Abnormal sample intelligent identification and reinspection system of platelet function analyzer
By combining a multi-physics field collaborative sensing module and an intelligent recognition center, and utilizing a neural symbolic hybrid reasoning algorithm and reinforcement learning decision-making, the platelet function analyzer achieves intelligent identification and automated re-examination of abnormal samples, solving the problem of insufficient identification capability in existing technologies and improving the accuracy and efficiency of test results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ANHUI JUNHE BIOTECHNOLOGY CO LTD
- Filing Date
- 2026-03-11
- Publication Date
- 2026-05-12
AI Technical Summary
Existing platelet function analyzers have limited ability to identify abnormal samples, making it difficult to distinguish complex abnormal reaction patterns caused by improper sample collection, pathological interferences, and rare lesions. Furthermore, they lack intelligent retesting decision-making and execution processes, resulting in insufficient accuracy and timeliness of results.
By employing a multi-physics field collaborative sensing module, an intelligent identification center, and an automated re-examination execution module, combined with a neural symbolic hybrid reasoning algorithm and reinforcement learning decision-making, the system achieves multi-dimensional heterogeneous response signal acquisition, anomaly identification, and automated re-examination of blood samples.
It can identify abnormal samples and infer the root cause, realize unmanned and standardized re-examination operations, improve the accuracy and efficiency of test results, and has the ability to continuously learn to adapt to changes in the disease spectrum.
Smart Images
Figure CN122017216A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of platelet function analysis technology, specifically to an intelligent identification and re-examination system for abnormal samples in a platelet function analyzer. Background Technology
[0002] Platelet function testing is crucial for assessing bleeding risk, monitoring the efficacy of antiplatelet drugs, and diagnosing related diseases. Currently, commonly used platelet function analysis techniques in clinical and laboratory settings include optical turbidimetry, impedance spectroscopy, rapid platelet function analysis, and flow cytometry. Optical turbidimetry, as the traditional gold standard, is significantly affected by plasma turbidity and sample processing, and is cumbersome and difficult to standardize. While impedance spectroscopy can perform whole blood testing, more closely mimicking the physiological environment, its signal is easily interfered with by factors such as hematocrit, platelet count, and non-platelet particles (e.g., pyrocytes, macroplatelets), leading to complex result interpretation. Rapid analysis methods improve the convenience of point-of-care testing, but provide limited information dimensions, making in-depth differential diagnosis difficult. Flow cytometry, while enabling multi-parameter, high-sensitivity analysis, requires highly skilled operators, is expensive, and is difficult to automate rapidly. During testing, various analyzers typically focus on analyzing "normal" aggregation curves and parameters. For atypical and complex abnormal reaction patterns caused by improper sample collection (e.g., partial activation), inappropriate storage conditions, the presence of pathological interfering substances (e.g., cryoglobulins, lipemia, hemolysis), or rare platelet abnormalities, their built-in algorithms have limited recognition capabilities, often categorizing them as "invalid results" or "errors," only suggesting repeat testing. This not only wastes samples and time but may also delay clinical diagnosis and treatment. Furthermore, current technologies underutilize the large amount of high-resolution raw data generated during testing, often limiting themselves to extracting a few peak parameters and failing to deeply mine the rich information contained in the complete kinetic curve. The identification and subsequent processing of abnormal samples heavily rely on operator experience, lacking standardized and intelligent retesting decision-making and execution processes, making it difficult to meet the growing clinical demands for accuracy, timeliness, and interpretability of test results.
[0003] Therefore, we propose an intelligent identification and re-examination system for abnormal samples in platelet function analyzers. Summary of the Invention
[0004] The purpose of this invention is to provide an intelligent identification and re-examination system for abnormal samples in a platelet function analyzer, so as to solve the problems mentioned in the background art.
[0005] To achieve the above objectives, the present invention provides the following technical solution: an intelligent identification and re-examination system for abnormal samples in a platelet function analyzer, comprising: A multi-physics collaborative sensing module is used to apply multiple physical field stimuli to a blood sample and simultaneously acquire the resulting heterogeneous response signals. The intelligent identification center receives heterogeneous response signals collected by the multi-physics field collaborative sensing module through a communication link, and executes a neural symbol hybrid reasoning algorithm to identify abnormal sample states and infer their root causes. The automated re-inspection execution module receives the re-inspection command issued by the intelligent recognition center and uses an adaptive decision-making algorithm based on reinforcement learning to control the integrated microfluidic chip laboratory to automatically reprocess and re-inspect the original sample. The system's self-evolutionary engine continuously collects the reasoning results of the intelligent recognition center, the verification results of the automated re-examination execution module, and external clinical feedback information. Through a federated learning framework and an abnormal pattern library dynamic update algorithm, it iteratively optimizes the model parameters and knowledge base in the intelligent recognition center. The prospective sample quality screening module uses a fast scanning sub-mode of the multi-physics field collaborative sensing module to acquire the preliminary terahertz absorption spectrum, low-frequency impedance baseline, and microscopic image texture of the sample within 30 seconds before the sample enters the main detection process.
