Kidney rejection monitoring system based on wireless implantable biosensor

Abstract

The invention relates to the technical field of biosensors and wireless health monitoring, and particularly discloses a renal rejection monitoring system based on a wireless implantable biosensor. The system comprises a wireless implantable biosensor array, an in-vivo signal preprocessing unit, an in-vitro data receiving and synchronizing device, a multi-modal heterogeneous data fusion engine and a space-time diagram neural network prediction model, immune and metabolism multi-dimensional biomarkers are collected around a transplanted kidney in real time, a dynamic'immunity-metabolism 'knowledge graph is constructed, and accurate prediction and graded early warning of rejection risks are achieved by means of space cascading and time evolution laws of time-space diagram neural network modeling signals. By the adoption of the technical scheme, sub-clinical rejection signals can be recognized before conventional indexes such as creatinine are abnormal, a window is intervened in advance, meanwhile, long-term non-invasive monitoring is achieved through wireless low-power-consumption design, and the accuracy and the intelligent level of postoperative management are improved.

Description

Technical Field

[0001] This invention belongs to the field of biosensor and wireless health monitoring technology, specifically relating to a kidney rejection monitoring system based on a wireless implantable biosensor. Background Technology

[0002] With the convergence of biomedical sensors and clinical monitoring technologies, wireless implantable monitoring systems have become a cutting-edge tool for post-organ transplant management. Real-time monitoring of the physiological function of transplanted organs is crucial for reducing post-operative complications and prolonging graft survival. In the process of kidney transplant rehabilitation, using high-precision sensing elements to capture biochemical signals within the body can provide direct and objective evidence for the early detection of kidney function impairment and rejection, driving a shift in clinical diagnosis and treatment from an experience-driven to a data-driven model.

[0003] Monitoring systems for renal rejection focus on constructing a comprehensive evaluation system covering metabolic pathways and immune responses through real-time sensing of multimodal biological information. The core objective of this research is to utilize miniaturized sensing matrices to collaboratively acquire key physiological parameters within the recipient and aggregate them to the analysis end via efficient data transmission protocols. The system typically requires preprocessing and feature alignment of these complex biological signals to accurately characterize the survival activity and immune rejection intensity of the transplanted kidney within the complex biological environment.

[0004] Current monitoring methods primarily rely on macroscopic clinical indicators such as elevated creatinine or changes in urine output. However, their diagnostic feedback often lags behind the actual occurrence of pathological damage, limiting the window for treatment intervention. Traditional analytical methods typically perform threshold matching only on single-dimensional data, lacking in-depth analysis of the heterogeneous relationships between metabolites and cytokines, making them prone to misdiagnosis when receptors experience infection or other complications.

[0005] Existing algorithms and models exhibit insufficient ability to capture spatiotemporal features when processing long-term, dynamically evolving immune signals, and are unable to simulate the nonlinear diffusion process of immune storms in biological networks. These shortcomings make it difficult to identify subtle pathological fluctuations at the subclinical level in their early stages, thereby affecting the accuracy and scientific rigor of early warning decisions. Therefore, a renal rejection monitoring system based on wireless implantable biosensors is desired. Summary of the Invention

[0006] The purpose of this invention is to provide a kidney rejection monitoring system based on a wireless implantable biosensor, which can solve the problems in the background art.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A kidney rejection monitoring system based on wireless implantable biosensors includes a wireless implantable biosensor array, an in vivo signal preprocessing unit, an in vitro data receiving and synchronization device, a multimodal heterogeneous data fusion engine, and a spatiotemporal graph neural network prediction model. The wireless implantable biosensor array is configured to collect multi-dimensional biomarker signals associated with immunology and metabolomics in real time in the tissue microenvironment surrounding the transplanted kidney, including dynamic changes in cytokine concentration and fluctuations in metabolite concentration. The in vivo signal preprocessing unit is electrically connected to the wireless implantable biosensor array and is configured to filter, amplify and convert the original biological signal into an analog-to-digital signal, and transmit the processed digital signal to the outside of the body through a wireless communication protocol. The in vitro data receiving and synchronization device is configured to receive wireless signals from the in vivo signal preprocessing unit and align different types of biomarker data on the time axis according to a unified timestamp to form a structured time-series data stream. The multimodal heterogeneous data fusion engine is configured to map aligned immunological and metabolomics data to a unified knowledge graph framework, constructing a dynamic heterogeneous graph structure that reflects the interactions between immune cells, renal tubular epithelial cells, and metabolic pathways. The spatiotemporal graph neural network prediction model is configured to model the cascading propagation characteristics of biological signals in the spatial dimension and the latent evolution law in the temporal dimension based on the dynamic heterogeneous graph structure, and output the risk prediction result of renal rejection reaction.

[0008] Preferably, the wireless implantable biosensor array includes multiple miniaturized biosensitive elements that specifically identify key biomarkers such as interleukin-6, tumor necrosis factor alpha, glucose, and lactate, and are integrated on a flexible substrate to adapt to the local anatomical structure of the transplanted kidney.

[0009] Furthermore, the in vivo signal preprocessing unit incorporates a low-power microcontroller and a radio frequency transceiver module. Its operating mode is dynamically controlled by external commands, activating sensing and transmission functions only within a preset sampling period to extend the battery life of the implanted device.

