A Clinical Quality Risk Assessment Method for Medical Devices Based on Multi-Source Data Fusion

By integrating multi-source data and analyzing risk propagation, the problem of duplicate recording of the same basic issues in the clinical quality risk assessment of medical devices was solved, thus achieving accuracy in risk assessment and the ability to update data after changes.

CN123120164APending Publication Date: 2026-11-17SUZHOU KANGTAI TIANNIAN TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611410511.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-11
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing clinical quality risk assessment methods for medical devices are unable to accurately identify duplicate records of the same underlying problem, and are difficult to effectively update risk assessment results after data changes, resulting in inaccurate assessment results and increased processing volume.

Method used

By fusing multi-source data, clinical trial data from different sources are transformed into unified clinical quality event objects, homologous evidence clusters are identified, risk propagation subgraphs are constructed, positive shadow evaluation and counterfactual ablation are performed, and combined with echo ablation and time alignment perturbation verification, risk clusters that are confirmed and pending review are located and locally recalculated.

Benefits of technology

It reduces the duplication of the same quality issues in risk assessment, preserves the original source and dissemination path of risk evidence, ensures the accuracy and traceability of assessment results, and enables targeted updates after data changes.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123120164A_ABST
    Figure CN123120164A_ABST
Patent Text Reader

Abstract

This invention discloses a method for assessing the clinical quality risk of medical devices based on multi-source data fusion, specifically relating to the field of clinical quality assessment technology. It involves acquiring multi-source data from clinical trials of medical devices and constructing a unified clinical quality event object. Risk evidence pointing to the same underlying problem is merging from the same source. A risk propagation subgraph is constructed by combining subjects, research centers, medical device instances, and key quality processes. Candidate risk states are formed based on propagation support and impact. Confirmed risk clusters are identified through positive shadow evaluation, counterfactual ablation, echo ablation, and time perturbation verification, and risk levels are established. When manual review or source data correction occurs, the affected risk subgraph is locally recalculated, achieving traceable assessment and dynamic updating of clinical quality risks.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of clinical quality assessment technology, and more specifically, to a method for clinical quality risk assessment of medical devices based on multi-source data fusion. Background Technology

[0002] Clinical trials of medical devices continuously generate data from various sources, including subject visit records, device allocation and usage records, device operation or malfunction logs, adverse event records, protocol deviation records, data challenges, and monitoring findings. Because this data is scattered across different operational systems, the recording time, data format, and information granularity are not entirely consistent. The same clinical quality issue may also appear simultaneously in case records, challenge records, and monitoring records. Current quality risk assessments, if based solely on the number of abnormal records or individual indicators, are prone to double-counting the same underlying issues and fail to accurately reconstruct the actual connection between abnormalities and subjects, medical device instances, and key quality processes.

[0003] On the other hand, not all clinical quality anomalies have the same impact. Some anomalies are simply duplicates of upstream problems in different systems, while others may continue to have an impact along clinical processes such as device use, key endpoint collection, or adverse event reporting. If risk levels are determined solely based on fixed thresholds or composite scores, it is often difficult to identify anomalies that truly contribute independently to the overall quality risk. Recalculating all clinical trial data after source data has been reviewed, corrected, or supplemented increases the workload of evaluation and affects already confirmed risk outcomes. Therefore, a method for assessing the clinical quality risk of medical devices is needed that can distinguish between duplicated evidence from the same source, preserve the actual risk propagation relationships, and re-evaluate the affected scope after data changes. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method for clinical quality risk assessment of medical devices based on multi-source data fusion.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for assessing the clinical quality risk of medical devices based on multi-source data fusion includes the following steps:

[0007] S1. Obtain multi-source clinical data from medical device clinical trials, determine key quality processes based on the clinical trial protocol, and convert multi-source clinical data into unified clinical quality event objects.

[0008] S2. Identify initial risk evidence from clinical quality events, merge initial risk evidence pointing to the same basic problem to form homogeneous evidence clusters, construct risk propagation subgraphs based on homogeneous evidence clusters and form candidate risk propagation clusters;

[0009] S3. Determine the candidate risk loads of critical quality processes based on the propagation support and impact of candidate risk propagation clusters, and form candidate risk states;

[0010] S4. Perform positive shadow evaluation and counterfactual ablation on candidate risk propagation clusters, determine the marginal risk contribution based on the change in the overall candidate evaluation quantity before and after ablation, and determine the confirmed risk clusters and risk clusters to be reviewed by combining echo ablation and time alignment perturbation verification.

[0011] S5. Based on the confirmed risk clusters and their risk levels, when manual review or source data correction causes changes in risk evidence, locate the affected risk subplots and perform local recalculation of the affected risk subplots.

[0012] Specifically, in S1, the multi-source clinical data includes data from electronic data acquisition systems, electronic source data or structured source data indexes, subject visit records, medical device allocation and usage records, device operation or malfunction logs, adverse event and serious adverse event records, protocol deviation records, laboratory or independent evaluation results, data questioning and closure records, monitoring findings, audit rectification records, and basic operational information of the research center.

[0013] The critical quality processes are defined at least from informed consent and enrollment eligibility, medical device allocation and use, collection of primary endpoints or key performance indicators, adverse event reporting, key follow-ups, source data traceability, and other critical steps specified in the clinical trial protocol.

