Intelligent monitoring method and system for blood purification

Abstract

The invention discloses an intelligent monitoring method and system for blood purification, and the method comprises the steps: constructing a monitoring subsystem comprising two sets of independent sensor networks, collecting and processing key physiological parameters in a blood purification process in parallel, and obtaining a parameter data flow subjected to preliminary verification; a blood purification system execution mechanism is controlled through a self-adaptive task space nonsingular terminal super-distortion sliding mode control algorithm, and a control instruction is obtained; redundancy detection and cross validation of parameter abnormity are realized through a dual-channel comparative analysis algorithm, and an abnormity detection result is obtained; analyzing data features of the patient in the treatment process through a deep learning model, predicting potential risks, and obtaining intelligent intervention suggestions; a data management and traceability system is constructed through a distributed database technology and a block chain verification mechanism. According to the invention, the accuracy and stability of blood purification monitoring are greatly improved; redundant anomaly detection and an intelligent aid decision-making system are adopted, so that the risk prediction and anomaly coping capability is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical equipment monitoring, and in particular to a blood purification intelligent monitoring method and system. Background Art

[0002] Blood purification technology is a key technology in the modern medical field for treating diseases such as renal failure and severe poisoning. Its core principle is to draw the patient's blood out of the body through extracorporeal circulation, remove toxins and excess water in the blood through a specific purification device, and then return it to the body to achieve a steady-state balance of the internal environment.

[0003] Currently, common blood purification systems in clinical practice primarily include traditional hemodialysis systems and continuous renal replacement therapy systems. These systems typically feature basic monitoring modules that monitor parameters such as blood flow, transmembrane pressure, and dialysate temperature. These systems also have simple alarm functions that alert medical staff when parameters exceed preset ranges.

[0004] Existing advanced blood purification monitoring systems utilize a single sensor network architecture. Integrated sensors collect key parameters during the blood purification process, such as blood pressure, oxygen saturation, and fluid balance, and control algorithms regulate the purification process. The system automatically adjusts parameters such as dialysate composition and ultrafiltration rate based on pre-set treatment plans, triggering alarms when abnormalities are detected.

[0005] However, such systems have obvious technical defects in actual applications: first, the risk of single-point failure of sensors is high. Once errors or failures occur in key sensors, the system may make incorrect judgments; second, there is a lack of redundant verification mechanisms, which cannot effectively identify and respond to abnormal sensor data; third, the control algorithm lacks adaptive capabilities and has difficulty responding to dynamic changes in patient status; fourth, the emergency response mechanism is imperfect, and the ability to automatically intervene in abnormal situations is limited. It is still highly dependent on manual operation and poses safety risks.

[0006] Therefore, there is an urgent need to design a blood purification monitoring system with a redundant verification mechanism, strong adaptive control capabilities, and intelligent abnormality handling functions to improve the safety and effectiveness of blood purification treatment. Summary of the Invention

[0007] The purpose of the present invention is to provide an intelligent blood purification monitoring method and system, aiming to solve the technical problems of high risk of single-point failure of sensors, lack of redundant verification mechanism, insufficient adaptability of control algorithm and imperfect emergency response mechanism in existing blood purification monitoring systems.

[0008] To achieve the above objectives, the present invention provides a blood purification intelligent monitoring method, comprising:

[0009] Obtain the core parameter requirements of the blood purification equipment and build a monitoring subsystem consisting of two independent sensor networks. By parallelly collecting and processing key physiological parameters during the blood purification process, a preliminarily verified parameter data stream is obtained.

[0010] Obtaining the parameter data stream that has been preliminarily verified, controlling the actuator of the blood purification system through an adaptive task space non-singular terminal super-twisted sliding mode control algorithm to obtain a control instruction;

[0011] Obtaining the preliminarily verified parameter data stream and the control instruction, and implementing redundant detection and cross-validation of parameter anomalies through a dual-channel comparative analysis algorithm to obtain an anomaly detection result;

[0012] Obtaining the preliminarily verified parameter data stream, the control instructions, and the abnormality detection results, analyzing the data characteristics of the patient's treatment process through the trained deep learning model, predicting potential risks, and obtaining intelligent intervention recommendations;

[0013] The parameter data stream that has undergone preliminary verification, the control instructions, the abnormality detection results and the intelligent intervention suggestions are obtained, and a data management and traceability system is constructed through distributed database technology and blockchain verification mechanism to achieve secure storage and retrospective analysis of data throughout the entire treatment process.

[0014] Preferably, the construction comprises a monitoring subsystem comprising two independent sensor networks, which acquires and processes key physiological parameters during the blood purification process in parallel to obtain a preliminarily verified parameter data stream, including:

[0015] Based on the monitoring requirements of key blood purification parameters, a main sensor network containing the key physiological parameters is constructed through sensor deployment to obtain the main sensor data stream;

[0016] Constructing an auxiliary sensor network for collecting the key physiological parameters, wherein the auxiliary sensor network is constructed using a different principle from the main sensor network and generates an auxiliary sensor data stream through an independent power supply and signal processing unit;

[0017] The main sensor data stream and the auxiliary sensor data stream are obtained, and data consistency verification is performed through a difference comparison algorithm and rationality verification to obtain the parameter data stream that has undergone preliminary verification.

[0018] Preferably, the step of obtaining a preliminarily verified parameter data stream by concurrently collecting and processing key physiological parameters during the blood purification process includes:

[0019] Acquiring the key physiological parameters collected by the primary sensor network and the auxiliary sensor network, and simultaneously collecting the key physiological parameters through two sets of independent sensors to generate dual-channel raw data; wherein the key physiological parameters include blood pressure, blood flow, dialysate temperature, conductivity, and ultrafiltration rate data;

[0020] Based on the dual-channel raw data, removing noise and interference in the sensor signal through digital filtering and noise reduction processing to generate pre-processed dual-channel raw data;

[0021] Based on the pre-processed dual-channel raw data, the consistency of the two sets of sensor data is compared by a statistical test method, and a data confidence index is calculated to obtain the parameter data stream that has been preliminarily verified and is marked with confidence.

