Hemodialysis multi-mode risk closed-loop control system

Abstract

The invention relates to the field of medical informatics, and provides a hemodialysis multi-mode risk closed-loop control system. Comprising a multi-source data acquisition module used for acquiring heterogeneous data in a dialysis process in real time; the space-time fusion engine module is used for carrying out space-time alignment and feature extraction of cross-modal data on the heterogeneous data by adopting a graph neural network architecture to generate a fusion feature vector; the dynamic risk prediction module is used for constructing a multi-dimensional risk assessment system based on the fusion feature vector, carrying out probability prediction on multiple risk types and outputting a prediction result; and the intelligent intervention control module evaluates the risk level according to the prediction result output by the dynamic risk prediction module, and generates an intervention strategy through a reinforcement learning algorithm. According to the invention, the timeliness and accuracy of intervention response are improved, and the incidence rate of dialysis complications is reduced.

Description

Technical Field

[0001] This invention relates to the field of medical informatics technology, and in particular to a multimodal risk closed-loop control system for hemodialysis. Background Technology

[0002] Traditional hemodialysis monitoring systems primarily rely on the built-in parameter monitoring functions of the dialysis machine, using fixed thresholds to monitor physiological indicators such as blood pressure, venous pressure, and transmembrane pressure in real time. When monitored parameters exceed preset ranges, the system triggers an alarm to prompt medical staff to intervene. This threshold-based alarm mechanism plays a fundamental role in ensuring safety during dialysis. However, existing monitoring systems typically employ a single-modality data acquisition approach, with various sensors and monitoring devices operating independently, lacking data fusion and collaborative analysis capabilities. For example, dialysis machine parameter monitoring systems cannot acquire multi-dimensional physiological information such as changes in the patient's facial images and electroencephalogram (EEG) activity, resulting in a limited monitoring perspective. Furthermore, existing systems often employ a passive monitoring mode, only issuing alarms when parameters are abnormal, lacking the ability to proactively predict dialysis complications. Medical staff often can only intervene after the patient has already developed obvious symptoms.

[0003] Existing technologies suffer from the following shortcomings: First, data silos are severe, with data sources separated and unable to achieve cross-modal information integration and correlation analysis, resulting in the system's inability to comprehensively assess the patient's real-time status. Second, alarm mechanisms exhibit significant lag; alarm strategies based on fixed thresholds are reactive, and by the time the system triggers an alarm, the patient has often already experienced complications such as hypotension or coagulation, missing the optimal intervention window. Third, clinical practice primarily relies on medical staff visually observing changes in dialyzer color or bleeding at the puncture site, lacking objective and quantitative real-time monitoring technology, which easily leads to missed detections or misdiagnoses. These technological deficiencies result in a persistently high incidence of dialysis complications, seriously affecting patient dialysis safety and quality. Summary of the Invention

[0004] This invention provides a multimodal risk closed-loop control system for hemodialysis to overcome the shortcomings of existing technologies.

[0005] This invention provides a multimodal risk closed-loop control system for hemodialysis, comprising: A multi-source data acquisition module is used to acquire heterogeneous data in real time during the dialysis process; The spatiotemporal fusion engine module uses a graph neural network architecture to perform spatiotemporal alignment and feature extraction on the heterogeneous data across modalities, generating a fused feature vector. The dynamic risk prediction module constructs a multi-dimensional risk assessment system based on the fused feature vector, performs probability predictions for various risk types, and outputs the prediction results. The intelligent intervention control module assesses the risk level based on the prediction results output by the dynamic risk prediction module and generates intervention strategies through reinforcement learning algorithms.

[0006] According to the present invention, a multimodal risk closed-loop control system for hemodialysis is provided, wherein the multi-source data acquisition module includes: The dialysis machine parameter acquisition unit parses the dialysis machine parameters output by the dialysis equipment in real time through a standardized protocol interface to obtain dialysis machine parameter data; A physiological signal acquisition unit is equipped with a millimeter-wave radar sensor to monitor the patient's heartbeat displacement and calculate the stroke volume variability and peripheral vascular resistance index based on the heartbeat displacement to obtain physiological signals. The hyperspectral imaging acquisition unit receives imaging data from the hyperspectral imaging device and scans the patient's puncture point and face based on the imaging data to capture the hemoglobin absorbance change characteristics. It also receives temperature distribution data around the puncture point acquired by the thermal imaging camera to obtain hyperspectral feature data. The neurocognitive data acquisition unit receives the patient's electroencephalogram (EEG) signals acquired by the dry electrode EEG probe array, extracts the alpha and beta waves from the EEG signals, and records the patient's responses to cognitive questions at fixed intervals and the response time to obtain neurocognitive data. A distributed piezoelectric thin film array sensor is deployed at dialysis tubing nodes to detect tubing connectivity.

[0007] According to the present invention, a multimodal risk closed-loop control system for hemodialysis is provided, wherein the spatiotemporal fusion engine module includes: Node construction unit: used to construct the dialysis machine parameters into multiple time-series nodes according to time windows, construct the physiological signals into multiple physiological nodes, construct the hyperspectral data into multiple feature nodes, construct the neurocognitive data into multiple cognitive nodes, and establish a node set including all time-series nodes, all physiological nodes, all feature nodes, and all cognitive nodes; Graph structure construction unit: Calculates the Pearson correlation coefficient between any two nodes in the node set. When the absolute value of the calculated correlation coefficient is greater than a preset threshold, an edge connection is established between the corresponding two nodes to obtain an association graph structure; wherein, the weight value of the edge connection is equal to the absolute value of the correlation coefficient between the corresponding two nodes. An aggregation unit, which is built based on a graph convolutional network, is used to process the association graph structure to aggregate the node features in the node set and output a fused feature vector.

[0008] According to the present invention, a multimodal risk closed-loop control system for hemodialysis is provided, wherein the dynamic risk prediction module includes: The coagulation and bleeding co-monitoring unit is used to process the fused feature vector through the coagulation identification model and the bleeding verification model respectively, and output the coagulation risk probability value and the bleeding risk probability value. The low blood pressure early warning unit is based on the spatiotemporal cross-attention prediction model. It is used to perform fusion processing on the fused feature vector through the spatiotemporal cross-attention prediction model and output the probability value of low blood pressure occurrence. The heart failure decompensation early warning unit is used to calculate the heart failure index based on the fused feature vector through the heart failure risk assessment model. The cognitive impairment assessment unit is used to generate a cognitive impairment risk score based on the fused feature vector through a cognitive state assessment model.

[0009] According to the present invention, a multimodal risk closed-loop control system for hemodialysis is provided, wherein the coagulation and bleeding co-monitoring unit includes: Fiber optic sensors are used to acquire real-time light intensity changes in the dialyzer fiber bundles. A coagulation recognition subunit, which is constructed based on a convolutional neural network, is used to perform pattern recognition based on the fused feature vector and the light intensity change signal, extract fiber bundle blockage features, and output a coagulation risk probability value based on the fiber bundle blockage features; wherein, when the degree of blockage corresponding to the fiber bundle blockage features is greater than or equal to a first preset threshold, a coagulation risk signal is output. The bleeding verification subunit processes the fused feature vector and outputs a bleeding risk probability value based on the abnormal temperature diffusion rate and hemoglobin absorbance mutation data in the fused feature vector; wherein a bleeding risk signal is output when the abnormal temperature diffusion rate is greater than or equal to a second preset threshold and the hemoglobin absorbance mutation data is greater than or equal to a third preset threshold. The anticoagulation optimization subunit is constructed based on a deep Q-network reinforcement learning algorithm. It uses the coagulation risk probability value and the bleeding risk probability value as state inputs, calculates the Q value corresponding to different heparin doses, and outputs the heparin dose adjustment strategy with the largest Q value.