[0010] Furthermore, the in vitro data receiving and synchronization device adopts a dual-channel time synchronization mechanism. On the one hand, it ensures the time consistency of each biomarker data through a high-precision real-time clock. On the other hand, it uses event-triggered markers to accurately label sudden physiological fluctuations, thereby improving the time-series fidelity of subsequent analysis.

[0011] Preferably, in the knowledge graph constructed by the multimodal heterogeneous data fusion engine, the node representation includes immune cell types, metabolite types, and kidney tissue functional units, while the edge weights are dynamically adjusted based on the correlation between biomedical prior knowledge and real-time data to reflect the changes in the intensity of intercellular signal interaction under different pathological states.

[0012] Furthermore, the spatial branch of the spatiotemporal graph neural network prediction model adopts a graph convolutional network architecture, which simulates the diffusion process of immune activation signals in the renal tissue microenvironment by aggregating the feature information of adjacent nodes; its temporal branch adopts a long short-term memory network structure to capture subclinical fluctuation patterns in the days to weeks before the rejection reaction occurs.

[0013] Furthermore, the output layer of the spatiotemporal graph neural network prediction model is equipped with a risk grading module, which classifies the rejection risk into three levels—low, medium, and high—based on the prediction score, and generates corresponding clinical intervention suggestions for doctors to make decisions.

[0014] Preferably, the system further includes a user interaction terminal, which is communicatively connected to the spatiotemporal graph neural network prediction model, for visually displaying the rejection risk trend curve, the dynamic change heat map of key biomarkers and the knowledge graph topology, and supports remote parameter configuration and alarm threshold adjustment.

[0015] Furthermore, the user interaction terminal is equipped with an abnormal event backtracking function. When the system determines that there is a high-risk rejection tendency, it can automatically retrieve all related biomarker data and map evolution snapshots within the past 7 days to assist clinical experts in causal inference and diagnostic verification.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. The renal rejection monitoring system based on a wireless implantable biosensor provided by this invention constructs an "immuno-metabolic" knowledge graph that dynamically characterizes the transplanted kidney's microenvironment by deeply integrating multi-source heterogeneous data from immunology and metabolomics. This overcomes the false-positive problem caused by confounding factors such as infection in traditional single biomarker monitoring methods. 2. This invention introduces a spatiotemporal graph neural network prediction model, which not only accurately depicts the signal cascade relationship between immune cells and renal parenchymal cells in the spatial dimension but also captures the subtle evolutionary features of the rejection latency period in the temporal dimension. This system can identify early signals of subclinical rejection before organ function indicators such as creatinine show obvious abnormalities, thus broadening the time window for clinical intervention. 3. Through wireless implantable sensing and low-power design, this invention enables long-term, continuous, and non-invasive in vivo monitoring, avoiding the burden on patients caused by frequent punctures or imaging examinations, and improving the intelligence and precision of postoperative management. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the spatiotemporal graph neural network prediction model in this invention; Figure 3 This is a logical flowchart of the multimodal heterogeneous data fusion and dynamic heterogeneous graph construction in this invention; Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow between in vivo signal preprocessing and in vitro data synchronous processing in this invention. Detailed Implementation

[0018] Example 1: Please refer to the appendix Figure 1 To be continued Figure 4 To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0019] A kidney rejection monitoring system based on wireless implantable biosensors includes a wireless implantable biosensor array, an in vivo signal preprocessing unit, an in vitro data receiving and synchronization device, a multimodal heterogeneous data fusion engine, a spatiotemporal graph neural network prediction model, and a user interaction terminal. The wireless implantable biosensor array is used to collect multi-dimensional biomarker signals associated with immunology and metabolomics in real time within the tissue microenvironment surrounding the transplanted kidney. These signals include dynamic changes in cytokine concentrations and fluctuations in metabolite concentrations. The wireless implantable biosensor array uses a flexible biocompatible substrate as its physical carrier. The substrate material is selected from polyimide or polydimethylsiloxane to ensure that the sensor array can closely adhere to the capsule surface of the transplanted kidney without causing significant foreign body reactions.

[0020] The flexible substrate integrates multiple miniaturized biosensors, including a specific immunosensor for interleukin-6, a specific immunosensor for tumor necrosis factor-alpha, an enzyme-catalyzed electrochemical sensor for glucose, and a mediator-modified electrochemical sensor for lactate.

[0021] The specific immunosensor targeting interleukin-6 employs a gold nanoparticle-modified electrode surface, with interleukin-6-specific antibodies covalently immobilized. When interleukin-6 molecules in the tissue fluid specifically bind to the antibody, a change in the charge transfer resistance at the electrode interface occurs, achieving an electrical signal conversion of cytokine concentration. The enzyme-catalyzed electrochemical sensor targeting glucose utilizes the catalytic action of glucose oxidase on the electrode surface, indirectly reflecting the real-time glucose concentration by detecting the oxidation current of hydrogen peroxide. All sensitive elements are encapsulated with polymer nanoporous membranes to filter out interference from large protein molecules, ensuring the selectivity of target analytes in complex interstitial fluids.

[0022] The in vivo signal preprocessing unit is electrically connected to the wireless implantable biosensor array and is configured to filter, amplify, and convert the raw biological signals to analog-to-digital signals, transmitting the processed digital signals to the outside via a wireless communication protocol. The in vivo signal preprocessing unit incorporates a precision low-noise instrumentation amplifier circuit, configured to amplify the microampere-level current signal or millivolt-level voltage signal output from the sensor array in one stage, followed by a second-order Butterworth low-pass filter to remove high-frequency physiological noise and electromagnetic interference. The preprocessing unit integrates a 16-bit resolution successive approximation analog-to-digital converter to convert the analog signal into a high-fidelity digital sequence.