[0014] The clinical quality event object includes at least the event identifier, research center identifier, subject trial number, device instance identifier, critical quality process identifier, event type, occurrence time or time interval, recording time, data value or status, source system, source record index, version number, and verification status.

[0015] Specifically, in S1, when the recording time of the same event differs in different systems, the actual occurrence time is used as the primary time anchor; when the actual occurrence time cannot be obtained, the source recording time closest to the actual business occurrence is used as the alternative time anchor, and the alternative time flag is retained.

[0016] The clinical quality event objects are subjected to structural consistency processing, and the units, date formats, enumeration values ​​and device identifiers are mapped according to the clinical trial data dictionary;

[0017] When a key field is missing, the original missing state is preserved and a key field missing event is generated, without using adjacent subject data, research center averages or model predictions to fill in the clinical facts.

[0018] For duplicate imports of the same version record, deduplication is performed based on the source system, record primary key, and version number; for new and old version records formed after legitimate correction, the versions before and after correction and the corresponding audit trail are retained.

[0019] Specifically, in step S2, the initial risk evidence is identified from the clinical quality event objects according to the rules corresponding to the key quality processes;

[0020] When two initial risk pieces of evidence have the same research center and subject trial number, point to the same medical device instance or the same critical quality process, the event occurs within the same allowable window of the corresponding critical quality process, and one of the initial risk pieces of evidence can be traced back to the other initial risk piece of evidence through audit trail, challenge number, protocol deviation number, monitoring discovery number or source record index, the two initial risk pieces of evidence will be grouped into the same cluster of homologous evidence.

[0021] For initial risk evidence that cannot be connected by explicit references, the corresponding initial risk evidence is marked as pending merging only when the event type is consistent, the core data items are consistent, and the time difference does not exceed the event alignment tolerance of the corresponding key quality process. Representative evidence is determined after the consistency verification is passed.

[0022] Specifically, in S2, the nodes of the risk propagation subgraph include the homologous evidence cluster, subject, research center, medical device instance, and key quality process;

[0023] Based on the association between the homologous evidence clusters and subjects, medical device instances and critical quality processes, the relationship of multiple subjects sharing the same medical device batch or the same center process, and the data source relationship of the previous critical quality process for the next critical quality process, a propagation edge is established. Each propagation edge records the relationship type, applicable time interval and relationship reliability.

[0024] Propagation is initiated from each of the aforementioned homologous evidence clusters to directly associated critical quality processes, and extended to subsequent critical quality processes in the order permitted by the clinical trial protocol. Propagation support is determined based on the path reliability and time matching of the propagation path.

[0025] When there is no effective transmission path that conforms to the business sequence of the clinical trial protocol, the corresponding transmission support is zero.

[0026] Specifically, in S3, the impact of the candidate risk propagation cluster is determined based on the type of critical quality process that the candidate risk propagation cluster ultimately points to, whether it affects subject protection, whether it affects the primary endpoint or critical performance data, whether it changes the actual use process of the medical device, whether it compromises the traceability of source data, and whether it forms a recurrence trend.

[0027] Using the homologous evidence clusters as the smallest scoring unit, the candidate risk loads of the corresponding critical quality processes are determined based on the propagation support and influence of the candidate risk propagation clusters that can reach the critical quality processes through effective propagation paths.

[0028] The candidate risk status is formed according to three levels: subjects, research centers, and key quality processes;

[0029] When multiple candidate risk propagation clusters are determined to originate from the same underlying problem through the merging of evidence from the same source, only the representative risk cluster is retained for one risk assessment, and the remaining records serve as supporting evidence for the source of the representative risk cluster.

[0030] For predefined unacceptable quality events, they are identified as hard risk events, which are not offset by the weighted results of other low-risk indicators.

[0031] Specifically, in S4, a positive shadow evaluation is performed on the candidate risk status in the risk assessment data copy, the current risk propagation subgraph and the corresponding total candidate evaluation quantity are saved, and the original clinical data remains unchanged;

[0032] For each candidate risk propagation cluster, a counterfactual ablation subgraph is constructed. Under the condition of only blocking the propagation edges of the candidate risk propagation cluster to be verified to the critical quality process, the original clinical quality event objects and other candidate risk propagation clusters remain unchanged, and the candidate clinical quality risk assessment is re-executed. The marginal risk contribution of the candidate risk propagation cluster to be verified is determined based on the change in the overall candidate assessment quantity before and after ablation.

[0033] For evidence clusters with multiple source records, keep the representative risk evidence and the direct source record corresponding to the representative risk evidence unchanged, and block the data questioning record, monitoring record or subsequent summary record derived from the representative risk evidence one by one.

[0034] When the marginal risk contribution remains essentially unchanged after masking the corresponding derived record, the corresponding derived record is identified as a repeating echo; when the marginal risk contribution decreases after masking the independent source record, the corresponding independent source record is retained as independent supporting evidence.

[0035] Specifically, in S4, a perturbation range is formed based on the data entry delay, timestamp error and cross-system synchronization error observed in the current clinical trials of medical devices. Under the condition of keeping the clinical fact value unchanged, the event alignment time and the presumed relationship edge are perturbed, and the marginal risk contribution is re-determined after each perturbation.