[0006] Preferably, the multiphysics collaborative sensing module includes: A quantum dot-encoded magnetoelectric coupling sensor array consists of multiple magnetoelectric nanoparticles with different biomolecular probes on their surfaces. Each particle is encoded by a unique quantum dot fluorescence spectrum and is used to simultaneously and in situ detect specific biomolecular binding events and changes in electromagnetic properties at the nanoscale during platelet aggregation. The terahertz time-domain spectroscopy unit includes a terahertz wave generating device and a detection device, which are used to transmit electromagnetic pulses in the radio frequency band of 0.1 to 10 terahertz to the sample and receive transmitted or reflected signals to analyze the characteristics of dielectric relaxation, water molecule dynamics and collective vibrational modes of biomacromolecules in the sample. The microfluidic acoustic tweezers unit, integrated in a microfluidic channel, generates surface acoustic waves through interdigital transducers to form a stable acoustic potential well within the channel. This is used to capture individual platelets and measure their displacement response under acoustic radiation force, thereby deduce their mechanical properties. The multi-frequency impedance spectroscopy unit measures the complex impedance of the sample at multiple discrete frequency points from kilohertz to megahertz, and obtains dispersion curves related to the conductivity of intracellular and extracellular media and the cell membrane capacitance.
[0007] Preferably, the heterogeneous response signals collected by the multi-physics collaborative sensing module are modeled as a high-order tensor in the intelligent recognition center, wherein different dimensions of the tensor correspond to time series, physical field stimulus type, spatial sensing point and signal feature mode, respectively; the intelligent recognition center adopts a fusion algorithm based on tensor decomposition and graph neural network to reduce the dimensionality of the high-order tensor and mine the correlation relationship.
[0008] Preferably, the neural-symbolic hybrid reasoning algorithm includes a neural network subsystem and a symbolic reasoning subsystem; The neural network subsystem is implemented by a hierarchical Transformer architecture, which includes a temporal Transformer encoder for processing time-series signals, a spectral Transformer encoder for processing spectral signals, and a cross-modal attention fusion layer responsible for extracting high-level features from the heterogeneous response signals and generating initial symbolic propositions. The symbolic reasoning subsystem includes an extensible domain knowledge base, which uses a probabilistic logic programming language to represent knowledge of platelet physiology, pathology, and interfering factors as weighted logical rules. The subsystem executes a reasoning algorithm that combines a probabilistic graphical model with a Markov logic network. It takes the symbolic propositions generated by the neural network subsystem as evidence input, performs uncertain reasoning under the rule constraints of the knowledge base, calculates the posterior probability distribution of the anomaly type and its root cause hypothesis, and generates an interpretable reasoning path graph. The key fusion step of the neural symbolic hybrid inference algorithm is characterized by the following formula: for the final decision variable Its probability is determined by the prior probability perceived by the neural network. Posterior probability of symbolic logic reasoning Weighted integration using Dirichlet distribution, i.e. in For the original multiphysics data, For a knowledge base, As a proposition of evidence, This is a hyperparameter that reflects the confidence level of the two subsystems.
[0009] Preferably, the reinforcement learning-based adaptive decision-making algorithm models the re-examination process as a partially observable Markov decision process; its state space is jointly defined by the anomaly hypothesis confidence vector output by the intelligent recognition center, the remaining sample quantity, the set of available re-examination methods, and historical operation records; the action space consists of all executable atomic re-examination operations and their parameter combinations; the reward function is defined as follows: ,in It is to perform an action The resulting information gain is calculated using the expected KL divergence. It is the cost of reagents consumed in the action. It is the time spent on the action. , , These are the weighting coefficients.
[0010] Preferably, the integrated microfluidic chip laboratory in the automated re-inspection execution module uses digital microfluidic technology as the core control platform.
[0011] Preferably, the federated learning framework in the system's self-evolutionary engine employs an asynchronous federated averaging algorithm based on contribution evaluation.
[0012] Preferably, the dynamic update algorithm for the abnormal pattern library is as follows: Feature extraction and representation: For each new input abnormal sample data, the feature extractor trained in the intelligent recognition center is used to map it into a low-dimensional semantic embedding space to obtain its feature vector. ; Similarity Calculation and Retrieval: Calculation Compared with the prototype vectors of all known anomaly patterns in the anomaly pattern library The cosine similarity, where the prototype vectors It is calculated from the mean of the feature vectors of all samples belonging to this category, i.e. , For feature extraction function, For category Support set; Novelty determination: If the maximum similarity Below the preset threshold If the sample is found to represent a new abnormal pattern, it will be temporarily stored in the "Pool of Emerging Patterns to be Confirmed". Expert review and confirmation: Cases in the "Pool of Emerging Patterns to be Confirmed" are regularly submitted to human experts for review and confirmation, along with their original multi-physics signal characteristics and preliminary analysis of the intelligent recognition center. Knowledge base expansion: For new patterns confirmed by experts, their feature vectors are added to the abnormal pattern library as new prototypes, and the knowledge base expansion process of the symbolic reasoning subsystem in the neural symbolic hybrid reasoning algorithm is triggered, allowing experts or natural language processing interfaces to add logical rules describing the new patterns, thus completing the forward expansion of system knowledge.
