Method and system for improving recording of adverse events related to laboratory abnormal values

Abstract

The invention discloses a method for improving recording of adverse events related to laboratory abnormal values, and relates to the technical field of internet data information services, and the method comprises the following steps: sensing multi-modal data in real time, and collecting equipment operation data, environment parameters and operation logs in real time; constructing a digital twinborn body to form a dynamic data portrait; performing double-layer dynamic anomaly detection to form an instant rule layer and an intelligent model layer, and updating a model according to the change of a detection standard; recording knowledge graph attribution, generating an attribution path, and automatically generating a structured record; and block chain closed-loop verification is carried out, a closed-loop state is automatically marked, and cross-mechanism privacy query is carried out. According to the method, federal learning and the knowledge graph are fused for the first time, the problems of cross-mechanism data islands and attribution fuzziness are solved, and a record closed loop is realized; static threshold limitation is broken through, detection methodology iteration is adapted, and model updating does not need manual intervention; the abnormal attribution accuracy is improved, and the existing limitation is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of Internet data information services, and in particular to a method and system for improving the recording of adverse events related to laboratory abnormal values. Background Art

[0002] Laboratory outliers are observations in experimental data that significantly deviate from the expected or normal range. They may be due to a variety of factors, including experimental error, equipment failure, environmental fluctuations, or sample abnormalities. Accurately recording and analyzing adverse events related to laboratory outliers is crucial for ensuring experimental data quality, tracing the root causes of problems, and optimizing experimental processes.

[0003] However, existing traditional methods for recording and analyzing adverse events associated with laboratory abnormal values ​​have many limitations and are unable to meet the growing data management and analysis needs of modern laboratories. Traditional methods rely on manual recording, which is prone to errors and has poor timeliness. Traditional laboratory information management systems (LIMS) often rely on manual entry of abnormal events, which is subject to recording delays, subjective biases, and error-proneness. Manual recording is difficult to perform in real time, resulting in delayed recording of abnormal events, which affects the efficiency of subsequent analysis and processing. Different personnel may have different judgment criteria for abnormal events, resulting in inconsistent recording results, affecting the objectivity and reliability of the data. Errors and omissions are prone to occur during manual entry, resulting in inaccurate data, which affects subsequent analysis and decision-making.

[0004] Furthermore, existing LIMS systems typically only record experimental data itself, lacking integration of relevant information such as equipment operation logs and environmental monitoring data. This single-data-source analysis model struggles to fully reflect the true nature of abnormal events. For example, failure to integrate equipment operation logs can make it impossible to accurately determine whether abnormal values ​​are caused by equipment failures; failure to integrate environmental monitoring data can make it impossible to accurately determine whether abnormal values ​​are caused by environmental fluctuations, resulting in insufficient accuracy in abnormal attribution. This makes existing traditional methods for recording and analyzing adverse events related to laboratory abnormal values ​​subject to numerous limitations, making it difficult to meet the growing data management and analysis needs of modern laboratories.

[0005] Therefore, it is urgent to develop a new method and system to solve the above problems. Summary of the Invention

[0006] The purpose of the present invention is to provide a method and system for improving the recording of adverse events related to laboratory abnormal values, so as to solve the problems raised in the above background technology.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for recording adverse events related to laboratory abnormal values, comprising the following steps:

[0008] Real-time multimodal data perception: using edge computing gateways to collect real-time laboratory equipment operating data, environmental parameters, and operation logs; building equipment digital twins, synchronously mapping the implicit parameters of laboratory equipment, and forming dynamic data portraits;

[0009] Two-layer dynamic anomaly detection, forming an instant rule layer and an intelligent model layer, and updating the model according to changes in detection standards;

[0010] Knowledge graph attribution records, building laboratory-related knowledge graphs, calculating the probability weights of outliers and associated entities, generating attribution paths, and automatically generating structured records that meet laboratory standards;

[0011] Blockchain closed-loop verification, abnormal records are immediately uploaded to the chain, the CAPA process is triggered by smart contracts, the disposal effect is verified based on subsequent batch quality control data, the closed-loop status is automatically marked, and cross-institutional privacy queries are performed.

