Real-world evidence oriented medical device risk monitoring and early warning method and system

Abstract

The application provides a medical device risk monitoring and early warning method and system facing real world evidence, and belongs to the technical field of medical device risk monitoring and early warning, and comprises the following steps: the method divides clinical application into a continuous monitoring stage, and collects multi-source heterogeneous real world data such as running state, clinical application, patient outcome and adverse events; through space-time synchronization registration and risk characteristic quantization correction, the risk correlation of the same batch of devices, patients with the same indication, institutions in the same scene and devices of the same risk type is identified, and comprehensive risk characteristics are fused and generated; a risk evolution time sequence is constructed, and an initial prediction sequence is generated; the target risk sequence meeting the regulatory requirements is obtained by combining historical and real-time data and dynamically correcting; and finally, a risk parameter set, a risk level and a graded early warning are output. The accuracy and timeliness of medical device risk monitoring are improved, and the needs of supervision and clinical prevention and control are met.

Description

Technical Field

[0001] This invention relates to the field of medical device risk monitoring and early warning technology, and in particular to a method and system for medical device risk monitoring and early warning based on real-world evidence. Background Technology

[0002] Currently, post-market risk monitoring of medical devices mainly relies on passive reporting, single-device threshold judgment, and simple statistical analysis, making it difficult to achieve accurate, dynamic, and full-cycle risk management based on real-world evidence. Multi-source heterogeneous data generated in clinical applications, including operational status, clinical application, patient outcomes, and adverse events, lack a unified collection and phased aggregation mechanism, resulting in severe data fragmentation. Existing technologies do not perform spatiotemporal synchronization registration of multi-source data, failing to eliminate collection errors and confidence level differences, and cannot quantitatively correct risk feature points, leading to insufficient accuracy in risk assessment. Furthermore, they can only monitor the status of the target device itself, failing to identify risk associations between devices from the same batch, patients with the same indication, medical institutions in the same setting, and devices of the same risk type, and cannot integrate multi-dimensional risk features, resulting in incomplete risk identification. In addition, existing methods struggle to construct time-series risk evolution sequences and lack trend prediction capabilities; the prediction models are fixed and cannot be dynamically corrected by combining historical and real-time data, failing to meet regulatory accuracy requirements; and a standardized risk parameter system has not been established, failing to output standardized risk levels and graded early warning signals, thus failing to meet the actual needs of medical device safety supervision and clinical risk prevention and control.

[0003] Therefore, this invention proposes a method and system for monitoring and warning of medical device risks based on real-world evidence. Summary of the Invention

[0004] This invention provides a method and system for monitoring and warning of medical device risks based on real-world evidence, in order to solve the aforementioned technical problems.

[0005] This invention provides a method for monitoring and early warning of medical device risks based on real-world evidence, comprising:

[0006] Step 1: Divide the clinical application process of the target medical device into multiple continuous monitoring stages, and collect multi-source heterogeneous real-world data corresponding to the target medical device in each monitoring stage;

[0007] Step 2: Perform spatiotemporal dimension synchronous registration on multi-source heterogeneous real-world data within the same monitoring phase to obtain a fused feature dataset. Calculate the baseline risk value corresponding to a single risk feature point in the fused feature dataset, and correct it based on the acquisition parameters, data confidence, and risk weight to obtain the actual contribution value of a single risk feature point.

[0008] Step 3: Using the target medical device as the core monitoring object, identify the associated monitoring objects with risk correlation, extract the individual risk characteristics of the core monitoring object, the associated risk characteristics of each associated monitoring object, and the real-time risk activity, and fuse them to generate the comprehensive risk characteristics of the target medical device at each monitoring stage;

[0009] Step 4: Collect the time-series risk quantification values ​​corresponding to the acceptance nodes of each monitoring stage, accumulate the time-series risk quantification values ​​to generate a risk evolution time-series sequence, and process the comprehensive risk characteristics of each monitoring stage to generate the initial prediction sequence of the target medical device.

[0010] Step 5: Based on real-world data from the historical monitoring phase that has been completed and accepted, and the corresponding risk prediction results, the initial prediction sequence is dynamically corrected in conjunction with real-time clinical application process data to obtain a target risk sequence that meets the preset regulatory requirements.

[0011] Step 6: Determine the risk parameter set of the target medical device based on the target risk sequence, and output the risk level and graded warning.

[0012] Preferably, the comprehensive risk characteristics of the target medical device at each monitoring stage are fused and generated, including:

[0013] For the fusion feature dataset of each monitoring stage, the core layer risk factors corresponding to the inherent risk characteristics of the target medical device itself and the associated layer risk factors corresponding to the risk characteristics of clinical application, patient group and environmental control are extracted. A risk factor library is constructed and each risk factor is uniquely encoded according to the risk source level. The initial total number of factor populations M is set according to the total number of factors in the risk factor library. The number of effective risk factor dimensions contained in each initial factor population is set to D. D matches the total dimension of the risk feature cluster in a single monitoring stage.

[0014] An initial factor population is randomly generated according to the risk factor hierarchical coding. Based on the real-world dataset of historical adverse events of medical devices, a multi-dimensional pre-evaluation of the fit of each individual in the initial factor population is performed. Invalid individuals with fit below a preset threshold are removed to obtain valid individuals.

[0015] All valid individuals under the same initial factor population are clustered according to the risk optimization objective to obtain multiple population clusters and transform them into a four-dimensional structured matrix to determine the risk optimization objective of the corresponding initial factor population;

[0016] According to the risk optimization objective, the risk contribution dominance levels of effective factors in the corresponding initial factor population are sorted in ascending order to generate a factor sequence and determine the crossover step size of adjacent effective factors.

[0017] Based on the cross step size analysis, the initial number of cross factors and the single adjustment position distribution are adjusted, and the adjustment frequency and overall position distribution are statistically analyzed to obtain the secondary clusters;

[0018] The population clusters and secondary clusters of each initial factor population are merged and divided to obtain global clusters and local clusters. The derived clusters are determined based on the first individual variable of the global cluster, the second individual variable of the local cluster, and the coverage coefficient, resulting in the cluster sequence of the corresponding initial factor population. The hierarchical fusion weight coefficients of the cluster sequences of all initial factor populations and each risk feature cluster are determined. Among them, the weight coefficients of the core layer risk factors and the weight coefficients of the related layer risk factors are positively correlated with the risk contribution of the core monitoring object and the related monitoring object, respectively.

[0019] The individual risk characteristics of the core monitoring object, the associated risk characteristics of each related monitoring object, and the real-time risk activity are weighted and fused according to the hierarchical fusion weight coefficient to generate the comprehensive risk characteristics of the target medical device at each monitoring stage.

[0020] Preferably, the core layer risk factor and the associated layer risk factor correspond to the risk characteristic dimensions of the core monitoring object and the associated monitoring object, respectively;

[0021] The risk optimization objectives include risk event differentiation, model generalization, and regulatory rule adaptability;

[0022] The four dimensions of the four-dimensional structured matrix are as follows:

[0023] The first dimension is the cluster dimension, and the number of rows in the matrix corresponds to the number of clusters.

[0024] The second dimension is the time-series stage dimension, and the number of columns in the matrix corresponds to the total number of monitoring stages in the entire clinical cycle of the target medical device; the third dimension is the factor stratification dimension, and the matrix depth corresponds to the number of risk levels of the risk factors.

[0025] The fourth dimension is the evaluation channel dimension, and the number of matrix channels corresponds to the number of evaluation dimensions in the comprehensive evaluation system.

[0026] Preferably, the formula for calculating the cross step length is: ,in, The dimension length of the risk factors in the association layer; This refers to the hierarchical risk weighting coefficient, with a value ranging from 0.4 to 1.6, representing the core layer risk factors. Values ​​higher than those of the correlation layer risk factors Values; This is the dynamic coefficient for the time series stage, with a value range of 0.7 to 1.3; The cross step size; This is the floor symbol.

[0027] Preferably, in the process of obtaining the actual contribution value of a single risk feature point, the scenarios for collecting target medical device data using a handheld data acquisition terminal include:

[0028] Timestamp synchronization and spatial coordinate matching are performed on multi-source heterogeneous real-world data within the same monitoring phase to obtain a fused feature dataset, and the clinical monitoring distance L0 between the current monitoring site and the target medical device and the subject patient is measured.

[0029] Rotate the multi-source heterogeneous real-world data acquisition terminal horizontally by a preset calibration angle. , The value range is 0.5° to 3°;

[0030] Based on preset calibration angle The raw data collected is subjected to pose correction to obtain a calibrated fused feature dataset. Simultaneously, the calibration monitoring distance L1 is measured and combined with the clinical monitoring distance L0 to determine the tilt correction angle of the acquisition terminal corresponding to the target medical device. ;

[0031] All risk feature points were identified based on the calibrated fused feature dataset;

[0032] According to the tilt correction angle Determine the corrected data confidence level and calculate the baseline risk quantification value S0 corresponding to a single risk feature point;

[0033] Based on the baseline risk quantification value S0 and the risk level weighting coefficient, the actual risk contribution value S1 corresponding to a single risk feature point is obtained. Preferably, generating the initial prediction sequence for the target medical device includes:

[0034] Collect the time-series risk quantification values ​​corresponding to the acceptance nodes of each monitoring stage, and decompose them into core layer time-series quantification values ​​and related layer time-series quantification values ​​according to the risk level;

[0035] Based on the acceptance and compliance results and historical risk contribution of each monitoring stage, dynamic time-series weighting coefficients are matched to the time-series quantification values ​​of the core layer and the related layer. The weighted time-series quantification values ​​are progressively accumulated according to the chronological order of the monitoring stages to generate the risk evolution time-series sequence of the target medical device throughout its entire clinical cycle. The progressive accumulation formula is as follows: ,in, Let be the cumulative value of risk evolution in the t-th monitoring phase. This represents the cumulative risk evolution value during the (t-1)th monitoring phase. Let be the core layer time-series weighting coefficient for the t-th monitoring phase. Let be the core layer time series quantization value for the t-th monitoring phase. Let be the correlation layer time series weight coefficient for the t-th monitoring stage. Let be the time-series quantization value of the associated layer in the t-th monitoring phase; The value is always greater than ;

[0036] The comprehensive risk characteristics of each monitoring stage are aligned with the risk evolution time series in terms of time dimension. Static risk distribution characteristics, trend change characteristics, and mutation risk characteristics are extracted and enhanced by multi-dimensional feature fusion to generate a time-enhanced risk feature set. The time-enhanced risk feature set of each monitoring stage is input into a preset double-constraint regularized risk prediction model to generate the initial prediction sequence of the target medical device.

[0037] Preferably, the target risk sequence that meets the preset regulatory requirements includes:

[0038] Based on real-world data from the historical monitoring phase that has been completed and accepted, the corresponding risk prediction results, and historical regulatory compliance conclusions, the historical prediction deviations are broken down into core inherent risk deviations and related risk deviations. Initial deviation correction coefficients are then assigned to the core layer and related layer risks, respectively. These initial deviation correction coefficients are subject to the deviation tolerance limits set by the medical device regulatory rules.

[0039] Collect real-world data from real-time clinical applications, extract temporal mutation features, risk clustering features, and regulatory sensitivity features of real-time risk characteristics, and dynamically update the initial deviation correction coefficients of the core layer and related layers.

[0040] Based on the updated bias correction coefficient, the initial prediction sequence is subjected to progressive bias compensation in the time series dimension and weighted correction in the risk level dimension to generate a pre-corrected risk sequence.