[0010] According to the multimodal risk closed-loop control system for hemodialysis provided by the present invention, the spatiotemporal cross-attention prediction model constructed by the hypotension early warning unit includes: A spatial encoder, which employs a visual Transformer architecture, is used to receive the fused feature vector, extract the spatial feature tensor, and quantize the degree of paleness. A time-series encoder, which employs a time-series convolutional network architecture, is used to receive time-series feature tensors from the fused feature vector, which extract the blood pressure fluctuation curve, ultrafiltration rate time-series data, and stroke volume variability trend. A cross-attention fusion layer is used to calculate the attention score matrix between the spatial feature tensor and the temporal feature tensor, and to perform weighted fusion of the spatial feature tensor and the temporal feature tensor through the attention score matrix to establish spatiotemporal collaborative features; The probability output layer receives the spatiotemporal collaborative features and outputs a low blood pressure occurrence probability value. When the low blood pressure occurrence probability value is greater than or equal to 85%, a low blood pressure warning signal is triggered.

[0011] According to the present invention, a multimodal risk closed-loop control system for hemodialysis is provided, wherein the intelligent intervention control module establishes a four-level risk level, including: Level 1 risk, which corresponds to the distributed piezoelectric thin film array sensor detecting a tubing dislodgement event or the coagulation and bleeding collaborative monitoring unit outputting a bleeding risk signal; Level 2 risk, wherein the low blood pressure early warning unit outputs a low blood pressure early warning signal; Level 3 risk, the Level 3 risk corresponds to the heart failure index output by the heart failure decompensation early warning unit being greater than or equal to the fourth preset threshold; Level 4 risk, where the level 4 risk corresponds to the cognitive impairment assessment unit outputting an impairment risk score greater than or equal to a fifth preset threshold, or the coagulation and bleeding co-monitoring unit outputting a coagulation risk signal.

[0012] According to the present invention, a multimodal risk closed-loop control system for hemodialysis is provided, wherein the intelligent intervention control module establishes a four-level intervention strategy, including: When the risk level is Level 1, a Level 1 intervention strategy is implemented, which is to clamp the dialysis tubing and activate the location call system. When the risk level is level 2, a level 2 intervention strategy is implemented, which involves reducing the ultrafiltration rate to zero and issuing a saline demand signal. When the risk level is level three, a level three intervention strategy is implemented, which is to switch the system to continuous venous-venous hemofiltration mode. When the risk level is level four, a level four intervention strategy is implemented, which involves issuing signals to adjust the heparin dosage requirement and the neuroprotective drug requirement.

[0013] A multimodal risk closed-loop control system for hemodialysis provided by the present invention further includes: Disease-customized monitoring module; For heart failure patients, the disease-customized monitoring module acquires the concentration trend data of N-terminal pro-brain natriuretic peptide, identifies the number of comet tail signs through lung ultrasound examination, and inputs the concentration trend data and the number of comet tail signs into the heart failure decompensation early warning unit of the dynamic risk prediction module for fusion evaluation. For diabetic patients, a non-invasive blood glucose monitoring device is configured to measure blood glucose levels in real time.

[0014] According to the present invention, a multimodal risk closed-loop control system for hemodialysis is provided, wherein the multimodal risk closed-loop control system for hemodialysis is implemented through a collaborative architecture of edge computing and cloud federated learning.

[0015] This invention provides a multimodal risk closed-loop control system for hemodialysis. By constructing a multi-source data acquisition module, it integrates dialysis machine parameters, physiological signals, hyperspectral imaging, and neurocognitive data, breaking the data silos of various independent sensing devices in traditional monitoring systems. This allows the system to comprehensively acquire real-time patient status information from multiple dimensions, providing a richer and more complete data foundation for subsequent risk assessment. Secondly, the spatiotemporal fusion engine module of this invention, employing a graph neural network architecture, can establish a relational graph structure between cross-modal data nodes. By calculating the correlation between nodes and establishing edge connections, it achieves deep fusion and collaborative analysis of heterogeneous data, overcoming the technical bottleneck of traditional systems' inability to correlate and analyze different data sources. This enables the system to capture potential risk features that are difficult to detect with single-modal data. Thirdly, through a dynamic risk prediction module, this invention, based on a multi-dimensional risk assessment system constructed from fused feature vectors, can simultaneously predict the probability of various dialysis complications such as coagulation, bleeding, hypotension, heart failure decompensation, and cognitive impairment. Compared to traditional fixed threshold alarm mechanisms, this invention achieves a shift from passive response to proactive early warning, significantly advancing the risk detection window and gaining valuable time for clinical intervention. The coagulation-bleeding co-monitoring unit of this invention identifies dialyzer fiber bundle blockage patterns using fiber optic sensors combined with convolutional neural networks, and performs dual verification through thermal imaging and hyperspectral imaging. This solves the problems of high subjectivity and missed detection rates caused by reliance on manual visual observation in traditional monitoring, achieving objective and quantitative monitoring of coagulation and bleeding risks. The hypotension early warning unit uses a spatiotemporal cross-attention mechanism to fuse facial image spatial features with blood pressure temporal features, enabling early warning before hypotension actually occurs. This avoids the lag problem of traditional systems triggering alarms only after blood pressure has already dropped, effectively reducing the risk of hypotension during dialysis. In addition, this invention also uses an intelligent intervention control module to generate graded intervention strategies based on risk levels using reinforcement learning algorithms, and links with the hardware execution layer to achieve closed-loop control. This closely integrates risk prediction with clinical intervention measures, changing the traditional system's disconnect between early warning and treatment, requiring manual judgment and intervention by medical staff. It achieves intelligent management of the entire process from risk identification to automatic intervention, significantly reducing the workload of medical staff, improving the timeliness and accuracy of intervention response, thereby effectively reducing the incidence of dialysis complications, improving the safety of hemodialysis treatment, and enhancing patients' quality of life. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the structure of a multimodal risk closed-loop control system for hemodialysis provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the spatiotemporal fusion engine module provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the dynamic risk prediction module provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, embodiments of this invention, and should not be construed as limiting the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. In the description of this invention, it should be understood that the terminology used is for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0019] The embodiments of the present invention are described below with reference to the figures.

[0020] like Figure 1 As shown, this invention provides a multimodal risk closed-loop control system for hemodialysis, comprising: a multi-source data acquisition module 100 for real-time acquisition of heterogeneous data during the dialysis process; a spatiotemporal fusion engine module 200, which uses a graph neural network architecture to perform spatiotemporal alignment and feature extraction of the heterogeneous data across modalities, generating a fused feature vector; a dynamic risk prediction module 300, which constructs a multi-dimensional risk assessment system based on the fused feature vector, performs probability predictions for various risk types, and outputs prediction results; and an intelligent intervention control module 400, which assesses the risk level based on the prediction results output by the dynamic risk prediction module, and generates intervention strategies through a reinforcement learning algorithm.