[0013] Preferred options also include: A multi-source confidence fusion and conflict resolution module receives multiple competing hypotheses and their confidence levels from the intelligent identification center for the same anomaly root cause, and also receives independent or partially related verification evidence that may be generated by different re-examination methods in the automated re-examination execution module. This module uses the Dempster-Shafer evidence theory algorithm for information fusion: each information source is regarded as an evidence body, and a basic probability is assigned to the set of hypotheses it supports; when there is a conflict between different evidence bodies, an improved combination rule that considers evidence distance and reliability is used for synthesis, and finally a comprehensive confidence function distribution is output to guide whether to terminate the re-examination or initiate a higher-order expert arbitration process. The detection process is monitored in real time and self-calibrated. During each main detection and re-detection process, this subsystem runs a dynamic state-space model based on a recurrent neural network in parallel. Taking the sensor readings and control commands from the previous moment as input, it predicts the sensor readings at the current moment. By comparing the residual sequence of the predicted values and the actual measured values and performing chi-square tests and sequential probability ratio tests, it detects process anomalies such as sensor drift, reagent failure, microchannel blockage, or bubble interference in real time. Once a significant deviation is detected, it immediately triggers a targeted self-calibration routine or sends an alarm to the system to ensure the reliability of data acquisition.
[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. By integrating quantum dot-coded magnetoelectric sensors, terahertz time-domain spectroscopy, microfluidic acoustic tweezers, and multi-frequency impedance spectroscopy, the system synchronously acquires multi-dimensional heterogeneous response signals of samples in physical, chemical, and biomechanical dimensions. This overcomes the limitations of single detection methods (such as optical methods which are susceptible to turbidity interference and impedance methods which are susceptible to hematocrit), providing an unprecedentedly rich information foundation for anomaly identification.
[0015] 2. The neural symbolic hybrid reasoning algorithm adopted by the intelligent recognition center not only uses deep Transformer networks to automatically mine complex patterns from high-dimensional data, but also performs logical verification and causal inference through symbolic reasoning that integrates domain knowledge. This enables the system to not only identify abnormal phenomena, but also to infer their root causes (such as distinguishing drug effects, sample collection artifacts or rare pathologies) and generate interpretable reasoning paths, solving the problem of existing technologies relying on experience and black-box decision-making for interpreting results.
[0016] 3. The adaptive decision-making algorithm based on reinforcement learning can dynamically plan the optimal re-inspection strategy according to real-time confidence and cost constraints. Combined with the fully automated digital microfluidic chip laboratory execution, it realizes unmanned and standardized operation from decision-making to execution, which completely changes the traditional re-inspection mode that relies on human experience, has cumbersome steps and low efficiency, and significantly improves laboratory throughput and result consistency.
[0017] 4. It endows the system with the ability to continuously learn and evolve: Through the federated learning framework and dynamic anomaly pattern library update algorithm, the system can continuously optimize the core model using multi-center data without sharing the original privacy data, and can automatically discover, verify and integrate new anomaly patterns; this enables the system to keep up with the changes in disease spectrum and diagnosis and treatment technology, overcoming the shortcomings of traditional equipment algorithms being rigid and unable to keep up with the times.
[0018] 5. The prospective sample quality screening module can quickly identify unqualified samples such as hemolysis, lipemia, and partial coagulation before the main test, and provide pretreatment suggestions or early warnings for resampling. This avoids the waste of samples and time caused by invalid tests from the source, and improves the efficiency and cost-effectiveness of the overall testing process. Attached Figure Description
[0019] Figure 1 This is a block diagram of an intelligent identification and re-examination system for abnormal samples in a platelet function analyzer. Detailed Implementation
[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0021] This invention provides a technical solution: an intelligent identification and re-examination system for abnormal samples in a platelet function analyzer, comprising: A multi-physics collaborative sensing module is used to apply multiple physical field stimuli to a blood sample and simultaneously acquire the resulting heterogeneous response signals. The intelligent identification center receives heterogeneous response signals collected by the multi-physics field collaborative sensing module through a communication link, and executes a neural symbol hybrid reasoning algorithm to identify abnormal sample states and infer their root causes. The automated re-inspection execution module receives re-inspection instructions from the intelligent recognition center and uses an adaptive decision-making algorithm based on reinforcement learning to control the integrated microfluidic chip laboratory to automatically reprocess and re-inspect the original sample. The system's self-evolutionary engine continuously collects the reasoning results of the intelligent recognition center, the verification results of the automated re-examination execution module, and external clinical feedback information. Through a federated learning framework and an abnormal pattern library dynamic update algorithm, it iteratively optimizes the model parameters and knowledge base in the intelligent recognition center. The prospective sample quality screening module, before the sample enters the main detection process, utilizes a fast scanning sub-mode of the multi-physics collaborative sensing module to acquire the preliminary terahertz absorption spectrum, low-frequency impedance baseline, and microscopic image texture of the sample within 30 seconds. A lightweight causal discovery algorithm analyzes the conditional independence relationships between these preliminary signals, constructing a Bayesian network regarding potential contaminating factors (such as hemolysis, lipemia, partial coagulation, and giant platelets). This network is used to calculate the prior probability that the sample is unsuitable for standard platelet function testing. If this probability exceeds a threshold, the system will directly suggest sample preprocessing or re-collection, thereby avoiding unnecessary main detection overhead. Among them, the multi-physics field collaborative sensing module, intelligent identification center, automated re-inspection execution module and system self-evolution engine are connected in sequence to form a complete closed loop from data perception, anomaly identification, decision execution to knowledge update.