[0041] The pre-corrected risk sequence is verified based on the preset regulatory risk threshold and prediction accuracy requirements. If the verification fails, the hierarchical deviation constraint term and correction coefficient are iteratively optimized until the preset regulatory requirements are met, and the target risk sequence is output.

[0042] This invention provides a medical device risk monitoring and early warning system based on real-world evidence, comprising:

[0043] The data acquisition module is used to divide the clinical application process of the target medical device into multiple continuous monitoring stages and collect multi-source heterogeneous real-world data corresponding to the target medical device in each monitoring stage.

[0044] The contribution value determination module is used to perform spatiotemporal dimension synchronous registration of multi-source heterogeneous real-world data within the same monitoring phase to obtain a fused feature dataset. It calculates the baseline risk value corresponding to a single risk feature point in the fused feature dataset and corrects it based on the acquisition parameters, data confidence and risk weight to obtain the actual contribution value of a single risk feature point.

[0045] The feature fusion module is used to identify associated monitoring objects with risk correlation, with the target medical device as the core monitoring object, extract the individual risk characteristics of the core monitoring object, the associated risk characteristics of each associated monitoring object and the real-time risk activity, and fuse them to generate the comprehensive risk characteristics of the target medical device at each monitoring stage.

[0046] The sequence generation module is used to collect the time-series risk quantification values ​​corresponding to the acceptance nodes of each monitoring stage, accumulate the time-series risk quantification values ​​to generate a risk evolution time-series sequence, and process the comprehensive risk characteristics of each monitoring stage to generate the initial prediction sequence of the target medical device.

[0047] The sequence correction module is used to dynamically correct the initial predicted sequence based on real-world data from the historical monitoring phase that has been accepted and the corresponding risk prediction results, combined with real-time clinical application process data, to obtain a target risk sequence that meets preset regulatory requirements.

[0048] The risk warning module is used to determine the risk parameter set of the target medical device based on the target risk sequence, and output the risk level and graded warning.

[0049] Compared with the prior art, the beneficial effects of this application are as follows:

[0050] Comprehensive risk identification with significantly reduced false negative rate: A two-layer risk system of core layer and related layer is constructed, which includes devices from the same batch, patients with the same indication, medical institutions in the same clinical setting, and related devices with the same risk type in the monitoring scope. The risk identification dimension is expanded from the inherent risk of a single device to the entire chain. The false negative rate of risk is reduced from more than 60% of the existing technology to less than 20%, and the identification rate of serious adverse events is improved.

[0051] The accuracy of data fusion is significantly improved and the confidence level is controllable: a spatiotemporal synchronization registration method with pose and distance correction is designed to eliminate the system error of handheld acquisition terminal, the time synchronization error of multi-source data is ≤1s, the spatial matching accuracy is improved, and the data confidence level is increased from 70% of the existing technology to more than 80%.

[0052] The model achieves both generalization and regulatory adaptability: an adaptive optimization method for risk factors using a four-dimensional structured matrix is ​​proposed, incorporating regulatory rule adaptability into the optimization objective. The accuracy fluctuation of the model is ≤5% when adapting across devices and scenarios, fully complying with the NMPA's regulatory requirements for adverse event monitoring of medical devices.

[0053] High prediction accuracy, low false alarm rate, and no cold start defects: The constructed dual-constraint regularized risk prediction model has a risk prediction deviation of ≤±12% and a false alarm rate of ≤8%, which is far superior to existing technologies; the designed hierarchical dynamic deviation correction mechanism completes cold start calibration through historical data of the same type of medical device, which can cover the full life cycle monitoring needs of newly launched medical devices from the first day of market launch.

[0054] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0055] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0056] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0057] Figure 1 This is a flowchart of a medical device risk monitoring and early warning method based on real-world evidence, as described in an embodiment of the present invention.

[0058] Figure 2 This is a structural diagram of a medical device risk monitoring and early warning system based on real-world evidence, as described in an embodiment of the present invention. Detailed Implementation

[0059] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0060] The data collection method of this invention adopts de-identification, de-identification, federated learning, and privacy computing to achieve cross-source data fusion. It does not directly transmit the original patient privacy data, and completes multi-source data fusion under the premise of data compliance, without breaking the boundaries of medical data ownership and privacy.

[0061] This invention provides a method for monitoring and early warning of medical device risks based on real-world evidence, such as... Figure 1 As shown, it includes:

[0062] Step 1: Divide the clinical application process of the target medical device into multiple continuous monitoring stages, and collect multi-source heterogeneous real-world data corresponding to the target medical device in each monitoring stage;

[0063] In this embodiment, the target medical device includes active implantable medical devices, active surgical instruments, in vitro diagnostic medical devices, passive implantable medical devices, etc.; the monitoring phase refers to a continuous and equally spaced monitoring cycle based on the rated service life of the device, clinical diagnosis and follow-up nodes, and regulatory monitoring requirements, according to the entire clinical application cycle of the target medical device. The duration of a single monitoring phase is 7 to 30 days, which can be dynamically adjusted according to the type and risk level of the target medical device.

[0064] In this embodiment, the multi-source heterogeneous real-world data includes four core types of data:

[0065] Operational status data: operating parameters, operation logs, fault records, calibration records, wear and tear data, equipment maintenance records, etc. of the target medical device;

[0066] Clinical application data: usage records of the target medical device, operator qualifications, implementation of clinical operating procedures, environmental parameters (temperature, humidity, electromagnetic interference, etc.), medical institution and department information, treatment pathway data, etc.

[0067] Patient outcome data: vital signs data, treatment outcomes, follow-up data, prognostic indicators, comorbidities and concomitant medication data, baseline characteristics, etc. of the subjects using the target medical device;

[0068] Adverse event data: adverse events, serious adverse events, recall information, regulatory warnings, etc. related to the target medical device.

[0069] In this embodiment, the rules for dividing the monitoring phases are as follows: Basic monitoring cycle: For Class III high-risk medical devices, the duration of a single monitoring phase is 7 days; for Class II medical devices, the duration of a single monitoring phase is 15 days; for Class I medical devices, the duration of a single monitoring phase is 30 days.

[0070] Follow-up alignment cycle: Automatically merges with 1 / 3 / 6 / 12-month follow-up nodes for acceptance and synchronous reporting, meeting regulatory follow-up requirements.

[0071] In this embodiment, heterogeneity is defined as including structural heterogeneity (structured, semi-structured, and unstructured data), sampling frequency heterogeneity (milliseconds to monthly), source heterogeneity (device, hospital, patient, and regulatory), and format heterogeneity (numerical, text, image, and log data).

[0072] Step 2: Perform spatiotemporal dimension synchronous registration on multi-source heterogeneous real-world data within the same monitoring phase to obtain a fused feature dataset. Calculate the baseline risk value corresponding to a single risk feature point in the fused feature dataset, and correct it based on the acquisition parameters, data confidence, and risk weight to obtain the actual contribution value of a single risk feature point.

[0073] In this embodiment, spatiotemporal synchronization registration refers to synchronizing the time dimension of multi-source heterogeneous RWD within the same monitoring phase based on the start and end timestamps of the monitoring phase, and matching the spatial dimension using the unique identifier (UDI) of the target medical device and the unique diagnosis ID of the subject patient as spatial anchors, thereby achieving a one-to-one correspondence between multi-source data and monitoring objects and monitoring time points.

[0074] In this embodiment, risk feature points refer to the smallest data feature units that have a causal relationship with the occurrence of adverse events of medical devices. They are divided into core layer risk feature points (related to the inherent risks of the target medical device itself, such as device failure, parameter deviation, excessive wear and tear, etc.) and related layer risk feature points (related to the risks of clinical application, patient groups, and environmental management, such as non-standard operation, abnormal patient baseline, excessive environmental parameters, etc.).

[0075] Step 3: Using the target medical device as the core monitoring object, identify the associated monitoring objects with risk correlation, extract the individual risk characteristics of the core monitoring object, the associated risk characteristics of each associated monitoring object, and the real-time risk activity, and fuse them to generate the comprehensive risk characteristics of the target medical device at each monitoring stage;

[0076] In this embodiment, risk association refers to the mutual influence relationship between different monitoring objects on the probability of risk occurrence caused by the correlation between device production batches, indications, clinical application scenarios, and risk failure modes.

[0077] In this embodiment, the associated monitoring object refers to an object that has a risk association with the core monitoring object (target medical device), including four categories:

[0078] Medical devices from the same batch: Extract all in-use medical devices with the same production batch number and model specifications as the target medical device, verify the causal relationship between their failure events and the failure events of the target medical device, and confirm the risk relationship;

[0079] Patients with the same indication: Based on the indication, age, gender, and baseline characteristics of the subjects, a patient population with the same baseline characteristics was constructed using propensity score matching (PSM) to verify the causal association between their adverse outcomes and the use of the target medical device;

[0080] Medical institutions with similar clinical application scenarios: Extract medical institutions of the same level, department, and operating procedures as the target medical device user and verify the correlation between the adverse event rate of similar devices and that of the target medical institution.

[0081] Medical devices with the same risk type: Based on the Medical Device Failure Mode and Effects Analysis (FMEA) database, similar medical devices with the same failure mode and the same risk points as the target medical device are extracted to verify the correlation of their risk events.

[0082] In this embodiment, real-time risk activity refers to the normalized quantitative value of the frequency and magnitude of abnormal fluctuations of the corresponding risk characteristic points of the associated monitoring object during the current monitoring phase. The value ranges from 0 to 1, and the higher the value, the higher the risk activity of the associated monitoring object.

[0083] In this embodiment, the comprehensive risk characteristics refer to the individual risk characteristics of the core monitoring object, the associated risk characteristics of each related monitoring object, and the real-time risk activity. After hierarchical weighted fusion, a multi-dimensional feature set that can comprehensively characterize the risk level of the target medical device in the current monitoring stage is obtained.

[0084] Step 4: Collect the time-series risk quantification values ​​corresponding to the acceptance nodes of each monitoring stage, accumulate the time-series risk quantification values ​​to generate a risk evolution time-series sequence, and process the comprehensive risk characteristics of each monitoring stage to generate the initial prediction sequence of the target medical device.

[0085] In this embodiment, the time-series risk quantification value refers to the normalized quantitative score of the overall risk of the target medical device at each monitoring stage, calculated based on the comprehensive risk characteristics of that stage at the acceptance node. The value ranges from 0 to 100, with a higher score indicating a higher risk level.

[0086] In this embodiment, the risk evolution time series refers to the time series that can characterize the risk evolution trend of the target medical device throughout the entire clinical cycle, generated by progressively accumulating the time series risk quantification values ​​of each monitoring stage according to the chronological order of the monitoring stages.

[0087] Step 5: Based on real-world data from the historical monitoring phase that has been completed and accepted, and the corresponding risk prediction results, the initial prediction sequence is dynamically corrected in conjunction with real-time clinical application process data to obtain a target risk sequence that meets the preset regulatory requirements.

[0088] Step 6: Determine the risk parameter set of the target medical device based on the target risk sequence, and output the risk level and graded warning.

[0089] In this embodiment, the initial prediction sequence refers to the time series of risk prediction values ​​for the next N monitoring stages, which are output by the dual-constraint regularized risk prediction model based on the time-series enhanced risk feature set of each monitoring stage. N is the preset prediction step size, which ranges from 1 to 12 monitoring stages.

[0090] In this embodiment, the target risk sequence refers to the final risk prediction sequence that meets the preset regulatory requirements after the initial prediction sequence has undergone dynamic correction of historical prediction deviations and verification of regulatory rule compliance.