[0036] When the lower quantile of the marginal risk contribution in the disturbance result is higher than the marginal contribution confirmation threshold, the corresponding candidate risk propagation cluster is determined as the confirmed risk cluster.

[0037] When a candidate risk propagation cluster generates a marginal risk contribution higher than the marginal contribution confirmation threshold only under a specific time alignment method, the corresponding candidate risk propagation cluster is identified as a risk cluster to be reviewed, and the corresponding disputed time anchor or propagation edge is output.

[0038] Specifically, in step S5, the risk load of key quality processes and the overall quality risk assessment quantity are re-determined based on the identified risk clusters, and low-risk, medium-risk, high-risk, or other risk levels pre-configured for medical device clinical trials are formed.

[0039] When monitors, data administrators, or researchers complete source data verification for the aforementioned risk clusters to be reviewed and return correction results, they locate the affected risk subplots that have a direct relationship or propagation relationship with the clinical quality events that have been corrected;

[0040] Keep the confirmed risk clusters that have been identified and have no effective relationship with the affected risk subgraph unchanged, and re-perform homology evidence merging and risk propagation subgraph construction, candidate clinical quality risk assessment, and counterfactual ablation and marginal risk verification on the affected risk subgraph to update the homology merging relationship, propagation path, marginal risk contribution, and risk level;

[0041] If a stable propagation relationship cannot be formed after two consecutive local reviews of the same risk cluster due to source conflict or time anchor conflict, the corresponding risk cluster will be identified as a manually judged state, and the data source, field, time point and propagation edge of the conflict will be output.

[0042] The technical effects and advantages of this invention are as follows:

[0043] This invention transforms clinical trial records from different sources into clinical quality event objects with unified identifiers and source information, and identifies key quality processes based on currently effective clinical trial protocols. For abnormal records pointing to the same underlying problem, they are grouped together based on source relationships, subjects, device instances, time windows, and audit information. Risk propagation relationships are then established around subjects, research centers, medical device instances, and key quality processes. This reduces the repeated inclusion of the same quality problem in risk assessments due to its recurrence in multiple systems, while preserving the original source and propagation path of various risk evidences.

[0044] After candidate risks are identified, this invention examines the marginal contribution of each risk cluster to the overall quality risk through positive shadow assessment and counterfactual ablation, and distinguishes between confirmed risks, duplicate echoes, and risks requiring further verification by combining echo ablation and time alignment perturbation. When manual review or source data correction occurs, only the affected risk subplots that are directly related to or propagated from the correction event are located. The relevant parts are then re-merged, propagated, and validated, while other confirmed risk results remain unchanged. This allows the risk assessment results to be updated in a targeted manner as clinical data changes, and maintains the traceability of the assessment basis and risk sources. Attached Figure Description

[0045] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0046] 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.

[0047] like Figure 1 As shown, the steps of the medical device clinical quality risk assessment method based on multi-source data fusion are as follows:

[0048] S1. Construction of Multi-Source Clinical Quality Event Objects: This is used to establish a unified data foundation and trial-specific evaluation boundaries required for subsequent risk assessment. Based on currently effective clinical trial protocols, key quality processes are identified. Clinical trial records from different sources are converted into clinical quality event objects with unified identification and traceable origins, and time, unit, version, and missing status are standardized. For key clinical facts that cannot be obtained, imputation values ​​are not used; the missing status itself is retained as a basis for subsequent quality risk identification. The process is as follows:

[0049] This method acquires the clinical trial protocol, monitoring plan, data management plan, and approved list of key quality elements for the target medical device clinical trial. It also reads data from authorized electronic data acquisition systems, electronic source data or structured source data indexes, subject visit records, medical device allocation and usage records, device operation or malfunction logs, adverse event and serious adverse event records, protocol deviation records, laboratory or independent evaluation results, data challenge and closure records, monitoring findings, audit rectification records, and basic operational information of the research center. Data involving subject identity is replaced with the trial subject number or irreversible mapping identifier before being included in the evaluation method. This method does not modify the original clinical records; it only preserves the data source identifier, version identifier, and reference location at the risk assessment layer.

[0050] Key quality processes are determined based on the clinical trial protocol and its currently effective revisions. Key quality processes must be identified from at least the following: informed consent and enrollment eligibility, medical device allocation and use, collection of primary endpoints or key performance indicators, adverse event reporting, key follow-ups, source data traceability, and other key steps specified in the protocol. For each key quality process, the applicable subjects, the required visits or timeframes, the required data items, the permissible sequence of procedures, the corresponding medical device example, and the permissible data modification paths must be recorded. Key quality processes are determined by the current trial protocol and its revisions, and general templates do not replace specific trial requirements; after the protocol revision takes effect, newly occurring events are evaluated according to the new version of the rules, while events occurring before the revision retain the original version of the rules.

[0051] Different data sources are converted into unified clinical quality event objects. Each event object includes at least the event identifier, research center identifier, subject trial number, device instance identifier, critical quality process identifier, event type, occurrence time or time interval, recording time, data value or status, source system, source record index, version number, and verification status. For cases where the recording time of the same event differs across systems, the actual occurrence time is used as the primary time anchor; if the actual occurrence time cannot be obtained, the source record time closest to the actual business occurrence is used as the alternative time anchor, and the "alternate time" flag is retained to avoid misinterpreting subsequent entry times as the clinical event occurrence time.