[0012] Preferably, in the real-time perception of multimodal data, the edge computing gateway supports the HL7 / IEC62304 protocol, the equipment operation data includes but is not limited to the speed fluctuation of the experimental equipment centrifuge and the temperature curve of the temperature control box, the environmental parameters include but are not limited to temperature, humidity, and air pressure, and the real-time sampling frequency is ≥100Hz.

[0013] Preferably, the multimodal data collection in the multimodal data real-time perception includes:

[0014] Digital twin: Build a 3D model of the equipment based on Unity3D, including real-time dimensional parameters such as synchronous motor current and bearing temperature.

[0015] Differential privacy aggregation: ε-differential privacy is used for cross-institutional data aggregation during cross-institutional transmission, protecting sensitive information such as device IDs and operators, with a data distortion rate of ≤2%;

[0016] Offline caching mechanism: Locally store 72 hours of data when the network is disconnected, and verify the integrity through device fingerprint after the connection is restored, with an omission rate of ≤0.1%.

[0017] Preferably, the dual-layer dynamic anomaly detection includes:

[0018] Instant rule layer: Based on statistical process SPC control, set limit-exceeding triggers and generate primary warnings within 1 second;

[0019] Intelligent model layer: The LSTM-Attention model trained with federated learning is used, with a sliding time window T of 15 minutes. The threshold is dynamically adjusted based on the experimental phase coefficient (threshold = mean + 3σ × phase coefficient) to identify trend anomalies.

[0020] Model update: When the detection standard changes, incremental training is performed with a delta parameter of ≤ 0.1%, and adaptation is performed within 2 hours.

[0021] Preferably, the dual-layer dynamic anomaly detection further includes:

[0022] Model interpretability: Generates a heat map of abnormal influencing factors through SHAP values ​​to assist manual review, supports one-click export of PDF reports, and generates explanation reports in ≤ 2 seconds;

[0023] Drift warning: Based on KL divergence monitoring, the predicted distribution is automatically triggered when the divergence is greater than 0.3. The PoC algorithm is used to select the laboratory samples with the top 20% contribution points. The contribution of the kth abnormal single sample uploaded by the laboratory is calculated as:

[0024]

[0025] n i is the historical cumulative sample number of the i-th type of anomaly in the global database; α=100, β=1 are the scarcity smoothing parameters; k i,k The consistency coefficient of the k-th sample annotation; acc i,k is the sample labeling accuracy; γ = 0.6 is the consistency weight; λ = 0.99 is the time decay factor; t now -t k The time difference between sample uploads.

[0026] Preferably, in the knowledge graph attribution record, the laboratory-related knowledge graph includes but is not limited to equipment ID, reagent batch number and personnel qualifications. The probability weights of outliers and associated entities are calculated by GNN graph neural network, and the weight control value is ≥0.7. The generated structured record includes but is not limited to the original data hash, impact range and disposal suggestions, and the record generation time is ≤10 seconds.

[0027] Preferably, the knowledge graph attribution in the knowledge graph attribution record includes:

[0028] Entity relationship enhancement: define different types of relationships and train relationship confidence through the TransE algorithm;

[0029] Smart reporting assistance: Use the mobile app to scan the device's NFC chip and automatically fill in the device model and current experimental stage.

[0030] Preferably, the blockchain closed loop in the blockchain closed loop verification includes:

[0031] Smart contract logic: When the abnormality level is ≥3, the handling task is automatically pushed to the responsible person through DingTalk / WeChat Enterprise. If the task response timeout occurs, a warning from the superior is triggered. If the mean deviation is ≤1.5% through comparison of subsequent batch quality control data, the smart contract automatically marks the CAPA as valid.

[0032] Privacy-preserving query: Regulators can verify the authenticity of specific events through zero-knowledge proofs without obtaining original data, with verification time ≤ 200ms;

[0033] Electronic signature solidification: Disposal records must be signed via USB-Key in accordance with security requirements standards, the signature timestamp is accurate to milliseconds, and the tamper-proof hash value is synchronized on the chain.

[0034] The present invention also provides a system for improving the recording of adverse events related to laboratory abnormal values, comprising:

[0035] The multimodal edge acquisition module deploys an edge computing gateway that supports HL7v2.8 and IEC62304 protocols. It uses digital twin technology to map the operating parameters and environmental data of different equipment such as centrifuges and temperature control boxes in real time, and aggregates cross-institutional data using differential privacy algorithms.