[0091] In this embodiment, the risk parameter set includes the predicted risk value, risk evolution trend, and probability of risk mutation for each future monitoring stage. Based on a preset risk level classification threshold, the risk level of the target medical device is determined, and a corresponding graded warning is output. The risk level classification thresholds are as follows: Low risk: predicted risk value 0-20 points, no warning, routine monitoring; Moderate risk: predicted risk value 20-40 points, blue warning, indicating attention to risk changes; Higher risk: predicted risk value 40-70 points, yellow warning, initiating risk investigation; High risk: predicted risk value 70-100 points, red warning, initiating emergency control measures, and reporting to regulatory agencies.

[0092] This invention provides a method for monitoring and early warning of medical device risks based on real-world evidence, which integrates and generates comprehensive risk characteristics of the target medical device at each monitoring stage, including:

[0093] For the fusion feature dataset of each monitoring stage, the core layer risk factors corresponding to the inherent risk characteristics of the target medical device itself and the associated layer risk factors corresponding to the risk characteristics of clinical application, patient group and environmental control are extracted. A risk factor library is constructed and each risk factor is uniquely encoded according to the risk source level. The initial total number of factor populations M is set according to the total number of factors in the risk factor library. The number of effective risk factor dimensions contained in each initial factor population is set to D. D matches the total dimension of the risk feature cluster in a single monitoring stage.

[0094] In this embodiment, the core layer risk factors correspond to the risk characteristic dimensions of the core monitoring objects, including device operation status risk factors, failure risk factors, wear and tear risk factors, calibration status risk factors, etc., which are directly related to the inherent risks of the target medical device itself, and the dimension length is D1.

[0095] Risk factors at the association layer: These are risk characteristic dimensions corresponding to the associated monitoring objects, including clinical operation risk factors, patient outcome risk factors, environmental control risk factors, and associated object risk factors. They are directly related to risk association and have a dimension length of D2.

[0096] Unique coding rules: An 8-digit fixed-length numeric coding rule is adopted, and the coding format is: hierarchical code (1 digit) + risk category code (2 digits) + factor type code (2 digits) + factor serial number (3 digits).

[0097] Hierarchical code assignment rules: 1 digit, 1 = core layer risk factor, 2 = related layer risk factor;

[0098] Risk category code assignment rules: 2 digits, core layer risk factors correspond to 01=operational status, 02=failure, 03=loss, 04=calibration status; related layer risk factors correspond to 01=clinical operation, 02=patient outcome, 03=environmental control, 04=related object;

[0099] Factor type code assignment rules: 2 digits, corresponding to the sub-type under the risk category, such as 01=abnormal pacing threshold, 02=excessive battery power decline rate, etc. under the core layer operation status category;

[0100] Factor number assignment rules: 3-digit numbers, sequentially encoded starting from 001, with no duplicate factor numbers within the same level, category, or type;

[0101] After coding is completed, a coding-risk factor mapping table is created and stored in the system database to achieve a one-to-one correspondence between coding and risk factors.

[0102] In this embodiment, taking the core layer risk factor "excessive battery power decline rate" of an implantable cardiac pacemaker as an example, its unique code is 10102003, where: 1 = core layer, 01 = operational status risk category, 02 = abnormal battery parameter type, 003 = the third factor under this type. Through this code, all attributes and historical data of the risk factor can be directly located.

[0103] An initial factor population is randomly generated according to the risk factor hierarchical coding. Based on the real-world dataset of historical adverse events of medical devices, a multi-dimensional pre-evaluation of the fit of each individual in the initial factor population is performed. Invalid individuals with fit below a preset threshold are removed to obtain valid individuals.

[0104] In this embodiment, an initial factor population is randomly generated according to the risk factor hierarchical coding, including:

[0105] Determine the population size and dimensions: Set the initial total number of factor populations M, where M is 5 to 10 times the total number of risk factors D, with a minimum value of 100; the dimension of each individual is D = D1 + D2, where D1 is the dimension length of the core layer risk factors and D2 is the dimension length of the related layer risk factors, which perfectly matches the total dimensions of the risk factor library.

[0106] Hierarchical random generation constraint: According to the hierarchical coding of risk factors, the weight vector is split into core layer weight segment and related layer weight segment. The value range of each element of the core layer weight segment is [0,1], and the value range of each element of the related layer weight segment is [0,1].

[0107] Random number generation rules: Pseudo-random numbers are generated using the linear congruential method to ensure that each weight value is evenly distributed within the range of values, and to avoid the initial weights being concentrated in the extreme value range;

[0108] Individual validity verification: For each randomly generated individual, verify that the ratio of the sum of the weights of the core layer to the sum of the weights of the related layers is not less than 1.2 (to ensure that the risk of the core layer is the core evaluation dimension). Individuals that do not meet the requirements are regenerated until M valid initial individuals that meet the constraints are generated to form the initial factor population.

[0109] For example, the total dimension of the implantable cardiac pacemaker risk factor database is D=60 (D1=24, D2=36), and the initial factor population is set to M=300 (5 times D). According to the hierarchical coding, the first 24 bits of each individual are the core layer risk factor weights, and the last 36 bits are the related layer risk factor weights. Each element is randomly generated in the range [0,1]. The sum of the core layer weights / the sum of the related layer weights of each individual is verified to be ≥1.2. Finally, 300 individuals that meet the requirements are generated to form the initial factor population.

[0110] In this embodiment, the construction rules for the real-world dataset of historical adverse events of medical devices are as follows: Time range: covering historical data of the past 5 years to ensure data timeliness; Sample size requirement: the effective sample size of similar medical devices is not less than 100,000, of which the proportion of positive adverse event samples is not less than 20% to avoid sample imbalance; Data dimension: completely consistent with the dimensions of the multi-source heterogeneous real-world data of this method, covering four major categories of data: operating status, clinical application, patient outcome, and adverse events.

[0111] Data preprocessing: Perform missing value imputation, outlier removal, and standardization / normalization on the raw data; perform text feature extraction on the unstructured data; and transform it into standardized structured data.

[0112] Dataset partitioning: The dataset is divided into training, validation, and test sets in a ratio of 7:1.5:1.5, for factor fit evaluation, model validation, and performance testing, respectively.

[0113] Data annotation: The data samples are annotated by professionals with medical device adverse event monitoring qualifications. The annotation content includes: whether an adverse event occurred, the level of the adverse event, the source of risk, and the correlation between risk factors. The annotation results are reviewed by two people to ensure that the annotation accuracy rate is ≥99%.

[0114] In this embodiment, the process of multi-dimensional adaptability pre-evaluation is as follows:

[0115] Evaluation dimensions and quantification methods: all three indicators use a normalized score of 0-1, with higher scores indicating better fit.

[0116] Dimension 1: Risk event discrimination, quantified using the area under the ROC curve (AUC value). The calculation formula is: AUC score = actual AUC value. The AUC value ranges from 0 to 1. The higher the score, the stronger the risk factor's ability to distinguish between adverse events and non-adverse events.

[0117] Dimension 2: Model generalization, quantified by the inverse normalization of the generalization error. The calculation formula is: Generalization score = 1 / (1 + generalization error), where generalization error = |validation set AUC value - training set AUC value|. The higher the score, the stronger the model's generalization ability across samples and scenarios.

[0118] Dimension 3: Regulatory rule adaptability, which is quantified by regulatory threshold compliance rate. The calculation formula is: Adaptability score = Number of predicted samples that meet regulatory requirements / Total number of samples. Regulatory requirements include serious risk event identification rate ≥95% and prediction deviation ≤15%. The higher the score, the more compliant it is with the medical device regulatory rules.

[0119] Overall adaptability score calculation: The weighting of the three indicators is as follows: risk event discrimination 40%, model generalization 30%, and regulatory rule adaptability 30%. The formula for calculating the overall score is: Overall score = 0.4 × AUC score + 0.3 × generalization score + 0.3 × regulatory adaptability score;

[0120] Pre-evaluation execution process:

[0121] The weight vector of each individual in the initial factor population is substituted into the risk assessment model, and risk prediction is completed based on the training set.

[0122] Calculate the suitability scores and overall scores for each of the three dimensions;

[0123] The overall scores of all individuals are ranked to complete the pre-evaluation.

[0124] For example, when a pre-evaluation is performed on an individual in the initial factor population of implantable cardiac pacemakers, the risk event discrimination index AUC is calculated to be 0.88, the generalization error is 0.05, and the regulatory rule compliance rate is 0.92; the corresponding scores are 0.88, 0.95, and 0.92, respectively; the overall score is 0.4×0.88+0.3×0.95+0.3×0.92=0.913, indicating good fit.

[0125] In this embodiment, the core rule for setting the preset threshold is: a dual threshold constraint is adopted, and only individuals that meet both threshold requirements at the same time are determined to be valid individuals;

[0126] Threshold 1: Overall fit score ≥ 0.75, ensuring that the individual's overall performance meets the basic requirements;

[0127] Threshold 2: Risk event discrimination AUC ≥ 0.75, ensuring that individuals have basic adverse event identification ability and avoiding situations where the overall score meets the standard but the core discrimination ability is insufficient;

[0128] Threshold dynamic adjustment rules: For Class III high-risk medical devices, the threshold can be adjusted to a comprehensive score ≥0.8 and AUC ≥0.8 to improve the monitoring accuracy of high-risk devices; for Class I low-risk medical devices, the threshold can be maintained at the basic value to balance monitoring efficiency and accuracy.

[0129] Threshold verification: Based on historical datasets, the threshold is backtested to ensure that the risk prediction performance of valid individuals after threshold screening meets the preset regulatory requirements. If the verification fails, the threshold is adjusted until it meets the requirements.

[0130] For example, the preset thresholds for implantable cardiac pacemakers are: overall fit score ≥ 0.8 and risk event discrimination AUC ≥ 0.8; the initial factor population consists of 300 individuals. After pre-evaluation, 84 individuals that do not meet the dual threshold requirements are removed, and 216 valid individuals are retained to form the effective factor population.

[0131] All valid individuals under the same initial factor population are clustered according to the risk optimization objective to obtain multiple population clusters and transform them into a four-dimensional structured matrix to determine the risk optimization objective of the corresponding initial factor population;

[0132] In this embodiment, all valid individuals under the same initial factor population are clustered according to the risk optimization objective, including:

[0133] Risk optimization objectives are prioritized as follows, from highest to lowest: maximizing risk event differentiation > maximizing regulatory rule adaptability > maximizing model generalization.

[0134] Clustering feature construction: The clustering feature vector is composed of three adaptability scores for each individual (risk event discrimination, generalization, and regulatory adaptability), with a dimension of 3.

[0135] Clustering algorithm selection: K-means clustering algorithm is adopted, and the optimal number of clusters K is determined by the elbow rule. The value of K is in the range of 5 to 20.

[0136] Clustering execution process:

[0137] Standardize the clustering feature vectors to eliminate the influence of units;

[0138] Based on the elbow rule, the sum of squares within groups is calculated for different K values ​​to determine the optimal K value;

[0139] K-means clustering is performed with the optimal K value, with an upper limit of 1000 iterations and a convergence threshold of 1e-6.

[0140] The clustering results are validated using the silhouette coefficient. If the silhouette coefficient is ≥0.5, the clustering results are considered valid; otherwise, the K value is re-determined.

[0141] For example, the effective population of implantable cardiac pacemakers consists of 216 effective individuals. Using three fitness scores as clustering features, the optimal number of clusters K=10 was determined by the elbow rule. After performing K-means clustering, 10 population clusters were obtained with a silhouette coefficient of 0.68. The clustering results are effective, and the fitness features of individuals within each cluster are highly similar, while the feature differences between different clusters are significant.