[0022] Furthermore, the multiphysics collaborative sensing module includes: A quantum dot-encoded magnetoelectric coupling sensor array consists of multiple magnetoelectric nanoparticles with different biomolecular probes on their surfaces. Each particle is encoded by a unique quantum dot fluorescence spectrum and is used to simultaneously and in situ detect specific biomolecular binding events and changes in electromagnetic properties at the nanoscale during platelet aggregation. The terahertz time-domain spectroscopy unit includes a terahertz wave generating device and a detection device, which are used to transmit electromagnetic pulses in the radio frequency band of 0.1 to 10 terahertz to the sample and receive transmitted or reflected signals to analyze the characteristics of dielectric relaxation, water molecule dynamics and collective vibrational modes of biomacromolecules in the sample. The microfluidic acoustic tweezers unit, integrated in a microfluidic channel, generates surface acoustic waves through interdigital transducers to form a stable acoustic potential well within the channel. This is used to capture individual platelets and measure their displacement response under acoustic radiation force, thereby deduce their mechanical properties. The multi-frequency impedance spectroscopy unit measures the complex impedance of the sample at multiple discrete frequency points from kilohertz to megahertz, and obtains dispersion curves related to the conductivity of intracellular and extracellular media and cell membrane capacitance. The outputs of the quantum dot-encoded magnetoelectric coupling sensor array, the terahertz time-domain spectroscopy unit, the microfluidic acoustic tweezers unit, and the multi-frequency impedance spectroscopy scanning unit together constitute a heterogeneous response signal.
[0023] Furthermore, the heterogeneous response signals collected by the multi-physics collaborative sensing module are modeled as a high-order tensor in the intelligent recognition center. The different dimensions of the tensor correspond to the time series, physical field stimulus type, spatial sensing point, and signal feature mode, respectively. The intelligent recognition center adopts a fusion algorithm based on tensor decomposition and graph neural network to reduce the dimensionality and mine the correlation of this high-order tensor. Its core steps include decomposing the original tensor into the product of the core tensor and multiple factor matrices through the Tucker decomposition algorithm to extract the latent variable features across physical fields. Then, these latent variable features are used as node attributes. Combined with time dependence and physical field correlation, a spatiotemporal heterogeneous graph is constructed. Finally, a graph attention network is used to aggregate and update the node information to generate the global state embedding vector of the sample. In the fusion algorithm based on tensor decomposition and graph neural networks, the construction rules for the spatiotemporal heterogeneous graph are as follows: each sensing unit under each physical field stimulus at each time point is taken as a node; the node attributes are composed of the energy entropy extracted by wavelet transform of the original signal of the unit at that time and the principal components obtained by singular spectrum analysis; it includes two types of edges: one is the temporal edge, which connects nodes of the same sensing unit at adjacent time points, and the edge weight is determined by the autocorrelation coefficient of the signal; the other is the field-related edge, which connects nodes of different sensing units at the same time point whose physical field theoretical coupling coefficient exceeds the threshold, and the edge weight is determined by the normalized product of the theoretical coupling coefficient and the mutual information value of the measured signal.
[0024] Furthermore, the neural-symbolic hybrid reasoning algorithm includes a neural network subsystem and a symbolic reasoning subsystem; The neural network subsystem is implemented by a hierarchical Transformer architecture, which includes a temporal Transformer encoder for processing time-series signals, a spectral Transformer encoder for processing spectral signals, and a cross-modal attention fusion layer responsible for extracting high-level features from heterogeneous response signals and generating initial symbolic propositions. The symbolic reasoning subsystem includes an extensible domain knowledge base that uses a probabilistic logic programming language to represent knowledge of platelet physiology, pathology, and interfering factors as weighted logical rules. The subsystem executes a reasoning algorithm that combines a probabilistic graphical model with a Markov logic network. It takes symbolic propositions generated by the neural network subsystem as evidence input, performs uncertain reasoning under the rule constraints of the knowledge base, calculates the posterior probability distribution of the anomaly type and its root cause hypothesis, and generates an interpretable reasoning path graph. The key fusion step of the neural symbolic hybrid reasoning algorithm is characterized by the following formula: for the final decision variable Its probability is determined by the prior probability perceived by the neural network. Posterior probability of symbolic logic reasoning Weighted integration using Dirichlet distribution, i.e. in For the original multiphysics data, For a knowledge base, As a proposition of evidence, The hyperparameters that reflect the confidence level of the two subsystems; The probabilistic logic programming implementation of the symbolic reasoning subsystem in the neural symbolic hybrid reasoning algorithm is as follows: a distributed semantic logic program with continuous random variables is used; knowledge in the platelet dysfunction domain is encoded into a set of weighted Horn clauses, where the weights are learned through maximum likelihood estimation on historical data; during reasoning, the propositions output by the neural network subsystem are used as evidence atoms, and the total weight of the query atom being true in all possible worlds is calculated through a weighted model counting algorithm, thereby obtaining its posterior probability; the interpretable reasoning path graph is automatically generated by the rule chains activated during the backtracking satisfaction solution process.
