Clinical trial intelligent follow-up management system and method based on artificial intelligence and big data

Abstract

The application provides an intelligent follow-up management system and method for clinical trials based on artificial intelligence and big data, and relates to the technical field of data analysis and processing. The system is composed of three parts: causal federal knowledge graph, risk perception follow-up scheduling and privacy compliance audit. Each center generates a causal summary without sharing original data and obtains a global graph through secure aggregation. A patient-level causal simulator is issued to generate individualized follow-up actions and causal evidence chains in combination with a constrained conservative strategy learning. Key events in the whole process are written into a permission chain in the form of hash and are governed by a smart contract for informed consent and withdrawal. This scheme improves the follow-up completion rate and early detection ability of adverse events and strengthens compliance auditability.

Description

Technical Field

[0001] This invention relates to the field of data analysis and processing technology, specifically to an intelligent follow-up management system and method for clinical trials based on artificial intelligence and big data. Background Technology

[0002] Clinical trial follow-ups have long relied on fixed schedules and manual reminders, resulting in inconsistent data standards and significant quality fluctuations across centers. Centralized aggregation of raw data poses privacy and compliance risks. Existing correlation-based predictive models struggle to distinguish confounding factors and answer how risks change after a particular follow-up action, leading to a lack of individualized strategies and interpretable evidence. In multi-center environments, significant differences in site resource constraints and patient acceptability make it difficult for traditional rules to simultaneously balance efficiency, safety, and fairness. Follow-up events and model versions lack verifiable evidence links, and data and model governance after informed consent withdrawal often lags behind, failing to meet stringent audit and regulatory requirements. To address these pain points, it is necessary to achieve causal-level cross-center knowledge sharing without disclosing raw data, output interpretable individualized follow-up strategies using constrained and conservative decision-making methods, and strengthen end-to-end auditing and withdrawal governance through technological means. Summary of the Invention

[0003] Technical problems to be solved To address the shortcomings of existing technologies, this invention provides an intelligent follow-up management system and method for clinical trials based on artificial intelligence and big data, thus solving the problems of existing technologies.

[0004] Technical solution

[0005] To achieve the above objectives, the present invention provides the following technical solution: an intelligent follow-up management system and method for clinical trials based on artificial intelligence and big data, comprising: The causal federated knowledge graph module is used to construct a local causal relationship skeleton based on local multi-source data at each clinical trial center, and upload the causal statistical information to the central aggregation node after encryption and differential privacy processing. The central aggregation node forms a global causal knowledge graph, and the global causal knowledge graph is then distributed to each clinical trial center. The risk perception follow-up scheduling module is used to construct a patient-level causal simulator based on the global causal knowledge graph, generate follow-up strategies under resource constraints, ethical constraints and patient compliance constraints using a conservative offline reinforcement learning method, and output causal evidence chains and explanatory information for each recommended action during follow-up execution. The privacy and compliance audit module is used to generate event summaries for key events such as patient informed consent, follow-up strategy generation, follow-up execution results, and model updates throughout the follow-up process. These summaries are then stored in a permissioned blockchain using hashing. Smart contracts are used to govern patient informed consent and withdrawal, ensuring that the system is auditable and tamper-proof.

[0006] Preferably, the causal federated knowledge graph module includes: Local causal structure learning unit, used to extract causal relationships and their confidence levels between variables from local clinical trial data; The side statistical summary generation unit is used to encrypt and perform differential privacy processing on the causal relationship strength and residual distribution; The Federation Aggregation Unit is used to merge causal statistical summaries from multiple clinical trial centers into a global causal skeleton via a secure aggregation protocol.

[0007] Preferably, the risk perception follow-up scheduling module includes: A causal simulator is used to generate potential outcome sequences for patients under different follow-up actions based on a global causal knowledge graph. The policy generation unit is used to perform constrained conservative reinforcement learning training on the causal simulator to obtain a follow-up action selection policy. The strategy interpretation unit is used to generate a causal evidence chain for each follow-up action and store it in the audit log.