[0023] The in vivo signal preprocessing unit also houses an ultra-low-power microcontroller configured to execute sampling logic control, with its operating mode dynamically adjusted by external commands. In sleep mode, the system maintains only the real-time clock circuit; in sampling mode, the microcontroller sequentially activates each sensor channel and completes data acquisition. The in vivo signal preprocessing unit also includes a radio frequency transceiver module that supports wireless transmission compliant with medical implantable communication service protocols, ensuring data integrity and accuracy during cross-skin transmission through a cyclic redundancy check mechanism.

[0024] The in vitro data receiving and synchronization device is configured to receive wireless signals from the in vivo signal preprocessing unit and align different types of biomarker data along the timeline based on a unified timestamp to form a structured time-series data stream. The in vitro data receiving and synchronization device consists of a portable receiving terminal and a base station, employing a dual-channel time synchronization mechanism. The first channel uses a high-precision real-time clock built into the receiving terminal to assign an absolute timestamp to each frame of received raw biomarker data; the second channel utilizes event-triggered markers to automatically record the start and end times of sudden physiological fluctuations when a biomarker concentration instantaneous change rate exceeds a preset threshold.

[0025] The in vitro data receiving and synchronization device internally stores a buffer array to address the issue of inconsistent sampling frequencies for different biomarkers. For high-frequency electrochemical signals and low-frequency immune signals, the device employs linear interpolation or spline interpolation algorithms for resampling, ensuring that the data for interleukin-6, tumor necrosis factor-alpha, glucose, and lactate are aligned on the same time axis, thus eliminating analytical errors caused by sampling lag.

[0026] The multimodal heterogeneous data fusion engine is configured to map aligned immunological and metabolomics data onto a unified knowledge graph framework, constructing a dynamic heterogeneous graph structure reflecting the interactions between immune cells, renal tubular epithelial cells, and metabolic pathways. Within this knowledge graph, three core node types are defined: immune nodes, metabolic nodes, and tissue function nodes. Immune nodes represent different types of cytokines and their corresponding immune cell activation states; metabolic nodes represent the concentration levels of metabolic substrates or products such as glucose and lactate; and tissue function nodes characterize the metabolic activity and permeability of renal tubular epithelial cells.

[0027] The multimodal heterogeneous data fusion engine establishes initial connections between nodes based on a biomedical prior knowledge base. For example, when the system detects an increase in tumor necrosis factor alpha concentration, it logically establishes a negatively weighted edge pointing to a node representing renal tubular tissue function, indicating potential damage to the renal parenchyma from pro-inflammatory factors. The edge weights are dynamically adjusted based on the correlation of real-time acquired heterogeneous data. If the curve of interleukin-6 changes shows a high degree of spatiotemporal consistency with the trend of increasing lactate concentration, the system automatically increases the edge weight connecting these two nodes. In this way, the multimodal heterogeneous data fusion engine transforms isolated sensor data into an interconnected "immuno-metabolic" biosignal network, achieving a system-level characterization of the pathophysiological state of the transplanted kidney microenvironment.

[0028] The spatiotemporal graph neural network prediction model is configured to model the cascading propagation characteristics of biological signals in the spatial dimension and the latent evolution law in the temporal dimension based on the dynamic heterogeneous graph structure, and output the risk prediction result of renal rejection. The spatiotemporal graph neural network prediction model consists of alternating stacked spatial feature extraction layers and temporal feature memory layers. The spatial feature extraction layer adopts a graph convolutional network architecture, configured to simulate the diffusion process of immune activation signals in the renal tissue microenvironment by aggregating the feature information of neighboring nodes in the dynamic heterogeneous graph. The hidden state update of each node depends on its own current feature vector and the weighted sum of the feature vectors of all neighboring nodes, with the weights calculated based on the strength of biological interactions between nodes.

[0029] The temporal feature memory layer employs a long short-term memory (LSTM) network structure, configured to capture subclinical fluctuation patterns occurring several days to weeks prior to rejection. The LSM network includes input gates, forget gates, and output gates, maintaining long-term cell state vectors to preserve crucial early immune warning signals while discarding irrelevant physiological fluctuation noise. The output layer of the spatiotemporal graph neural network prediction model is equipped with a risk grading module. This module classifies the current risk into three levels: low, medium, and high risk, based on the rejection probability score calculated by the model. The probability score is a normalized value obtained after processing with a flexible maximization function. The system generates corresponding clinical intervention recommendations based on different risk levels; for example, it suggests increasing the frequency of immunosuppressant concentration monitoring when identified as medium risk, and issues a biopsy warning when identified as high risk.

[0030] The user interaction terminal communicates with the spatiotemporal graph neural network prediction model, and is used to visualize the rejection risk trend curve, the dynamic change heatmap of key biomarkers, and the knowledge graph topology. It also supports remote parameter configuration and alarm threshold adjustment. The user interaction terminal is equipped with an abnormal event backtracking function; when the system determines that there is a high-risk rejection tendency, it automatically retrieves all related biomarker data and graph evolution snapshots from the past seven days. Clinical experts can observe the abnormal evolution path of edge weights in the knowledge graph through the terminal interface, trace the initial source of the immune storm, and assist in causal inference and diagnostic verification.