[0025] Furthermore, the reinforcement learning-based adaptive decision-making algorithm models the re-examination process as a partially observable Markov decision process; its state space is jointly defined by the anomaly hypothesis confidence vector output by the intelligent recognition center, the remaining sample quantity, the set of available re-examination methods, and historical operation records; the action space consists of all executable atomic re-examination operations and their parameter combinations; the reward function is defined as... ,in It is to perform an action The resulting information gain is calculated using the expected KL divergence. It is the cost of reagents consumed in the action. It is the time spent on the action. , , The weight coefficients are used. The algorithm uses a proximal policy optimization algorithm to train a policy network. The network outputs a probability distribution in the action space based on the current state to seek the optimal balance between information acquisition, cost and time, thereby dynamically generating a personalized re-inspection workflow. In reinforcement learning-based adaptive decision-making algorithms, information gain The specific calculation method is as follows: Let the current belief state regarding the root cause of the anomaly be a probability distribution. Execute actions After selecting a specific re-examination item, the observation results will be obtained. This observation result is derived from the likelihood function. Characterization; then the expected information gain is defined as the information gain about the hypothesis before and after the action is performed. The expected value of the KL divergence of the belief distribution: The calculation is solved online using Monte Carlo sampling and particle filtering algorithms.
[0026] Furthermore, the integrated microfluidic chip laboratory in the automated re-inspection execution module adopts digital microfluidics technology as the core control platform. The chip includes a substrate composed of a two-dimensional array of electrodes, on which sample and reagent droplets are independently manipulated through dielectric wetting effect. The chip also integrates a subset of quantum dot-encoded magnetoelectric coupling sensor arrays and a miniature optical detection unit. The atomic re-inspection operation output by the reinforcement learning-based adaptive decision algorithm is translated into the activation sequence, voltage amplitude, and duration of specific electrodes to drive nanoscale droplets to complete a series of operations such as transport, splitting, merging, mixing, and dispensing to the sensing area, realizing the on-chip fully automated execution of the multi-step re-inspection process. The construction and signal analysis method of quantum dot-encoded magnetoelectric coupling sensor array includes: selecting CdSe / ZnS core-shell quantum dots of different sizes, and covalently bonding them to Fe3O4@Au core-shell magnetoelectric nanoparticles with specific biomolecular probes on their surface through ligand exchange reaction to form spectrally encoded sensing units; during detection, frequency and intensity modulation of the applied magnetic field are performed, and the time-varying signal of fluorescence intensity of each quantum dot and AC magnetic susceptibility of the nanoparticles are measured simultaneously; a decoupling algorithm is used to separate the specific bio-binding signal in fluorescence changes from the non-specific background signal caused by the rotation and aggregation of nanoparticles. The decoupling algorithm is based on the differential response of each sensing unit to the magnetic field modulation frequency to establish a system of linear equations for solution.
[0027] Furthermore, the federated learning framework in the system's self-evolutionary engine employs an asynchronous federated averaging algorithm based on contribution evaluation. This framework uses local models deployed in intelligent recognition centers across different medical institutions as clients, with each client calculating model parameter updates on its local private data. Before aggregation, the central server evaluates the contribution of each client's update, using criteria including the direct impact of the update on the global model's performance on the common validation set and the similarity between the client's data distribution and the global distribution. During aggregation, the global model's update formula is... ,in It is the first Data volume per client It is the contribution weight after evaluation, which is used to suppress the impact of low-quality or malicious updates and achieve safe and efficient distributed model evolution; Client-side contribution weight updates in the federated learning framework The calculation algorithm is as follows: a central server maintains a dynamic baseline dataset, which consists of a small amount of non-sensitive feature data contributed by each client under privacy protection conditions (such as differential privacy); in the... During round aggregation, the server first evaluates the global model using a benchmark dataset. performance Then evaluate the client Update Temporarily updated model Performance on benchmark datasets Contribution weight Defined as the normalized value of performance gain ,in This is the smoothing constant.