[0022] Preferably, the actuator of the blood purification system is controlled by an adaptive task space non-singular terminal super-distortion sliding mode control algorithm to obtain a control instruction, including:

[0023] Obtaining the parameter data stream that has been preliminarily verified, constructing a task space model including key parameters of the blood purification process through a multi-parameter state space equation, and obtaining a dynamic task space model;

[0024] Based on the dynamic task space model, singular states that may cause control instability are identified through matrix condition number analysis and eigenvalue decomposition, and a preprocessed non-singular control space mapping is obtained;

[0025] Based on the preprocessed non-singular control space mapping, a sliding mode structure is constructed and a second-order derivative term is introduced to design a control law with fast convergence and strong robustness to obtain the control instruction.

[0026] Preferably, the control instruction is obtained by constructing a sliding mode structure and introducing a second-order derivative term to design a control law with fast convergence and strong robustness, including:

[0027] Based on the preprocessed non-singular control space mapping, a sliding surface is constructed by defining a deviation function between the system state and the desired state to generate an initial control law;

[0028] Based on the initial control law, a super-twisted controller is designed by introducing high-order derivative terms with fractional exponents to eliminate the chattering phenomenon of the control signal and generate a smooth control signal;

[0029] Based on the smooth control signal and system operation status feedback, parameters are optimized and adjusted online through Lyapunov stability analysis to obtain the control instruction.

[0030] Preferably, redundant detection and cross-validation of parameter anomalies are achieved through a dual-channel comparative analysis algorithm to obtain anomaly detection results, including:

[0031] Based on clinical experience and historical case analysis data sets, through data mining and expert knowledge extraction, a parameter abnormality pattern library that may occur during blood purification is established, and a structured abnormal pattern feature set is obtained;

[0032] Based on the preliminarily verified parameter data stream, the control instruction, and the structured abnormal pattern feature set, a real-time parameter deviation analysis is performed using a multi-dimensional anomaly detection algorithm to obtain an abnormal event candidate set;

[0033] Based on the abnormal event candidate set, the abnormal event is reconfirmed through cross comparison and temporal continuity analysis of different sensor network data to obtain the abnormal event detection result.

[0034] Preferably, the redundant detection and cross-validation of parameter anomalies are achieved by using a dual-channel comparative analysis algorithm to obtain anomaly detection results, including:

[0035] Obtaining the preliminarily verified parameter data stream and the control instruction, performing sensor redundancy verification by calculating data consistency indexes of the primary sensor network and the auxiliary sensor network, and generating a sensor state assessment;

[0036] Obtaining historical parameter change patterns, combining them with the sensor status assessment, and performing temporal continuity analysis by calculating the autocorrelation function and entropy value of the parameter changes to distinguish between persistent anomalies and occasional interference;

[0037] Based on the temporal continuity analysis results and clinical knowledge rules, the severity of the abnormal event is determined by a risk level assessment algorithm to obtain the abnormality detection result;

[0038] Based on the abnormality detection result, the corresponding level of response measures are activated according to the preset emergency processing process to obtain the emergency processing instructions.

[0039] Preferably, the deep learning model is used to analyze data features during the patient's treatment process, predict potential risks, and obtain intelligent intervention recommendations, including:

[0040] Based on the preliminarily verified parameter data stream, the control instruction, and the abnormality detection result, key features of the patient's treatment process are extracted through a time-frequency domain feature extraction algorithm and physiological signal processing technology to obtain a standardized feature vector;

[0041] Obtaining historical treatment results, combining them with the standardized feature vectors, and constructing and training a deep neural network including a temporal feature processing module and a static feature processing module through a hybrid network architecture design and transfer learning method to obtain a trained deep learning model;

[0042] The current treatment status data is obtained, combined with the trained deep learning model, and through forward reasoning and sensitivity analysis, possible risk conditions in the future are predicted to obtain the intelligent intervention recommendation.

[0043] Preferably, the deep learning model is used to analyze data features during the patient's treatment process, predict potential risks, and obtain intelligent intervention recommendations, including:

[0044] Obtain the current treatment status features and continuously extract features from the most recent observation window using sliding window technology to generate a real-time feature set;

[0045] Based on the real-time feature set, a trained deep learning model is used to predict the changing trends of key physiological parameters in the next 10-30 minutes and generate a multidimensional risk score.

[0046] Based on the multidimensional risk score, a complete causal reasoning chain is constructed through causal chain tracing and evidence level assessment to generate the intelligent intervention suggestion including parameter adjustment suggestion, intervention timing suggestion and graded intervention plan.

[0047] The present invention also provides a blood purification intelligent monitoring system, comprising:

[0048] The dual-sensor monitoring module is used to obtain the core parameter requirements of the blood purification equipment. It builds a monitoring subsystem consisting of two independent sensor networks. By parallelly collecting and processing key physiological parameters during the blood purification process, it obtains a preliminarily verified parameter data stream;

[0049] A controller module is used to obtain the parameter data stream that has been preliminarily verified, and control the actuator of the blood purification system through an adaptive task space non-singular terminal super-distortion sliding mode control algorithm to obtain a control instruction;

[0050] An anomaly detection module is used to obtain the parameter data stream that has been preliminarily verified and the control instruction, and to implement redundant detection and cross-validation of parameter anomalies through a dual-channel comparative analysis algorithm to obtain an anomaly detection result;

[0051] An intelligent diagnosis module is used to obtain the preliminarily verified parameter data stream, the control instructions, and the abnormality detection results, analyze the data characteristics of the patient's treatment process through the trained deep learning model, predict potential risks, and obtain intelligent intervention recommendations;

[0052] The data management module is used to obtain the parameter data stream that has been preliminarily verified, the control instructions, the abnormality detection results and the intelligent intervention suggestions, and to build a data management and traceability system through distributed database technology and blockchain verification mechanism to achieve secure storage and retrospective analysis of data throughout the treatment process.