[0021] This invention provides a multimodal risk closed-loop control system for hemodialysis. Through the collaborative work of a multi-source data acquisition module, a spatiotemporal fusion engine module, a dynamic risk prediction module, and an intelligent intervention control module, it enables real-time monitoring, proactive early warning, and automatic intervention of various risks during the dialysis process.

[0022] For the multi-source data acquisition module, heterogeneous data during dialysis is acquired in real time. This heterogeneous data covers multiple dimensions, including dialysis equipment operating parameters, patient physiological signals, hyperspectral imaging information, and neurocognitive data. Subsequently, the acquired heterogeneous data is transmitted to the spatiotemporal fusion engine module via a data transmission interface. The spatiotemporal fusion engine module uses a graph neural network architecture to perform cross-modal spatiotemporal alignment and feature extraction on this heterogeneous data from different sensors and monitoring devices, integrating the originally independent data streams into a unified fusion feature vector. This fusion feature vector contains spatiotemporal correlation information between data from various dimensions during dialysis, providing a more comprehensive reflection of the patient's real-time status. The dynamic risk prediction module receives the fusion feature vector output by the spatiotemporal fusion engine module and constructs a multi-dimensional risk assessment system based on this vector. It performs probability predictions for various dialysis complications such as coagulation, bleeding, hypotension, heart failure decompensation, and cognitive impairment, outputting prediction results for each risk level. The intelligent intervention control module assesses the risk level based on the prediction results output by the dynamic risk prediction module and generates corresponding intervention strategies through reinforcement learning algorithms, linking the hardware execution unit to implement specific intervention actions.

[0023] The entire system forms a complete closed loop from data collection, feature fusion, risk prediction to intelligent intervention, overcoming the shortcomings of traditional dialysis monitoring systems such as isolated data, delayed alarms, and passive intervention. It realizes intelligent and proactive dialysis safety monitoring, significantly improves the safety of dialysis treatment, reduces the incidence of complications, and alleviates the workload of medical staff.

[0024] The multi-source data acquisition module 100 includes: a dialysis machine parameter acquisition unit, which analyzes the dialysis machine parameters output by the dialysis equipment in real time through a standardized protocol interface to obtain dialysis machine parameter data; a physiological signal acquisition unit, which is equipped with a millimeter-wave radar sensor to monitor the patient's heart rate displacement and calculate the stroke volume variability and peripheral vascular resistance index based on the heart rate displacement to obtain physiological signals; a hyperspectral imaging acquisition unit, which receives imaging data from a hyperspectral imaging device and scans the patient's puncture point and face based on the imaging data to capture the hemoglobin absorbance change characteristics and receives temperature distribution data around the puncture point acquired by a thermal imaging camera to obtain hyperspectral feature data; a neurocognitive data acquisition unit, which receives the patient's electroencephalogram (EEG) signals acquired by a dry electrode EEG probe array, extracts alpha and beta waves from the EEG signals, and records the patient's responses to cognitive questions at fixed intervals and response times to obtain neurocognitive data; and a distributed piezoelectric thin-film array sensor, which is deployed at dialysis tubing nodes to detect tubing connectivity.

[0025] Furthermore, the multi-source data acquisition module, serving as the data input terminal for the entire system, is responsible for acquiring various heterogeneous data during the dialysis process from multiple sensors and monitoring devices. The multi-source data acquisition module includes a dialysis machine parameter acquisition unit, a physiological signal acquisition unit, a hyperspectral image acquisition unit, a neurocognitive data acquisition unit, and a distributed piezoelectric thin-film array sensor. These multiple acquisition units establish connections with the dialysis equipment and patient monitoring equipment through their respective sensors and interfaces, acquiring multi-dimensional monitoring data in real time. Specifically, the dialysis machine parameter acquisition unit communicates with the control system of the dialysis equipment through a standardized protocol interface. This standardized protocol interface is compatible with the communication protocols of various brands of dialysis machines and can analyze at least fifteen dialysis machine parameters, such as ultrafiltration rate, transmembrane pressure, venous pressure, and blood flow velocity, output by the dialysis equipment in real time to obtain dialysis machine parameter data. The obtained parameters reflect the operating status of the dialysis process and the working condition of the dialyzer, and are an important basis for evaluating the dialysis effect and identifying equipment abnormalities.

[0026] The physiological signal acquisition unit is equipped with a millimeter-wave radar sensor. This millimeter-wave radar sensor adopts a non-contact monitoring method. By emitting millimeter-wave signals and receiving reflected signals, it can monitor the patient's heartbeat displacement with a displacement accuracy of 0.5 millimeters. After the heartbeat displacement data is analyzed by signal processing algorithms, key physiological parameters such as stroke volume variability and peripheral vascular resistance index can be calculated. These parameters can reflect the patient's cardiovascular function status and blood volume change trend, which is of great value for predicting dialysis complications such as hypotension.

[0027] The physiological signal acquisition unit outputs the calculated stroke volume variability and peripheral vascular resistance index as physiological signal data.

[0028] The hyperspectral imaging acquisition unit receives imaging data from a hyperspectral imaging device. This device can perform spectral scanning of the patient's puncture site and face at multiple wavelengths, capturing image information at different wavelengths. Based on the received imaging data, the hyperspectral imaging acquisition unit scans and analyzes the patient's puncture site and face. By analyzing the absorbance characteristics of hemoglobin at different wavelengths, it captures the changes in hemoglobin absorbance, reflecting the bleeding around the puncture site and the blood perfusion status of the patient's face. Simultaneously, the hyperspectral imaging acquisition unit also receives temperature distribution data around the puncture site from a thermal imaging camera. The thermal imaging camera can acquire images of the temperature distribution on the skin surface around the puncture site in a non-contact manner. Abnormal temperature changes can serve as early indicators of bleeding or infection. Finally, the hyperspectral imaging acquisition unit integrates the hemoglobin absorbance change characteristics and temperature distribution data into hyperspectral feature data for output.

[0029] The neurocognitive data acquisition unit receives patient EEG signals from a dry electrode EEG probe array. This miniaturized array allows for signal acquisition without the use of conductive gel, improving clinical usability. The unit performs frequency domain analysis on the received EEG signals, extracting alpha and beta wave components. Changes in the power ratio of alpha and beta waves reflect the patient's state of consciousness and cognitive function. Furthermore, the unit initiates cognitive question-and-answer sessions with the patient every 20 minutes via a voice interaction agent. These questions include questions about date, location, and simple calculations. The system records the patient's responses and response times, analyzing accuracy, fluency, and reaction speed to assess cognitive status. Finally, the unit integrates the extracted alpha and beta wave data, along with the responses and response times, into a comprehensive neurocognitive data output.

[0030] Distributed piezoelectric thin-film array sensors are deployed at critical nodes in the dialysis tubing, including arterial connections, dialyzer inlet and outlet, and venous connections—areas prone to dislodgement. The piezoelectric thin-film sensors can detect pressure changes at tubing connections. When the tubing becomes dislodged or loose, the pressure at the connection changes abruptly. The sensor can detect this pressure anomaly within 0.3 seconds and output tubing connectivity data.