[0031] In the manufacturing process of the wireless implantable biosensor array, the electrode surface undergoes multi-layer functionalization modification. First, a self-assembled monolayer is modified onto the gold electrode surface. This self-assembled monolayer consists of short-chain alkanes with thiol groups, providing stable chemical anchoring sites. Subsequently, a capture antibody is covalently linked to the terminal carboxyl group of the self-assembled monolayer using a carbodiimide chemical reaction. To reduce non-specific adsorption, the electrode surface is further blocked using bovine serum albumin. This multi-layer modified structure ensures that the sensor exhibits extremely low background current and extremely high sensitivity in complex biological fluids, capable of detecting cytokine fluctuations at concentrations as low as picograms per milliliter.

[0032] In the low-power management mechanism of the in vivo signal preprocessing unit, the microcontroller is configured to execute task scheduling based on an adaptive sampling strategy. When a biomarker is detected to be in a stable state, the system automatically extends the sampling period to once every 4 hours to maximize battery conservation. Once the instantaneous rate of change of interleukin-6 or lactate is detected to exceed a preset safety slope, the microcontroller immediately triggers a high-frequency monitoring mode, increasing the sampling frequency to once every 10 minutes until the signal stabilizes again. This dynamic adjustment mechanism allows the system to extend the operating time of the implanted device without missing any potential rejection signals.

[0033] The multimodal heterogeneous data fusion engine introduces an attention mechanism to calculate edge weights when constructing a dynamic heterogeneous graph. This attention mechanism is configured to calculate the mutual information between any two related nodes and assign attention scores based on the magnitude of the mutual information. As the co-evolutionary characteristics between immune and metabolic nodes increase, the attention score increases, reflecting the intensified coupling between immune responses and metabolic disorders. The engine also incorporates external clinical benchmark data as global biases into the graph structure, such as patient age, postoperative days, and baseline blood concentrations of immunosuppressants, making the constructed dynamic graph structure more consistent with the individualized pathological background of the patient.

[0034] The training process of the spatiotemporal graph neural network prediction model employs a pre-training scheme based on contrastive learning. The system first learns general characteristic representations of biomarker fluctuations on a large historical kidney transplant dataset, and then fine-tunes and optimizes it for real-time patient data. The model loss function consists of a classification prediction loss and a spatiotemporal consistency constraint term. The classification prediction loss measures the consistency between the prediction result and the actual biopsy diagnosis; the spatiotemporal consistency constraint term requires the model to output risk predictions smoothly within adjacent time periods, preventing warning jitter caused by instantaneous sensor errors.

[0035] The user interaction terminal also integrates a remote medical collaboration module, supporting the encrypted transmission of exclusion risk reports and associated knowledge graph snapshots to a cloud-based expert system. The cloud-based expert system is equipped with a large-scale language model-assisted decision support tool, capable of automatically generating detailed pathology analysis reports by combining real-time monitoring data with medical literature databases, providing in-depth decision support for frontline clinicians.

[0036] Example 2: As a further extension or alternative architecture to Example 1, this example describes a kidney rejection monitoring system based on a wireless implantable biosensor employing a distributed edge computing architecture. In this distributed edge computing architecture, data processing and risk prediction tasks are dynamically load-balanced among the implanted device, edge terminal, and cloud server to meet the demands for higher real-time performance and extended battery life.

[0037] A kidney rejection monitoring system based on wireless implantable biosensors includes a distributed implantable sensing unit, an edge computing gateway, a cloud-based big data analysis platform, and a medical monitoring terminal. The distributed implantable sensing unit consists of multiple independently packaged miniature sensing modules, deployed at the renal hilum, renal cortex surface, and surrounding lymph nodes of the transplanted kidney. Each miniature sensing module includes a biosensitive element, a microprocessor, and a short-range near-field communication chip. The short-range near-field communication chip is used for self-assembled network communication between the miniature sensing modules, enabling collaborative acquisition of local microenvironment data.

[0038] The edge computing gateway receives data from distributed implantable sensing units. Deployed on the patient's skin in a flexible, wearable form, the edge computing gateway integrates a high-performance, low-power neural network acceleration unit. It performs preliminary signal denoising and feature extraction logic and runs a simplified spatiotemporal graph neural network model. This simplified model focuses on capturing high-risk physiological mutation signals; when suspected acute rejection features are detected, an alarm is triggered immediately without waiting for results from the cloud.

[0039] The cloud-based big data analytics platform is connected to the edge computing gateway via an encrypted cellular network. The cloud-based big data analytics platform runs a complete, high-parameter multimodal heterogeneous data fusion engine and a deep spatiotemporal graph neural network prediction model. The cloud model is configured to handle historical trend analysis over long time spans, extracting subtle feature patterns reflecting subclinical rejection by performing population modeling on tens of thousands of kidney transplant case data.

[0040] The medical monitoring terminal is used to provide medical staff with a multi-dimensional monitoring view. Compared with Embodiment 1, the medical monitoring terminal in this embodiment adds a group risk profile comparison function. The system matches the current patient's "immune-metabolic" knowledge graph evolution path with various known rejection reaction standard templates in the database in real time and calculates the trajectory similarity score. If the current trajectory has a similarity of more than 85% with the typical evolution path of acute cellular rejection reaction, the system will automatically lock the diagnostic tendency and list supporting evidence.