[0053] The beneficial effects of the present invention are:

[0054] 1. This invention effectively solves the problem of system misjudgment caused by single-point sensor failure by building two independent sensor networks and performing data cross-validation, greatly improving the reliability of system monitoring data;

[0055] 2. The present invention adopts an adaptive task space non-singular terminal super-distorted sliding mode control algorithm, which effectively overcomes the singular state in the control process and improves the robustness and anti-interference ability of the control system;

[0056] 3. This invention has designed a multi-level anomaly detection and graded response mechanism that can accurately identify abnormal events and automatically activate emergency response measures based on risk levels, reducing false alarm rates while improving emergency response efficiency.

[0057] 4. This invention uses a deep learning model to analyze treatment data, enabling forward-looking predictions of potential risks, providing intelligent decision-making support for medical staff and improving the safety and effectiveness of treatment.

[0058] 5. This invention uses distributed database and blockchain technology to achieve secure storage and integrity verification of data throughout the entire treatment process, supports panoramic reproduction and audit analysis of the treatment process, and provides strong support for clinical quality management. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0060] Figure 1 Schematic diagram of the process of the intelligent blood purification monitoring method provided by the present invention;

[0061] Figure 2 This is a schematic diagram of the overall architecture of the blood purification intelligent monitoring system provided by an embodiment of the present invention;

[0062] Figure 3 is a schematic structural diagram of a dual sensing monitoring subsystem provided by an embodiment of the present invention;

[0063] Figure 41 is a flow chart of a dual exception determination and redundant processing mechanism provided by an embodiment of the present invention;

[0064] Figure 5 It is a flowchart of artificial intelligence-assisted diagnosis and intervention provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0065] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0066] Figure 1 The figure shows a flow chart of the blood purification intelligent monitoring method provided by the present invention. Figure 1 As shown, the method includes the following steps:

[0067] Step S1: Obtain the core parameter requirements of the blood purification equipment and build a monitoring subsystem consisting of two independent sensor networks. By parallelly collecting and processing key physiological parameters in the blood purification process, a preliminarily verified parameter data stream is obtained.

[0068] Specifically, the core parameter requirements of the blood purification equipment are first determined, and the key physiological and equipment parameters that need to be monitored are identified. Based on these requirements, a monitoring subsystem consisting of two independent sensor networks is constructed. These two sensor networks utilize different technical principles, with independent power supply and signal processing units, to concurrently collect and process the same set of parameters, thereby avoiding the risks associated with single points of failure. The dual sensor networks simultaneously collect key physiological parameters such as blood pressure, blood flow, dialysate temperature, conductivity, and ultrafiltration rate, generating dual-channel raw data. After digital filtering and noise reduction, this raw data is then compared using statistical tests to verify the consistency of the two sensor data sets. The confidence index for each parameter is calculated, ultimately generating a validation parameter data stream with confidence markers. For sensors with low confidence levels, automatic status assessment and online calibration are performed to ensure data reliability. This dual verification mechanism significantly improves the accuracy and stability of the system's monitoring data, laying a solid foundation for subsequent control and anomaly detection.

[0069] Step S2: Obtain the parameter data stream that has been preliminarily verified, and control the actuator of the blood purification system through the adaptive task space non-singular terminal super-warp sliding mode control algorithm to obtain a control instruction;

[0070] Specifically, the system receives the preliminarily verified parameter data stream from step S1 and precisely controls the actuators of the blood purification system using an adaptive task-space non-singular terminal super-warped sliding mode control algorithm. This algorithm first uses multi-parameter state-space equations to construct a task-space model containing key parameters of the blood purification process, generating a dynamic task-space model. This model not only describes the system's current state but also predicts the system's dynamic response under given control inputs. Matrix condition number analysis and eigenvalue decomposition are then used to identify singular states that could lead to control instability. Preprocessing is performed using techniques such as singular value damping and task-space remapping to obtain a non-singular control space mapping. Based on this, the system defines a deviation function between the system state and the desired state, constructs a sliding mode surface, and generates an initial control law. By introducing high-order derivatives with fractional exponents, a super-warped controller is designed to effectively eliminate chattering in traditional sliding mode control and generate a smooth control signal. Finally, based on the smooth control signal and operational state feedback, Lyapunov stability analysis is used to perform online parameter optimization and output the final control command. This adaptive control mechanism enables the system to automatically adjust control parameters based on individual patient differences and state changes during treatment, achieving stable and precise control.

[0071] Step S3: Obtain the parameter data stream that has undergone preliminary verification and the control instruction, implement redundant detection and cross-validation of parameter anomalies through a dual-channel comparative analysis algorithm, and obtain an anomaly detection result;

[0072] Specifically, the system obtains preliminarily verified parameter data streams and control instructions, and implements redundant detection and cross-validation of parameter anomalies using a dual-channel comparative analysis algorithm. This process first establishes a library of parameter anomaly patterns that may occur during the blood purification process through data mining and expert knowledge extraction, drawing on clinical experience and historical case analysis datasets. This library then forms a structured abnormality pattern feature set. This feature set includes characteristic descriptions, triggering conditions, temporal characteristics, and severity classifications for various abnormal conditions. Subsequently, based on the verified parameter data streams, control instructions, and abnormal pattern feature set, a multi-dimensional anomaly detection algorithm performs real-time parameter deviation analysis. Anomaly detection employs a multi-level, multi-model strategy, including dynamic threshold detection, multi-parameter correlation analysis, temporal pattern matching, and control deviation analysis, to comprehensively assess the likelihood of anomalies and generate a set of abnormal event candidates. For each anomaly candidate, the system performs sensor redundancy verification by calculating the data consistency index of the primary and auxiliary sensor networks. Furthermore, temporal coherence analysis is performed by calculating the autocorrelation function and entropy of parameter changes to distinguish persistent anomalies from sporadic interference. Finally, based on the results of temporal continuity analysis and clinical knowledge rules, a risk assessment algorithm determines the severity of the abnormal event, generates an abnormality detection result, and activates the corresponding level of response measures according to the preset emergency response process. This multi-level verification mechanism significantly reduces the false alarm rate while improving the system's ability to identify complex abnormal patterns and the efficiency of emergency response.