[0142] In this embodiment, the multiple population clusters obtained from clustering are transformed into a four-dimensional structured matrix according to preset dimensional rules, thereby realizing the structured and standardized management of risk factor populations. The process includes:

[0143] The definition and assignment rules for the dimensions of a four-dimensional structured matrix are as follows: the matrix elements are the adaptation normalization scores for the corresponding dimension, with values ​​ranging from 0 to 1.

[0144] First dimension: cluster dimension, the number of rows in the matrix = the number of population clusters K obtained by clustering, and each row corresponds to one population cluster;

[0145] The second dimension is the time-series stage dimension. The number of matrix columns equals the total number of monitoring stages T in the entire clinical cycle of the target medical device. Each column corresponds to one monitoring stage.

[0146] The third dimension is the factor stratification dimension. The matrix depth equals the number of risk levels of the risk factors, which is fixed at 2 levels. The first level corresponds to the core risk factors, and the second level corresponds to the related risk factors.

[0147] The fourth dimension is the evaluation channel dimension. The number of matrix channels equals the number of evaluation dimensions in the comprehensive evaluation system, which is fixed at 3 channels. The first channel corresponds to the risk event discrimination, the second channel corresponds to the model generalization, and the third channel corresponds to the regulatory rule adaptability.

[0148] Matrix transformation execution flow:

[0149] Initialize a four-dimensional zero matrix with dimensions [K,T,2,3];

[0150] For each population cluster, calculate the average fit score of all individuals within the cluster at the corresponding monitoring stage, corresponding risk level, and corresponding evaluation dimension;

[0151] The average fit score is assigned to the corresponding position in the matrix to complete the matrix element filling;

[0152] Normalize the matrix elements to ensure that all elements take values ​​between 0 and 1, thus completing the construction of the four-dimensional structured matrix.

[0153] For example, if the number of population clusters for implantable cardiac pacemakers is K=10, the total number of monitoring phases is T=52, the number of factor stratifications is 2, and the number of evaluation channels is 3, the dimensions of the constructed four-dimensional structured matrix are [10,52,2,3]. Taking the matrix elements corresponding to the first population cluster, the first monitoring phase, the core layer, and the risk event discrimination as an example, the value is assigned to the average AUC value of all individuals in the cluster under this scenario, which is 0.92. After filling all matrix elements, a standardized four-dimensional structured matrix is ​​obtained.

[0154] According to the risk optimization objective, the risk contribution dominance levels of effective factors in the corresponding initial factor population are sorted in ascending order to generate a factor sequence and determine the crossover step size of adjacent effective factors.

[0155] In this embodiment, the ascending order sorting process is as follows:

[0156] Risk contribution calculation: The marginal contribution of each risk factor to the risk prediction result is calculated using the SHAP (SHapley Additive exPlanations) value. The larger the absolute value of the SHAP value, the higher the risk contribution of that factor.

[0157] Dominance hierarchy rules:

[0158] The risk factors are sorted from largest to smallest based on their absolute SHAP values. The factor ranked first has a dominance level of 1; the factor ranked second has a dominance level of 2, and so on.

[0159] The smaller the dominance level value, the higher the risk contribution of the factor, and the higher its priority in optimization;

[0160] Execution flow for ascending order sorting:

[0161] Based on a four-dimensional structured matrix and historical datasets, the average SHAP value of each effective risk factor is calculated.

[0162] Assign a corresponding dominance level to each factor according to the absolute value of the SHAP value, from largest to smallest.

[0163] The effective factors are sorted in ascending order according to their dominance level, generating an ordered list of factors.

[0164] For example, among the 60 risk factors for implantable cardiac pacemakers, the absolute value of the SHAP value for "excessive rate of battery power decline" is the largest, with a dominance level of 1; the absolute value of the SHAP value for "abnormal pacing threshold" is the second largest, with a dominance level of 2; and so on. Finally, the factors are arranged in ascending order according to the dominance level from 1 to 60, resulting in an ordered list of factors with high contribution factors first and low contribution factors last.

[0165] In this embodiment, the factor sequence is composed as follows: sequence length = total number of risk factors D. Each element in the sequence contains four core pieces of information: unique code of the risk factor, dominance level, SHAP contribution value, and weight range. The sequence is constrained by the following rules: the sequence is arranged in ascending order of dominance level. The first 20% of the sequence elements are high-priority core factors, and the last 80% are low-priority auxiliary factors. For example, the factor sequence length for an implantable cardiac pacemaker is 60. The first 12 elements of the sequence are high-priority core factors with dominance levels of 1-12, which are all core-level risk factors. The last 48 elements of the sequence are low-priority auxiliary factors with dominance levels of 13-60, which include core-level key factors and related-level factors. During cross-optimization, the high-priority factors in the first half of the sequence are adjusted first.

[0166] Based on the cross step size analysis, the initial number of cross factors and the single adjustment position distribution are adjusted, and the adjustment frequency and overall position distribution are statistically analyzed to obtain the secondary clusters;

[0167] In this embodiment, the rule for determining the initial number of cross factors is as follows:

[0168] Total number of cross factors = Total length of factor sequences / Cross step size ∆, rounded up;

[0169] The number of high-priority factors = the total number of cross factors × 60%, and is selected from the top 20% of high-priority factors in the factor sequence;

[0170] The number of low-priority factors = the total number of cross factors × 40%, selected from the last 80% of low-priority factors in the factor sequence;

[0171] Single adjustment of position distribution rules:

[0172] The adjustment positions of high-priority factors are evenly distributed within the first 20% of the factor sequence to ensure that core high-contribution factors are fully optimized.

[0173] The adjustment positions of low-priority factors are randomly distributed in the last 80% of the factor sequence to avoid optimization blind spots;

[0174] The interval between two adjacent adjustment positions should not be less than the cross step size ∆, so as to avoid excessive concentration of adjustment positions and insufficient optimization granularity.

[0175] Adjustment process:

[0176] According to the above rules, determine the initial number and adjustment position of the cross factors;

[0177] For the factor weights corresponding to the adjusted positions, perform a simulated binary crossover operation to generate a new weight vector;

[0178] The newly generated weight vector is validated to ensure it complies with the core layer weight ratio constraints.

[0179] For example, in an implantable cardiac pacemaker, the factor sequence length D=60, and the crossover step size... =1, total number of crossover factors = 60 / 1 = 60; of which the number of high-priority factors = 60 × 60% = 36, the adjustment positions are distributed in the high-priority interval of the first 12 positions of the sequence; the number of low-priority factors = 24, the adjustment positions are distributed in the interval of the last 48 positions of the sequence; the interval between adjacent adjustment positions is 1, which meets the crossover step size requirement, and new factor individuals are generated after the crossover operation is performed.

[0180] In this embodiment, the secondary cluster refers to the set of individual factors whose fitness improvement did not meet the preset requirements during the cross-optimization process. By statistically analyzing the adjustment frequency and position distribution of cross factors, individuals with poor optimization effects are selected and formed into secondary clusters for subsequent local optimization to supplement the blind spots of global optimization. The adjustment frequency statistics rule is as follows: count the number of times each individual factor is adjusted during the cross-optimization process, that is, the total frequency of the individual being selected to perform cross-operation; count the selection frequency of each adjustment position, that is, the total number of times the factor at that position is adjusted.

[0181] The criteria for determining secondary clusters are as follows: Individuals that meet any of the following conditions are included in the secondary cluster: After cross-optimization, the individual's overall fitness score increases by ≤5%; the adjustment frequency is less than 30% of the average adjustment frequency, indicating that the individual has not been sufficiently optimized; the adjustment position is concentrated in the low priority interval of the last 50% of the factor sequence, indicating that the core factors have not been optimized.

[0182] Secondary cluster generation process: After cross-optimization is completed, the adjustment frequency, fitness improvement, and adjustment position distribution of all individuals are counted; individuals that meet the criteria are selected according to the secondary cluster determination criteria; the selected individuals are deduplicated to form secondary clusters.

[0183] After the cross-optimization of implantable cardiac pacemakers was completed, a total of 216 optimized individuals were generated. Statistics showed that the fit improvement of 32 individuals was ≤5%, and the adjustment frequency of 18 individuals was less than 30% of the average frequency. After deduplication, a total of 42 individuals were formed into a minor cluster for subsequent local weight optimization.

[0184] The population clusters and secondary clusters of each initial factor population are merged and divided to obtain global clusters and local clusters. The derived clusters are determined based on the first individual variable of the global cluster, the second individual variable of the local cluster, and the coverage coefficient, thus obtaining the cluster sequence of the corresponding initial factor population.

[0185] In this embodiment, the merging process rule is as follows: individuals from all population clusters obtained by clustering are merged with individuals from minor clusters to obtain a complete optimized set of individuals, and duplicate individuals are removed.

[0186] Global and local cluster partitioning rules: The merged set of individuals is sorted from high to low according to the comprehensive fitness score; the top 20% of individuals form the global cluster, which is the set of individuals with the best fitness and is used to determine the global optimal weight direction; the remaining 80% of individuals form the local cluster, which includes individuals from the population cluster with moderate fitness and all minor cluster individuals, and is used to supplement the blind spots of global optimization and complete local detail optimization.

[0187] Post-partition verification: Verify that the average comprehensive fitness score of the global cluster is ≥0.85 and the average comprehensive fitness score of the local cluster is ≥0.7 to ensure that the partitioned clusters meet the optimization requirements. If they do not meet the requirements, readjust the partition ratio.

[0188] For example, the population of implantable cardiac pacemakers consists of 216 individuals in the primary cluster and 42 individuals in the secondary cluster, totaling 248 individuals after merging and deduplication. After sorting by comprehensive fitness score, the top 20% (50 individuals) form the global cluster with an average comprehensive score of 0.92; the remaining 198 individuals form the local cluster with an average comprehensive score of 0.78, which meets the classification requirements.

[0189] In this embodiment, the first volume variable refers to the risk factor weight vector corresponding to the individual with the highest comprehensive fitness score in the global cluster. It is the optimal solution obtained by global optimization, representing the weight allocation scheme with the best fitness in the current population, and is the core benchmark variable for generating the derived cluster.

[0190] The first volume variable selection rule is as follows: For all individuals in the global cluster, sort them from high to low according to their comprehensive fitness scores; the D-dimensional weight vector corresponding to the individual ranked first is the first volume variable; if there are individuals with the same comprehensive score, select the individual with the higher risk event discrimination AUC value as the first volume variable.

[0191] Validation of the first entity variable: Validate that the sum of the core layer weights / the sum of the related layer weights of the first entity variable is ≥1.2, which meets the weight constraint requirements; validate that its risk prediction performance meets the preset regulatory requirements, and if it does not, re-screen it.

[0192] Update of the first volume variable: After each monitoring phase, the population is re-optimized based on the latest monitoring data, and the first volume variable is updated synchronously to ensure that it is always the current optimal solution.

[0193] For example, in the global cluster of 50 individuals for implantable cardiac pacemakers, the top-ranked individual has a comprehensive fit score of 0.96, an AUC value of 0.94, and a regulatory fit score of 0.95, making it the globally optimal individual. Its corresponding 60-dimensional weight vector is the first volume variable, and the sum of the weights of the core layer / the sum of the weights of the related layers = 2.33, which meets the constraint requirements.

[0194] In this embodiment, the second volume variable refers to the risk factor weight vector corresponding to the individual with the highest comprehensive fitness score in the local cluster. It is the optimal solution obtained by local optimization, representing the optimal weight allocation scheme in the local feature space. It complements the first volume variable and is a supplementary benchmark variable for the generation of the derived cluster.