[0041] The distributed implantable sensing unit utilizes wireless power transmission technology for power supply. The edge computing gateway, acting as the power transmitter, provides power to the miniature sensing module at the implantation site via inductive coupling, eliminating the implantation site's reliance on lithium batteries, reducing the size of the implantation device, and lowering the difficulty and trauma of surgical implantation.

[0042] When the multimodal heterogeneous data fusion engine is executed in the cloud, it incorporates a reinforcement learning optimization operator. This operator is configured to automatically adjust the weight contribution rate of different biomarker nodes in the knowledge graph based on feedback from clinicians regarding the accuracy of alarms. For example, if clinical feedback indicates that a patient's lactate fluctuations are significantly affected by non-rejection factors, the reinforcement learning operator will automatically reduce the influence factor of the lactate node in risk prediction, achieving personalized self-evolution of the model.

[0043] In this embodiment, the spatiotemporal graph neural network prediction model employs a multi-scale convolutional architecture. Spatially, the model extracts microscopic signal cascade features from local renal segments and macroscopic physiological response features from the entire renal region. Temporally, the model uses a parallel dilated causal convolutional network to simultaneously analyze real-time fluctuation trends at the minute level and chronic evolution patterns at the monthly level. This multi-scale feature fusion strategy further enhances the system's ability to differentiate and diagnose different types of rejection reactions.

[0044] Example 3: Based on Examples 1 and 2, this example proposes a kidney rejection monitoring system based on a wireless implantable biosensor with redundancy verification and self-repair capabilities, aiming to improve system reliability and data integrity in long-term implantation environments.

[0045] A kidney rejection monitoring system based on a wireless implantable biosensor includes a self-calibrating sensor array, an intelligent relay unit, a multi-model integrated prediction platform, and a fault-tolerant interactive terminal. The self-calibrating sensor array is equipped with a primary and backup dual-channel sensor at each biomarker detection site. The self-calibrating sensor array integrates an impedance monitoring circuit configured to assess the contact impedance between the sensing electrode and the tissue interface in real time. When the impedance value of the primary sensor deviates from the preset range due to protein contamination or bio-encapsulation failure, the self-calibrating sensor array automatically switches to the backup sensor channel and sends a self-repair status report to the system via a heartbeat data packet.

[0046] The intelligent relay unit is configured within the patient's portable device. It not only handles wireless data forwarding but also executes data consistency verification logic. The intelligent relay unit utilizes an autoencoder model to compress and represent multi-dimensional biosignals. If the reconstruction error exceeds a preset confidence interval, it determines that the sensor data has an abnormal offset and triggers a local compensation algorithm to calibrate the offset data.

[0047] The multi-model ensemble prediction platform runs on a hospital LAN server. This platform integrates a rule-based expert system, a random forest regression model, and the spatiotemporal graph neural network model, the core of this invention. The platform uses a weighted voting mechanism to derive the final risk classification result. When discrepancies arise between the prediction results of different models, the platform automatically activates the uncertainty estimation module to calculate the credibility index of the current prediction result. If the credibility is lower than a preset threshold, the system does not directly output a rejection conclusion but instead prompts clinicians to manually review the prediction.

[0048] The fault-tolerant interactive terminal features an offline monitoring mode. In the event of a wireless network signal interruption, the terminal can utilize the temporal characteristics of the last 500 locally cached sampling points to execute basic alarm logic. Furthermore, the fault-tolerant interactive terminal supports handheld QR code scanning access, allowing nurses or patients to quickly synchronize local monitoring data to a mobile client by scanning a dynamic QR code on the terminal, ensuring continuous data transmission.

[0049] The multimodal heterogeneous data fusion engine employs a generative adversarial network-based completion algorithm when handling missing data. When a marker in the sensor array experiences data loss due to a physical failure, the generative adversarial network is configured to utilize the spatiotemporal correlations between the remaining markers to generate highly physically realistic simulated supplementary data. This maintains the integrity of the knowledge graph structure and ensures that the prediction model can continue to run without interruption.

[0050] In this embodiment, the spatial branch of the spatiotemporal graph neural network prediction model incorporates a dynamic attention pooling layer. This dynamic attention pooling layer is configured to automatically identify the "critical regions" in the transplanted kidney tissue most sensitive to rejection and assign higher computational weights to the nodes corresponding to these critical regions. For example, in the early postoperative period, the model automatically focuses on immune nodes near the renal hilum vascular anastomosis; in the later recovery period, the model's focus shifts to functional metabolic nodes in the renal cortex. This dynamic overlap mechanism improves the predictive model's warning sensitivity at different pathological stages.

[0051] Example 4: This example describes a kidney rejection monitoring system based on wireless implantable biosensors driven by digital twins. By constructing a virtual image of the patient's transplanted kidney in digital space, it achieves higher-dimensional simulation and prediction of rejection risk.

[0052] A kidney rejection monitoring system based on wireless implantable biosensors includes a high-fidelity sensing module, a digital twin engine, a spatiotemporal neural network simulation platform, and a clinical decision inference terminal. In addition to collecting biomarkers such as interleukin-6 and tumor necrosis factor-alpha, the high-fidelity sensing module also integrates a miniature piezoelectric hemodynamic sensor to monitor the pulsating pressure and blood flow velocity of the transplanted renal artery. These high-dimensional physical and biochemical signals together constitute the real-time driving input for the digital twin.