[0073] Step S4: Obtain the preliminarily verified parameter data stream, the control instructions, and the abnormality detection results, analyze the data characteristics of the patient's treatment process through the trained deep learning model, predict potential risks, and obtain intelligent intervention suggestions;

[0074] Specifically, the system obtains preliminarily validated parameter data streams, control instructions, and anomaly detection results. A trained deep learning model analyzes data features from the patient's treatment process, predicts potential risks, and generates intelligent intervention recommendations. This process first uses time-frequency domain feature extraction algorithms and physiological signal processing techniques to extract key features of the patient's treatment process from historical data, including time-domain features, frequency-domain features, time-frequency domain features, morphological features, correlation features, and clinical event features, generating standardized feature vectors. The system then constructs and trains a deep neural network consisting of both time-series and static feature processing modules through a hybrid network architecture design and transfer learning methods. To address the sample imbalance issue in medical data, strategies such as weighted loss functions and generative adversarial networks to assist in generating minority class samples are employed to ensure model robustness. In practical applications, the system continuously extracts features from the most recent observation window using a sliding window technique to generate a real-time feature set. The trained deep learning model is then used to predict the changing trends of key physiological parameters over the next 10-30 minutes, generating a multidimensional risk score. Based on these risk scores, the system constructs a complete causal chain through causal chain tracing and evidence level assessment, generating intelligent intervention recommendations that include parameter adjustment suggestions, intervention timing recommendations, and graded intervention plans. These recommendations are presented to medical staff in the form of graded warnings, providing clear decision-making support information and significantly improving the safety and effectiveness of treatment.

[0075] Step S5: Obtain the parameter data stream that has undergone preliminary verification, the control instructions, the abnormality detection results, and the intelligent intervention suggestions, and build a data management and traceability system through distributed database technology and blockchain verification mechanism to achieve secure storage and retrospective analysis of data throughout the treatment process.

[0076] Specifically, in step S5, the data generated by all the preceding steps is acquired and a data management and traceability system is constructed using distributed database technology and blockchain verification mechanisms to enable secure storage and retrospective analysis of data throughout the entire treatment process. This process first utilizes sharded storage and time-series database technology to design a highly reliable, high-throughput distributed data storage architecture based on the characteristics and traffic patterns of various data types. Subsequently, blockchain technology and cryptographic algorithms are used to verify the data writing process is tamper-proof and securely desensitize patient privacy information, generating an encrypted dataset with integrity proof. Furthermore, the system utilizes multimodal data indexing and distributed query optimization to construct an analysis engine that supports multi-dimensional retrieval of time, parameters, and events, providing an interactive data analysis interface. Finally, based on this interface and the complete historical data, the system uses time series reconstruction and parameter correlation analysis to enable a comprehensive retrospective and reconstructive view of the treatment process, generating audit reports and quality improvement recommendations. This blockchain-based data management mechanism not only ensures data security and integrity but also supports comprehensive auditing and quality control of the treatment process, providing strong support for the continuous improvement of medical institutions.

[0077] In step S1, the monitoring subsystem comprising two independent sensor networks is constructed to obtain a preliminarily verified parameter data stream by concurrently collecting and processing key physiological parameters during the blood purification process, including:

[0078] Step S1.1: Based on the monitoring requirements of key blood purification parameters, high-precision sensors are deployed to build a main sensor network including parameters such as blood pressure, blood flow, dialysate temperature, conductivity, ultrafiltration rate, etc., and output the main sensor data stream.

[0079] Step S1.2: Based on a technical principle different from that of the main sensor network, an auxiliary sensor network with the same functional parameters but different implementation methods is constructed through independent power supply and signal processing units, and an auxiliary sensor data stream is output.

[0080] Step S1.3: Based on the main sensor data stream and the auxiliary sensor data stream, preliminary data consistency verification is performed through difference comparison algorithm and rationality verification, and a verification parameter data stream with confidence mark is output.

[0081] Step S1.4: Based on the confidence mark in the verification parameter data stream, the state of the sensor with low confidence is evaluated and calibrated online through the automatic diagnosis program, and a calibrated sensor state report and an updated verification parameter data stream are output.

[0082] More specifically, step S1 includes: obtaining the key physiological parameters collected by the main sensor network and the auxiliary sensor network, and simultaneously collecting the key physiological parameters through two sets of independent sensors to generate dual-channel raw data; wherein, the key physiological parameters include blood pressure, blood flow, dialysate temperature, conductivity and ultrafiltration rate data; based on the dual-channel raw data, removing noise and interference in the sensor signal through digital filtering and noise reduction processing to generate preprocessed dual-channel raw data; based on the preprocessed dual-channel raw data, comparing the consistency of the two sets of sensor data through statistical test methods, calculating the data confidence index, and obtaining the parameter data stream that has been preliminarily verified with a confidence mark.

[0083] In step S2, the control of the actuator of the blood purification system by the adaptive task space non-singular terminal super-warp sliding mode control algorithm to obtain the control instruction includes:

[0084] Step S2.1: Based on the verification parameter data stream output in step S1, a task space model including key parameters of the blood purification process is constructed through a multi-parameter state space equation, and a dynamic task space model is output.

[0085] Step S2.2: Based on the dynamic task space model, identify the singular states that may cause control instability through matrix condition number analysis and eigenvalue decomposition, and output the preprocessed non-singular control space mapping.

[0086] Step S2.3: Based on the non-singular control space mapping, by constructing a sliding mode structure and introducing the second-order derivative term, a control law with fast convergence and strong robustness is designed, and the initial parameters of the super-twisted sliding mode controller are output.