[0008] Preferably, the privacy and compliance audit module includes: The differential privacy budget management unit is used to record and display the privacy loss when each clinical trial center uploads causal statistical summaries; A blockchain-based evidence storage unit is used to write follow-up event summaries into a permissioned blockchain using hash values ​​and timestamps. Smart contract units are used to automatically trigger governance processes such as data isolation, model weight rollback, or retraining when a patient withdraws informed consent.

[0009] Preferably, the follow-up strategy includes SMS reminders, telephone follow-ups, video consultations, on-site follow-ups, and follow-up extensions. When outputting the strategy, the system selects the optimal action based on the prediction results of the causal simulator.

[0010] A preferred method for intelligent follow-up management of clinical trials based on artificial intelligence and big data includes the following steps: Sp1. At each clinical trial center, a causal framework is constructed based on local multi-source data, a causal statistical summary is generated, and after encryption and differential privacy processing, it is uploaded to the central aggregation node. Sp2. At the central aggregation node, causal statistical summaries from each clinical trial center are merged through a secure multi-party computation mechanism to generate a global causal knowledge graph and distribute it to each clinical trial center. Sp3. Patient-level causal simulators are constructed at each clinical trial center based on the global causal knowledge graph, and constrained conservative offline reinforcement learning training is performed on the simulators to obtain follow-up strategies. Sp4. During the actual follow-up process, follow-up actions and causal evidence chains are generated based on the follow-up strategy and output to researchers or executed automatically by the system. Sp5. Throughout the follow-up process, key events are summarized and stored via a permissioned blockchain, and patient informed consent and withdrawal are managed based on smart contracts.

[0011] Preferably, the constrained reinforcement learning training in Sp3 uses follow-up costs, patient burden, and ethical constraints as optimization constraints to ensure that the generated follow-up strategy operates within safe boundaries.

[0012] Preferably, the smart contract in Sp5 includes informed consent writing logic, withdrawal triggering logic, and audit access control logic to ensure that the authenticity of follow-up events is verifiable and that the patient's original data is not leaked.

[0013] Preferably, the global causal knowledge graph is updated according to a preset cycle, and the graph is calibrated after each round of updates to ensure its consistency with the latest clinical trial data.

[0014] Preferably, the causal evidence chain includes the reasons for the follow-up action selection, key causal paths, and simulation prediction results, which are used to support regulatory audits and scientific research interpretations.

[0015] Beneficial effects This invention provides an intelligent follow-up management system and method for clinical trials based on artificial intelligence and big data. It has the following beneficial effects: 1. This invention achieves cross-center causal knowledge fusion without sharing original data, supports intervention and counterfactual reasoning, significantly shortens the latency of adverse event detection, and improves strategy stability and external generalizability. Based on constrained conservative strategy learning, it generates individualized follow-up actions by integrating resource, ethical, and compliance boundaries, improving follow-up completion rates and reducing unit follow-up costs while maintaining safety boundaries.

[0016] 2. This invention achieves verifiable end-to-end auditing from informed consent to policy execution by combining differential privacy and security aggregation with blockchain notarization and smart contracts, enabling immediate governance upon withdrawal and meeting the needs of multi-party supervision and compliance evidence collection. Attached Figure Description

[0017] Figure 1This is a system architecture diagram of the present invention; Figure 2 This is a system flowchart of the present invention; Figure 3 This is a screenshot of the system of the present invention; Figure 4 This is a screenshot of the system of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Specific Implementation Example 1: like Figures 1 to 4 As shown, an intelligent follow-up management system and method for clinical trials based on artificial intelligence and big data are presented. The system operates in a multi-center environment. All raw data undergoes cleaning, standardization, and semantic mapping within the site before entering the local modeling process. The site only outputs privacy-processed statistical summaries and does not output any original records that can be traced back to individuals. The variable dictionary should be frozen at the time of project initiation, covering subject baseline attributes, comorbidities and diagnostic codes, laboratory items and units, medication and dosage and adherence, symptoms and signs, wearable time-series indicators, questionnaire scales and subjective discomfort scores, follow-up action types and execution result labels. The time base is unified to millisecond-level timestamps and uses a unified time zone. Cross-source conflicts are resolved through priority and confidence fusion. Unit conversion, threshold truncation, anomaly removal, and missing data imputation need to form a stable and replayable set of rules within the site and be recorded in the site audit metadata with each batch.