[0052] For standardized clinical quality event objects, a consistent structure is applied, and no fabricated clinical data is used. Units, date formats, enumerated values, and device identifiers are mapped according to the clinical trial data dictionary. When a critical field is missing, instead of using adjacent subjects, center averages, or model predictions to fill in the gaps, a "critical field missing" event is generated, and the original missing state is preserved. For duplicate imports of the same version of the record, deduplication is performed based on the source system, record primary key, and version number. For new and old version records formed after legitimate corrections, both the original and revised versions, as well as the audit trail, are retained, enabling subsequent steps to distinguish between genuine data changes and duplicate uploads.

[0053] S2. Homogeneous Evidence Consolidation and Risk Propagation Subgraph Construction: Initial risk evidence is identified based on standardized clinical quality event objects, and records from different systems that point to the same underlying problem are consolidated using homogeneous evidence clusters as the smallest risk unit. A risk propagation subgraph is constructed by combining the actual relationships between subjects, medical device instances, research centers, and key quality processes, and then candidate risk propagation clusters are formed along relationships that conform to the protocol's business sequence. This process preserves the source and propagation path of risk while avoiding duplicate scoring of the same underlying problem due to repeated records in multiple systems. The process is as follows:

[0054] Initial risk evidence is identified from standardized event objects according to the key quality process rules. Initial risk evidence includes, but is not limited to: missing required records at specified time points; events occurring in a sequence that violates protocol constraints; mismatch between device instances and subjects or visits; values ​​of the same clinical fact exceeding permissible consistency boundaries in two sources; serious adverse event reports exceeding protocol or management deadlines; unexplained modifications to key endpoint records; repeated initiation of similar data challenges; inconsistencies between device malfunction records and device status in case report forms; and unclosed discrepancies between monitoring findings and data records. Each initial risk piece of evidence references the aforementioned standardized clinical quality event object and is not stored separately from its original source.

[0055] For risk evidence that may describe the same underlying issue, homogeneous merging is performed. If two pieces of risk evidence have the same research center and subject trial number, point to the same device instance or the same critical quality process, and the events occurred within the same allowable window of the corresponding process, and one record can be traced back to the other through audit trails, challenge numbers, protocol deviation numbers, monitoring discovery numbers, or source record indexes, then they are grouped into the same homogeneous evidence cluster. For evidence that cannot be linked by explicit citations, it only enters the pending merging state if the event type is consistent, the core data items are consistent, and the time difference does not exceed the event alignment tolerance of the critical quality process; the pending merging state does not directly delete any evidence, but rather determines representative evidence after subsequent consistency verification.

[0056] For each cluster of evidence from the same source, an evidence credibility is determined. This is used to reflect whether the evidence has a clear source, cross-source consistency support, and the completeness of necessary fields: In the formula, Indicates a cluster of evidence from the same source; This indicates the traceability of the source. A higher value is taken when the evidence can be located to a specific source system, record the primary key and version, and a lower value is taken when the source index is missing. It represents cross-source consistency support, which is obtained based on the proportion of support for the same basic fact from independent data sources within the same evidence cluster. The completeness of the evidence is indicated by the actual usable proportion of the fields required for risk assessment. , and The weights are non-negative and sum to 1. Weights are determined using historically reviewed quality issues as the calibration sample, labeled with "whether it is a genuine quality issue or not." Constrained regression or ranking is used to assign higher credibility to confirmed issues. When historical samples are insufficient, equal weights are used, only those issues are weighted. The degree of reliability of the evidence does not determine the risk level independently.

[0057] A risk propagation subgraph is constructed based on common evidence clusters. Graph nodes include at least common evidence clusters, subjects, research centers, medical device instances, and critical quality processes. Edges are established based on relationships such as "evidence belongs to a subject," "evidence is associated with a device instance," "evidence triggers a critical quality process," "multiple subjects share the same device batch or the same center process," and "the preceding critical quality process is the data source for the following critical quality process." Each edge records the relationship type, applicable time interval, and relationship reliability. Relationship reliability is primarily determined directly by the trial protocol, device allocation records, and audit trails; relationships established only through temporal proximity or field similarity are marked as presumptive relationships and assigned a lower reliability than direct relationships.

[0058] Propagation proceeds from each cluster of evidence originating from the same source to directly related critical quality processes, and extends to subsequent critical quality processes within the business sequence allowed by the solution, forming candidate risk propagation clusters. Evidence Cluster Key quality processes dissemination support Obtained according to the following relationship: In the formula, Indicates from the same source evidence cluster To key quality processes The reliability of the relationship path is determined by selecting the most reliable effective path when multiple paths exist. The path reliability is obtained by multiplying the reliability of each edge on the path. Therefore, the support will naturally decrease after propagation through presumed relationships or too many intermediate nodes. This indicates the time matching degree. Within a specified time window for the same critical quality process, it takes the value 1. Beyond the time window, it decreases from 0 to 1 based on the actual interval and the allowable delay range for that process. If no valid path conforms to the business sequence of the solution, then... Setting it to 0 prevents the risk from spreading across processes with no business relationship simply because of field relevance.