[0031] Finally, the multi-source data acquisition module collects dialysis machine parameter data, physiological signals, hyperspectral feature data, neurocognitive data, and tubing connectivity data from the aforementioned acquisition units as a heterogeneous data set, and transmits it in real time to the spatiotemporal fusion engine module for subsequent processing via a data transmission bus. This invention's synchronous acquisition method of multi-source heterogeneous data overcomes the limitations of traditional monitoring systems with a single data source, providing a rich data foundation for comprehensively assessing patient status and accurately predicting dialysis risks.

[0032] like Figure 2 As shown, the spatiotemporal fusion engine module 200 includes: Node construction unit 210: used to construct the dialysis machine parameters into multiple time-series nodes according to time windows, the physiological signals into multiple physiological nodes, the hyperspectral data into multiple feature nodes, and the neurocognitive data into multiple cognitive nodes, establishing a node set including all time-series nodes, all physiological nodes, all feature nodes, and all cognitive nodes; Graph structure construction unit 220: calculates the Pearson correlation coefficient between any two nodes in the node set. When the absolute value of the calculated correlation coefficient is greater than a preset threshold, an edge connection is established between the corresponding two nodes to obtain an association graph structure; wherein, the weight value of the edge connection is equal to the absolute value of the correlation coefficient between the corresponding two nodes; Aggregation unit 230, the aggregation unit 230 is established based on a graph convolutional network, used to process the association graph structure to aggregate the node features in the node set and output a fused feature vector.

[0033] Furthermore, the spatiotemporal fusion engine module receives heterogeneous data transmitted from the multi-source data acquisition module and employs a graph neural network architecture to perform cross-modal spatiotemporal alignment and feature extraction on this heterogeneous data from different sensors and monitoring devices. Since the data sampling frequencies, data formats, and physical meanings of different acquisition units vary, direct data fusion would lead to information mismatch and feature conflicts. Therefore, the spatiotemporal fusion engine module of this invention first performs time alignment processing on the heterogeneous data, unifying data with different sampling frequencies to the same time reference, ensuring the synchronization of data across different modalities in the time dimension. Subsequently, the spatiotemporal fusion engine module of this invention includes a node construction unit, a graph structure construction unit, and an aggregation unit, which work collaboratively to complete the fusion of heterogeneous data.

[0034] Specifically, the node construction unit constructs dialysis machine parameters into multiple time-series nodes according to time windows. Each time-series node contains statistical characteristics of parameters such as ultrafiltration rate, transmembrane pressure, and venous pressure within a time window. Subsequently, physiological signals are constructed into multiple physiological nodes, each containing physiological parameters such as stroke volume variability and peripheral vascular resistance index at the corresponding time moment. Similarly, hyperspectral feature data are constructed into multiple feature nodes, each containing hemoglobin absorbance variation characteristics and temperature distribution characteristics. Neurocognitive data are constructed into multiple cognitive nodes, each containing the power ratio of alpha and beta waves and the scoring results of cognitive question answering. Finally, the node construction unit establishes a node set including all time-series nodes, all physiological nodes, all feature nodes, and all cognitive nodes. Each node in the node set represents the feature representation of a certain modality at a certain time moment or time window.

[0035] The graph structure construction unit receives the node set output by the node construction unit and calculates the Pearson correlation coefficient between any two nodes in the node set. This correlation coefficient reflects the degree of linear correlation between the data features represented by the two nodes. When the absolute value of the calculated correlation coefficient is greater than a preset threshold, it indicates that there is a strong correlation between the features represented by the two nodes. Then, the graph structure construction unit of this invention establishes an edge connection between the corresponding two nodes. The weight value of the edge connection is equal to the absolute value of the correlation coefficient between the corresponding two nodes. The larger the weight value, the stronger the correlation between the two nodes. By establishing a graph structure, the graph structure construction unit of this invention organizes the originally independent modal data nodes into a relational graph structure through edge connections. The obtained relational graph structure explicitly expresses the spatiotemporal correlation between different modal data. For example, when the degree of pallor in a patient increases, it is often accompanied by a decrease in blood pressure and an increase in stroke volume variability. Strong edge connections will be established between the nodes corresponding to these features.

[0036] The aggregation unit is built upon a graph convolutional network and is used to process the relational graph structure output by the graph structure construction unit. The graph convolutional network performs convolution operations on the graph structure data, updating the feature representation of the current node by aggregating the feature information of adjacent nodes layer by layer. The aggregation unit of this invention performs multi-layer aggregation processing on the node features in the node set. Each layer of graph convolution operation weights and aggregates the features of adjacent nodes according to the weight values ​​of the edge connections, ensuring that the feature representation of each node not only includes its own original features but also incorporates the feature information of other related nodes. After multi-layer graph convolution operations, the node feature representation already contains cross-modal spatiotemporal interaction information. Finally, the aggregation unit concatenates or pools the final feature representations of all nodes to output a fused feature vector.

[0037] The obtained fusion feature vector is a high-dimensional vector that integrates spatiotemporal correlation information from dialysis machine parameters, physiological signals, hyperspectral features, and neurocognitive data, comprehensively reflecting the patient's overall condition during dialysis. The spatiotemporal fusion engine module of this invention transmits the generated fusion feature vector to the dynamic risk prediction module via a data channel, providing a unified feature input for subsequent risk assessment. This invention's graph neural network-based spatiotemporal fusion method overcomes the limitations of traditional feature fusion methods that rely on simple concatenation or weighted averaging, effectively capturing the complex correlations between different modalities of data and significantly improving the quality of feature representation and the accuracy of risk prediction.

[0038] like Figure 3As shown, the dynamic risk prediction module 300 includes: a coagulation and bleeding collaborative monitoring unit 310, used to process the fused feature vector through a coagulation identification model and a bleeding verification model respectively, and output coagulation risk probability value and bleeding risk probability value; a hypotension early warning unit 320, established based on a spatiotemporal cross-attention prediction model, used to perform fusion processing on the fused feature vector through the spatiotemporal cross-attention prediction model, and output hypotension occurrence probability value; a heart failure decompensation early warning unit 330, used to calculate the heart failure index based on the fused feature vector through a heart failure risk assessment model; and a cognitive impairment assessment unit 340, used to generate a cognitive impairment risk score based on the fused feature vector through a cognitive state assessment model.

[0039] Furthermore, the dynamic risk prediction module of the present invention receives the fusion feature vector output by the spatiotemporal fusion engine module, and constructs a multi-dimensional risk assessment system based on the fusion feature vector, including a coagulation-bleeding collaborative monitoring unit, a hypotension early warning unit, a heart failure decompensation early warning unit, and a cognitive impairment assessment unit. The four units are respectively responsible for predicting the risk of different types of dialysis complications.

[0040] The heart failure decompensation early warning unit receives N-terminal pro-brain natriuretic peptide (NTNP) concentration data and lung ultrasound data from the fused feature vector, and calculates a heart failure index using a heart failure risk assessment model. NTNP is a specific biomarker for heart failure; elevated NTNP concentration indicates increased cardiac workload. Lung ultrasound can identify the comet tail sign that appears in pulmonary edema, and the number of comet tail signs is correlated with the degree of pulmonary effusion. The heart failure risk assessment model of this invention comprehensively analyzes the NNTNP concentration trend and the number of comet tail signs to calculate a heart failure index, which reflects the degree of heart failure decompensation risk.