[0036] Federated dynamic detection framework: This framework uses a sliding time window LSTM-Attention model, integrates the SHAP interpretability framework to generate anomaly attribution heatmaps, and combines it with the SPC statistical process control rule engine for two-layer detection. Model parameters are updated across institutions through federated learning, supporting incremental training when detection standards change.

[0037] Knowledge graph attribution engine: Builds a knowledge graph containing equipment ID, reagent batch number, operator qualifications, and experimental stage, calculates the probability weights of outliers and related entities through the GNN graph neural network, and automatically generates structured reports;

[0038] Blockchain closed-loop tracking system: Based on Hyperledger Fabric's consortium chain evidence storage, it drives the CAPA process through smart contracts, integrates electronic signatures for tamper-proof traceability of the disposal process, and supports penetrating queries by regulators.

[0039] Preferably, the edge computing gateway in the multimodal edge acquisition module includes:

[0040] Protocol conversion submodule, which supports real-time conversion from RS485 and ModbusTCP to MQTT, with a data acquisition frequency of ≥100Hz;

[0041] A digital twin engine, which builds a 3D model of the device based on Unity3D and synchronizes device parameters in real time with an error rate of ≤0.5%;

[0042] The offline cache module stores data locally for 72 hours when the network is disconnected, and verifies the data integrity through device fingerprint after the connection is restored.

[0043] Technical effects and advantages of the present invention:

[0044] (1) This invention integrates federated learning and knowledge graphs for the first time, solving the problems of cross-institutional data silos and attribution ambiguity, and realizing a closed loop of "local detection → global optimization → accurate recording"; breaking through the static threshold limit, adapting to the iteration of detection methodology, and updating the model without manual intervention; achieving regulatory penetration while protecting data privacy, realizing the integration of device-related information, comprehensively reflecting the true situation of abnormal events, improving the accuracy of abnormal attribution, and reducing the limitations of existing adverse event records;

[0045] (2) The present invention utilizes a combination of multimodal data real-time perception and dual-layer dynamic anomaly detection to collect laboratory equipment and personnel-related data in real time. Through dynamic data profiling, the stability of data information recording is improved. Furthermore, through drift warning, the time consumption of model prediction retraining is reduced, decision-making efficiency is improved, and the growing data management and analysis needs of modern laboratories are met.

[0046] (3) The method of the present invention takes ≤200ms from the occurrence of an abnormality to the generation of a record. When a key abnormality occurs, the mobile terminal vibrates and speaks to provide dual reminders. Through clinical validation data set testing, the accuracy of multi-dimensional attribution is improved. Blockchain evidence storage supports FDA21 CFR Part 11 audits, and the time for retrieving the chain of evidence is shortened from 4 hours to 10 minutes, improving the efficiency of subsequent analysis and decision-making after the occurrence of laboratory abnormal values. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a step diagram of the method of the present invention.

[0048] Figure 2 This is a diagram of the multimodal data collection steps of the present invention.

[0049] Figure 3 This is a diagram of the steps of double-layer dynamic anomaly detection of the present invention.

[0050] Figure 4 Schematic diagram of the system framework of the present invention. DETAILED DESCRIPTION

[0051] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0052] The present invention provides Figure 1-4 The method for recording adverse events associated with abnormal laboratory values ​​shown includes the following steps:

[0053] Real-time multimodal data perception uses an edge computing gateway that supports the HL7 / IEC62304 protocol to collect laboratory equipment operating data, environmental parameters, and operation logs in real time. Equipment operating data includes but is not limited to centrifuge speed fluctuations and temperature curves of temperature control boxes. Environmental parameters include but are not limited to temperature, humidity, and air pressure. The real-time sampling frequency is ≥100Hz. Build a digital twin of the equipment, synchronously map the hidden parameters of the laboratory equipment, and form a dynamic data portrait.