[0195] The second individual variable selection rule is as follows: For all individuals in the local cluster, sort them from high to low according to their comprehensive fitness scores; the D-dimensional weight vector corresponding to the individual ranked first is the second individual variable; if there are individuals with the same comprehensive score, select the individual with the higher model generalization score as the second individual variable.

[0196] Verification of the second individual variable: Verify that the cosine similarity between the second individual variable and the first individual variable is ≤0.8 to ensure that the two variables have feature differences and can complement each other; verify that it meets the weight constraint requirements, and if it does not, re-select.

[0197] The second individual variable is updated synchronously with the first individual variable to ensure that it is always the optimal solution for the local cluster.

[0198] For example, in the local cluster of implantable cardiac pacemakers, there are 198 individuals. The individual ranked first has a comprehensive fit score of 0.86 and a generalization score of 0.93. Its corresponding 60-dimensional weight vector is the second volume variable. The cosine similarity between the second and first volume variables is 0.72, indicating significant feature differences and complementarity.

[0199] In this embodiment, the coverage coefficient = the number of feature intersection dimensions between the local cluster and the global cluster / the total number of feature dimensions of the global cluster, with a value range of 0 to 1. The number of feature intersection dimensions refers to the number of overlapping dimensions between the high contribution factors of the local cluster and the high contribution factors of the global cluster.

[0200] Derivative cluster generation process: Calculate the coverage coefficient and determine the step size of Gaussian mutation: step size = 1 - coverage coefficient. The lower the coverage coefficient, the larger the step size, and the stronger the ability to supplement blind spots; Using the first individual variable as the baseline mean and the second individual variable as the variance baseline, combined with the step size, perform Gaussian mutation to generate N new individuals, where N = the total number of risk factors D; Evaluate the fit of the newly generated individuals and remove invalid individuals with a composite score < 0.75; Deduplicate the valid individuals to form a derivative cluster;

[0201] Derivative cluster verification: The number of individuals in the derived cluster is ≥10 and the average comprehensive fitness score is ≥0.8 to ensure the optimization effect of the derived cluster. If it does not meet the requirements, the mutation parameters are readjusted and the cluster is generated again.

[0202] For example, the feature intersection dimension of the local cluster and the global cluster of an implantable cardiac pacemaker is 18, the total feature dimension of the global cluster is 60, and the coverage coefficient is 18 / 60=0.3; the variable length is 1-0.3=0.7; using the first individual variable as the mean and the second individual variable as the variance benchmark, Gaussian mutation is performed to generate 60 new individuals; after removing invalid individuals, 48 ​​valid individuals are retained to form a derived cluster, with an average comprehensive fitness score of 0.88, which meets the requirements.

[0203] In this embodiment, the cluster sequence is composed of three elements in a fixed order: global cluster, derived cluster, and local cluster. The sequence sorting rule is as follows: the average comprehensive fit score of each cluster is calculated; the three clusters are sorted in ascending order of average score from highest to lowest, with the highest-scoring cluster at the beginning of the sequence and having the highest priority. The general sorting result is: global cluster (1st position) > derived cluster (2nd position) > local cluster (3rd position). If the average score of the derived cluster exceeds that of the global cluster, the order is adjusted. The sequence weight mapping rule is as follows: the priority of each cluster in the sequence corresponds to its contribution percentage in the calculation of the hierarchical fusion weight coefficient; the higher the priority, the greater the contribution percentage. The sequence is updated as follows: after each monitoring phase, the individual and average scores of the three clusters are updated synchronously, and a new cluster sequence is generated by reordering. For example, the average score of the global cluster for an implantable cardiac pacemaker is 0.92, the average score of the derived cluster is 0.88, and the average score of the local cluster is 0.78. Sorted from highest to lowest score, the generated cluster sequence is: [global cluster, derived cluster, local cluster], with the global cluster having the highest priority and the largest contribution percentage in the weight calculation.

[0204] Determine the hierarchical fusion weight coefficients of the cluster sequences of all initial factor populations and each risk feature cluster, wherein the weight coefficients of the core layer risk factors and the weight coefficients of the associated layer risk factors are positively correlated with the risk contribution of the core monitoring objects and the associated monitoring objects, respectively.

[0205] In this embodiment, based on the fitness scores of all valid individuals in the cluster sequence, the average fitness score S_core of the core layer risk factors and the average fitness score S_assoc of the associated layer risk factors are calculated.

[0206] The total weight coefficient of the core layer is W_core = S_core / (S_core + S_assoc), and the total weight coefficient of the related layer is W_assoc = S_assoc / (S_core + S_assoc), where W_core > W_assoc;

[0207] The weight of a single core layer risk factor = the SHAP contribution of that factor / the sum of the SHAP contributions of all core layer factors × W_core;

[0208] The weight of a single associated layer risk factor = the factor's SHAP contribution × the real-time risk activity of the corresponding associated monitoring object / the sum of the SHAP contributions of all associated layer factors × W_assoc.

[0209] The individual risk characteristics of the core monitoring object, the associated risk characteristics of each related monitoring object, and the real-time risk activity are weighted and fused according to the hierarchical fusion weight coefficient to generate the comprehensive risk characteristics of the target medical device at each monitoring stage.

[0210] In this embodiment, the composition of individual risk characteristics includes four core categories, each corresponding to a core layer risk factor:

[0211] Operating status characteristics: instrument operating parameters, operation log, wear status, battery status, etc.

[0212] Fault risk characteristics: Features such as instrument fault records, abnormal alarm records, and parameter deviation records;

[0213] Calibration status characteristics: Features such as instrument calibration records, calibration deviations, and compliance of calibration cycles;

[0214] Maintenance status characteristics: Features such as instrument maintenance records, maintenance cycles, and component replacement records;

[0215] Feature extraction process:

[0216] Extract the actual contribution value of the core layer risk feature points corresponding to the target medical device from the corrected fusion feature dataset;

[0217] Based on the unique coding order of risk factors, the actual contribution values ​​are arranged into a multi-dimensional vector, with each element of the vector corresponding to the actual contribution value of a core layer risk factor.

[0218] The feature vector is normalized to ensure that each element takes a value between 0 and 1, thus obtaining the final individual risk characteristics.

[0219] For example, the core layer risk factor dimension D1=24 for implantable cardiac pacemakers. The actual contribution values ​​of 24 core layer risk feature points are extracted from the fusion feature dataset, arranged into a 24-dimensional vector according to the encoding order, and normalized to obtain individual risk features. Among them, the feature value corresponding to "excessive battery power decline rate" is 0.98, which means that the risk level of this risk point is relatively high.

[0220] In this embodiment, the associated risk feature refers to a multi-dimensional feature vector composed of associated layer risk feature points of each associated monitoring object. It is a quantitative representation of the risk brought about by risk association and a supplementary component of the comprehensive risk feature. The dimension is completely consistent with the dimension length D2 of the associated layer risk factor. For example, the dimension D2 of the associated layer risk factor of implantable cardiac pacemaker is 36. The actual contribution values ​​of 36 associated layer risk feature points of 4 types of associated monitoring objects are extracted from the fusion feature dataset, arranged into a 36-dimensional vector according to the coding order, and normalized to obtain the associated risk feature. Among them, the feature value corresponding to "battery failure rate of the same batch of devices" is 0.85, which means that the risk level of this risk point is relatively high.

[0221] In this embodiment, the real-time risk activity quantification calculation formula is: Act_i1=0.5×(F_i1+N_i1), where: Act_i1 is the real-time risk activity of the i1th associated monitoring object, with a value range of 0~1; F_i1 is the normalized value of the abnormal amplitude of the risk characteristics of the associated monitoring object, F_i1=min(abnormal amplitude / preset abnormal threshold, 1); N_i1 is the normalized value of the frequency of abnormal fluctuations of the risk characteristics of the associated monitoring object, N_i1=min(number of abnormal fluctuations / total number of samplings in the monitoring phase, 1).

[0222] In this embodiment, the activity level classification rule is as follows:

[0223] Low activity: Act_i1 < 0.3, risk impact is negligible;

[0224] Medium activity level: 0.3 ≤ Act_i1 < 0.7, indicating a medium risk impact;

[0225] High activity level: Act_i1≥0.7 indicates a strong risk impact and requires close monitoring.

[0226] In this embodiment, F_combine = W_core × F_core + sum(W_assoc_i1 × F_assoc_i1 × Act_i1), where F_combine is the comprehensive risk feature vector, F_core is the individual risk feature vector of the core monitoring object, F_assoc_i1 is the associated risk feature vector of the i1th associated monitoring object, Act_i1 is the real-time risk activity of the i1th associated monitoring object, W_core is the total weight coefficient of the core layer, and W_assoc_i1 is the weight coefficient of the risk factor of the i1th associated layer.

[0227] Preferably, the core layer risk factor and the associated layer risk factor correspond to the risk characteristic dimensions of the core monitoring object and the associated monitoring object, respectively;

[0228] The risk optimization objectives include risk event discrimination, model generalization, and regulatory rule adaptability. It should be noted that, in order of priority, they are: maximizing risk event discrimination, maximizing model generalization, and maximizing regulatory rule adaptability.

[0229] The four dimensions of the four-dimensional structured matrix are as follows:

[0230] The first dimension is the cluster dimension: the number of rows in the matrix corresponds to the number of clusters K obtained from the clustering.

[0231] The second dimension is the time-series stage dimension: the number of matrix columns corresponds to the total number of monitoring stages T in the full clinical cycle of the target medical device;

[0232] The third dimension is the factor stratification dimension: the matrix depth corresponds to the number of risk levels of the risk factor, which is fixed at 2 levels (core layer and related layer).

[0233] The fourth dimension is the evaluation channel dimension: the number of matrix channels corresponds to the number of evaluation dimensions in the comprehensive evaluation system, which is fixed at 3 channels (risk event discrimination, model generalization, and regulatory rule adaptability).

[0234] In this embodiment, the three-dimensional evaluation system for the optimization effect of risk factors corresponds one-to-one with the adaptability pre-evaluation index, and the quantitative value range of each dimension is 0 to 1 point.

[0235] In this embodiment, the priority of the three optimization objectives, from highest to lowest, is as follows: risk event differentiation > regulatory rule adaptability > model generalization.

[0236] In this embodiment, the element values ​​of the four-dimensional structured matrix are determined as follows: the element values ​​in the k-th row, t-th column, l-th layer, and c-th channel of the matrix represent the normalized fit score of the k-th cluster, t-th monitoring stage, l-th risk level, and c-th evaluation dimension, with a value range of 0 to 1. The matrix is ​​used in the following way: subsequent cross-optimization, cluster partitioning, and weight coefficient calculation are all based on the fit score of each element in the matrix, thus achieving full-process traceability and supervision of risk factor optimization.

[0237] The four-dimensional structured matrix enables comprehensive structured management of risk factors, including cluster classification, time-series adaptation, hierarchical management, and multi-dimensional evaluation. Subsequent cross-optimization and weight calculation are all based on this matrix, solving the problems of chaotic risk factor management and lack of optimization objectives in existing technologies.