[0020] The causal federated knowledge component completes causal structure learning and quality assessment at the site level. The input is the standardized structured and time-series data mentioned above. The processing objective is to produce directed dependencies and their direction, strength, significance, edge uncertainty, residuals for robustness in subsequent rounds, and fit quality metrics. Simultaneously, it records algorithm configuration summaries, sample coverage, data validity windows, and fingerprints of key preprocessing parameters. In the statistical summary generation stage, privacy pruning and random perturbation are performed on each edge and each type of local residual metric. The privacy budget is accumulated per round and recorded at two levels: site-wide and global. The summary uses a single edge and single segment as the smallest unit, along with the site quality score and sample size range. After digital signature, the summary enters the reporting queue. Federated aggregation is completed at the central coordination level. The input is a multi-site encrypted digest. During merging, the site weights are weighted by three factors: sample size is the primary factor, and data quality score and historical consistency score are used for adjustment. Edges with conflicting directions or strength differences exceeding the threshold are included in the dispute set. Disputed edges are not written into the current version of the global graph but enter the next round of evidence collection. After aggregation, a global causal skeleton and parameter digest, version number, version digest hash, list of participating sites, cumulative privacy budget, and edge stability statistics are produced. A lightweight index is then distributed to each site to support inference services. To address heterogeneity and time drift, the edge stability assessment and site deviation correction of the previous version must be completed before each round of aggregation. Sites with insufficient stability or excessive deviation for two consecutive rounds are suspended from aggregation and enter a quality rectification process. They can only be reinstated after rectification is completed and regression testing is passed.

[0021] Risk perception follow-up scheduling operates within the site, inputting three types of information: the first type is the individual subject profile and latest dynamics, including but not limited to fixed or low-frequency baseline attributes and medical history, laboratory trends and vital signs in the most recent assessment window, medication adherence time series, questionnaire scores and symptom changes, and completion status and quality score of the last follow-up; the second type is the global causal knowledge graph and its version description; and the third type is resource constraints, ethical constraints, patient acceptability constraints, and site workload limits. Causal simulation first generates short-term and medium-term outcome distributions for candidate actions such as SMS reminders, telephone follow-ups, video consultations, on-site follow-ups, and follow-up extensions under the joint conditions of the global graph and individual profiles. The output includes adverse event risk, follow-up completion probability, adherence change trend, patient burden score, and site resource consumption estimate, and provides the interval uncertainty and data support label for each outcome for subsequent conservative decision-making. The strategy generation aims to minimize risk and cost while maximizing completion rate and fairness. Hard constraints include adherence to legal, ethical, and safety bottom lines, while soft constraints include target ranges for costs and workload within the evaluation period. When multiple actions simultaneously satisfy constraints and have similar utility, a stability-first or lowest-burden-first selection rule should be adopted to avoid decreased compliance due to frequent action switching. The strategy output should include not only the current action but also alternative actions and fallback paths triggered by conditions, along with trigger thresholds. The interpretation output integrates the core causal path, key variable contributions, and simulated evidence to form a causal evidence chain. This chain includes the reasons for action selection, the dominant causal pathway, sensitive variables and their thresholds and value ranges, differences from alternative actions and switching costs, confidence levels and data support, corresponding graph version numbers, and time information. The evidence chain is written as a structured summary in the audit log and is available for researchers to review quickly and for regulatory spot checks.