[0053] The digital twin engine is used to construct a four-dimensional spatiotemporal model of the transplanted kidney in virtual space based on the collected multimodal heterogeneous data. The engine internally stores the geometric and topological parameters of the kidney's anatomical structure, as well as a set of partial differential equations for physiological and biochemical reactions. The engine uses real-time sensor data as boundary conditions for the equations and reconstructs the concentration gradient distribution of immune molecules within the interstitial spaces of the kidney tissue through numerical simulation methods.

[0054] The spatiotemporal neural network simulation platform is bidirectionally connected to the digital twin engine. The simulation platform receives input from real sensors as well as "pseudo-marker" data generated by the digital twin engine. The spatiotemporal graph neural network model is configured to perform inference on an augmented dataset composed of a mixture of real and simulated data. This structure allows the system to simulate forward-looking scenarios such as "how the risk of rejection will evolve in the next 24 hours if the immunosuppressant concentration decreases by 10%."

[0055] The clinical decision simulation terminal is configured to allow physicians to perform "simulated drug administration" operations on the interface. After the physician inputs the proposed medication regimen on the terminal, the digital twin engine and collaborative prediction model automatically deduce the expected response curve of the renal immune status under that medication regimen. By comparing the predicted rejection scores under different regimens, the system provides physicians with the optimal medication regimen recommendation.

[0056] In the multimodal heterogeneous data fusion engine, a cross-modal alignment strategy based on contrastive learning is introduced. This strategy is configured to map biochemical concentration signals, hemodynamic signals, and clinical laboratory indicators to the same high-dimensional feature space. By calculating the cosine similarity between feature vectors of different modalities, the system can identify data points that "appear normal but have inconsistent internal logic," thus capturing subtle physiological imbalances caused by subclinical rejection reactions at an early stage.

[0057] The spatiotemporal graph neural network prediction model employs a domain-adaptive algorithm during training. This algorithm is configured to extract common rejection features across patients and suppress noise interference from individual differences. The weight matrix within the model is regularized using orthogonalization constraints to ensure good generalization ability and interpretability. The hidden layer activation heatmap, visualized on the clinical decision-making terminal, clearly reveals which abnormal combinations of markers triggered the rejection warning, enhancing the system's clinical transparency.

[0058] In this embodiment, the wireless implantable biosensor array also features a self-cleaning function. By integrating a miniature ultrasonic transducer on the electrode surface, the system periodically generates mechanical vibrations with micron-level amplitudes to remove fibrin and cell debris adhering to the sensing membrane surface. This self-cleaning cycle is dynamically triggered by an in vivo signal preprocessing unit based on changes in the signal-to-noise ratio, extending the sensor's in vivo operating life from the traditional three months to over a year.

[0059] Example 5: This example details the specific operating logic and internal parameter configuration of the system when detecting "subclinical rejection". Subclinical rejection is defined as a transplanted kidney with normal functional indicators, but histological infiltration of immune cells.

[0060] The multimodal heterogeneous data fusion engine sets the sampling window of the knowledge graph to a sliding window mode with a window length of 48 hours. The engine is configured to calculate the cumulative rate of change of the association weights between immune nodes and metabolic nodes within this sampling window. If the association weight between tumor necrosis factor alpha and lactate shows a monotonically increasing trend within three consecutive windows, the fusion engine will issue an early warning enable signal to the prediction model even if the concentration values ​​at any single point do not exceed the alarm threshold.

[0061] The temporal branch of the spatiotemporal graph neural network prediction model is configured as a multi-level memory structure. The first level of memory consists of short-cycle cyclic units to capture hourly physiological fluctuations; the second level of memory consists of long-cycle cyclic units to capture chronic evolutionary features spanning one week. When outputting a risk score, the model introduces an intermediate variable called the "pathological offset index." The pathological offset index is equal to the Euclidean distance between the current atlas topological feature vector and the healthy baseline feature vector.

[0062] When the pathological deviation index exceeds a preset sensitivity threshold, the system determines that there is a risk of subclinical rejection. At this time, the user terminal automatically switches to the "deep analysis view". In this deep analysis view, the system uses mapping technology to highlight the top three biomarkers that contribute the most to risk prediction and displays their joint evolutionary trajectory over the past seven days.

[0063] The spatiotemporal graph neural network prediction model also integrates a differential privacy protection module. When interacting with the cloud, this module is configured to add controlled Laplace noise to the feature vectors of sensitive biomarkers. This approach ensures that the patient's genetic-level privacy data is not leaked while maintaining a loss of less than 1 / 100 in model inference accuracy.

[0064] The system of this invention is also equipped with automatic calibration logic, which is executed every 24 hours. During self-calibration, the in vivo signal preprocessing unit generates a set of standard microcurrent pulses injected into the sensing circuit to detect the gain linearity of the entire signal conditioning chain. If a hardware gain drift is detected, the microcontroller automatically updates the quantization compensation coefficients of the analog-to-digital converter. This mechanism ensures that the output data remains comparable throughout the implantation period of several months, and does not produce false "signal evolution" due to the aging of electronic components.

[0065] The multimodal heterogeneous data fusion engine supports custom ontology extensions. Researchers can add new marker nodes to the knowledge graph, such as novel chemokines like CXCL9 and CXCL10, through the user interface. The system is equipped with an automatic ontology reconstruction algorithm that can automatically find logical connection points between new nodes and existing immune-metabolic networks after their addition, and use historically retained original sampling signals for retrospective learning, enabling smooth upgrades to system functions.