[0059] S3. Candidate Clinical Quality Risk Assessment: The quality risk of candidate risk propagation clusters is quantified, without directly substituting the number of abnormal records for the degree of risk. The system combines risk propagation support, risk cluster impact, and the importance of key quality processes to form candidate risk loads for key quality processes, and establishes candidate risk statuses at three levels: subjects, research centers, and key quality processes. For predefined unacceptable quality events, they are directly retained as hard risk events without being offset by weighted results from other low-risk indicators. The process is as follows:

[0060] Determine the impact for each candidate risk propagation cluster. Impact is not determined directly by the number of anomalies, but rather by mapping the risk cluster to the type of critical quality process it ultimately points to, whether it affects subject protection, whether it affects the primary endpoint or key performance data, whether it alters the actual use of the medical device, whether it compromises source data traceability, and whether a recurrence trend has emerged. The impact mapping table is developed during the trial initiation phase based on the clinical trial protocol, risk management documents, and monitoring plan, and is confirmed by quality management personnel; the same rule remains unchanged within the same protocol version.

[0061] Calculate candidate risk loads for each critical quality process. First, the clusters of homogeneous evidence obtained from the risk propagation subgraph construction step are used as the smallest scoring unit to avoid the same basic issue being scored repeatedly in electronic case report forms, data challenge, and monitoring records; then, only those that can reach key quality processes through effective propagation paths are accumulated. Risk cluster contribution: In the formula, the summation object is the critical quality process. Different clusters of evidence of homology; To increase support for dissemination; Impact degree of risk clusters; This value is used to limit the results to a range of 0 to 1. It is used solely as a measure of quality risk and is not interpreted as the probability of a clinical event or medical outcome. If multiple risk clusters are determined to originate from the same underlying problem through merging of similar evidence, only the representative risk cluster is retained in the summation once, and the remaining records serve as supporting evidence for the source of that risk cluster.

[0062] Candidate risk statuses are established at three levels: subjects, research centers, and key quality processes. The subject level is used to locate individual process anomalies; the research center level is used to identify whether the same quality process is repeated in multiple subjects; and the key quality process level is used to determine whether the risk is concentrated in areas such as informed consent, device use, key endpoints, serious adverse event reporting, or data traceability. The overall candidate evaluation count for research centers is also considered. The risk load of four key quality processes—subject protection, integrity of critical endpoints, compliance with device use, and data traceability—is summarized separately. The weights of each category are determined by the degree of criticality or the impact level of historically confirmed issues in the trial risk management document. If any predefined unacceptable quality event is triggered, it is directly marked as a hard risk event and cannot be offset by the weighted results of other low-risk indicators.

[0063] Key quality processes exceeding the corresponding candidate thresholds and their risk propagation clusters are passed to the counterfactual ablation and marginal risk verification steps. Candidate thresholds are formed from confirmed issues accumulated from historically manually reviewed clinical trials of similar medical devices or from the ongoing rolling accumulation of current trials: issues requiring confirmed quality action are addressed accordingly. The distribution takes the low quantile boundary as the candidate threshold and uses the sample of problems without quality measures to check for false triggering. When the sample is insufficient, the numerical threshold is not arbitrarily set. Instead, all objects involving hard risk events or at least having a high-impact and traceable risk cluster are entered into counterfactual ablation and marginal risk verification.

[0064] S4. Counterfactual Ablation and Marginal Risk Validation: This step verifies whether candidate risk clusters have an independent and stable contribution to overall clinical quality risk. While preserving the original clinical data, a positive shadow assessment is first performed. Then, the propagation effect of each candidate risk cluster is masked to construct counterfactual ablation results. The marginal risk contribution is determined by changes in the assessment quantity. Simultaneously, echo ablation and time alignment perturbation validation are combined to distinguish between duplicate echoes, relationship conflicts, risk clusters awaiting review, and confirmed risk clusters. The process is as follows:

[0065] A positive shadow assessment is performed on the candidate risk statuses obtained from the candidate clinical quality risk assessment steps. The positive shadow assessment only runs on the risk assessment data copy and does not delete or overwrite the original clinical data. The system saves the current risk propagation subgraph. and the corresponding overall candidate evaluation quantity And record the key quality processes, subjects, and research centers affected by each candidate risk cluster.

[0066] For each candidate risk cluster Constructing counterfactual ablation subgraphs .exist Only risk clusters are blocked in the middle. The edges propagate to critical quality processes, while preserving the original event objects and other risk clusters, and then the candidate clinical quality risk assessment is repeated. Risk clusters are then calculated. Marginal risk contribution : In the formula, Indicates the retention of risk clusters The overall candidate evaluation quantity at that time Indicates ablation risk cluster Evaluation quantity after the dissemination effect. A value greater than 0 indicates that the risk cluster makes a positive contribution to the current quality risk; A value close to 0 indicates that although the risk cluster exhibits anomalies, its risk information has been covered by other risk clusters of the same origin or upstream. When the value is less than 0, the risk cluster is marked as a relational conflict and the evidence of the same origin is returned and merged into the risk propagation subgraph construction process to verify the propagation path. It is not directly used as evidence to reduce risk.