[0022] Privacy and compliance are integrated throughout the entire process. Differential privacy budget management updates the site-level and global budget ledgers after each round of site summary generation and reporting. Ledger items include round number, noise intensity, number and category of affected statistics, estimated site sample size, and risk stratification overview. Budget thresholds are set by the trial manager at the time of project initiation and reviewed at key milestones. Blockchain notarization forms a minimum necessary information set for key events throughout the process and hashes and timestamps them, writing them to the permissioned blockchain. Key events include informed consent signing and modification, follow-up strategy generation and adoption, follow-up execution and result backfilling, model release and rollback, graph version changes, and site eligibility status changes. The written content must not contain identifiable personal information or raw data and must be expressed in the form of a summary and index reference. Write permissions for chain nodes are jointly managed by the system provider and the trial organizer. The smart contract maintains the agreed status and terms version. When a withdrawal occurs, an irreversible withdrawal event is broadcast to all sites and the central coordination service. Each site should immediately mark the subject as unusable for any aggregation and training, and perform model weight rollback and recalculation in the first feasible period. The central coordination automatically removes the relevant summary in the next round of aggregation. Every step of the entire governance process must enter the audit queue and be written into the on-chain proof. The time base is unified, and cross-time zone writes adopt delay tolerance and sequential reordering to ensure that the audit timeline is monotonous and consistent.

[0023] The set of follow-up actions needs to be defined in action blueprints before the system goes live. The availability prerequisites, execution time limits, resource consumption estimates, patient burden estimates, and compliance boundaries for each action must be clearly quantified. For example, the availability of on-site follow-up depends on geographical distance, the availability of home visit resources, and the subject's home visit permission. The execution time limit is within the specified time window after issuance. Resource consumption is broken down into staff hours, transportation costs, and consumable budgets. Patient burden is comprehensively expressed by the number of disturbances, contact duration, and psychological stress scores. Compliance boundaries set additional restrictions for vulnerable populations. Strategy selection must be based on the joint decision-making of full evidence from causal simulation and the constraint set. Any extreme improvement of a single indicator must not come at the expense of other indicators. When there are differences in historical responses, priority should be given to matching the action combination that has the best historical response and lowest burden for an individual. Researchers retain the right to manually overwrite, and the system will forcibly generate a causal summary and write it into the audit during the overwriting process.

[0024] The engineering requirements for the method steps are as follows. Sp1's input is multi-source data processed by ETL within the site, and the output is a privacy-protected causal statistical summary and audit metadata. It must ensure that the original data does not leave the site and the summary privacy budget is within the threshold. In case of anomalies such as excessively high missing rates or severe unit conflicts, the generation of that batch of summaries should be blocked and a rectification push should be initiated. Sp2's input is multi-site anonymized summaries, and the output is a global causal knowledge graph, version metadata, and edge stability statistics. Individual site contributions must not be exposed during the aggregation process. Conflicting edges should use consistent and replayable adjudication rules. The aggregation results must pass regression testing and consistency verification before being distributed. Sp3's input is the graph version, individual data, and constraint configuration, and the output is an individualized strategy, a set of alternative actions, and a causal evidence chain. Before generating the strategy, an offline security assessment and robustness assessment must be completed, and assessment evidence must be retained within the site. Strategies that fail to meet the assessment standards cannot be deployed. Sp4 takes policy and patient communication capabilities and site execution resources as inputs, and outputs action execution records, failure and retry records, result backfilling, and policy effectiveness evaluation. The system needs to trigger alarms and automatically switch to alternative actions when action failures reach a threshold, and update individual profiles and the priors for the next policy after backfilling. Sp5 takes key events and model and graph change information as inputs, and outputs on-chain evidence records and governance actions triggered by contracts and completion confirmations. Only summaries are stored on-chain, while detailed evidence that can be authorized for review is retained off-chain. Withdrawal of governance requires an end-to-end traceability list and completion status.