[0041] The cognitive impairment assessment unit provided by this invention receives EEG power ratio data and speech response data from a fusion feature vector, and generates a cognitive impairment risk score through a cognitive state assessment model. The EEG power ratio data reflects the patient's brain electrical activity patterns, while the speech response data includes indicators such as accuracy, fluency, and reaction time. Specifically, the cognitive state assessment model of this invention comprehensively assesses the patient's cognitive function based on the above indicators, and outputs a cognitive impairment risk score when cognitive function declines.

[0042] The dynamic risk prediction module uses the coagulation risk probability value, bleeding risk probability value, hypotension occurrence probability value, heart failure index, and cognitive impairment risk score output by each unit as prediction results, and transmits them to the intelligent intervention control module through the data interface. The established multi-dimensional risk assessment system can comprehensively monitor various complications that may occur during dialysis, providing rich risk information for clinical decision-making. Compared with traditional single-parameter alarm systems, it has higher sensitivity and stronger predictability.

[0043] The coagulation and bleeding co-monitoring unit 310 includes: a fiber optic sensor for real-time acquisition of light intensity change signals from the dialyzer fiber bundles; a coagulation identification subunit, constructed based on a convolutional neural network, for pattern recognition based on the fused feature vector and the light intensity change signals, extracting fiber bundle blockage features, and outputting a coagulation risk probability value based on the fiber bundle blockage features; wherein, a coagulation risk signal is output when the blockage degree corresponding to the fiber bundle blockage features is greater than or equal to a first preset threshold; a bleeding verification subunit, which processes the fused feature vector and outputs a bleeding risk probability value based on abnormal temperature diffusion rate and hemoglobin absorbance mutation data in the fused feature vector; wherein, a bleeding risk signal is output when the abnormal temperature diffusion rate is greater than or equal to a second preset threshold and the hemoglobin absorbance mutation data is greater than or equal to a third preset threshold; and an anticoagulation optimization subunit, constructed based on a deep Q-network reinforcement learning algorithm, using the coagulation risk probability value and the bleeding risk probability value as state inputs, calculating the Q value corresponding to different heparin doses, and outputting a heparin dose adjustment strategy with the largest Q value.

[0044] Furthermore, the coagulation-bleeding co-monitoring unit is used to receive fiber optic sensing data and image data from the fused feature vector, process them through the coagulation identification model and the bleeding verification model respectively, and output the coagulation risk probability value and the bleeding risk probability value.

[0045] Specifically, the coagulation recognition subunit is built on a convolutional neural network. The model receives light intensity change signals and dialysis machine parameter data from the fused feature vector and performs pattern recognition on the light intensity change signals. The fiber optic sensor can collect light intensity change signals from the dialyzer fiber bundle in real time. When the fiber bundle is blocked, the propagation path of light in the fiber bundle will change, resulting in a specific change pattern in the light intensity signal at the receiving end. The convolutional neural network extracts the blockage features from the light intensity signal through multi-layer convolution and pooling operations, and combines them with blood flow velocity and transmembrane pressure information from the fused feature vector for comprehensive judgment. When the degree of blockage corresponding to the extracted blockage features exceeds a first preset threshold, the coagulation recognition subunit outputs a coagulation risk signal.

[0046] The bleeding verification subunit receives temperature distribution data of the area surrounding the puncture point acquired by a thermal imaging camera and hemoglobin absorbance mutation data acquired by a hyperspectral imaging device from the fused feature vector. The thermal imaging camera can detect local temperature changes caused by bleeding. After acquiring temperature change data, the bleeding verification subunit calculates the abnormal temperature diffusion rate, i.e., the rate at which the abnormal temperature region expands over time. Simultaneously, the bleeding verification subunit analyzes mutations in hemoglobin absorbance at specific wavelengths in the hyperspectral data; sudden changes in hemoglobin concentration are a direct indicator of bleeding. When both the abnormal temperature diffusion rate and the hemoglobin absorbance mutation data simultaneously exceed their respective second and third preset thresholds, the bleeding verification model outputs a bleeding risk signal. This dual verification mechanism of the bleeding verification subunit effectively reduces the false alarm rate and improves the accuracy of bleeding detection.

[0047] The coagulation and bleeding co-monitoring unit also includes an anticoagulation optimization subunit. This subunit is built based on a deep Q-network reinforcement learning algorithm. It takes coagulation risk signals, bleeding risk signals, and platelet count and prothrombin time in the fused feature vector as state inputs. By learning the long-term rewards of taking different heparin doses under different states, it can find the optimal balance between coagulation risk and bleeding risk. Then, the anticoagulation optimization subunit calculates the Q value corresponding to different heparin doses, which represents the expected cumulative reward that can be obtained by taking this dose under the current state. Finally, the heparin dose with the largest Q value is selected as the output to form a heparin dose adjustment strategy.

[0048] The spatiotemporal cross-attention prediction model constructed by the hypotension warning unit 320 includes: a spatial encoder, which adopts a visual Transformer architecture to receive the fused feature vector, extract spatial feature tensors, and quantify the degree of pallor; a temporal encoder, which adopts a temporal convolutional network architecture to receive the fused feature vector, extracting temporal feature tensors of blood pressure fluctuation curves, ultrafiltration rate time-series data, and stroke volume variability trends; a cross-attention fusion layer, which calculates the attention score matrix between the spatial feature tensor and the temporal feature tensor, and performs weighted fusion of the spatial feature tensor and the temporal feature tensor through the attention score matrix to establish spatiotemporal collaborative features; and a probability output layer, which receives the spatiotemporal collaborative features and outputs a hypotension occurrence probability value, triggering a hypotension warning signal when the hypotension occurrence probability value is greater than or equal to 85%.

[0049] Furthermore, the hypotension early warning unit receives facial image features and blood pressure temporal features from the fused feature vector, performs fusion processing through a spatiotemporal cross-attention prediction model, and outputs the probability value of hypotension occurring within a preset future time. The spatiotemporal cross-attention prediction model includes a spatial encoder, a temporal encoder, a cross-attention fusion layer, and a probability output layer. The spatial encoder employs a visual Transformer architecture, extracting patient facial image data from the fused feature vector, capturing the feature relationships between different facial regions through a self-attention mechanism, obtaining a spatial feature tensor, and quantifying the degree of pallor. For example, pallor is an important clinical manifestation of hypotension; quantitative analysis of skin color changes in facial images can provide reliable visual features for hypotension prediction. The temporal encoder employs a temporal convolutional network architecture, extracting blood pressure fluctuation curves, ultrafiltration rate temporal data, stroke volume variability trends, and peripheral vascular resistance index from the fused feature vector, capturing dynamic change patterns in the temporal data through causal convolution, and generating a temporal feature tensor. Subsequently, this invention provides a cross-attention fusion layer to calculate an attention score matrix between spatial and temporal feature tensors. This matrix quantifies the correlation between each element of the spatial feature and each element of the temporal feature. Then, the spatial and temporal feature tensors are weighted and fused using the attention score matrix, and a spatiotemporal co-feature is established by combining heart rate variability data from the fused feature vector. The obtained spatiotemporal co-feature can simultaneously reflect the correlation between changes in the patient's facial images and the temporal evolution of blood pressure, exhibiting stronger predictive power than using image features or temporal features alone. Finally, a probability output layer receives the spatiotemporal co-feature and outputs the probability value of hypotension occurring within the next fifteen minutes through a fully connected layer and an activation function. When the probability value of hypotension exceeds 85%, a hypotension warning signal is triggered.