[0238] Preferably, the formula for calculating the cross step length is: Where D2 is the dimension length of the risk factor in the association layer; The risk weight coefficients are tiered, ranging from 0.4 to 1.6, and are pre-determined based on the historical risk contribution of the corresponding risk factors. The core tier risk factors... Values ​​higher than those of the correlation layer risk factors Value retrieval, core layer Fixed at 1.2~1.6, associated layer Fixed at 0.4~1.0; This is a dynamic coefficient for the time series stage, with a value ranging from 0.7 to 1.3. The closer to the current monitoring stage, the higher the coefficient. The larger the value, the more historical the monitoring period. =0.7, current monitoring phase =1.3, Predictive Monitoring Stage =1.0; This is the crossover step size, which, after rounding, represents the number of crossover points between adjacent factors. The value range is 1 to D, with a fixed minimum value of 1 to avoid invalid crossovers. This is the rounding up symbol; it should be noted that... The value is determined by the level (core layer / related layer) of the first factor among the adjacent effective factors.

[0239] In this embodiment, the core layer risk factors Calibration rules: The value ranges from 1.2 to 1.6, where Lg is the correlation degree of the factor's historical adverse events; Hzg is the highest historical correlation degree of the core layer factor.

[0240] Related layer risk factors Calibration rules: The value ranges from 0.4 to 1.0, where Lcd is the historical risk transmission coefficient of the factor and Gzg is the highest historical transmission coefficient of the correlation layer factor.

[0241] In this embodiment, the secondary cluster is a cluster composed of factors whose fitness improvement during the cross-optimization process is less than a preset threshold, and is used for subsequent local optimization.

[0242] This invention provides a method for monitoring and warning of medical device risks based on real-world evidence. In obtaining the actual contribution value of a single risk feature point, the method is tailored to scenarios where a handheld data acquisition terminal collects data on the target medical device, including:

[0243] Timestamp synchronization and spatial coordinate matching are performed on multi-source heterogeneous real-world data within the same monitoring phase to obtain a fused feature dataset, and the clinical monitoring distance L0 between the current monitoring site and the target medical device and the subject patient is measured.

[0244] In this embodiment, the data acquisition terminal refers to a handheld mobile terminal used to collect medical device operation data and patient vital sign data, including pacemaker programmers, handheld in vitro diagnostic devices, portable vital sign monitors, etc.

[0245] In this embodiment, time dimension synchronization is achieved by using the start timestamp T0 and end timestamp T1 of the monitoring phase as a reference, converting all data timestamps to UTC standard time, and using linear interpolation to align the time dimension of non-continuously collected data, ensuring that the time dimension error of all data is ≤1s.

[0246] Spatial dimension matching: Using the UDI code of the target medical device as the first anchor point and the diagnosis ID of the subject patient as the second anchor point, the operational status data is bound to the target medical device one by one, and the clinical application data, patient outcome data, adverse event data are matched with the corresponding subject patient and target medical device one by one, so as to realize the spatial dimension alignment of multi-source data and obtain the registered fusion feature dataset.

[0247] Rotate the multi-source heterogeneous real-world data acquisition terminal horizontally by a preset calibration angle. , The value range is 0.5° to 3°;

[0248] In this embodiment, the system error correction for the acquisition scene is performed on the data acquired by the handheld mobile acquisition terminal (such as a pacemaker programmer or a handheld in vitro diagnostic device) to perform pose and distance correction of the acquisition scene and eliminate system errors in the data acquisition process.

[0249] In this embodiment, the calibrated fusion feature dataset is decomposed into a core layer data subset and an associated layer data subset according to the hierarchical architecture of the prior knowledge base, ensuring that each data subset corresponds one-to-one with the feature dimensions of the prior knowledge base;

[0250] Standardized feature matching: The decomposed data subset is mapped one-to-one with the risk feature points in the prior knowledge base at the field level. The feature points that match successfully are directly included in the set of risk feature points to be verified.

[0251] Abnormal Feature Point Supplementation and Identification: For abnormal data segments in the calibrated fused feature dataset that do not match the prior knowledge base, the Z-score normalization method and box plot method are used to identify abnormal fluctuation data, extract the feature dimensions corresponding to the abnormal data, and add them as new risk feature points to be verified to the set to be verified.

[0252] For each feature point in the set of risk feature points to be verified, the validity of the risk association is verified. Only feature points that pass the verification are ultimately determined as valid risk feature points. The verification rules are as follows:

[0253] Based on the calibrated fusion feature dataset, the Pearson correlation coefficient between the feature point and the occurrence of adverse events of similar medical devices was calculated, and the absolute value of the correlation coefficient was ≥0.6;

[0254] After controlling confounding factors through propensity score matching (PSM), risk attribution analysis is completed, and the risk attribution score of this feature point ≥ 0.7 (the value range of the attribution score is 0 - 1, and the higher the score, the stronger the causal association strength between this feature point and the adverse event);

[0255] Feature points that meet both of the above requirements are determined as effective risk feature points; feature points that do not meet the requirements are excluded from the set and not included in the subsequent risk quantification calculation, thus obtaining all risk feature points based on the calibrated integrated feature dataset.

[0256] For example, to pre - construct a hierarchical risk feature prior knowledge base for an implantable cardiac pacemaker, the core layer contains 24 prior features, and the associated layer contains 36 prior features; based on the integrated feature dataset after spatio - temporal registration and pose correction, it is disassembled into data subsets of the core layer and the associated layer, and field matching is completed with the prior knowledge base to obtain 58 successfully matched feature points to be verified; based on the box - plot method, the abnormal feature of "battery impedance fluctuation amplitude" in the dataset is identified and supplemented as a new feature point to be verified, and the set to be verified contains a total of 59 feature points; the effectiveness of the 59 feature points to be verified is verified, and finally 57 feature points meet the requirements of the correlation coefficient ≥ 0.6 and the attribution score ≥ 0.7, and are determined as effective risk feature points; unique codes are assigned to the 57 effective risk feature points, and the corresponding correction coefficients are matched to complete the final version for the calculation of subsequent benchmark risk values and actual contribution values.

[0257] In this embodiment, for fixed - installation devices, environmental temperature and humidity, electromagnetic interference, and power supply stability correction are enabled; for implantable devices, impedance, threshold, and battery parameter self - correction are enabled.

[0258] In this embodiment, for handheld acquisition terminals, pose correction is adopted, which specifically includes:

[0259] Using the built - in laser ranging module of the acquisition terminal, measure the straight - line distance L01 from the center of the acquisition terminal sensor to the implant site / action site of the target medical device, and the straight - line distance L02 to the lesion site of the test patient, and take the arithmetic mean of the two as L0, with the unit of meter, that is, L0 = (L01 + L02) / 2.

[0260] Calibrate the pose parameters of the acquisition terminal to obtain the preset calibration angle of the acquisition terminal , is the horizontal angle deviation between the optical axis / sensor axis of the acquisition terminal and the target monitoring plane, which is pre - calibrated by the installation pose of the acquisition terminal and the type of sensor. The specific calibration rule is: acquisition terminal for implantable medical devices , in vitro diagnostic equipment , acquisition terminal for passive medical devices ;

[0261] Horizontal rotation reference: Using the sensor axis of the initial pose of the acquisition terminal as a reference, rotate in a direction parallel to the target monitoring plane to ensure that the sensor is still aligned with the target monitoring point after rotation.

[0262] Based on preset calibration angle The raw data collected is subjected to pose correction to obtain a calibrated fused feature dataset. Simultaneously, the calibration monitoring distance L1 is measured and combined with the clinical monitoring distance L0 to determine the tilt correction angle of the acquisition terminal corresponding to the target medical device. ;

[0263] In this embodiment, the same measurement method as L0 is used to measure the arithmetic mean L1 of the straight-line distance from the acquisition terminal to the two target points after rotation, in meters.

[0264] In this embodiment, The calculation formula is:

[0265] , Tilt correction angle Where k is the clinical monitoring scenario adaptation coefficient, with a value range of 0.8 to 1.2, and is pre-calibrated according to the type of target medical device and the clinical application scenario: implantable medical devices k=1.2, in vitro diagnostic medical devices k=1.0, and passive medical devices k=0.8; This is the average of L0 and L1, in meters. ,when At °, °, no pose correction required; The value range is -85° to 85°, when If the data is found to be in an abnormal pose, the abnormal data will be removed. The initial tilt correction angle is the same as the calculated final tilt correction angle. Intermediate results; It is an angle measurement. The result after conversion to radians;

[0266] All risk feature points were identified based on the calibrated fused feature dataset;

[0267] According to the tilt correction angle Determine the corrected data confidence level and calculate the baseline risk quantification value S0 corresponding to a single risk feature point;

[0268] The calculation formula is: ,in, The corrected data confidence coefficient ranges from 0.5 to 1.0. , The confidence coefficient for the raw data is pre-calibrated based on the reliability of the data source corresponding to the risk characteristics: data originating from the adverse event monitoring system. =1.0, clinical data from the medical institution's HIS system. =0.9, data from the instrument operation log. =0.85, derived from data reported during patient follow-up. =0.7, data from a third-party organization =0.5; R0 is the basic risk score for a single feature point, which is pre-calibrated by the historical risk level of the same type of risk feature, and the value range is 0~10 points; m is the quantitative calibration coefficient of the risk feature point, which is pre-calibrated by the type of target medical device and the monitoring accuracy requirements, and the value range is 0.8~1.2. The tilt correction factor is the tilt angle. The larger the value, the lower the signal-to-noise ratio and the smaller the confidence correction coefficient, thus achieving... The angle's corrective effect on risk values.

[0269] Based on the benchmark risk quantification value S0 and the risk level weighting coefficient, the actual risk contribution value S1 corresponding to a single risk feature point is obtained, and the calculation formula is as follows: ,in, This is the risk level weighting coefficient, with a value ranging from 0.5 to 1.5, determined by the preset risk level corresponding to a single risk feature point: severe risk feature. =1.5, higher risk characteristics =1.2, general risk characteristics =1.0, low-risk characteristic =0.5; the value of S1 ranges from 0 to 10.

[0270] This invention provides a method for monitoring and early warning of medical device risks based on real-world evidence, generating an initial prediction sequence for the target medical device, including:

[0271] Collect the time-series risk quantification values ​​corresponding to the acceptance nodes of each monitoring stage, and decompose them into core layer time-series quantification values ​​and related layer time-series quantification values ​​according to the risk level;

[0272] Based on the acceptance and compliance results and historical risk contribution of each monitoring stage, dynamic time-series weighting coefficients are matched to the time-series quantification values ​​of the core layer and the related layer. The weighted time-series quantification values ​​are progressively accumulated according to the chronological order of the monitoring stages to generate the risk evolution time-series sequence of the target medical device throughout its entire clinical cycle. The progressive accumulation formula is as follows: ,in, Let be the cumulative value of risk evolution in the t-th monitoring phase. This represents the cumulative risk evolution value during the (t-1)th monitoring phase. Let be the core layer time-series weighting coefficient for the t-th monitoring phase. Let be the core layer time series quantization value for the t-th monitoring phase. Let be the correlation layer time series weight coefficient for the t-th monitoring stage. Let be the time-series quantization value of the associated layer in the t-th monitoring phase; The value is always greater than ;

[0273] Align the comprehensive risk characteristics of each monitoring stage with the risk evolution time series in the time dimension, extract static risk distribution characteristics, trend change characteristics and mutation risk characteristics, and perform multi-dimensional feature fusion enhancement to generate a time-series enhanced risk feature set;

[0274] The time-series enhanced risk feature sets of each monitoring stage are input into a preset dual-constraint regularized risk prediction model to generate the initial prediction sequence of the target medical device.

[0275] In this embodiment, the time-series risk quantification value .