[0067] For clusters of evidence from the same source with records from multiple sources, "echo ablation" is performed. The representative risk evidence and its direct source records remain unchanged; only the questioning records, monitoring records, or subsequent summary records derived from the representative evidence are individually masked. If these derived records are masked... If the fundamentals remain unchanged, the corresponding derived records are identified as repeating echoes, used only to enhance traceability without increasing the risk load; if the marginal contribution decreases significantly after masking a particular independent source, that source is retained as independent supporting evidence. This avoids mechanically increasing the risk score for the same issue due to an increase in the number of records.

[0068] Robustness verification of marginal risk contribution was conducted. A perturbation range was established using data entry delays, timestamp errors, and cross-system synchronization errors observed in the current trial. Multiple perturbations were applied to the event alignment time and the presumed relationship edge without altering the clinical fact values; recalculation was performed after each perturbation. .like If the low quantile value in the disturbance results is still higher than the marginal contribution confirmation threshold, then the risk cluster will be... Marked as a confirmed risk cluster; if a high contribution is only generated under a specific time alignment, it is marked as a risk cluster to be reviewed and the disputed time anchor or relationship edge is output. The marginal contribution confirmation threshold is formed by the corresponding samples between historical manually confirmed risk clusters and final quality measures; if no historical samples are available, a different threshold is used. Furthermore, there is no relationship conflict as an initial confirmation condition, and the result is marked as pending calibration.

[0069] Only risk clusters with stable marginal contributions after counterfactual ablation are included in the final risk classification. For For risk clusters with a near-zero repeating echo, their original records are not deleted, but are displayed as "covered by upstream risks" in the final report; for risk clusters that need to be reviewed due to relationship conflicts or uncertain time anchors, the specific source, affected key quality processes, and fields to be checked are sent to the risk classification and partial review rollback steps.

[0070] S5. Risk Classification and Partial Review Rollback: Based on the confirmed risk clusters, the final risk level and traceable risk assessment results are formed, and partial rollback is implemented for changes caused by manual review or source data correction. After the review information is returned, only the affected risk subgraphs that have a direct relationship or propagation relationship with the corrected event are located, and the relevant risks are recalculated while keeping other confirmed risk clusters unchanged. For objects that still cannot form a stable propagation relationship after continuous partial reviews, they are transferred to manual judgment, and the confirmation results that meet the data integrity conditions are written to the historical calibration database. The process is as follows:

[0071] Based on the identified risk clusters, the risk load of key quality processes and the overall quality risk assessment are recalculated, resulting in low-risk, medium-risk, high-risk, or other pre-configured risk levels for the trial. Risk level boundaries are primarily defined using samples of clinical quality issues that have been manually graded and addressed in the past, ensuring that high-level samples cover issues requiring immediate quality intervention or special monitoring. When the current trial lacks sufficient samples, the critical event levels already defined in the trial risk management document are used as level boundaries, without creating custom clinical severity standards.

[0072] The results should be traceable and accountable. Each output should include at least the research center, subject trial number, medical device instance, key quality process, timeframe of risk occurrence, identified risk cluster, source of representative evidence, transmission path, marginal risk contribution, final risk level, and fields requiring manual verification. For research center-level risks, the number of subjects and common key quality processes contributing to the risk at that center should also be output, but the number of subjects alone should not replace the risk contribution assessment.

[0073] When monitors, data administrators, or researchers complete source data verification for the risk clusters to be reviewed and return correction results, they only locate risk subgraphs that have a direct relationship or propagation relationship with the object of the correction. Risk clusters that have been identified and have no valid relationship with the subgraph remain unchanged; for the affected subgraphs, they re-perform homology evidence merging and risk propagation subgraph construction, candidate clinical quality risk assessment, and counterfactual ablation and marginal risk verification, updating homology merging relationships, propagation paths, and marginal risk contributions to avoid recalculating the risk of all research centers due to a single data correction.

[0074] Local review is performed according to preset termination conditions. If a stable propagation relationship cannot be established after two consecutive local reviews of the same risk cluster due to source conflict or time anchor conflict, automatic confirmation is stopped, its status is set to "manual judgment," and the conflicting data source, fields, time points, and relationship edges are output; the manual judgment result is written back, and a local recalculation is performed again. If the manual judgment still cannot determine the clinical facts, the uncertain status is retained, and a high or low risk conclusion is not forced to be generated.

[0075] Risk clusters, marginal contributions, final risk levels, and actual quality measures taken that have been manually verified and have complete data sources are written into the historical calibration database. This database is used to update the evidence credibility weights during the homogeneous evidence merging process, the candidate thresholds during the candidate clinical quality risk assessment process, and the marginal contribution confirmation thresholds during the counterfactual ablation and marginal risk verification process. Data that is only affected by system synchronization failures, missing sources, or temporary manual modifications without complete audit reasons is not included in the automatic calibration sample to avoid erroneous records that could contaminate subsequent risk assessments.

[0076] Critical quality processes are not guessed by algorithms, but are extracted from current medical device clinical trial protocols, risk management documents, monitoring plans, and data management plans, and are linked to the protocol version. If the primary endpoints, usage procedures, or follow-up requirements of different medical device trials differ, corresponding sets of critical quality processes should be developed for each trial.