[0066] It should be noted that although the above embodiments provide detailed logical and hardware descriptions of the various modules, units, and models of the system, in actual engineering implementation, the physical boundaries of these components can be merged or further separated according to the needs of integrated circuit design. For example, the function of the in vivo signal preprocessing unit can be partially integrated into the flexible circuit board of the sensor array.

[0067] Those skilled in the art will understand that various modifications and transformations can be made to the above-described system architecture, algorithm logic, and hardware configuration within the scope of the claims of this invention. For example, the type of sensor can be changed, the number of layers in the neural network can be adjusted, or different wireless communication protocols can be introduced. As long as these modifications are still based on the core technical ideas of "immune-metabolic heterogeneous data knowledge graph" and "spatiotemporal graph neural network," and aim to achieve early monitoring of renal rejection through multimodal data fusion, they should all fall within the protection scope of this invention.

[0068] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A renal rejection monitoring system based on a wireless implantable biosensor, characterized in that, include: A wireless implantable biosensor array is configured to acquire in real time multidimensional biomarker signals associated with immunology and metabolomics in the tissue microenvironment surrounding the transplanted kidney. The multidimensional biomarker signals include dynamic changes in cytokine concentrations and fluctuations in metabolite concentrations. An in vivo signal preprocessing unit is electrically connected to the wireless implantable biosensor array and is configured to filter, amplify, and convert the raw biological signals into analog and digital signals, and transmit the processed digital signals to the outside of the body via a wireless communication protocol. An in vitro data receiving and synchronization device is configured to receive wireless signals from the in vivo signal preprocessing unit and align different types of biomarker data on the time axis according to a unified timestamp to form a structured time-series data stream. A multimodal heterogeneous data fusion engine is configured to map aligned immunological and metabolomics data to a unified knowledge graph framework, constructing a dynamic heterogeneous graph structure that reflects the interactions between immune cells, renal tubular epithelial cells, and metabolic pathways. The spatiotemporal graph neural network prediction model is configured to model the cascading propagation characteristics of biological signals in the spatial dimension and the latent evolution law in the temporal dimension based on the dynamic heterogeneous graph structure, and output the risk prediction results of renal rejection reaction.

2. The renal rejection monitoring system based on a wireless implantable biosensor according to claim 1, characterized in that: The wireless implantable biosensor array uses a flexible biocompatible substrate as a physical carrier. The substrate is made of polyimide or polydimethylsiloxane and is used to adhere to the membrane surface of the transplanted kidney. The flexible biocompatible substrate integrates multiple miniaturized biosensitive elements, including a specific immunosensor for interleukin-6, a specific immunosensor for tumor necrosis factor alpha, an enzyme-catalyzed electrochemical sensor for glucose, and a mediator-modified electrochemical sensor for lactate. The specific immunosensor includes a gold electrode and a self-assembled monolayer modified on the surface of the gold electrode. The self-assembled monolayer is composed of short-chain alkanes with thiol groups, and capture antibodies are covalently linked to the terminal carboxyl groups of the self-assembled monolayer via a carbodiimide chemical reaction. The enzyme-catalyzed electrochemical sensor detects the oxidation current of hydrogen peroxide by utilizing the enzyme-catalyzed reaction on the electrode surface. The outer layer of each miniaturized biosensor is encapsulated with a polymer nanoporous membrane to filter out interference from macromolecular proteins.

3. The renal rejection monitoring system based on a wireless implantable biosensor according to claim 1, characterized in that: The in vivo signal preprocessing unit incorporates a precision low-noise instrumentation amplifier circuit, a second-order Butterworth low-pass filter, a successive approximation analog-to-digital converter, and a microcontroller that executes an adaptive sampling strategy. The precision low-noise instrumentation amplifier circuit is configured to amplify the microampere-level current signal or millivolt-level voltage signal output by the wireless implantable biosensor array in one stage. The second-order Butterworth low-pass filter is connected to the output of the precision low-noise instrumentation amplifier circuit to remove high-frequency physiological noise and electromagnetic interference. The microcontroller is configured with a sleep mode and a sampling mode. In the sampling mode, each sensor channel is activated sequentially, and the microcontroller is configured to execute task scheduling logic based on the signal change rate. When the biomarker concentration is found to be stable, the sampling period will be extended to the preset first duration. When the instantaneous rate of change of a biomarker is detected to exceed the preset safety slope, the sampling frequency is immediately increased to a preset second duration, wherein the first duration is longer than the second duration.

4. The renal rejection monitoring system based on a wireless implantable biosensor according to claim 1, characterized in that: The external data receiving and synchronization device adopts a dual-channel time synchronization mechanism, including a built-in high-precision real-time clock, an event triggering marker module, and a buffer array; The high-precision real-time clock configuration is used to mark an absolute timestamp for each frame of raw biomarker data received. The event triggering marker module is configured to automatically record the start and end times of physiological fluctuations when the instantaneous rate of change in the concentration of a biomarker exceeds a preset threshold. The buffer array is configured to store heterogeneous data with different sampling frequencies and to perform resampling logic; The resampling logic is configured to process high-frequency electrochemical signals and low-frequency immune signals using linear interpolation or spline interpolation algorithms to align interleukin-6 data, tumor necrosis factor alpha data, glucose data, and lactate data on the same time axis to eliminate sampling lag.