[0276] In this embodiment, , The value range is from 0 to 100, and the initial value is... , Hys, the original risk score of the core layer, is the sum of the actual contribution values ​​of all risk factors in the core layer and the corresponding factor weights. HLmax is the theoretical maximum value of the original risk score of the core layer. The original risk score of the associated layer needs to be multiplied by the real-time risk activity of the corresponding associated monitoring object, and then weighted and summed. Here, Gys is the original risk score of the associated layer, and GLmax is the theoretical maximum value of the original risk score of the associated layer.

[0277] ,and ,and Where Hgd represents the contribution of historical risk to the core layer; Zgd represents the contribution of total historical risk.

[0278] In this embodiment, the static risk distribution characteristics are as follows: the kernel density estimation (KDE) method is used to extract the probability distribution characteristics of the comprehensive risk characteristics, including mean, median, variance, quantile, kurtosis, and skewness.

[0279] Trend change characteristics: Using linear regression, a trend line is fitted to the risk evolution time series, and the trend slope, coefficient of determination, and volatility are extracted;

[0280] Mutation risk characteristics: The sliding window standard deviation and cumulative sum (CUSUM) algorithm is used to extract the mutation point location, mutation magnitude, and mutation frequency of the risk sequence;

[0281] Feature fusion enhancement method: A multi-head attention mechanism is adopted to perform weighted fusion of three types of features, highlighting features that contribute highly to risk prediction, and generating a time-series enhanced risk feature set with fixed dimensions.

[0282] In this embodiment, the dual-constraint regularized risk prediction model has the following specific structure:

[0283] The base network layer uses a 3-layer LSTM network to extract long-term dependencies of temporal features. The hidden layer has a dimension of 64 and a dropout rate of 0.2 to prevent overfitting. The input layer is a temporal enhancement risk feature set, and the output layer is a 64-dimensional temporal feature vector.

[0284] Fully connected layer: 2 fully connected layers. The first layer has a dimension of 32 and uses ReLU as the activation function. The second layer has a dimension of N (prediction step size) and uses Sigmoid as the activation function. It outputs a risk prediction value of 0 to 100.

[0285] Double-constraint regularization layer: A double-regularization constraint term is added to the loss function. The formula for calculating the loss function is as follows: ,in, Mean squared error loss is used to measure the deviation between the predicted value and the true value. ,and Let be the true value of the i-th sample used for model prediction. This is the predicted value for the i-th sample used in the model prediction; As a risk event discrimination constraint, a contrastive loss function is used to maximize the feature distance between risk events and non-risk events. ,and The batch size is the number of samples contained in the current training batch. This is the sample label for the b-th sample in the current training batch (1 = risk event, 0 = non-risk event). For feature distance, The boundary threshold is fixed at 1.0; This is a regulatory compliance constraint used to ensure that the prediction results meet the threshold requirements of medical device regulatory rules, and Ensure that the prediction deviation is ≤15%, which meets the regulatory tolerance requirements for deviation; , These are regularization coefficients, with values ​​of 0.5, 0.3, and 0.2 respectively. , To ensure the theoretical maximum value of the corresponding loss term, all terms are dimensionless values ​​in the range of 0 to 1. To participate and The total number of samples calculated is the number of samples used for model prediction.

[0286] Training dataset: Real-world dataset of adverse events for all types of medical devices, with a sample size of ≥100,000, of which risk events account for ≥20%;

[0287] Dataset partitioning: 70% training set, 15% validation set, and 15% test set;

[0288] Training parameters: The optimizer used is Adam, the learning rate is 0.001, the batch size is 32, the number of training rounds is 100, and the early stopping policy is to stop training if the validation set loss does not decrease for 10 consecutive rounds.

[0289] Model qualification criteria: test set prediction accuracy ≥ 85%, AUC ≥ 0.85, serious risk event identification rate ≥ 95%, false positive rate ≤ 10%, meeting regulatory requirements.

[0290] This invention provides a method for monitoring and early warning of medical device risks based on real-world evidence, obtaining a target risk sequence that meets preset regulatory requirements, including:

[0291] Based on real-world data from the historical monitoring phase that has been completed and accepted, the corresponding risk prediction results, and historical regulatory compliance conclusions, the historical prediction deviations are broken down into core inherent risk deviations and related risk deviations. Initial deviation correction coefficients are then assigned to the core layer and related layer risks, respectively. These initial deviation correction coefficients are subject to the deviation tolerance limits set by the medical device regulatory rules.

[0292] In this embodiment, the core inherent risk deviation is the relative deviation between the predicted value and the actual value of the core layer risk, which originates from the prediction error of the inherent risk of the device itself; the associated risk deviation is the relative deviation between the predicted value and the actual value of the associated layer risk, which originates from the prediction error of the risk.

[0293] In this embodiment, the cold start calibration scheme for newly launched medical devices is as follows: For newly launched medical devices without historical monitoring data, the historical dataset of already launched medical devices with the same principle, risk level, and applicable scenario is used to complete the cold start calibration of the initial deviation correction coefficient; at the same time, the dynamic learning rate is set to 0.3, and after the acceptance of every 3 monitoring stages, the correction coefficient is updated once based on the newly generated monitoring data to gradually complete the personalized adaptation of the model.

[0294] Collect real-world data from real-time clinical applications, extract temporal mutation features, risk clustering features, and regulatory sensitivity features of real-time risk characteristics, and dynamically update the initial deviation correction coefficients of the core layer and related layers.

[0295] Based on the updated bias correction coefficient, the initial prediction sequence is subjected to progressive bias compensation in the time series dimension and weighted correction in the risk level dimension to generate a pre-corrected risk sequence.

[0296] The pre-corrected risk sequence is verified based on the preset regulatory risk threshold and prediction accuracy requirements. If the verification fails, the hierarchical deviation constraint term and correction coefficient are iteratively optimized until the preset regulatory requirements are met, and the target risk sequence is output.

[0297] In this embodiment, the formula for calculating the initial deviation correction coefficient is:

[0298] , ;

[0299] in, This is the initial deviation correction coefficient for the core layer. The initial deviation correction factor for the correlation layer is T0, where T0 is the total number of historical monitoring phases that have been completed and accepted. , These represent the actual value and the predicted value of the core layer time-series quantization value at the t-th historical stage, respectively. , These are the actual and predicted values ​​of the time-series quantization value of the association layer at the t-th historical stage, respectively. , The minimum value set;

[0300] It should be noted that the initial deviation correction coefficient is constrained by the deviation tolerance preset in the medical device regulatory rules. The deviation tolerance is ±20%. When the correction coefficient exceeds the range of 0.8 to 1.2, it is fixed as the boundary value.

[0301] In this embodiment, the dynamic update formula for the deviation correction coefficient is: , ,in, , This is the correction factor for the deviation between the updated core layer and related layers. The real-time anomaly magnitude of the core layer risk characteristics ranges from -0.2 to 0.2, and is calculated as the relative deviation between the current monitoring phase's core layer risk average and the historical average. The real-time anomaly magnitude of the associated layer risk characteristics is measured in the range of -0.2 to 0.2. It is calculated as the relative deviation between the average risk value of the associated layer during the current monitoring phase and the historical average. It should be noted that the updated correction coefficient still needs to meet the boundary constraint of 0.8 to 1.2.

[0302] In this embodiment, the formula for calculating the pre-corrected risk value is: ,in, Let t be the pre-corrected risk value for the prediction stage. This is the predicted value for the t-th stage in the initial prediction sequence.

[0303] In this embodiment, the preset regulatory requirements include: a tolerance for deviations in risk prediction values ​​of ≤ ±15%; an identification rate of serious risk events of ≥ 95%; and a false alarm rate of risk warnings of ≤ 10%.

[0304] In this embodiment, the iterative optimization rule is as follows:

[0305] Verification passed: The pre-corrected risk sequence meets all three regulatory requirements and is directly output as the target risk sequence;

[0306] Verification Failure: Using the minimization of prediction bias as the objective function, the gradient descent method is used to iteratively optimize the hierarchical bias constraint term and correction coefficient. The iteration step size is 0.01, and the maximum number of iterations is 50. If the requirement is still not met after reaching the maximum number of iterations, the optimal iteration result is output and a manual review prompt is triggered.

[0307] Taking implantable pacemaker risk monitoring as an example:

[0308] Target device: Class III implantable cardiac pacemaker; Monitoring period: 12 months; Monitoring phase: 7 days / phase, 52 phases in total; Monitoring subjects: 120 implanted patients.

[0309] Step 1: Data Collection and Phase Division:

[0310] Data on operational status, clinical application, patient outcomes, adverse events, equipment maintenance, environmental parameters, and follow-up are collected and monitored in 7-day / phase cycles, with timestamps standardized to UTC.

[0311] Step 2: Spatiotemporal registration and data correction:

[0312] Spatiotemporal registration: Using UDI codes and patient treatment IDs as anchors, with a time error of ≤1s, multi-source data alignment is completed.

[0313] Handheld terminal calibration: The data acquisition distance of the programmable controller is L0=0.5m, the rotation is α=1°, and after calibration L1=0.52m; k=1.2, Lavg=0.51m, calculate θ=reasonable range; calibrate the confidence level of the data, and calculate the actual contribution value of the risk feature point.

[0314] Step 3: Risk Association Identification and Comprehensive Feature Fusion:

[0315] Related entities: 1200 units in the same batch, 1200 patients with the same indication, 20 hospitals of the same level, and 5 pacemakers of the same risk type.

[0316] Factor library: 24 core layers and 36 related layers, 8-bit encoding; M=300, D=60.

[0317] Clustering optimization: K-means clustering of 10 clusters, four-dimensional matrix [10,52,2,3]; cross step size ∆ is calculated according to the formula, simulating binary cross optimization; core layer weight 0.7, association layer weight 0.3, generating comprehensive risk features.

[0318] Step 4: Time Series and Initial Prediction:

[0319] Evolutionary sequence: Accumulated progressively according to a formula, core layer weights =0.7, related layer =0.3.

[0320] Initial prediction: After feature enhancement, the model is input into a 3-layer LSTM double-constraint model to predict the next 4 stages: [28, 32, 35, 42].

[0321] Step 5: Dynamic Correction and Target Sequence:

[0322] Initial correction coefficients: 1.0417 for the core layer and 0.95 for the related layer, both within the range of 0.8 to 1.2.

[0323] Cold start adaptation: The new device is calibrated using historical data from the same model pacemaker, and the coefficients are updated every 3 stages.

[0324] Real-time correction: After updating the coefficients, the pre-corrected sequence [30.8,35.2,38.5,46.2] is verified to meet regulatory requirements, and the target sequence is output.

[0325] Step 6: Risk Level and Early Warning:

[0326] The highest predicted score is 46.2, which is considered a relatively high risk. A yellow alert has been issued, and risk screening and enhanced follow-up have been initiated.

[0327] This invention provides a medical device risk monitoring and early warning system based on real-world evidence, such as... Figure 2 As shown, it includes:

[0328] The data acquisition module is used to divide the clinical application process of the target medical device into multiple continuous monitoring stages and collect multi-source heterogeneous real-world data corresponding to the target medical device in each monitoring stage.

[0329] The contribution value determination module is used to perform spatiotemporal dimension synchronous registration of multi-source heterogeneous real-world data within the same monitoring phase to obtain a fused feature dataset. It calculates the baseline risk value corresponding to a single risk feature point in the fused feature dataset and corrects it based on the acquisition parameters, data confidence and risk weight to obtain the actual contribution value of a single risk feature point.

[0330] The feature fusion module is used to identify associated monitoring objects with risk correlation, with the target medical device as the core monitoring object, extract the individual risk characteristics of the core monitoring object, the associated risk characteristics of each associated monitoring object and the real-time risk activity, and fuse them to generate the comprehensive risk characteristics of the target medical device at each monitoring stage.