[0077] Event alignment tolerances are determined based on the allowable time range of the corresponding clinical procedure. For example, visit-related events use the visit window specified in the protocol; device operation-related events use the allowable time precision of the device operation record; data entry and system synchronization delays use the time difference distribution in the current trial's historical audit trajectory to form a high quantile boundary. Different event types do not share a fixed time tolerance.

[0078] , , , , , All are dimensionless evaluation quantities ranging from 0 to 1. , and Describe the reliability of the evidence; and Describe the dissemination of evidence to key quality processes; Describe the key quality impacts. The above evaluation metrics are not intended to be interpreted as probabilities of medical diagnosis, device malfunction, or subject safety events.

[0079] Historical sample calibration should only use quality issues that have been manually reviewed and have clear handling conclusions. When using historical samples across trials, at least the device category, clinical usage procedures, and key quality processes should be comparable; if comparability is not possible, only current trial samples should be used for rolling calibration to avoid directly transferring the risk boundaries of other device trials.

[0080] All hard risk events are configured into clear rules by quality management personnel during the trial initiation phase, based on the protocol and applicable management requirements. Examples include missing informed consent, failure to meet key entry / exit criteria despite acceptance of test equipment, and missing or untraceable records of primary endpoint sources. Once a hard risk event is identified, it cannot be offset by averaging it out with other low-risk dimensions.

[0081] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, ATA hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state ATA hard disk.

[0082] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0083] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0084] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0085] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0086] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0087] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable ATA hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0088] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for clinical quality risk assessment of medical devices based on multi-source data fusion, characterized in that, Includes the following steps: S1. Obtain multi-source clinical data from medical device clinical trials, determine key quality processes based on the clinical trial protocol, and convert multi-source clinical data into unified clinical quality event objects. S2. Identify initial risk evidence from clinical quality events, merge initial risk evidence pointing to the same basic problem to form homogeneous evidence clusters, construct risk propagation subgraphs based on homogeneous evidence clusters and form candidate risk propagation clusters; S3. Determine the candidate risk loads of critical quality processes based on the propagation support and impact of candidate risk propagation clusters, and form candidate risk states; S4. Perform positive shadow evaluation and counterfactual ablation on candidate risk propagation clusters, determine the marginal risk contribution based on the change in the overall candidate evaluation quantity before and after ablation, and determine the confirmed risk clusters and risk clusters to be reviewed by combining echo ablation and time alignment perturbation verification. S5. Based on the confirmed risk clusters and their risk levels, when manual review or source data correction causes changes in risk evidence, locate the affected risk subplots and perform local recalculation of the affected risk subplots.

2. The method for clinical quality risk assessment of medical devices based on multi-source data fusion according to claim 1, characterized in that, In S1, the multi-source clinical data includes data from electronic data acquisition systems, electronic source data or structured source data indexes, subject visit records, medical device allocation and usage records, device operation or malfunction logs, adverse event and serious adverse event records, protocol deviation records, laboratory or independent evaluation results, data questioning and closure records, monitoring findings, audit rectification records, and basic operational information of the research center. The critical quality processes are defined at least from informed consent and enrollment eligibility, medical device allocation and use, collection of primary endpoints or key performance indicators, adverse event reporting, key follow-ups, source data traceability, and other critical steps specified in the clinical trial protocol. The clinical quality event object includes at least the event identifier, research center identifier, subject trial number, device instance identifier, critical quality process identifier, event type, occurrence time or time interval, recording time, data value or status, source system, source record index, version number, and verification status.

3. The method for clinical quality risk assessment of medical devices based on multi-source data fusion according to claim 2, characterized in that, In S1, when the recording time of the same event differs in different systems, the actual occurrence time is used as the primary time anchor; when the actual occurrence time cannot be obtained, the source recording time closest to the actual business occurrence is used as the alternative time anchor, and the alternative time flag is retained. The clinical quality event objects are subjected to structural consistency processing, and the units, date formats, enumeration values ​​and device identifiers are mapped according to the clinical trial data dictionary; When a key field is missing, the original missing state is preserved and a key field missing event is generated, without using adjacent subject data, research center averages or model predictions to fill in the clinical facts. For duplicate imports of the same version record, deduplication is performed based on the source system, record primary key, and version number; for new and old version records formed after legitimate correction, the versions before and after correction and the corresponding audit trail are retained.

4. The method for clinical quality risk assessment of medical devices based on multi-source data fusion according to claim 1, characterized in that, In step S2, the initial risk evidence is identified from the clinical quality event objects according to the rules corresponding to the key quality processes. When two initial risk pieces of evidence have the same research center and subject trial number, point to the same medical device instance or the same critical quality process, the event occurs within the same allowable window of the corresponding critical quality process, and one of the initial risk pieces of evidence can be traced back to the other initial risk piece of evidence through audit trail, challenge number, protocol deviation number, monitoring discovery number or source record index, the two initial risk pieces of evidence will be grouped into the same cluster of homologous evidence. For initial risk evidence that cannot be connected by explicit references, the corresponding initial risk evidence is marked as pending merging only when the event type is consistent, the core data items are consistent, and the time difference does not exceed the event alignment tolerance of the corresponding key quality process. Representative evidence is determined after the consistency verification is passed.