[0025] The quality control and measurement system requires standards to be frozen at the project initiation stage and maintained throughout the entire trial cycle. Data quality thresholds include upper limits for site-level missing rate, unit consistency error, duplicate event rate, and time conflict rate. Sites exceeding these thresholds are suspended from aggregation and undergo remediation. Model and strategy performance indicators are divided into two categories: clinical and operational effectiveness, and governance and safety effectiveness. The former includes the relative improvement in follow-up completion rate, median lead time for early detection of adverse events, relative reduction in loss to follow-up rate, change in follow-up cost per subject, and improvement in patient experience scores. The latter includes privacy budget consumption, on-chain audit coverage, withdrawal response latency, constraint violation rate, strategy stability, and explainability coverage. A phased rollout strategy is adopted for deployment with set safety thresholds, including constraint violation rate, model drift index, and audit anomaly rate. Any exceeding these thresholds will automatically roll back to the previous stable version and trigger manual review. Anomaly and boundary handling need to cover situations such as sudden data gaps, long-term site offline, loss of patient communication, wearable data drift, and regulatory inspections. The system provides a degradation strategy, prioritizing switching to historically stable and less burdensome action combinations, while recording the reasons for degradation and recovery time on the blockchain. Specific Implementation Example 2: like Figures 1 to 4As shown, based on the technical solution of Specific Implementation 1, homomorphic encryption is used to complete the digest merging in the federated aggregation stage and reduce the dependence on differential privacy noise. This can improve the accuracy and interpretability of the global graph in high-security scenarios, but it brings higher computational latency and key governance complexity, and puts higher requirements on site computing power and network. It is suitable for alliances with large sample sizes and dedicated security operation and maintenance teams. Specific Implementation Example 3: like Figures 1 to 4 As shown, based on the technical solution of Specific Implementation Example 1, a hierarchical strategy is adopted instead of a single-layer strategy in the strategy generation stage. The high layer determines the major categories of follow-up frequency and reach methods, while the low layer selects specific actions and execution plans based on individual profiles and real-time status. This solution has better scalability and stability under multiple objectives and strict resource constraints, but it requires the simultaneous construction of two sets of causal simulations and two-level evaluation links, resulting in a longer development and acceptance cycle. It is more suitable for long-term follow-up projects across disease subgroups and multi-institutional collaborative research scenarios. Specific Implementation Example 4: Based on the technical solution of Specific Implementation Example 1, further use cases are given: The project aims to shorten the latency of adverse event detection and improve follow-up completion rates in a multi-center phase III clinical trial, covering five provincial-level tertiary hospitals and two oncology centers. Regarding data and access, the site-side unified variable dictionary includes baseline tumor stage, past treatment history, immune-related laboratory indicators, concomitant medications, imaging assessment conclusions, symptom entries, medication adherence, key points of intelligent voice follow-up transcription, follow-up action types, and result tags. The time base is a unified timezone millisecond timestamp. All raw data undergoes unit conversion and consistency verification within the site before being written to the local repository and generating site audit metadata. In the causal federation construction process, the site identifies directed relationships such as hormone replacement therapy with rash occurrence and liver enzyme fluctuations in its local causal structure learning unit. It outputs edge direction, strength, confidence, residual summary, data validity window, and algorithm configuration summary. Subsequently, after privacy pruning and noise perturbation, an irreversible statistical summary is