[0087] Step S2.4: Based on the initial parameters and operating state feedback of the super-twisted sliding mode controller, the control parameters are optimized and adjusted online through Lyapunov stability analysis and adaptive law design, and adaptive control instructions are output.

[0088] Now let's look at step S2.1 in detail: Modeling the blood purification control task space is the foundation of the entire control system. Its core is to abstract the blood purification process into a quantifiable and controllable mathematical model. First, the system receives the validation parameter data stream output from step S1. This data includes real-time values and reliability indicators for key parameters such as blood pressure, blood flow, dialysate temperature, conductivity, and ultrafiltration rate.

[0089] In the modeling process, multi-parameter state space equations are used to describe the dynamic characteristics of the system. Specifically, the system state vector x(t) can be expressed as:

[0090] x(t) = [blood pressure, blood flow, dialysate temperature, conductivity, ultrafiltration rate, blood purification efficiency...] T

[0091] The dynamic model of the system can be expressed as follows:

[0092]

[0093] Where u(t) is the control input vector (such as pump speed, heating power, mixing ratio, etc.), and d(t) is the external interference vector (such as changes in the patient's physiological state, environmental influences, etc.).

[0094] The modeling process specifically considers the coupling relationships between parameters, such as the impact of changes in ultrafiltration rate on blood pressure and the effect of temperature changes on solute diffusion coefficients. By introducing nonlinear terms and coupling matrices, the model more accurately reflects the complex dynamic characteristics of the actual blood purification process. Furthermore, the model incorporates individual patient parameters, enabling parameter identification and adjustment based on individual patient characteristics.

[0095] Ultimately, this step outputs a complete dynamic task space model that not only describes the current state of the system but also predicts the dynamic response of the system under given control inputs, providing the necessary mathematical basis for subsequent controller design.

[0096] The specific mathematical expression of the super-twisted sliding mode controller design is as follows:

[0097] The sliding surface S(x) is defined as:

[0098]

[0099] where x e is the desired state, and λ is a positive design parameter.

[0100] The super-twisted sliding mode control law is designed as:

[0101] u=-k1|S| α ·sign(S)-k2∫|S| β ·sign(S)dt

[0102] Where k1 and k2 are positive control gains, α and β are fractional exponents in the range (0, 1), and typical values are α = 0.5 and β = 0.25.

[0103] The adaptive adjustment law of the controller parameters is:

[0104]

[0105] Where γ1 and γ2 are positive adaptive gain parameters.

[0106] The stability proof of the controller is based on constructing the Lyapunov function:

[0107]

[0108] Proof It can ensure that the system converges to the sliding mode S=0 within a finite time.

[0109] Step S2.2: Identifying and preprocessing singular states is a key challenge in control system design. These are special configurations where the system Jacobian matrix becomes irreversible. Conventional controllers may fail in these states. In blood purification systems, singular states can occur when certain parameter combinations exhibit specific relationships (e.g., when the ratio of ultrafiltration rate to blood flow reaches a critical value).

[0110] This step first receives the dynamic task space model output by step S2.1 and then identifies potential singular states by calculating the condition number κ(J) of the model Jacobian matrix J:

[0111] κ(J)=‖J‖·‖J-1‖;

[0112] When the condition number κ(J) exceeds the preset threshold, it indicates that the system is approaching a singular state. The system also analyzes the Jacobian matrix through the eigenvalue decomposition method and calculates its minimum eigenvalue λ min , when λ min When it approaches zero, it indicates that the controllability of the system in that direction is weakened.

[0113] To pre-process these singular states, this step uses two key techniques:

[0114] Singular value damping: Based on the singular value decomposition, the singular values smaller than the threshold are damped to avoid excessive control output in the singular direction.

[0115] Task space remapping: By introducing auxiliary variables, an extended task space is constructed to transform the original singular configuration into a non-singular form in the extended space.

[0116] This process can be expressed as:

[0117] J*(q)=J(q)+αI (when a near singular state is detected)

[0118] Where α is the damping factor that is adaptively adjusted according to the system state, and I is the identity matrix.

[0119] Through these techniques, the system successfully transforms the original task space, which may contain singular states, into a stable, non-singular control space mapping, ensuring that the controller consistently generates effective control commands throughout the blood purification process and maintaining system stability even when parameters approach critical values. This preprocessing is a key step in achieving robust control and lays the foundation for the subsequent design of a super-twisted sliding mode controller.

[0120] More specifically, step S2.3 includes: based on the preprocessed non-singular control space mapping, by defining the deviation function between the system state and the desired state, constructing a sliding surface and generating an initial control law; based on the initial control law, by introducing high-order derivative terms with fractional exponents, designing a super-distortion controller to eliminate the control signal chattering phenomenon and generate a smooth control signal; based on the smooth control signal and system operating state feedback, performing online parameter optimization adjustment through Lyapunov stability analysis to obtain the control instruction.

[0121] In step S3, the redundant detection and cross-validation of parameter anomalies are implemented by the dual-channel comparative analysis algorithm to obtain anomaly detection results, including:

[0122] Step S3.1: Based on clinical experience and historical case analysis, through data mining and expert knowledge extraction, a parameter abnormality pattern library that may occur during the blood purification process is established, and a structured abnormal pattern feature set is output.

[0123] Step S3.2: Based on the verification parameter data stream output from step S1 and the control instruction output from step S2, a multi-dimensional anomaly detection algorithm is used to match the abnormal pattern feature set, perform real-time parameter deviation analysis, and output an abnormal event candidate set.

[0124] Step S3.3: Based on the abnormal event candidate set, the abnormal events are reconfirmed through cross-comparison and temporal continuity analysis of different sensor network data, and a confirmation list of abnormal events with confidence levels is output.

[0125] Step S3.4: Based on the abnormal event confirmation list, the severity of the abnormal event is determined through the risk level assessment algorithm, and the corresponding level of response measures are activated according to the preset emergency response process, and emergency response instructions and operation suggestions are output.