5. The method for clinical quality risk assessment of medical devices based on multi-source data fusion according to claim 4, characterized in that, In S2, the nodes of the risk propagation subgraph include the homologous evidence cluster, subject, research center, medical device instance, and key quality process; Based on the association between the homologous evidence clusters and subjects, medical device instances and critical quality processes, the relationship of multiple subjects sharing the same medical device batch or the same center process, and the data source relationship of the previous critical quality process for the next critical quality process, a propagation edge is established. Each propagation edge records the relationship type, applicable time interval and relationship reliability. Propagation is initiated from each of the aforementioned homologous evidence clusters to directly associated critical quality processes, and extended to subsequent critical quality processes in the order permitted by the clinical trial protocol. Propagation support is determined based on the path reliability and time matching of the propagation path. When there is no effective transmission path that conforms to the business sequence of the clinical trial protocol, the corresponding transmission support is zero.

6. The method for clinical quality risk assessment of medical devices based on multi-source data fusion according to claim 5, characterized in that, In S3, the impact of the candidate risk propagation cluster is determined based on the type of critical quality process that the candidate risk propagation cluster ultimately points to, whether it affects subject protection, whether it affects the primary endpoint or critical performance data, whether it changes the actual use process of the medical device, whether it undermines the traceability of source data, and whether it forms a recurrence trend. Using the homologous evidence clusters as the smallest scoring unit, the candidate risk loads of the corresponding critical quality processes are determined based on the propagation support and influence of the candidate risk propagation clusters that can reach the critical quality processes through effective propagation paths. The candidate risk status is formed according to three levels: subjects, research centers, and key quality processes; When multiple candidate risk propagation clusters are determined to originate from the same underlying problem through the merging of evidence from the same source, only the representative risk cluster is retained for one risk assessment, and the remaining records serve as supporting evidence for the source of the representative risk cluster. For predefined unacceptable quality events, they are identified as hard risk events, which are not offset by the weighted results of other low-risk indicators.

7. The method for clinical quality risk assessment of medical devices based on multi-source data fusion according to claim 6, characterized in that, In step S4, a positive shadow evaluation is performed on the candidate risk status in the risk assessment data copy, the current risk propagation subgraph and the corresponding total candidate evaluation quantity are saved, and the original clinical data remains unchanged. For each candidate risk propagation cluster, a counterfactual ablation subgraph is constructed. Under the condition of only blocking the propagation edges of the candidate risk propagation cluster to be verified to the critical quality process, the original clinical quality event objects and other candidate risk propagation clusters remain unchanged, and the candidate clinical quality risk assessment is re-executed. The marginal risk contribution of the candidate risk propagation cluster to be verified is determined based on the change in the overall candidate assessment quantity before and after ablation. For evidence clusters with multiple source records, keep the representative risk evidence and the direct source record corresponding to the representative risk evidence unchanged, and block the data questioning record, monitoring record or subsequent summary record derived from the representative risk evidence one by one. When the marginal risk contribution remains essentially unchanged after masking the corresponding derived record, the corresponding derived record is identified as a repeating echo; when the marginal risk contribution decreases after masking the independent source record, the corresponding independent source record is retained as independent supporting evidence.

8. The method for clinical quality risk assessment of medical devices based on multi-source data fusion according to claim 7, characterized in that, In S4, a perturbation range is formed based on the data entry delay, timestamp error and cross-system synchronization error observed in the current clinical trials of medical devices. Under the condition of keeping the clinical fact value unchanged, the event alignment time and the presumed relationship edge are perturbed, and the marginal risk contribution is re-determined after each perturbation. When the lower quantile of the marginal risk contribution in the disturbance result is higher than the marginal contribution confirmation threshold, the corresponding candidate risk propagation cluster is determined as the confirmed risk cluster. When a candidate risk propagation cluster generates a marginal risk contribution higher than the marginal contribution confirmation threshold only under a specific time alignment method, the corresponding candidate risk propagation cluster is identified as a risk cluster to be reviewed, and the corresponding disputed time anchor or propagation edge is output.

9. The method for clinical quality risk assessment of medical devices based on multi-source data fusion according to claim 8, characterized in that, In step S5, the risk load of key quality processes and the overall quality risk assessment quantity are re-determined based on the confirmed risk clusters, and low-risk, medium-risk, high-risk or other risk levels pre-configured for medical device clinical trials are formed. When monitors, data administrators, or researchers complete source data verification for the aforementioned risk clusters to be reviewed and return correction results, they locate the affected risk subplots that have a direct relationship or propagation relationship with the clinical quality events that have been corrected; Keep the confirmed risk clusters that have been identified and have no effective relationship with the affected risk subgraph unchanged, and re-perform homology evidence merging and risk propagation subgraph construction, candidate clinical quality risk assessment, and counterfactual ablation and marginal risk verification on the affected risk subgraph to update the homology merging relationship, propagation path, marginal risk contribution, and risk level; If a stable propagation relationship cannot be formed after two consecutive local reviews of the same risk cluster due to source conflict or time anchor conflict, the corresponding risk cluster will be identified as a manually judged state, and the data source, field, time point and propagation edge of the conflict will be output.