrastive learning loss function as the core of optimization, and combining the drug resistance biology interpretability constraint loss, unsupervised pre-training is performed on learners of different temporal heterogeneous bases. After pre-training is completed, the biological mechanism matching of the pre-trained features is verified, noisy features without clear biological significance of drug resistance are removed, and the weights of the underlying common features of each base learner are frozen.

6. The integrated learning-mediated early warning method for tumor drug resistance risk according to claim 1, characterized in that, Dynamically adjust the fusion weights of the outputs of each temporally heterogeneous base learner, including: By integrating learning-mediated modules, the modal completeness of input data, the number of patient treatment lines, the type of anti-tumor drugs, and the tumor stage are identified to determine the priority of core features in the current early warning scenario. Based on the prior medical constraints of the panoramic knowledge graph of tumor drug resistance, the initial fusion weights of the primary prediction results of each temporal heterogeneous base learner are calculated through the attention mechanism, and the minimum basic weight thresholds are set for the base learners corresponding to the core drug resistance biomarkers specified in clinical guidelines. Based on the core drug resistance mechanism of anti-tumor drugs, the drug-specific adaptation of fusion weights is completed, the dominant weights are assigned to the base learners to which the corresponding core drug resistance features belong, and the evidence-based medicine evidence level of the fusion weight adjustment results is verified. Based on the final fusion weights after adaptation, the initial prediction results of each temporal heterogeneous base learner are weighted and fused to output the final comprehensive prediction result of drug resistance risk.

7. The integrated learning-mediated early warning method for tumor drug resistance risk according to claim 1, characterized in that, Evidence-based clinical management protocols that are causally associated with core driver drug resistance characteristics include: Based on the probability of drug resistance, core driving drug resistance characteristics, and drug resistance occurrence time window output by the converged ensemble learning-mediated drug resistance risk warning model, the individualized drug resistance risk level is divided into three levels: low risk, medium risk, and high risk according to evidence-based medicine standards. For low-risk patients, maintain routine monitoring synchronized with the medication cycle and cancel imaging examinations; For patients at medium risk, increase the frequency of blood sampling and monitoring, conduct basic compatibility tests that are radiation-free and non-invasive, and track the changing trends of indicators corresponding to the core drivers of drug resistance. For high-risk patients, the clinical review of the early warning results is completed within a preset time limit. Based on the causal association matching mechanism between the core driving drug resistance characteristics and the examination items, the targeted examination items are customized for the patients, including the examination of the target points and lesion sites corresponding to the driving characteristics, in order to confirm the drug resistance status.

8. The integrated learning-mediated early warning method for tumor drug resistance risk according to claim 1, characterized in that, Incremental iterative optimization of the interconnected knowledge graph of tumor drug resistance, including: By integrating the learning-mediated module, the elastic weight consolidation algorithm is used to calculate the importance score of the core weights related to the common characteristics of drug resistance in the converged early warning model and apply protective restrictions. We screened representative drug-resistant and sensitive samples covering multiple tumor types, drug types, and populations, and constructed a core sample memory bank with drug resistance biological annotations. When new clinical data is input, the task-specific branch layer of each time-series heterogeneous base learner and the dynamic weight fusion meta-learner of evidence-based medicine constraints are adjusted, and the core samples in the memory bank are replayed simultaneously to complete joint training. At the same time, the tumor drug resistance panoramic knowledge graph is linked to complete the autonomous learning of new drug resistance mechanisms and the dynamic update of the core drug resistance causal feature set. After each incremental training is completed, model performance is validated and clinical compliance is verified through an independent validation set, and version updates and full-process traceability records are maintained.

9. The integrated learning-mediated early warning method for tumor drug resistance risk according to claim 1, characterized in that, By integrating the learning-mediated module and feeding back into the early warning model, incremental iterative optimization of compliance is achieved in conjunction with the comprehensive knowledge graph of tumor drug resistance, including: By integrating learning-mediated modules, novel blood biomarkers strongly associated with tumor drug resistance are discovered, and the corresponding drug resistance driving mechanisms and potential drug targets are analyzed. High-throughput virtual screening and drug repositioning analysis of drugs for drug resistance reversal were completed based on the core drug resistance causal feature set. By using biomarkers corresponding to the core drug resistance causal feature set, we can enrich the subjects in clinical trials of anti-tumor drugs and construct a prospective stratification tool for predicting drug resistance risk in clinical trial subjects. Using the dynamic changes of blood biomarkers corresponding to the core drug resistance causal feature set as a surrogate efficacy endpoint in clinical trials of anti-tumor drugs, a drug resistance biological model is constructed based on the core characteristics of drug-resistant patients, providing real-world data support for the in vitro validation of drug resistance reversal drugs.

10. The integrated learning-mediated early warning method for tumor drug resistance risk according to claim 1, characterized in that, It also includes an algorithmic fairness optimization process mediated by ensemble learning to achieve performance balancing across the entire population, including: Construct a set of two-dimensional sensitive attributes specifically for early warning of tumor drug resistance, and complete the causal irrelevance verification of all sensitive attributes; A hierarchical fairness optimization architecture matching the two-layer heterogeneous model is constructed, consisting of a causal-fair joint constraint unit at the base learner level and a fairness attention fusion unit at the meta learner level. During the training of the two-layer heterogeneous model, a causal-fair joint loss function corresponding to the feature dimension is constructed through a causal-fair joint constraint unit to eliminate spurious causal associations between sensitive attributes and drug resistance causal features, and to eliminate the feature dependence of the base learner on sensitive attributes. The weight allocation mechanism of the meta-learner is optimized by using a fair attention fusion unit, and the lower limit of the basic weight of the corresponding base learner is increased for subgroups with insufficient sample size. For drug-resistant small sample populations, a cross-scenario feature enhancement fairness adaptation branch is constructed to complete cross-subgroup feature transfer enhancement and generate a dedicated lightweight early warning branch. Establish a continuous fairness verification and update mechanism, complete the performance balance verification of all subgroups before each incremental update, and optimize the fairness constraint parameters simultaneously during the incremental update process; After the two-layer heterogeneous model is trained and updated, it is validated through a clinically oriented, multi-dimensional fairness verification system, enabling different subgroups of people to receive drug resistance warnings.

Images