[0028] Furthermore, the specific algorithm for dynamically updating the exception pattern library is as follows: Feature extraction and representation: For each new input anomalous sample data, the feature extractor trained in the intelligent recognition center is used to map it into a low-dimensional semantic embedding space to obtain its feature vector. ; Similarity Calculation and Retrieval: Calculation Compared with the prototype vectors of all known anomaly patterns in the anomaly pattern library The cosine similarity, where the prototype vectors It is calculated from the mean of the feature vectors of all samples belonging to this category, i.e. , For feature extraction function, For category Support set; Novelty determination: If the maximum similarity Below the preset threshold If the sample is found to represent a new abnormal pattern, it will be temporarily stored in the "Pool of Emerging Patterns to be Confirmed". Expert review and confirmation: Cases in the "Pool of Emerging Patterns to be Confirmed" are regularly submitted to human experts for review and confirmation, along with their original multi-physics signal characteristics and preliminary analysis of the intelligent recognition center. Knowledge base expansion: For new patterns confirmed by experts, their feature vectors are added to the abnormal pattern library as new prototypes, and the knowledge base expansion process of the symbolic reasoning subsystem in the neural symbolic hybrid reasoning algorithm is triggered, allowing experts or natural language processing interfaces to add logical rules describing the new patterns, thus completing the forward expansion of system knowledge. In the dynamic update algorithm of the abnormal pattern library, the novelty determination threshold It is not a fixed value, but is dynamically adjusted through an adaptive algorithm: this algorithm monitors the intra-class average distance of feature vectors of samples from each category in the abnormal pattern library. Average inter-class distance between each class prototype Threshold According to the formula Periodic updates are performed, among which and A coefficient is set to balance the desired recall and precision, thereby ensuring that the expansion of the pattern library has consistent sensitivity across different data density regions.
[0029] Furthermore, it also includes: The multi-source confidence fusion and conflict resolution module receives multiple competing hypotheses and their confidence levels from the intelligent identification center for the same anomaly root cause, and also receives independent or partially related verification evidence that may be generated by different review methods in the automated review execution module. This module uses the Dempster-Shafer evidence theory algorithm for information fusion: each information source is treated as an evidence body, and a basic probability is assigned to the set of hypotheses it supports. When there is a conflict between different evidence bodies, an improved combination rule that considers evidence distance and reliability is used for synthesis, and finally a comprehensive confidence function distribution is output to guide whether to terminate the review or initiate a higher-order expert arbitration process. The detection process is monitored in real time and self-calibrated. During each main detection and re-detection process, this subsystem runs a dynamic state-space model based on a recurrent neural network in parallel. Taking the sensor readings and control commands from the previous moment as input, it predicts the sensor readings at the current moment. By comparing the residual sequence of the predicted values and the actual measured values and performing chi-square tests and sequential probability ratio tests, it detects process anomalies such as sensor drift, reagent failure, microchannel blockage, or bubble interference in real time. Once a significant deviation is detected, it immediately triggers a targeted self-calibration routine (such as performing standard product testing or rinsing the flow channel) or sends an alarm to the system to ensure the reliability of data acquisition. The specific process is as follows: Step S1: Multi-physics field data synchronous acquisition: Load the blood sample to be tested into the multi-physics field collaborative sensing module. While applying the standard platelet agonist, simultaneously start the quantum dot encoded magnetoelectric coupling sensor array, terahertz time-domain spectroscopy unit, microfluidic acoustic tweezers unit and multi-frequency electrical impedance spectroscopy scanning unit to acquire the multimodal heterogeneous response signal of the entire aggregation process with high time resolution. Step S2: Heterogeneous Data Fusion and Anomaly Identification: The raw signals collected in Step S1 are transmitted to the intelligent identification center; first, a fusion algorithm based on tensor decomposition and graph neural network is executed to extract a unified feature representation across physical fields; then, this feature representation is input into the neural symbolic hybrid reasoning algorithm; the neural network subsystem generates preliminary anomaly propositions, and the symbolic reasoning subsystem performs probabilistic logical reasoning under the constraints of the domain knowledge base, outputting a comprehensive analysis report including anomaly type classification, root cause hypothesis, confidence score, and interpretable reasoning path; Step S3: Adaptive Re-inspection Decision and Execution: The intelligent identification center determines whether a re-inspection needs to be initiated based on the confidence score and anomaly type output in Step S2. If so, it calls the reinforcement learning-based adaptive decision algorithm to calculate the optimal re-inspection strategy sequence based on the current system state (sample balance, time, and cost constraints). The automated re-inspection execution module receives this strategy sequence and compiles it into precise control instructions for the digital microfluidic chip laboratory, driving the sample droplets to complete automated reagent addition, mixing, incubation, and re-inspection operations. The digital microfluidic chip laboratory contains an embedded, low-power field-programmable gate array control unit. This unit pre-stores standardized voltage pulse sequence waveforms for all atomic droplet operations (transportation, splitting, merging, etc.). When it receives the re-inspection workflow instructions composed of atomic operations from the automated re-inspection execution module, the FPGA