[0050] The intelligent intervention control module 400 establishes a four-level risk level, including: Level 1 risk, which corresponds to the distributed piezoelectric thin film array sensor detecting a tube dislodgement event or the coagulation and bleeding collaborative monitoring unit outputting a bleeding risk signal; Level 2 risk, which corresponds to the hypotension warning unit 320 outputting a hypotension warning signal; Level 3 risk, which corresponds to the heart failure index output by the heart failure decompensation warning unit being greater than or equal to a fourth preset threshold; and Level 4 risk, which corresponds to the cognitive impairment assessment unit outputting an impairment risk score greater than or equal to a fifth preset threshold, or the coagulation and bleeding collaborative monitoring unit outputting a coagulation risk signal.

[0051] The intelligent intervention control module 400 establishes a four-level intervention strategy, including: when the risk level is level one, executing the level one intervention strategy, which involves clamping the dialysis tubing and activating the positioning and calling system; when the risk level is level two, executing the level two intervention strategy, which involves reducing the ultrafiltration rate to zero and issuing a saline demand signal; when the risk level is level three, executing the level three intervention strategy, which involves switching the system to continuous venous-venous hemofiltration mode; and when the risk level is level four, executing the level four intervention strategy, which involves issuing a heparin dosage adjustment signal and a neuroprotective drug demand signal.

[0052] Furthermore, the intelligent intervention control module receives the prediction results output by the dynamic risk prediction module, assesses the risk level based on the probability values ​​and indices of various risks, and generates corresponding intervention strategies through reinforcement learning algorithms. The intelligent intervention control module of this invention establishes a four-level hierarchical intervention protocol and adopts intervention measures of different intensities and types for different risk levels.

[0053] Specifically, a Level 1 risk corresponds to the distributed piezoelectric thin-film array sensor detecting a tubing dislodgement event or the bleeding verification module outputting a major bleeding risk signal. Tubing dislodgement and major bleeding are emergency situations that endanger the patient's life and require immediate measures to prevent excessive blood loss. When a Level 1 risk is detected, the intelligent intervention control module immediately sends a control command to the dialysis tubing clamping device. Upon receiving the command, the dialysis tubing clamping device clamps the dialysis tubing within milliseconds, blocking blood flow and preventing further blood loss. Simultaneously, the module activates the location call system, which locates the dialysis bed where the abnormality occurred based on the sensor signals and notifies medical staff via audible and visual alarms and mobile terminal push notifications. Because Level 1 risk involves the automatic control of dialysis tubing closure, to ensure operational safety and compliance with medical regulations, the system requires nurses to confirm their identity via fingerprint authentication before finally performing the clamping action.

[0054] Level 2 risk corresponds to a probability of hypotension exceeding 85%. When the hypotension warning unit predicts that the probability of a patient experiencing hypotension exceeds 85% within the next 15 minutes, it indicates that the patient is about to experience hypotension and requires immediate preventative measures. Upon detecting Level 2 risk, the intelligent intervention control module of this invention automatically reduces the ultrafiltration rate of the dialysis equipment to zero, stopping fluid ultrafiltration to prevent further decrease in blood volume and subsequent hypotension. Simultaneously, it prompts the need to administer 50 ml of normal saline via a drug infusion pump to quickly replenish blood volume and raise blood pressure. While implementing intervention measures, the intelligent intervention control module projects a three-dimensional risk heat map generated based on fused feature vectors onto medical staff through augmented reality glasses. The three-dimensional risk heat map visually displays the degree of abnormality and spatial distribution of various physiological parameters of the patient. Medical staff wearing augmented reality glasses can intuitively see the patient's risk status, assisting them in quickly assessing the situation and making further clinical decisions.

[0055] Level 3 risk corresponds to a heart failure index exceeding 0.7, indicating a higher risk of heart failure decompensation and necessitating adjustment of the dialysis mode and initiation of interdisciplinary consultation. When Level 3 risk is triggered, the intelligent intervention control module of this invention switches the system to continuous venous-venous hemofiltration mode. This mode offers a gentler fluid clearance rate compared to conventional hemodialysis, reducing cardiac load and making it suitable for patients with heart failure. Simultaneously, a cardiology consultation process is triggered, sending a consultation request to a cardiologist via the hospital information system. The cardiologist assesses the patient's cardiac function and develops a further treatment plan. Because switching dialysis modes involves adjusting important treatment parameters, voice verification is required before execution. Medical staff confirm the switching decision via voice commands, increasing operational safety.

[0056] Level 4 risk corresponds to a coagulation risk probability value exceeding 0.8 or a pupillary light reflex delay exceeding 2 seconds detected by the cognitive impairment assessment unit. An excessively high coagulation risk probability value suggests impending severe coagulation within the dialyzer, while a pupillary light reflex delay exceeding 2 seconds indicates potential severe neurological dysfunction. Both situations require timely pharmacological intervention. When a Level 4 risk is triggered, the intelligent intervention control module of this invention adjusts the heparin dosage according to the heparin dosage adjustment strategy output by the anticoagulation optimization module, precisely controlling the anticoagulation intensity to prevent coagulation events. Simultaneously, it prompts the user to prepare neuroprotective medications for potential neuroprotective treatment. Because medication dosage adjustment is directly related to patient safety, the system requires dual-factor authentication (fingerprint and voice) before execution. Medical staff must simultaneously confirm via fingerprint recognition and voice to ensure that the intervention decision has undergone sufficient human review.

[0057] The intelligent intervention control module of this invention realizes closed-loop control from risk identification to automatic intervention through the above-mentioned hierarchical intervention protocol. The intelligent intervention control module uses reinforcement learning algorithm to continuously optimize the intervention strategy. By recording the effect of each intervention and the changes in the patient's state, it learns the optimal intervention plan under different risk conditions, so that the intervention strategy can be continuously improved as the system's usage experience accumulates.

[0058] The intelligent intervention and control method of this invention changes the traditional monitoring system's passive alarm and manual decision-making mode, realizing proactive early warning and automatic intervention. It significantly shortens the time delay from risk discovery to taking measures, improves the timeliness and accuracy of intervention, effectively reduces the incidence of dialysis complications, and enhances the safety of dialysis treatment and the efficiency of medical and nursing work.

[0059] This also includes: Disease-customized monitoring module; For heart failure patients, the disease-customized monitoring module acquires the concentration trend data of N-terminal pro-brain natriuretic peptide (BNP), identifies the number of comet tail signs through lung ultrasound examination, and inputs the concentration trend data and the number of comet tail signs into the heart failure decompensation early warning unit of the dynamic risk prediction module for fusion evaluation; for diabetic patients, a non-invasive blood glucose monitoring device is configured to measure blood glucose levels in real time.

[0060] This invention also includes a disease-customized monitoring module, which provides personalized monitoring plans for dialysis patients with different disease subtypes. This overcomes the limitations of traditional monitoring systems that use a homogeneous monitoring strategy for all patients. The disease-customized monitoring module works in conjunction with the multi-source data acquisition module and the dynamic risk prediction module to obtain disease-specific biomarker data and clinical examination results, providing more accurate input information for risk assessment.