[0126] In step S3.1, constructing a parameter anomaly pattern library is fundamental to anomaly detection and processing. Its purpose is to establish a structured knowledge base containing various abnormal patterns and their characteristic descriptions that may occur during the blood purification process. This step systematically analyzes and constructs a comprehensive set of abnormal pattern features based on clinical experience, historical data, and expert knowledge.

[0127] First, by analyzing historical treatment data, we collected a large number of abnormal event cases, including scenarios such as equipment failure, unexpected patient reactions, and operational errors. These cases were categorized and organized, and the time series characteristics and parameter distribution characteristics of each abnormal event were extracted.

[0128] Secondly, data mining techniques, such as cluster analysis, association rule mining, and time series pattern discovery, are applied to automatically identify potential abnormal patterns from large-scale historical data. For example, the K-means clustering algorithm can be used to classify abnormal points in the parameter space into different categories; and sequential pattern mining can be used to identify abnormal sequential patterns in parameter changes.

[0129] Third, clinical experts are invited to participate in the knowledge extraction process. Through structured interviews and the Delphi method, expert experience is collected to transform implicit knowledge into explicit judgment rules. For example, an expert might provide an empirical judgment such as "When the rate of blood pressure decrease exceeds X mmHg / min and is accompanied by increased blood flow fluctuations, it indicates a possible vascular access problem."

[0130] Finally, all the above information is integrated into a unified abnormal pattern description framework. Each abnormal pattern contains at least the following elements:

[0131] Abnormal pattern ID and name (e.g. "AM001-Dialex Clotting")

[0132] Trigger conditions (e.g., "transmembrane pressure rise rate > X and blood flow decrease > Y")

[0133] Temporal characteristics (typical time patterns of abnormal development)

[0134] Severity level (1-5, corresponding to different risk levels)

[0135] Related parameter set (other parameters affected)

[0136] Recommended treatment plan (initial response measures)

[0137] These abnormal patterns are organized into a hierarchical structure, ranging from high-level categories (such as "blood-related issues," "dialysis fluid-related issues," and "patient-related issues") to specific abnormality types. At the same time, a correlation network is established between abnormal patterns, reflecting their causal and temporal relationships.

[0138] In order to deal with fuzziness and uncertainty, each anomaly pattern is also equipped with a confidence function, which describes the credibility of judging the anomaly under different parameter conditions.

[0139] This step ultimately outputs a structured anomaly pattern feature set, which serves as a reference knowledge base for subsequent real-time anomaly detection. This greatly improves the system's ability to identify complex anomalies and lays a knowledge foundation for safety redundancy design.

[0140] In step S3.2, real-time parameter anomaly detection is a key step in applying the anomaly pattern library constructed in step S3.1 to actual monitoring. This step receives two key inputs: the validation parameter data stream output from step S1 (containing the real-time values of various physiological and device parameters and their reliability indicators) and the control instructions output from step S2 (reflecting the desired system operating state). Using a multi-dimensional anomaly detection algorithm, the system can identify when parameters deviate from the normal range and output a set of possible anomaly events.

[0141] In terms of specific implementation, anomaly detection adopts a multi-level, multi-model detection strategy:

[0142] First, basic threshold detection is performed to monitor whether each key parameter exceeds its safe range. Unlike traditional systems with fixed thresholds, dynamic threshold technology is used here to automatically adjust the warning range based on individual patient characteristics, treatment stage, and historical performance:

[0143] τ lower ((t)=μ i -k i σ i f(phase,patient_profile)

[0144] τ upper (t) = μ i +k i ·σ i f(phase,patient_profile)

[0145] where μ i and σ i are the expected value and standard deviation of parameter i, and f() is the adjustment function.

[0146] Secondly, multi-parameter correlation analysis is performed to detect anomalies in the relationship between parameters. By building a parameter correlation network, the changes in the correlation strength between parameters are monitored:

[0147] C ij (t) = corr(P i (t-Δt:t), P j (t-Δt:t)) When the correlation coefficient C ij When it deviates significantly from its normal value, the system determines it as an associated abnormality.

[0148] Third, implement temporal pattern matching, perform dynamic time warping (DTW) matching on the current parameter change sequence with the temporal features in the abnormal pattern library, and calculate the similarity score:

[0149] DTW_score=DTW(current_sequence, pattern_sequence)

[0150] A high similarity indicates that the current state may match a known anomaly pattern.

[0151] Fourth, apply control deviation analysis to compare the difference between the actual system response and the expected output of the controller:

[0152] δ_control=‖actual_response-expected_response‖

[0153] A continuously increasing control deviation may indicate abnormal system behavior.

[0154] Finally, the above multi-dimensional detection results are integrated to comprehensively evaluate the possibility of anomaly through Dempster-Shafer evidence theory or Bayesian network:

[0155] P(anomaly_k)=fusion(threshold_evidence, correlation_evidence, pattern_evidence, control_evidence)

[0156] For each possible anomaly detected, the system records the following information: anomaly type ID and description, trigger time and duration, related parameters and their changing trends, preliminary confidence score, and possible impact range.

[0157] This information forms a candidate set of abnormal events, which is then passed on to the next step for further cross-validation. Through this multi-dimensional, multi-layered anomaly detection strategy, the system can sensitively capture complex abnormal patterns that are easily missed by traditional single-parameter monitoring, significantly improving the accuracy and timeliness of anomaly detection.

[0158] In step S3.3, cross-validation and anomaly confirmation are key steps in the in-depth verification of the abnormal event candidate set. This step fully utilizes the redundant design features of the dual-sensor monitoring subsystem, cross-comparing different data sources to effectively reduce the false alarm rate while maintaining a high detection rate. First, sensor redundancy verification is implemented. For each abnormal event candidate, the data consistency of the primary and secondary sensor networks is checked, and the consistency index (CI) is calculated. When the CI falls below a preset threshold, it is determined that a sensor fault may exist rather than a true anomaly. The confidence level of the anomaly is accordingly reduced, and the anomaly is marked as requiring sensor diagnosis.