[0054] Among them, multimodal data collection in real-time perception of multimodal data includes digital twins, differential privacy aggregation and offline caching mechanism. Digital twins are based on Unity3D to build a three-dimensional model of the equipment, which includes but is not limited to the dimensional parameters of synchronous motor current and bearing temperature in real time; differential privacy aggregation adopts the ε-differential privacy method (ε=0.3) when transmitting across institutions, aggregates cross-institutional data, protects sensitive information of equipment ID and operators, and the data distortion rate is ≤2%; the offline caching mechanism is to locally store 72 hours of data when the network is disconnected, and verify the integrity through device fingerprint (MAC+timestamp hash) after the connection is restored, and the omission rate is ≤0.1%.

[0055] Two-layer dynamic anomaly detection, forming an instant rule layer and an intelligent model layer, and updating the model according to changes in detection standards;

[0056] Two-layer dynamic anomaly detection includes instant rule layer, intelligent model layer, model update, model explainability and drift warning,

[0057] Instant rule layer: Based on statistical process SPC control, set limit-exceeding triggers (such as 9 consecutive points exceeding the mean line), and generate primary warnings within 1 second;

[0058] Intelligent model layer: Utilizes an LSTM-Attention model trained using federated learning, with a sliding time window T of 15 minutes. The threshold is dynamically adjusted based on experimental phase coefficients (e.g., logarithmic phase / plateau phase of cell culture). The threshold is calculated as: mean + 3σ × phase coefficient. This helps identify trend anomalies (e.g., progressive equipment aging).

[0059] Model update: When the test standard changes (such as CLSI EP17-A3), incremental training with a delta parameter of ≤0.1% and self-adaptation within 2 hours can reduce the false positive rate from 42% to 7%;

[0060] Model interpretability: The SHAP value evaluates the importance of each laboratory equipment feature by calculating its marginal contribution to the model prediction. The SHAP value is used to generate a heat map of abnormal influencing factors (e.g., "temperature deviation of the temperature control box contributes 68%"). This assists manual review and supports one-click export of PDF reports. The report generation time is ≤ 2 seconds.

[0061] Drift warning: Based on KL divergence monitoring, the predicted distribution is automatically triggered when the divergence is greater than 0.3. The PoC algorithm is used to select the laboratory samples with the top 20% contribution points, reducing the retraining time by 70%. The contribution of the kth abnormal single sample uploaded by the laboratory is calculated as:

[0062]

[0063] n i is the historical cumulative number of samples of the i-th type of anomaly (such as “pipette aspiration deviation”) in the global database; α = 100, β = 1 are scarcity smoothing parameters (to avoid high-frequency samples contributing to 0); k i,k is the consistency coefficient of the kth sample annotation (with the Kappa value of the industry expert annotation, range [-1, 1], ≥ 0.7 is considered valid); acc i,k is the sample annotation accuracy (the proportion of correct labels after secondary verification by experts, such as 0.9 means 90% correct); γ = 0.6 is the consistency weight (focusing on industry consensus); λ = 0.99 is the time decay factor (the newer the sample, the higher the contribution, half-life ≈ 69 days); t now -t k The time difference between sample uploads (unit: day)

[0064] Example: A laboratory uploaded a rare case of "mass spectrometer ion source contamination" (n i =5), labeling consistency k = 0.85, accuracy rate is 100%, and it was uploaded on the same day:

[0065] integral.

[0066] Knowledge graph attribution records, building a laboratory-related knowledge graph, which includes but is not limited to equipment ID, reagent batch number, and personnel qualifications. Calculate the probability weights of outliers and associated entities using a GNN graph neural network, with a weight control value ≥ 0.7. Generate an attribution path (e.g., "pipette batch number 202503 → aspiration deviation → data anomaly," with a confidence level of 0.89) and automatically generate structured records that comply with laboratory standards (ISO 15189). The generated structured records include but are not limited to the original data hash (SHA-256, accurate to milliseconds), the scope of impact (number of affected samples, experimental stage), and disposal recommendations. Record generation time is ≤ 10 seconds.

[0067] Knowledge graph attribution in knowledge graph attribution records includes entity relationship enhancement and intelligent reporting assistance. Entity relationship enhancement is to define different types of relationships such as "equipment-calibration record" and "reagent-batch expiration date", and train the relationship confidence through the TransE algorithm, with a calculated accuracy of ≥92%; intelligent reporting assistance is to use the mobile APP to scan the device NFC chip and automatically fill in the device model and current experimental stage. The field completion rate is increased from 65% to 98%.