[0061] For heart failure patients, the disease-customized monitoring module is specifically configured with heart failure-related monitoring equipment and data acquisition interfaces. Specifically, the module acquires N-terminal pro-brain natriuretic peptide (NTNP) concentration trend data through a biochemical analyzer. NTNP is a neurohormonal precursor secreted by cardiomyocytes when ventricular load increases, and its plasma concentration is closely related to the severity of heart failure. This invention's disease-customized monitoring module not only collects single NTNP concentration values ​​but also tracks its changes over dialysis time, calculating the rate of increase or decrease in concentration, as a rapid increase in concentration often indicates an increased risk of acute decompensation of heart failure. Simultaneously, the module identifies the number of comet tail artifacts through lung ultrasound examination. The comet tail artifact is a characteristic finding in lung ultrasound examinations; when fluid accumulates in the pulmonary interstitium, ultrasound waves create multiple reflections between the alveoli and pleura, appearing as a comet tail-like streak on the ultrasound image. The number of comet tail artifacts is positively correlated with the severity of pulmonary edema and is an important imaging indicator for assessing volume overload in heart failure patients. The disease-customized monitoring module inputs the acquired N-terminal pro-brain natriuretic peptide concentration trend data and comet tail sign count into the heart failure decompensation early warning unit of the dynamic risk prediction module via the data transmission interface for fusion evaluation. Upon receiving this disease-specific data, the heart failure decompensation early warning unit combines it with parameters such as blood pressure change trends, stroke volume variability, and ultrafiltration rate from the fusion feature vector provided by the spatiotemporal fusion engine module. Through a comprehensive analysis using a heart failure risk assessment model, it can simultaneously consider the complex correlations between biomarker levels, imaging findings, and hemodynamic parameters to calculate a heart failure index. When the heart failure index exceeds 0.7, the system determines that the patient has a high risk of heart failure decompensation and triggers a corresponding early warning signal, which is transmitted to the intelligent intervention control module. Upon receiving the early warning signal, the intelligent intervention control module, according to the three-level risk intervention protocol, controls the dialysis system to switch to continuous venous-venous hemofiltration mode. This mode uses a slower fluid clearance rate and lower ultrafiltration pressure to avoid the impact of rapid changes in blood volume on the heart. Simultaneously, it triggers a cardiology consultation process, where a cardiologist further evaluates the patient's cardiac function and adjusts the heart failure treatment plan.

[0062] For diabetic patients, the disease-customized monitoring module is equipped with a non-invasive blood glucose monitoring device to measure blood glucose levels in real time. This device uses near-infrared spectroscopy or electromagnetic wave technology, analyzing the absorption characteristics of glucose molecules in skin tissue to specific wavelengths of light or electromagnetic waves to achieve non-invasive, continuous monitoring of blood glucose levels. The non-invasive blood glucose monitoring device transmits the real-time measured blood glucose data to the disease-customized monitoring module via a wireless communication module. The disease-customized monitoring module performs time-series analysis on the blood glucose data to identify fluctuation patterns and trends. During dialysis, patients' blood glucose levels are prone to fluctuation. On the one hand, the glucose concentration in the dialysate affects the patient's blood glucose level; on the other hand, changes in volume and hormone levels caused by dialysis also affect glucose metabolism. Therefore, the disease-customized monitoring module subsequently transmits blood glucose fluctuation trend data to the intelligent intervention control module. The intelligent intervention control module uses a deep deterministic gradient algorithm to dynamically adjust the glucose concentration ratio in the dialysate based on real-time changes in blood glucose levels. The deep deterministic strategy gradient algorithm uses the current blood glucose level, rate of change of blood glucose, and dialysate glucose concentration as state inputs, and the adjustment amount of dialysate glucose concentration as action output. By learning the optimal strategy for adjusting dialysate concentration under different blood glucose states, it maintains the patient's blood glucose level within the target range, avoiding hyperglycemia or hypoglycemia. Simultaneously, the disease-customized monitoring module is equipped with a plantar microcirculation imaging device to monitor the microcirculation perfusion status of the diabetic patient's foot. Diabetic patients are prone to peripheral vascular disease and neuropathy, leading to insufficient blood supply and decreased sensation in the foot, increasing the risk of diabetic foot ulcers. When the plantar microcirculation imaging device detects a drop in microcirculation density corresponding to the microcirculation perfusion status below a preset second threshold, it indicates a significant reduction in blood supply to the patient's foot, posing a risk of diabetic foot. The disease-customized monitoring module transmits the abnormal foot microcirculation signal to the intelligent intervention control module, triggering a diabetic foot risk warning, prompting medical staff to conduct a detailed examination of the patient's foot and take preventative measures.

[0063] The disease-customized monitoring module of this invention achieves personalized risk monitoring by designing specialized monitoring schemes and data acquisition equipment for different disease subtypes. This enables the system to more accurately identify the special risks of patients with specific diseases during dialysis, providing a basis for precise intervention and significantly improving the quality and safety of monitoring patients with diseases such as heart failure and diabetes.

[0064] The hemodialysis multimodal risk closed-loop control system utilizes a collaborative architecture of edge computing and cloud-based federated learning.

[0065] The multimodal risk closed-loop control system for hemodialysis achieves an organic combination of real-time response and cross-center knowledge sharing through a collaborative architecture of edge computing and cloud federated learning, overcoming the dual limitations of large response latency in traditional cloud computing architecture and the inability to share knowledge in traditional local computing architecture.

[0066] The cloud-based federated learning layer aggregates data and model updates from multiple dialysis centers, enabling cross-center knowledge transfer and model optimization through the federated learning framework. This cross-center knowledge transfer mechanism allows even small dialysis centers with fewer dialysis cases to leverage data and knowledge from other centers to improve the predictive performance of their local models. This achieves 90% predictive accuracy with only five samples per center, significantly reducing the sample size requirements for model training. The collaborative architecture of edge computing and cloud-based federated learning also supports continuous model learning and updates. As clinical practice accumulates, new dialysis data and intervention feedback are constantly generated. The system can use this new data for incremental model training, allowing the model's predictive ability to continuously improve over time. The cloud periodically distributes updated global model parameters to the edge, and the lightweight model at the edge is updated accordingly, ensuring that the edge always uses the latest knowledge for risk assessment.

[0067] Furthermore, the entire collaborative architecture provided by this invention complies with the General Data Protection Regulation (GDPR) and health insurance circulation and privacy protection standards, achieving a balance between data privacy protection and knowledge sharing. The edge computing and cloud-based federated learning collaborative architecture not only improves the system's real-time response capabilities and risk prediction accuracy but also promotes the flow and accumulation of medical knowledge among different institutions, driving the overall advancement of hemodialysis safety monitoring technology and possessing significant clinical application value and social significance.

[0068] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0069] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multimodal risk closed-loop control system for hemodialysis, characterized in that, include: A multi-source data acquisition module is used to acquire heterogeneous data in real time during the dialysis process; The spatiotemporal fusion engine module uses a graph neural network architecture to perform spatiotemporal alignment and feature extraction on the heterogeneous data across modalities, generating a fused feature vector. The dynamic risk prediction module constructs a multi-dimensional risk assessment system based on the fused feature vector, performs probability predictions for various risk types, and outputs the prediction results. The intelligent intervention control module assesses the risk level based on the prediction results output by the dynamic risk prediction module and generates intervention strategies through reinforcement learning algorithms.