[0159] Second, a temporal coherence analysis is performed to distinguish persistent anomalies from sporadic disturbances by calculating the autocorrelation function and entropy of parameter changes. The entropy of parameter changes is calculated. A lower entropy indicates a deterministic pattern in parameter changes, making it more likely to be a true anomaly rather than random noise. Third, the system applies model prediction verification. Using the dynamic model established in step S2.1, the expected trajectory of the parameter under the current control input is predicted and compared with the actual trajectory. The deviation integral (Deviation_score) is calculated as ∫|actual_trajectory - predicted_trajectory|dt. Significant and persistent deviations strengthen the evidence of an anomaly, further improving the reliability of the anomaly judgment. Fourth, the system incorporates clinical knowledge rule verification to check whether the anomaly candidate conforms to known physiological and pathological patterns. For example, the system knows that if a decrease in blood pressure is accompanied by an increase in heart rate, this is consistent with the pattern of hypovolemia, increasing its credibility. Conversely, if the dialysate temperature increases while the patient's body temperature remains unchanged, it may be a sensor issue rather than a fever. Through these multiple validation steps, the system calculates a comprehensive confidence score for each anomaly candidate: Confidence = w1·CI + w2·Entropy_score + w3·Deviation_score + w4·Knowledge_score, where w1-w4 are weighting coefficients determined based on system operational experience. Based on the confidence score, the system filters out low-confidence candidates and consolidates high-confidence anomalies into a confidence-based confirmed anomaly event list. Anomalies with intermediate confidence levels are marked as "under observation" and will continue to be evaluated as subsequent data arrives. Furthermore, the system analyzes temporal and logical connections between anomalies to identify possible causal chains. Through this multi-layered cross-validation mechanism, the system accurately distinguishes true anomalies from false alarms, ensuring patient safety while avoiding unnecessary interventions and alarm fatigue for medical staff. The final output, a confirmed anomaly event list, contains fully verified anomaly information, providing a reliable basis for subsequent emergency response.

[0160] More specifically, step S3 includes: obtaining the parameter data stream and the control instruction that have been preliminarily verified, performing sensor redundancy verification by calculating the data consistency index of the main sensor network and the auxiliary sensor network, and generating a sensor status assessment; obtaining the historical parameter change pattern, combining it with the sensor status assessment, performing temporal continuity analysis by calculating the autocorrelation function and entropy value of the parameter change, and distinguishing between persistent abnormalities and occasional interferences; based on the temporal continuity analysis results and clinical knowledge rules, determining the severity of the abnormal event through a risk level assessment algorithm to obtain the abnormality detection result; and based on the abnormality detection result, activating the corresponding level of response measures according to the preset emergency processing process to obtain the emergency processing instruction.

[0161] Among them, during the anomaly detection process, historical parameter change patterns are obtained through a multidimensional time series database, which records the complete parameter change trajectory of each patient's previous treatments. Data acquisition adopts a segmented storage strategy, dividing each treatment process into four stages: preparation period, initial period, stable period and termination period, and setting different parameter sampling frequencies and focus points for each stage. For example, in the initial stage of treatment, the system collects hemodynamic parameters such as blood pressure and heart rate at a high frequency of 5 times per second to capture the rapid changes that may occur in this stage; in the stable period, the sampling frequency is reduced to 1-2 times per minute, focusing on monitoring the slow change trend of dialyzer performance and solute clearance indicators. The extraction of historical parameter change patterns adopts a multi-scale analysis method. At the microscale, the system uses signal processing techniques such as fast Fourier transform and wavelet analysis to extract the frequency characteristics, periodicity, and mutation points of parameter fluctuations. At the mesoscale, a sliding window technique is used to calculate statistical features (mean, variance, skew, kurtosis, etc.) and trend indicators (linear slope, curvature, etc.). At the macroscale, a dynamic time warping algorithm is applied to extract morphological features from parameter curves throughout the treatment cycle and identify typical patterns of variation. This multi-scale analysis approach enables the system to simultaneously capture short-term fluctuations, medium-term trends, and long-term patterns, comprehensively understanding the dynamic nature of parameter variation. This adaptive and personalized approach addresses individual patient differences. Through case similarity calculation, the system builds a personalized reference model for each patient, integrating the patient's own historical data and the collective characteristics of clinically similar patients. This similarity calculation is based on multiple factors, including demographic characteristics (age, gender, etc.), underlying diseases (primary disease, comorbidities, etc.), treatment regimen (dialysis mode, dialyzer type, etc.), and historical response patterns (sensitivity to changes in ultrafiltration rate, etc.). This personalized reference model makes anomaly detection more accurate, significantly reducing the false alarm rate while increasing the true positive detection rate. It also continuously updates the historical parameter change pattern library through an online learning mechanism. After each treatment, the parameter change characteristics during that treatment are automatically evaluated for conformity with the expected pattern, and the individual patient model is updated based on the evaluation results. For newly detected abnormal patterns, the relevant contextual information and processing results are recorded and, after clinical verification, incorporated into the global abnormal pattern library. This continuous learning mechanism enables the system to adapt to long-term changes in patient status and the continuous development of clinical practice, maintaining the advancement and adaptability of anomaly detection performance.

[0162] In step S4, the deep learning model is used to analyze the data characteristics of the patient's treatment process, predict potential risks, and obtain intelligent intervention suggestions, including:

[0163] Step S4.1: Based on the historical data stream from step S1 to step S3, the key features of the patient's treatment process are extracted through time-frequency domain feature extraction algorithm and physiological signal processing technology, and a standardized feature vector is output.