[0068] Blockchain closed-loop verification: abnormal records (≥Level 2) are instantly uploaded to the blockchain. CAPA processes (such as equipment calibration and reagent recalls) are triggered through smart contracts. The treatment effect is verified based on subsequent quality control data of, for example, three batches (mean deviation ≤ 1.5%). The closed-loop status is automatically marked, and the entire process traceability time is ≤ 3 seconds. Cross-institutional privacy queries are also performed (verification time ≤ 200ms), meeting GDPR / CLIA compliance requirements.

[0069] In the blockchain closed-loop verification, the blockchain closed loop includes smart contract logic, privacy protection query and electronic signature solidification. When the smart contract logic shows that the abnormality level is ≥3, the disposal task is automatically pushed to the responsible person. The push tool is DingTalk / Enterprise WeChat. If the task response timeout (>30 minutes) triggers the superior warning, and the mean deviation is ≤1.5% through the comparison of subsequent batch quality control data, the smart contract automatically marks the CAPA as valid; privacy protection query is for the regulatory agency to verify the authenticity of a specific event through zero-knowledge proof, without obtaining the original data, and the verification time is ≤200ms; electronic signature solidification is that the disposal record must be signed by USB-Key to meet the security requirement standard, the signature timestamp is accurate to milliseconds, and the tamper-proof hash value is synchronized on the chain; the recording method takes ≤200ms from the occurrence of the abnormality to the generation of the record, while the traditional system takes 15 min, the abnormal data recording effect is greatly improved, and the dual reminders of vibration and voice on the mobile terminal for key abnormalities (such as a sudden temperature rise of the PCR instrument) can achieve a rapid response to laboratory abnormalities. Through laboratory clinical validation data set testing, the multi-dimensional attribution accuracy has been increased from 62% to 89%. The improved attribution accuracy reduces the limitations in recording and analyzing adverse events related to laboratory abnormal values, and meets the growing data management and analysis needs of modern laboratories; and blockchain evidence storage supports FDA21CFRPart11 audits, and the evidence chain retrieval time is shortened from 4 hours to 10 minutes, improving compliance. Compared with traditional LIMS systems, it integrates federated learning and knowledge graphs for the first time to solve the problems of cross-institutional data silos and attribution ambiguity, and realize the closed loop of "local detection → global optimization → accurate recording".

[0070] The present invention also provides a system for improving the recording of adverse events related to laboratory abnormal values, comprising:

[0071] Multi-modal edge collection module, deploy edge computing gateway supporting HL7v2.8, IEC62304 protocol, through digital twin technology, real-time mapping of centrifuge, temperature control box, different equipment running parameters (speed, temperature, vibration frequency) and environmental data (temperature and humidity, air pressure), and through differential privacy algorithm (epsilon <= 0.5) to aggregate cross-institutional data;

[0072] The edge computing gateway includes a protocol conversion submodule, a digital twin engine, and an offline cache module. The protocol conversion submodule supports real-time conversion from RS485 and ModbusTCP to MQTT, and the data collection frequency is greater than or equal to 100Hz. The digital twin engine is based on Unity3D to construct a three-dimensional model of the device, and synchronizes the device parameters in real time, with an error rate less than or equal to 0.5%. The offline cache module stores 72 hours of data locally when the network is disconnected, and checks the data integrity through device fingerprinting after the connection is restored.

[0073] Federal dynamic detection framework: adopt LSTM-Attention model with sliding time window (window size T = 15min), integrate SHAP explainability framework to generate abnormal attribution heat map, combine SPC statistical process control rule engine for double-layer detection, model parameters are updated across institutions through federated learning (FedAvg algorithm), support incremental training (delta threshold <= 0.1%) when detection standard changes.