2. The multimodal risk closed-loop control system for hemodialysis according to claim 1, characterized in that, The multi-source data acquisition module includes: The dialysis machine parameter acquisition unit parses the dialysis machine parameters output by the dialysis equipment in real time through a standardized protocol interface to obtain dialysis machine parameter data; A physiological signal acquisition unit is equipped with a millimeter-wave radar sensor to monitor the patient's heartbeat displacement and calculate the stroke volume variability and peripheral vascular resistance index based on the heartbeat displacement to obtain physiological signals. The hyperspectral imaging acquisition unit receives imaging data from the hyperspectral imaging device and scans the patient's puncture point and face based on the imaging data to capture the hemoglobin absorbance change characteristics. It also receives temperature distribution data around the puncture point acquired by the thermal imaging camera to obtain hyperspectral feature data. The neurocognitive data acquisition unit receives the patient's electroencephalogram (EEG) signals acquired by the dry electrode EEG probe array, extracts the alpha and beta waves from the EEG signals, and records the patient's responses to cognitive questions at fixed intervals and the response time to obtain neurocognitive data. A distributed piezoelectric thin film array sensor is deployed at dialysis tubing nodes to detect tubing connectivity.

3. The multimodal risk closed-loop control system for hemodialysis according to claim 1, characterized in that, The spatiotemporal fusion engine module includes: Node construction unit: used to construct the dialysis machine parameters into multiple time-series nodes according to time windows, construct the physiological signals into multiple physiological nodes, construct the hyperspectral data into multiple feature nodes, construct the neurocognitive data into multiple cognitive nodes, and establish a node set including all time-series nodes, all physiological nodes, all feature nodes, and all cognitive nodes; Graph structure construction unit: Calculates the Pearson correlation coefficient between any two nodes in the node set. When the absolute value of the calculated correlation coefficient is greater than a preset threshold, an edge connection is established between the corresponding two nodes to obtain an association graph structure; wherein, the weight value of the edge connection is equal to the absolute value of the correlation coefficient between the corresponding two nodes. An aggregation unit, which is built based on a graph convolutional network, is used to process the association graph structure to aggregate the node features in the node set and output a fused feature vector.

4. The multimodal risk closed-loop control system for hemodialysis according to claim 1, characterized in that, The dynamic risk prediction module includes: The coagulation and bleeding co-monitoring unit is used to process the fused feature vector through the coagulation identification model and the bleeding verification model respectively, and output the coagulation risk probability value and the bleeding risk probability value. The low blood pressure early warning unit is based on the spatiotemporal cross-attention prediction model. It is used to perform fusion processing on the fused feature vector through the spatiotemporal cross-attention prediction model and output the probability value of low blood pressure occurrence. The heart failure decompensation early warning unit is used to calculate the heart failure index based on the fused feature vector through the heart failure risk assessment model. The cognitive impairment assessment unit is used to generate a cognitive impairment risk score based on the fused feature vector through a cognitive state assessment model.

5. A multimodal risk closed-loop control system for hemodialysis according to claim 4, characterized in that, The coagulation and bleeding co-monitoring unit includes: Fiber optic sensors are used to acquire real-time light intensity changes in the dialyzer fiber bundles. A coagulation recognition subunit, which is constructed based on a convolutional neural network, is used to perform pattern recognition based on the fused feature vector and the light intensity change signal, extract fiber bundle blockage features, and output a coagulation risk probability value based on the fiber bundle blockage features; wherein, when the degree of blockage corresponding to the fiber bundle blockage features is greater than or equal to a first preset threshold, a coagulation risk signal is output. The bleeding verification subunit processes the fused feature vector and outputs a bleeding risk probability value based on the abnormal temperature diffusion rate and hemoglobin absorbance mutation data in the fused feature vector; wherein a bleeding risk signal is output when the abnormal temperature diffusion rate is greater than or equal to a second preset threshold and the hemoglobin absorbance mutation data is greater than or equal to a third preset threshold. The anticoagulation optimization subunit is constructed based on a deep Q-network reinforcement learning algorithm. It uses the coagulation risk probability value and the bleeding risk probability value as state inputs, calculates the Q value corresponding to different heparin doses, and outputs the heparin dose adjustment strategy with the largest Q value.

6. The multimodal risk closed-loop control system for hemodialysis according to claim 4, characterized in that, The spatiotemporal cross-attention prediction model constructed by the hypotension early warning unit includes: A spatial encoder, which employs a visual Transformer architecture, is used to receive the fused feature vector, extract the spatial feature tensor, and quantize the degree of paleness. A time-series encoder, which employs a time-series convolutional network architecture, is used to receive time-series feature tensors from the fused feature vector, which extract the blood pressure fluctuation curve, ultrafiltration rate time-series data, and stroke volume variability trend. A cross-attention fusion layer is used to calculate the attention score matrix between the spatial feature tensor and the temporal feature tensor, and to perform weighted fusion of the spatial feature tensor and the temporal feature tensor through the attention score matrix to establish spatiotemporal collaborative features; The probability output layer receives the spatiotemporal collaborative features and outputs a low blood pressure occurrence probability value. When the low blood pressure occurrence probability value is greater than or equal to 85%, a low blood pressure warning signal is triggered.

7. The multimodal risk closed-loop control system for hemodialysis according to claim 1, characterized in that, The intelligent intervention control module establishes a four-level risk level, including: Level 1 risk, which corresponds to the distributed piezoelectric thin film array sensor detecting a tubing dislodgement event or the coagulation and bleeding collaborative monitoring unit outputting a bleeding risk signal; Level 2 risk, wherein the low blood pressure early warning unit outputs a low blood pressure early warning signal; Level 3 risk, the Level 3 risk corresponds to the heart failure index output by the heart failure decompensation early warning unit being greater than or equal to the fourth preset threshold; Level 4 risk, where the level 4 risk corresponds to the cognitive impairment assessment unit outputting an impairment risk score greater than or equal to a fifth preset threshold, or the coagulation and bleeding co-monitoring unit outputting a coagulation risk signal.

8. A multimodal risk closed-loop control system for hemodialysis according to claim 7, characterized in that, The intelligent intervention control module establishes a four-level intervention strategy, including: When the risk level is Level 1, a Level 1 intervention strategy is implemented, which is to clamp the dialysis tubing and activate the location call system. When the risk level is level 2, a level 2 intervention strategy is implemented, which involves reducing the ultrafiltration rate to zero and issuing a saline demand signal. When the risk level is level three, a level three intervention strategy is implemented, which is to switch the system to continuous venous-venous hemofiltration mode. When the risk level is level four, a level four intervention strategy is implemented, which involves issuing signals to adjust the heparin dosage requirement and the neuroprotective drug requirement.

9. A multimodal risk closed-loop control system for hemodialysis according to claim 1, characterized in that, Also includes: Disease-customized monitoring module; For heart failure patients, the disease-customized monitoring module acquires the concentration trend data of N-terminal pro-brain natriuretic peptide, identifies the number of comet tail signs through lung ultrasound examination, and inputs the concentration trend data and the number of comet tail signs into the heart failure decompensation early warning unit of the dynamic risk prediction module for fusion evaluation. For diabetic patients, a non-invasive blood glucose monitoring device is configured to measure blood glucose levels in real time.

10. A multimodal risk closed-loop control system for hemodialysis according to claim 1, characterized in that, The hemodialysis multimodal risk closed-loop control system utilizes a collaborative architecture of edge computing and cloud-based federated learning.

Images