generated, signed, and uploaded. The central coordination side performs a three-factor weighted aggregation based on sample size, site data quality score, and historical consistency score. Edges with conflicting strength or opposing directions are temporarily excluded from the dispute set. A global graph and version metadata are generated, and a lightweight index is distributed for inference. Simultaneously, cumulative privacy budget and edge stability statistics are recorded and written to the chain. For strategy training and deployment, each site's causal simulator receives the global graph and individual profiles. It outputs short-term and medium-term outcome distributions for five types of actions: SMS reminders, telephone follow-ups, video consultations, on-site follow-ups, and follow-up extensions. These distributions include the risk of immune-related adverse events, the probability of follow-up completion, the trend of compliance changes, patient burden scores, and site resource consumption estimates. The uncertainty range and data support for each outcome are also provided. The strategy generation unit aims to minimize adverse event risk and follow-up costs while maximizing completion rates. Hard constraints include legal, ethical, and safety bottom lines, maximum number of home visits, and holiday visit restrictions. Soft constraints include monthly costs and the target range for site workload. Training employs a conservative offline method, outputting individualized action and condition trigger alternatives and fallback paths, as well as trigger thresholds. The deployment adopts a small-scale pilot phase followed by gradual rollout. The strategy interpretation unit generates a causal evidence chain for each action, providing the reasons for action selection, the dominant causal path, the threshold range of key variables, differences from alternative actions and turnover costs, confidence level, and data support. The evidence chain summary, graph version number, and time information are written to the site audit log and stored on-chain. Regarding auditing and governance, informed consent signing and modification, strategy generation and adoption, follow-up execution and result backfilling, model release and rollback, graph version changes, and site qualification status changes all form a minimum necessary information set written to the permissioned chain. The smart contract maintains the consent status and terms version and, upon broadcasting a withdrawal event, requires the site to immediately exclude the subject from any aggregation and training. Model weight rollback and recalculation are completed within the first feasible timeframe. Central coordination automatically removes relevant summaries in the next round, and all actions generate verifiable on-chain proofs.Regarding evaluation metrics and results, the primary endpoints were the median lead time for adverse event detection and the follow-up completion rate. Secondary endpoints included the false positive rate, cost per follow-up subject, patient experience score, and loss to follow-up rate. Governance metrics included privacy budget consumption, on-chain audit coverage, withdrawal response latency, and constraint violation rate. Two consecutive months of comparative analysis showed that the experimental group significantly improved follow-up completion rate, shortened median lead time, kept the false positive rate within the preset threshold, reduced unit cost, increased patient experience score, and achieved a closed-loop withdrawal governance process from triggering to full-chain completion within one hour, with 100% on-chain event coverage. For anomalies and risk control, degradation strategies were implemented for scenarios such as sudden data gaps, site offline events, loss of patient communication, wearable device drift, and regulatory inspections. Priority was given to switching to historically stable and less burdensome action combinations, and the reasons for degradation and recovery time were recorded on-chain. If the constraint violation rate exceeded the threshold or the model drift index exceeded the limit, the system automatically rolled back to the previous stable version and triggered manual review. Reusable deliverables include site variable dictionaries and site audit metadata templates, edge statistics summary protocols and aggregation adjudication rules, graph version difference reports and strategy evaluation templates, and withdrawal governance operation guidelines.