[0074] Knowledge graph attribution engine: build a knowledge graph containing device ID, reagent batch number, operator qualification, and experiment stage, calculate the probability weight (weight threshold >= 0.7) of abnormal values and associated entities through GNN graph neural network, automatically generate (comply with ISO 15189) structured report (including original data hash value, hash algorithm SHA-256), fill in abnormal time (accurate to seconds), impact range (number of affected samples), and disposal suggestions based on template engine, report hash value is synchronized to the chain;

[0075] Blockchain closed-loop tracking system: based on Hyperledger Fabric's consortium chain storage, driven by smart contract CAPA process (trigger condition: abnormal level >= 3), abnormal level >= 3 or 2 consecutive 2 abnormal, automatically push the disposal task to the responsible person (Dingding / WeChat interface), integrated electronic signature (comply with Esign Act) to realize the non-tamperable traceability of disposal process, signature timestamp accurate to millisecond level, support penetration query of regulatory agencies (zero-knowledge proof verification time <= 200ms), that is, regulatory agencies verify the authenticity of a specific event through zero-knowledge proof, without obtaining the original data, verification time <= 200ms; Compared with traditional LIMS system, the system first integrates federated learning and knowledge graph to solve the problem of cross-institutional data island and attribution ambiguity related to laboratory abnormal values, realize the closed loop of "local detection -> global optimization -> accurate record"; Break through the limitation of static threshold, adapt to the iteration of detection methodology (such as the upgrade of new crown detection from fluorescent PCR to digital PCR), model update without manual intervention; Realize the penetration of supervision under the premise of protecting data privacy, solve the industry problem of "data available but invisible".

[0076] Finally, it should be noted that: the above only for the preferred embodiments of the present application, and not for limiting the present application, although the present application is described in detail with reference to the foregoing embodiments, for those skilled in the art, it still can modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features, any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application, should be included in the protection scope of the present application.

Claims

1. Improve the recording method of adverse events related to laboratory abnormal values, characterized by: The following steps are involved: Real-time multimodal data perception: using edge computing gateways to collect real-time laboratory equipment operating data, environmental parameters, and operation logs; building equipment digital twins, synchronously mapping the implicit parameters of laboratory equipment, and forming dynamic data portraits; Two-layer dynamic anomaly detection, forming an instant rule layer and an intelligent model layer, and updating the model according to changes in detection standards; Knowledge graph attribution records, building laboratory-related knowledge graphs, calculating the probability weights of outliers and associated entities, generating attribution paths, and automatically generating structured records that meet laboratory standards; Blockchain closed-loop verification, abnormal records are immediately uploaded to the chain, the CAPA process is triggered by smart contracts, the disposal effect is verified based on subsequent batch quality control data, the closed-loop status is automatically marked, and cross-institutional privacy queries are performed.

2. The method for improving the recording of adverse events related to laboratory abnormal values ​​according to claim 1, characterized in that: In the real-time perception of multimodal data, the edge computing gateway supports the HL7 / IEC62304 protocol. The equipment operation data includes but is not limited to the speed fluctuation of the experimental equipment centrifuge and the temperature curve of the temperature control box. The environmental parameters include but are not limited to temperature, humidity, and air pressure. The real-time sampling frequency is ≥100Hz.

3. The method for improving the recording of adverse events related to laboratory abnormal values ​​according to claim 1, characterized in that: The multimodal data collection in the multimodal data real-time perception includes: Digital twin: Build a 3D model of the equipment based on Unity3D, including real-time dimensional parameters such as synchronous motor current and bearing temperature. Differential privacy aggregation: ε-differential privacy is used for cross-institutional data aggregation during cross-institutional transmission, protecting sensitive information such as device IDs and operators, with a data distortion rate of ≤2%; Offline caching mechanism: Locally store 72 hours of data when the network is disconnected, and verify the integrity through device fingerprint after the connection is restored, with an omission rate of ≤0.1%.

4. The method for improving the recording of adverse events related to laboratory abnormal values ​​according to claim 1, characterized in that: The dual-layer dynamic anomaly detection includes: Instant rule layer: Based on statistical process SPC control, set limit-exceeding triggers and generate primary warnings within 1 second; Intelligent model layer: The LSTM-Attention model trained with federated learning is used, with a sliding time window T of 15 minutes. The threshold is dynamically adjusted based on the experimental phase coefficient (threshold = mean + 3σ × phase coefficient) to identify trend anomalies. Model update: When the detection standard changes, incremental training is performed with a delta parameter of ≤ 0.1%, and adaptation is performed within 2 hours.