5. The renal rejection monitoring system based on a wireless implantable biosensor according to claim 1, characterized in that: When constructing the dynamic heterogeneous graph structure, the multimodal heterogeneous data fusion engine defines three core node types: immune nodes representing cytokines and their immune cell activation states, metabolic nodes representing metabolite concentration levels, and tissue function nodes characterizing the metabolic activity of renal parenchymal cells. The multimodal heterogeneous data fusion engine is equipped with an attention mechanism module, which is configured to calculate the mutual information between any two related nodes and assign attention scores as edge weights based on the magnitude of the mutual information. The attention score increases as the co-evolutionary characteristics between immune nodes and metabolic nodes are enhanced. The multimodal heterogeneous data fusion engine is also equipped with a global bias introduction module, which is used to introduce external clinical benchmark data as a global bias term into the dynamic heterogeneous graph structure. The external clinical benchmark data includes patient age, postoperative days, and baseline blood drug concentration of immunosuppressants.

6. The renal rejection monitoring system based on a wireless implantable biosensor according to claim 1, characterized in that: The spatiotemporal graph neural network prediction model is composed of alternating stacks of spatial feature extraction layers and temporal feature memory layers; The spatial feature extraction layer adopts a graph convolutional network architecture, which is configured to simulate the diffusion process of immune activation signals in the kidney tissue microenvironment by aggregating the feature information of neighboring nodes in the dynamic heterogeneous graph. The hidden state update of each node depends on its own feature vector at the current moment and the weighted sum of the feature vectors of all neighboring nodes. The calculation of the weights takes into account the strength of biological interactions between nodes. The time-feature memory layer adopts a long short-term memory network structure and is configured to capture subclinical fluctuation patterns during the latent period of rejection reaction. The Long Short-Term Memory (LSTM) network contains an input gate, a forget gate, and an output gate. It retains immune warning signals and discards physiological fluctuation noise by maintaining long-term cell state vectors.

7. The renal rejection monitoring system based on a wireless implantable biosensor according to claim 1, characterized in that: The spatiotemporal graph neural network prediction model is equipped with a pre-training module and a risk classification module based on contrastive learning. The pre-training module is configured to learn general feature representations of biomarker fluctuations on historical datasets and fine-tune them based on real-time data of current patients. Its model loss function includes classification prediction loss and spatiotemporal consistency constraint term. The spatiotemporal consistency constraint term requires that the difference between the risk prediction values ​​output by the model in adjacent time periods be less than a preset smoothing threshold. The risk grading module is connected to the output layer of the spatiotemporal graph neural network prediction model. It is configured to normalize the rejection probability score and classify the risk into three levels: low risk, medium risk, and high risk based on the normalized value. It also generates corresponding clinical intervention recommendations for different risk levels.

8. The renal rejection monitoring system based on a wireless implantable biosensor according to claim 7, characterized in that: The system also includes a user interaction terminal, which is equipped with an abnormal event traceability module and a remote medical collaboration module. The abnormal event backtracking module is configured to automatically retrieve all related biomarker data within the past 7 days and the evolution snapshot of the dynamic heterogeneous graph structure when the spatiotemporal graph neural network prediction model determines that there is a high-risk rejection tendency, and visualize the abnormal evolution path of the edge weights in the dynamic heterogeneous graph structure. The remote medical collaboration module is configured to encrypt and transmit rejection risk reports and associated map snapshots to the cloud, and call a large-scale language model combined with a medical literature database to automatically generate pathological analysis reports.

9. The renal rejection monitoring system based on a wireless implantable biosensor according to claim 8, characterized in that: The system also includes a self-verification circuit and a self-cleaning module; The wireless implantable biosensor array is equipped with a primary and backup dual-channel sensor at each detection site. The self-calibration circuit is integrated in the in vivo signal preprocessing unit and is configured to evaluate the contact impedance between the sensing electrode and the tissue interface in real time. When the contact impedance value corresponding to the main sensor deviates from the preset range, the control system automatically switches to the backup sensor channel. The self-cleaning module includes a miniature ultrasonic transducer integrated on the electrode surface, configured to generate mechanical vibrations with micron-level amplitudes to remove fibrin and cell debris attached to the sensing membrane surface. The triggering period of the mechanical vibrations is dynamically determined by the in vivo signal preprocessing unit based on the degree of decrease in the signal-to-noise ratio.

10. The renal rejection monitoring system based on a wireless implantable biosensor according to claim 9, characterized in that: The system also includes a digital twin engine and a clinical decision deduction terminal; The digital twin engine is configured to construct a four-dimensional spatiotemporal model of the transplanted kidney in a virtual space based on the collected multimodal heterogeneous data, the geometric and topological parameters of the kidney's anatomical structure, and the partial differential equations of physiological and biochemical reactions. The digital twin engine uses real-time collected multi-dimensional biomarker signals as boundary conditions of the partial differential equation system, and reconstructs the concentration gradient distribution of immune molecules in the renal tissue interstitial space through numerical simulation, and generates simulation data which is then input into the spatiotemporal neural network prediction model. The clinical decision simulation terminal is configured to receive simulated medication plans input by physicians and invoke the digital twin engine and the spatiotemporal graph neural network prediction model to deduce the expected response curve of the renal immune status under the simulated medication plan.

Images