[0331] The sequence generation module is used to collect the time-series risk quantification values ​​corresponding to the acceptance nodes of each monitoring stage, accumulate the time-series risk quantification values ​​to generate a risk evolution time-series sequence, and process the comprehensive risk characteristics of each monitoring stage to generate the initial prediction sequence of the target medical device.

[0332] The sequence correction module is used to dynamically correct the initial predicted sequence based on real-world data from the historical monitoring phase that has been accepted and the corresponding risk prediction results, combined with real-time clinical application process data, to obtain a target risk sequence that meets preset regulatory requirements.

[0333] The risk warning module is used to determine the risk parameter set of the target medical device based on the target risk sequence, and output the risk level and graded warning.

[0334] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for monitoring and early warning of medical device risks based on real-world evidence, characterized in that, include: Step 1: Divide the clinical application process of the target medical device into multiple continuous monitoring stages, and collect multi-source heterogeneous real-world data corresponding to the target medical device in each monitoring stage; Step 2: Perform spatiotemporal dimension synchronous registration on multi-source heterogeneous real-world data within the same monitoring phase to obtain a fused feature dataset. Calculate the baseline risk value corresponding to a single risk feature point in the fused feature dataset, and correct it based on the acquisition parameters, data confidence, and risk weight to obtain the actual contribution value of a single risk feature point. Step 3: Using the target medical device as the core monitoring object, identify the associated monitoring objects with risk correlation, extract the individual risk characteristics of the core monitoring object, the associated risk characteristics of each associated monitoring object, and the real-time risk activity, and fuse them to generate the comprehensive risk characteristics of the target medical device at each monitoring stage; Step 4: Collect the time-series risk quantification values ​​corresponding to the acceptance nodes of each monitoring stage, accumulate the time-series risk quantification values ​​to generate a risk evolution time-series sequence, and process the comprehensive risk characteristics of each monitoring stage to generate the initial prediction sequence of the target medical device. Step 5: Based on real-world data from the historical monitoring phase that has been completed and accepted, and the corresponding risk prediction results, the initial prediction sequence is dynamically corrected in conjunction with real-time clinical application process data to obtain a target risk sequence that meets the preset regulatory requirements. Step 6: Determine the risk parameter set of the target medical device based on the target risk sequence, and output the risk level and graded warning.

2. The method for medical device risk monitoring and early warning based on real-world evidence according to claim 1, characterized in that, The comprehensive risk characteristics of the target medical device at each monitoring stage are generated by fusion, including: For the fusion feature dataset of each monitoring stage, the core layer risk factors corresponding to the inherent risk characteristics of the target medical device itself and the associated layer risk factors corresponding to the risk characteristics of clinical application, patient group and environmental control are extracted. A risk factor library is constructed and each risk factor is uniquely encoded according to the risk source level. The initial total number of factor populations M is set according to the total number of factors in the risk factor library. The number of effective risk factor dimensions contained in each initial factor population is set to D. D matches the total dimension of the risk feature cluster in a single monitoring stage. An initial factor population is randomly generated according to the risk factor hierarchical coding. Based on the real-world dataset of historical adverse events of medical devices, a multi-dimensional pre-evaluation of the fit of each individual in the initial factor population is performed. Invalid individuals with fit below a preset threshold are removed to obtain valid individuals. All valid individuals under the same initial factor population are clustered according to the risk optimization objective to obtain multiple population clusters and transform them into a four-dimensional structured matrix to determine the risk optimization objective of the corresponding initial factor population; According to the risk optimization objective, the risk contribution dominance levels of effective factors in the corresponding initial factor population are sorted in ascending order to generate a factor sequence and determine the crossover step size of adjacent effective factors. Based on the cross step size analysis, the initial number of cross factors and the single adjustment position distribution are adjusted, and the adjustment frequency and overall position distribution are statistically analyzed to obtain the secondary clusters; The population clusters and secondary clusters of each initial factor population are merged and divided to obtain global clusters and local clusters. The derived clusters are determined based on the first individual variable of the global cluster, the second individual variable of the local cluster, and the coverage coefficient, thus obtaining the cluster sequence of the corresponding initial factor population. Determine the hierarchical fusion weight coefficients of the cluster sequences of all initial factor populations and each risk feature cluster, wherein the weight coefficients of the core layer risk factors and the weight coefficients of the associated layer risk factors are positively correlated with the risk contribution of the core monitoring objects and the associated monitoring objects, respectively. The individual risk characteristics of the core monitoring object, the associated risk characteristics of each related monitoring object, and the real-time risk activity are weighted and fused according to the hierarchical fusion weight coefficient to generate the comprehensive risk characteristics of the target medical device at each monitoring stage.

3. The method for monitoring and early warning of medical device risks based on real-world evidence according to claim 2, characterized in that, The core layer risk factors and the associated layer risk factors respectively correspond to the risk characteristic dimensions of the core monitoring object and the associated monitoring object; The risk optimization objectives include risk event differentiation, model generalization, and regulatory rule adaptability; The four dimensions of the four-dimensional structured matrix are as follows: The first dimension is the cluster dimension, and the number of rows in the matrix corresponds to the number of clusters. The second dimension is the time-series stage dimension, and the number of columns in the matrix corresponds to the total number of monitoring stages in the entire clinical cycle of the target medical device; the third dimension is the factor stratification dimension, and the matrix depth corresponds to the number of risk levels of the risk factors. The fourth dimension is the evaluation channel dimension, and the number of matrix channels corresponds to the number of evaluation dimensions in the comprehensive evaluation system.

4. The method for medical device risk monitoring and early warning based on real-world evidence according to claim 2, characterized in that, The formula for calculating the cross step length is: ,in, The dimension length of the risk factors in the association layer; This refers to the hierarchical risk weighting coefficient, with a value ranging from 0.4 to 1.6, representing the core layer risk factors. Values ​​higher than those of the correlation layer risk factors Values; This is the dynamic coefficient for the time series stage, with a value range of 0.7 to 1.3; The cross step size; The rounding up symbol.

5. The method for monitoring and early warning of medical device risks based on real-world evidence according to claim 1, characterized in that, In obtaining the actual contribution value of a single risk feature point, scenarios for collecting target medical device data using a handheld data acquisition terminal include: Timestamp synchronization and spatial coordinate matching are performed on multi-source heterogeneous real-world data within the same monitoring phase to obtain a fused feature dataset, and the clinical monitoring distance L0 between the current monitoring site and the target medical device and the subject patient is measured. Rotate the multi-source heterogeneous real-world data acquisition terminal horizontally by a preset calibration angle. , The value range is 0.5° to 3°; Based on preset calibration angle The raw data collected is subjected to pose correction to obtain a calibrated fused feature dataset. Simultaneously, the calibration monitoring distance L1 is measured and combined with the clinical monitoring distance L0 to determine the tilt correction angle of the acquisition terminal corresponding to the target medical device. ; All risk feature points were identified based on the calibrated fused feature dataset; Based on tilt correction angle Determine the corrected data confidence level and calculate the baseline risk quantification value S0 corresponding to a single risk feature point; Based on the benchmark risk quantification value S0 and the risk level weighting coefficient, the actual risk contribution value S1 corresponding to a single risk feature point is obtained.

6. The method for monitoring and early warning of medical device risks based on real-world evidence according to claim 1, characterized in that, Generating an initial prediction sequence for the target medical device includes: Collect the time-series risk quantification values ​​corresponding to the acceptance nodes of each monitoring stage, and decompose them into core layer time-series quantification values ​​and related layer time-series quantification values ​​according to the risk level; Based on the acceptance and compliance results and historical risk contribution of each monitoring stage, dynamic time-series weighting coefficients are matched to the time-series quantification values ​​of the core layer and the related layer. The weighted time-series quantification values ​​are progressively accumulated according to the chronological order of the monitoring stages to generate the risk evolution time-series sequence of the target medical device throughout its entire clinical cycle. The progressive accumulation formula is as follows: ,in, Let be the cumulative value of risk evolution in the t-th monitoring phase. This represents the cumulative risk evolution value during the (t-1)th monitoring phase. Let be the core layer time-series weighting coefficient for the t-th monitoring phase. Let be the core layer time series quantization value for the t-th monitoring phase. Let be the correlation layer time series weight coefficient for the t-th monitoring stage. Let be the time-series quantization value of the associated layer in the t-th monitoring phase; The value is always greater than ; Align the comprehensive risk characteristics of each monitoring stage with the risk evolution time series in the time dimension, extract static risk distribution characteristics, trend change characteristics and mutation risk characteristics, and perform multi-dimensional feature fusion enhancement to generate a time-series enhanced risk feature set; The time-series enhanced risk feature sets of each monitoring stage are input into a preset dual-constraint regularized risk prediction model to generate the initial prediction sequence of the target medical device.

7. The method for monitoring and early warning of medical device risks based on real-world evidence according to claim 1, characterized in that, The target risk sequence that meets the preset regulatory requirements is obtained, including: Based on real-world data from the historical monitoring phase that has been completed and accepted, the corresponding risk prediction results, and historical regulatory compliance conclusions, the historical prediction deviations are broken down into core inherent risk deviations and related risk deviations. Initial deviation correction coefficients are then assigned to the core layer and related layer risks, respectively. These initial deviation correction coefficients are subject to the deviation tolerance limits set by the medical device regulatory rules. Collect real-world data from real-time clinical applications, extract temporal mutation features, risk clustering features, and regulatory sensitivity features of real-time risk characteristics, and dynamically update the initial deviation correction coefficients of the core layer and related layers. Based on the updated bias correction coefficient, the initial prediction sequence is subjected to progressive bias compensation in the time series dimension and weighted correction in the risk level dimension to generate a pre-corrected risk sequence. The pre-corrected risk sequence is verified based on the preset regulatory risk threshold and prediction accuracy requirements. If the verification fails, the hierarchical deviation constraint term and correction coefficient are iteratively optimized until the preset regulatory requirements are met, and the target risk sequence is output.

8. A medical device risk monitoring and early warning system based on real-world evidence, characterized in that, include: The data acquisition module is used to divide the clinical application process of the target medical device into multiple continuous monitoring stages and collect multi-source heterogeneous real-world data corresponding to the target medical device in each monitoring stage. The contribution value determination module is used to perform spatiotemporal dimension synchronous registration of multi-source heterogeneous real-world data within the same monitoring phase to obtain a fused feature dataset. It calculates the baseline risk value corresponding to a single risk feature point in the fused feature dataset and corrects it based on the acquisition parameters, data confidence and risk weight to obtain the actual contribution value of a single risk feature point. The feature fusion module is used to identify associated monitoring objects with risk correlation, with the target medical device as the core monitoring object, extract the individual risk characteristics of the core monitoring object, the associated risk characteristics of each associated monitoring object and the real-time risk activity, and fuse them to generate the comprehensive risk characteristics of the target medical device at each monitoring stage. The sequence generation module is used to collect the time-series risk quantification values ​​corresponding to the acceptance nodes of each monitoring stage, accumulate the time-series risk quantification values ​​to generate a risk evolution time-series sequence, and process the comprehensive risk characteristics of each monitoring stage to generate the initial prediction sequence of the target medical device. The sequence correction module is used to dynamically correct the initial predicted sequence based on real-world data from the historical monitoring phase that has been accepted and the corresponding risk prediction results, combined with real-time clinical application process data, to obtain a target risk sequence that meets preset regulatory requirements. The risk warning module is used to determine the risk parameter set of the target medical device based on the target risk sequence, and output the risk level and graded warning.

Images