5. The method for improving the recording of adverse events related to laboratory abnormal values ​​according to claim 1, characterized in that: The dual-layer dynamic anomaly detection further includes: Model interpretability: Generates a heat map of abnormal influencing factors through SHAP values ​​to assist manual review, supports one-click export of PDF reports, and generates explanation reports in ≤ 2 seconds; Drift warning: Based on KL divergence monitoring, the predicted distribution is automatically triggered when the divergence is greater than 0.

3. The PoC algorithm is used to select the laboratory samples with the top 20% contribution points. The contribution of the kth abnormal single sample uploaded by the laboratory is calculated as: n i is the historical cumulative sample number of the i-th type of anomaly in the global database; α=100, β=1 are the scarcity smoothing parameters; k i,k The consistency coefficient of the k-th sample annotation; acc i,k is the sample labeling accuracy; γ = 0.6 is the consistency weight; λ = 0.99 is the time decay factor; t now -t k The time difference between sample uploads.

6. The method for improving the recording of adverse events related to laboratory abnormal values ​​according to claim 1, characterized in that: In the knowledge graph attribution record, the laboratory-related knowledge graph includes but is not limited to equipment ID, reagent batch number and personnel qualifications. The probability weights of outliers and associated entities are calculated through the GNN graph neural network, and the weight control value is ≥0.

7. The generated structured record includes but is not limited to the original data hash, impact scope and disposal suggestions, and the record generation time is ≤10 seconds.

7. The method for improving the recording of adverse events related to laboratory abnormal values ​​according to claim 1, characterized in that: The knowledge graph attribution in the knowledge graph attribution record includes: Entity relationship enhancement: define different types of relationships and train relationship confidence using the TransE algorithm; Smart reporting assistance: Use the mobile app to scan the device's NFC chip and automatically fill in the device model and current experimental stage.

8. The method for improving the recording of adverse events related to laboratory abnormal values ​​according to claim 1, characterized in that: The blockchain closed loop in the blockchain closed loop verification includes: Smart contract logic: When the abnormality level is ≥3, the handling task is automatically pushed to the responsible person through DingTalk / WeChat Enterprise. If the task response timeout occurs, a warning from the superior is triggered. If the mean deviation is ≤1.5% through comparison of subsequent batch quality control data, the smart contract automatically marks the CAPA as valid. Privacy-preserving query: Regulators can verify the authenticity of specific events through zero-knowledge proofs without obtaining original data, with verification time ≤ 200ms; Electronic signature solidification: Disposal records must be signed via USB-Key in accordance with security requirements standards, the signature timestamp is accurate to milliseconds, and the tamper-proof hash value is synchronized on the chain.

9. The system for improving records of adverse events related to laboratory abnormal values ​​according to any one of claims 1 to 8, characterized in that: include: The multimodal edge acquisition module deploys an edge computing gateway that supports HL7v2.8 and IEC62304 protocols. It uses digital twin technology to map the operating parameters and environmental data of different equipment such as centrifuges and temperature control boxes in real time, and aggregates cross-institutional data using differential privacy algorithms. Federated dynamic detection framework: This framework uses a sliding time window LSTM-Attention model, integrates the SHAP interpretability framework to generate anomaly attribution heatmaps, and combines it with the SPC statistical process control rule engine for two-layer detection. Model parameters are updated across institutions through federated learning, supporting incremental training when detection standards change. Knowledge graph attribution engine: Builds a knowledge graph containing equipment ID, reagent batch number, operator qualifications, and experimental stage, calculates the probability weights of outliers and related entities through the GNN graph neural network, and automatically generates structured reports; Blockchain closed-loop tracking system: Based on Hyperledger Fabric's consortium chain evidence storage, it drives the CAPA process through smart contracts, integrates electronic signatures for tamper-proof traceability of the disposal process, and supports penetrating queries by regulators.

10. The system for improving the recording of adverse events related to abnormal laboratory values ​​according to claim 9, characterized in that: The edge computing gateway in the multimodal edge acquisition module includes: Protocol conversion submodule, which supports real-time conversion from RS485 and ModbusTCP to MQTT, with a data acquisition frequency of ≥100Hz; A digital twin engine, which builds a 3D model of the device based on Unity3D and synchronizes device parameters in real time with an error rate of ≤0.5%; The offline cache module stores data locally for 72 hours when the network is disconnected, and verifies the data integrity through device fingerprint after the connection is restored.

Images