control unit dynamically schedules and triggers the corresponding electrode sequence through real-time table lookup and state machine logic, and synchronously reads the feedback signals from the integrated sensor to achieve sub-millisecond closed-loop droplet control, ensuring the precise timing and reliability of the re-inspection steps. Between steps S2 and S3, there is also a confidence threshold judgment and process branching step: setting a high confidence threshold. With a low confidence threshold If the confidence score of the main anomalous hypothesis output by the intelligent recognition center is higher than 100%, then the confidence score of the main anomalous hypothesis is higher than 100%. If the confidence score is lower than 1, proceed directly to step S4 to generate the report; If the confidence score is [not specified], it will be judged as "uncertain," and the report will clearly indicate that a comprehensive judgment needs to be made in conjunction with other clinical examinations, and record the possible sources of uncertainty; only when the confidence score is [not specified] will it be considered as "uncertain." , When the test is within the specified range, step S3 is initiated to start the adaptive re-inspection process, thereby optimizing the overall detection efficiency and resource consumption. Step S4: Result Integration and Report Generation: Integrate the main detection analysis results of Step S2 with the re-examination verification results of Step S3, and make the final decision by the intelligent recognition center to generate an enhanced test report. This report not only includes the final platelet function parameters, but also includes sample quality assessment, abnormal identification basis, re-examination process performed and its results, and other diagnostic trajectory information. Step S5: System Closed-Loop Learning and Update: The complete data stream of this detection (from raw signal, intermediate inference to final report) and the subsequent potential clinical diagnostic gold standard are submitted to the system's self-evolution engine. The engine optimizes the global model through a federated learning framework and evaluates whether new abnormal patterns have been discovered through an algorithm that dynamically updates the abnormal pattern library, thereby completing a closed-loop iteration from data to knowledge and then to system performance improvement.
[0030] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0031] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. Intelligent identification and retesting system for abnormal samples in platelet function analyzers. Its characteristics include: A multi-physics collaborative sensing module is used to apply multiple physical field stimuli to a blood sample and simultaneously acquire the resulting heterogeneous response signals. The intelligent identification center receives heterogeneous response signals collected by the multi-physics field collaborative sensing module through a communication link, and executes a neural symbol hybrid reasoning algorithm to identify abnormal sample states and infer their root causes. The automated re-inspection execution module receives the re-inspection command issued by the intelligent recognition center and uses an adaptive decision-making algorithm based on reinforcement learning to control the integrated microfluidic chip laboratory to automatically reprocess and re-inspect the original sample. The system's self-evolutionary engine continuously collects the reasoning results of the intelligent recognition center, the verification results of the automated re-examination execution module, and external clinical feedback information. Through a federated learning framework and an abnormal pattern library dynamic update algorithm, it iteratively optimizes the model parameters and knowledge base in the intelligent recognition center. The prospective sample quality screening module uses a fast scanning sub-mode of the multi-physics field collaborative sensing module to acquire the preliminary terahertz absorption spectrum, low-frequency impedance baseline, and microscopic image texture of the sample within 30 seconds before the sample enters the main detection process.
2. The intelligent identification and re-examination system for abnormal samples in a platelet function analyzer according to claim 1, characterized in that: The multiphysics collaborative sensing module includes: A quantum dot-encoded magnetoelectric coupling sensor array consists of multiple magnetoelectric nanoparticles with different biomolecular probes on their surfaces. Each particle is encoded by a unique quantum dot fluorescence spectrum and is used to simultaneously and in situ detect specific biomolecular binding events and changes in electromagnetic properties at the nanoscale during platelet aggregation. The terahertz time-domain spectroscopy unit includes a terahertz wave generating device and a detection device, which are used to transmit electromagnetic pulses in the radio frequency band of 0.1 to 10 terahertz to the sample and receive transmitted or reflected signals to analyze the characteristics of dielectric relaxation, water molecule dynamics and collective vibrational modes of biomacromolecules in the sample. The microfluidic acoustic tweezers unit, integrated in a microfluidic channel, generates surface acoustic waves through interdigital transducers to form a stable acoustic potential well within the channel. This is used to capture individual platelets and measure their displacement response under acoustic radiation force, thereby deduce their mechanical properties. The multi-frequency impedance spectroscopy unit measures the complex impedance of the sample at multiple discrete frequency points from kilohertz to megahertz, and obtains dispersion curves related to the conductivity of intracellular and extracellular media and the cell membrane capacitance.
3. The intelligent identification and re-examination system for abnormal samples in a platelet function analyzer according to claim 2, characterized in that: The heterogeneous response signals collected by the multi-physics field collaborative sensing module are modeled as a high-order tensor in the intelligent recognition center, where the different dimensions of the tensor correspond to the time series, physical field stimulus type, spatial sensing point and signal feature mode, respectively. The intelligent recognition center employs a fusion algorithm based on tensor decomposition and graph neural networks to reduce the dimensionality of the high-order tensor and mine its correlations.