[0029] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising a reference structure" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0030] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A clinical trial intelligent follow-up management system based on artificial intelligence and big data, characterized in that: include: The causal federated knowledge graph module is used to construct a local causal relationship skeleton based on local multi-source data at each clinical trial center, and upload the causal statistical information to the central aggregation node after encryption and differential privacy processing. The central aggregation node forms a global causal knowledge graph, and the global causal knowledge graph is then distributed to each clinical trial center. The risk perception follow-up scheduling module is used to construct a patient-level causal simulator based on the global causal knowledge graph, generate follow-up strategies under resource constraints, ethical constraints and patient compliance constraints using a conservative offline reinforcement learning method, and output causal evidence chains and explanatory information for each recommended action during follow-up execution. The privacy and compliance audit module is used to generate event summaries for key events such as patient informed consent, follow-up strategy generation, follow-up execution results, and model updates throughout the follow-up process. These summaries are then stored in a permissioned blockchain using hashing. Smart contracts are used to govern patient informed consent and withdrawal, ensuring that the system is auditable and tamper-proof.

2. The intelligent follow-up management system for clinical trials based on artificial intelligence and big data according to claim 1, wherein the causal federated knowledge graph module comprises: Local causal structure learning unit, used to extract causal relationships and their confidence levels between variables from local clinical trial data; The side statistical summary generation unit is used to encrypt and perform differential privacy processing on the causal relationship strength and residual distribution; The Federation Aggregation Unit is used to merge causal statistical summaries from multiple clinical trial centers into a global causal skeleton via a secure aggregation protocol.

3. The intelligent follow-up management system for clinical trials based on artificial intelligence and big data according to claim 1, wherein the risk perception follow-up scheduling module comprises: A causal simulator is used to generate potential outcome sequences for patients under different follow-up actions based on a global causal knowledge graph. The policy generation unit is used to perform constrained conservative reinforcement learning training on the causal simulator to obtain a follow-up action selection policy. The strategy interpretation unit is used to generate a causal evidence chain for each follow-up action and store it in the audit log.

4. The intelligent follow-up management system for clinical trials based on artificial intelligence and big data as described in claim 1, wherein the privacy and compliance audit module includes: The differential privacy budget management unit is used to record and display the privacy loss when each clinical trial center uploads causal statistical summaries; A blockchain-based evidence storage unit is used to write follow-up event summaries into a permissioned blockchain using hash values ​​and timestamps. Smart contract units are used to automatically trigger governance processes such as data isolation, model weight rollback, or retraining when a patient withdraws informed consent.

5. The intelligent follow-up management system for clinical trials based on artificial intelligence and big data as described in claim 1, wherein the follow-up strategies include SMS reminders, telephone follow-ups, video consultations, on-site follow-ups, and follow-up extensions, and the system selects the optimal action based on the prediction results of the causal simulator when outputting the strategy.

6. A method for intelligent follow-up management of clinical trials based on artificial intelligence and big data, characterized in that: Includes the following steps: Sp1. At each clinical trial center, a causal framework is constructed based on local multi-source data, a causal statistical summary is generated, and after encryption and differential privacy processing, it is uploaded to the central aggregation node. Sp2. At the central aggregation node, causal statistical summaries from each clinical trial center are merged through a secure multi-party computation mechanism to generate a global causal knowledge graph and distribute it to each clinical trial center. Sp3. Patient-level causal simulators are constructed at each clinical trial center based on the global causal knowledge graph, and constrained conservative offline reinforcement learning training is performed on the simulators to obtain follow-up strategies. Sp4. During the actual follow-up process, follow-up actions and causal evidence chains are generated based on the follow-up strategy and output to researchers or executed automatically by the system. Sp5. Throughout the follow-up process, key events are summarized and stored via a permissioned blockchain, and patient informed consent and withdrawal are managed based on smart contracts.

7. The intelligent follow-up management method for clinical trials based on artificial intelligence and big data as described in claim 6, wherein the constrained reinforcement learning training in Sp3 uses follow-up costs, patient burden and ethical constraints as optimization constraints to ensure that the generated follow-up strategy operates within a safe boundary.

8. The intelligent follow-up management method for clinical trials based on artificial intelligence and big data as described in claim 6, wherein the smart contract in Sp5 includes informed consent writing logic, withdrawal triggering logic, and audit access control logic, which are used to ensure that the authenticity of follow-up events is verifiable and that the patient's original data is not leaked.

9. The intelligent follow-up management method for clinical trials based on artificial intelligence and big data according to claim 6, wherein the global causal knowledge graph is updated according to a preset cycle, and the graph is calibrated after each round of updates to ensure its consistency with the latest clinical trial data.

10. The intelligent follow-up management system for clinical trials based on artificial intelligence and big data according to claim 1 or the intelligent follow-up management method for clinical trials based on artificial intelligence and big data according to claim 6, wherein the causal evidence chain includes the reasons for the selection of follow-up actions, key causal paths, and simulation prediction results, used to support regulatory audits and scientific research interpretation.

Images