4. The intelligent identification and re-examination system for abnormal samples in a platelet function analyzer according to claim 1, characterized in that: The neural-symbolic hybrid reasoning algorithm includes a neural network subsystem and a symbolic reasoning subsystem; The neural network subsystem is implemented by a hierarchical Transformer architecture, which includes a temporal Transformer encoder for processing time-series signals, a spectral Transformer encoder for processing spectral signals, and a cross-modal attention fusion layer responsible for extracting high-level features from the heterogeneous response signals and generating initial symbolic propositions. The symbolic reasoning subsystem includes an extensible domain knowledge base, which uses a probabilistic logic programming language to represent knowledge of platelet physiology, pathology, and interfering factors as weighted logical rules. The subsystem executes a reasoning algorithm that combines a probabilistic graphical model with a Markov logic network. It takes the symbolic propositions generated by the neural network subsystem as evidence input, performs uncertain reasoning under the rule constraints of the knowledge base, calculates the posterior probability distribution of the anomaly type and its root cause hypothesis, and generates an interpretable reasoning path graph. The key fusion step of the neural symbolic hybrid inference algorithm is characterized by the following formula: for the final decision variable Its probability is determined by the prior probability perceived by the neural network. Posterior probability of symbolic logic reasoning Weighted integration using Dirichlet distribution, i.e. in For the original multiphysics data, For a knowledge base, As a proposition of evidence, This is a hyperparameter that reflects the confidence level of the two subsystems.
5. The intelligent identification and re-examination system for abnormal samples in a platelet function analyzer according to claim 1, characterized in that: The reinforcement learning-based adaptive decision-making algorithm models the re-examination process as a partially observable Markov decision process. Its state space is jointly defined by the anomaly hypothesis confidence vector output by the intelligent recognition center, the remaining sample size, the set of available re-examination methods, and historical operation records; the action space consists of all executable atomic re-examination operations and their parameter combinations; the reward function is defined as follows: ,in It is to perform an action The resulting information gain is calculated using the expected KL divergence. It is the cost of reagents consumed in the action. It is the time spent on the action. , , These are the weighting coefficients.
6. The intelligent identification and re-examination system for abnormal samples in a platelet function analyzer according to claim 5, characterized in that: The integrated microfluidic chip laboratory in the automated re-inspection execution module uses digital microfluidic technology as the core control platform.
7. The intelligent identification and re-examination system for abnormal samples of the platelet function analyzer according to claim 1, characterized in that: The federated learning framework in the system's self-evolutionary engine employs an asynchronous federated averaging algorithm based on contribution evaluation.
8. The intelligent identification and re-examination system for abnormal samples of the platelet function analyzer according to claim 1, characterized in that: The specific algorithm for dynamically updating the anomaly pattern library is as follows: Feature extraction and representation: For each new input abnormal sample data, the feature extractor trained in the intelligent recognition center is used to map it into a low-dimensional semantic embedding space to obtain its feature vector. ; Similarity Calculation and Retrieval: Calculation Compared with the prototype vectors of all known anomaly patterns in the anomaly pattern library The cosine similarity, where the prototype vectors It is calculated from the mean of the feature vectors of all samples belonging to this category, i.e. , For feature extraction function, For category Support set; Novelty determination: If the maximum similarity Below the preset threshold If the sample is found to represent a new abnormal pattern, it will be temporarily stored in the "Pool of Emerging Patterns to be Confirmed". Expert review and confirmation: Cases in the "Pool of Emerging Patterns to be Confirmed" are regularly submitted to human experts for review and confirmation, along with their original multi-physics signal characteristics and preliminary analysis of the intelligent recognition center. Knowledge base expansion: For new patterns confirmed by experts, their feature vectors are added to the abnormal pattern library as new prototypes, and the knowledge base expansion process of the symbolic reasoning subsystem in the neural symbolic hybrid reasoning algorithm is triggered, allowing experts or natural language processing interfaces to add logical rules describing the new patterns, thus completing the forward expansion of system knowledge.
9. The intelligent identification and re-examination system for abnormal samples in a platelet function analyzer according to claim 1, characterized in that: Also includes: Multi-source confidence fusion and conflict resolution module; This module receives multiple competing hypotheses and their confidence levels from the intelligent identification center regarding the same anomaly root cause, and also receives independent or partially related verification evidence that may be generated by different re-examination methods in the automated re-examination execution module. This module uses the Dempster-Shafer evidence theory algorithm for information fusion: each information source is regarded as an evidence body, and basic probability values are assigned to the set of hypotheses it supports; when there is a conflict between different evidence bodies, an improved combination rule that considers evidence distance and reliability is used for synthesis, and finally a comprehensive confidence function distribution is output to guide whether to terminate the re-examination or initiate a higher-order expert arbitration process. The detection process is monitored and self-calibrated in real time. During each main detection and re-detection process, the subsystem runs a dynamic state-space model based on a recurrent neural network in parallel, using the sensor readings and control commands from the previous moment as inputs to predict the sensor readings at the current moment. By comparing the residual sequence of predicted values with actual measured values and performing chi-square and sequential probability ratio tests, the system can detect process anomalies such as sensor drift, reagent failure, microchannel blockage, or bubble interference in real time. Once a significant deviation is detected, a targeted self-calibration routine is immediately triggered or an alarm is sent to the system to ensure the reliability of data acquisition.