[0164] In this step, the historical data stream from steps S1 to S3 is preprocessed and cleaned, including outlier detection and processing, missing value interpolation, and noise filtering. Specifically, the system uses a modified Z-score method to identify anomalous data points outside the 3σ range, handles missing values using time series interpolation or K-nearest neighbor interpolation, and removes high-frequency noise through wavelet transform or adaptive Kalman filtering. After preprocessing, the system extracts features from multiple dimensions. For time-domain features, it calculates statistical characteristics such as the mean, median, standard deviation, interquartile range, kurtosis, and skewness of each physiological parameter, as well as variability indicators for parameters such as blood pressure. For frequency-domain features, it uses fast Fourier transforms or autoregressive power spectrum analysis to extract features such as energy distribution, dominant frequency, and spectral slope across different frequency bands. For time-frequency domain features, it uses wavelet transforms or Hilbert-Huang transforms to analyze the signal's variation in both the time and frequency domains. For morphological features, it identifies key morphological features such as peaks, valleys, plateaus, rates of rise, and rates of fall in parameter variation curves. For correlation features, it calculates metrics such as correlation coefficients, mutual information, and transfer entropy between different parameters to construct a parameter association network. For clinical event features, it extracts information such as event frequency, temporal distribution, and severity distribution from the anomaly detection results in step S3. For feature selection, the system employs a two-stage strategy: initial screening using methods such as Lasso regression and mutual information analysis, followed by further refinement of the feature subset using wrapper methods such as recursive feature elimination. The extracted features are then normalized, including min-max scaling, Z-score normalization, one-hot encoding of categorical features, and alignment and length normalization of time-series features. Furthermore, based on clinical expert recommendations, features are weighted and combined to construct composite features with enhanced clinical interpretability. Ultimately, this step outputs a standardized set of feature vectors, each representing a comprehensive feature representation of a treatment process or episode, laying the data foundation for the subsequent construction and training of deep learning models.

[0165] Step S4.2: Based on the standardized feature vectors and historical treatment results, a prediction model capable of identifying treatment trends is constructed and trained through deep neural network architecture design and transfer learning methods, and the trained deep learning model is output.

[0166] In this step, a hybrid network architecture was first designed based on the temporal and multimodal characteristics of blood purification data. The temporal feature processing module utilizes a bidirectional long short-term memory (LSTM) network, consisting of three layers with 128 hidden units per layer, capable of simultaneously considering the contextual relationship of parameter changes. The static feature processing module uses a three-layer fully connected neural network with 256, 128, and 64 units, employing the ReLU activation function and batch normalization to improve training efficiency. The feature fusion module uses an attention mechanism to weightedly fuse the outputs of the temporal and static modules, enabling the model to focus on the most relevant feature combinations based on different contexts. The model design specifically considers the characteristics of the blood purification field, including: a multi-task learning framework that simultaneously predicts multiple relevant risk indicators and leverages the correlation between tasks to improve overall prediction performance; a hierarchical prediction structure that first predicts intermediate physiological indicators and then infers clinical risk based on these predictions; and medical knowledge constraints that introduce a physical constraint layer based on expert knowledge into the network design to ensure that the prediction results conform to physiological laws. To address the sample imbalance issue in medical data, strategies such as a weighted loss function, generative adversarial networks (GANs) to assist in generating minority class samples, and focal loss were employed. The training process employed transfer learning, pre-training the underlying network on a general medical dataset before fine-tuning on a specific blood purification dataset. The specific training strategy included time-segmented datasets to ensure validation and testing effectiveness, the Adam optimizer with a cosine annealing schedule to dynamically adjust the learning rate, various regularization techniques to prevent overfitting, and 5-fold time-series cross-validation to ensure model stability. During model training, multiple metrics were regularly evaluated, including accuracy, precision, recall, F1 score, AUC, and calibration curves. For time-series predictions, the prediction lead was also calculated. To meet the interpretability requirements of medical AI, the model integrated interpretability techniques such as SHAP value calculation, gradient-weighted class activation mapping, and prototype network design. After training, model performance was evaluated through independent test set performance testing, adversarial testing using simulated clinical scenarios, and consistency comparison with clinical expert judgment. Finally, this step outputs the trained deep learning model, including model weights, architecture definition, performance evaluation report, and interpretable analysis results, providing intelligent support for subsequent risk prediction and intervention plan generation.

[0167] Step S4.3: Based on the trained deep learning model and current treatment status data, predict possible future risk conditions through forward reasoning and sensitivity analysis, and output a risk prediction report and intervention recommendation plan.

[0168] In this step, a sliding window technique is first used to extract features from the current treatment status. Unlike the training phase, this requires processing real-time data inflow. Therefore, optimization techniques such as incremental feature updates are employed to ensure feature updates are completed within 100 milliseconds, meeting real-time monitoring requirements. The system then utilizes a deep learning model to perform multi-level risk prediction: First, the changing trends of key physiological parameters over the next 10-30 minutes are predicted, with predicted values and confidence intervals presented in descending time windows (1 minute, 5 minutes, 15 minutes, and 30 minutes). The system then assesses the risk of multiple clinical events, including the risk of hypotension (graded as low, moderate, or high) and the time window of potential occurrence; the coagulation risk score (0-100) and prediction of high-risk sites; the risk of electrolyte imbalance (classified as sodium, potassium, calcium, and magnesium) and its severity; the risk of inadequate dialysis and the potential impact on toxin clearance; and the probability of patient discomfort symptoms. Finally, the system integrates the individual risks, taking into account interactions between risks, and calculates a normalized composite risk score (0-100), which is then interpreted clinically. To gain a deeper understanding of the risk drivers, the system performs sensitivity analysis, simulating changes in various parameters to observe risk changes and identify key factors with the greatest impact, such as a quantitative relationship like "every 10 ml / h decrease in ultrafiltration rate reduces the risk of hypotension by 15%." Based on this risk assessment, the system generates personalized intervention recommendations, including: parameter adjustment suggestions, such as ultrafiltration rate, dialysate temperature, and sodium concentration, each accompanied by an expected effect estimate and confidence level; intervention timing recommendations, assessing the pros and cons of "immediate intervention" versus "watchful waiting" and recommending the optimal intervention time; a tiered intervention plan, providing multiple options ranging from the most conservative to the most aggressive, each with expected effects and risk assessments; and emergency preparedness recommendations, providing emergency preparedness advice for situations where the risk is predicted to be high but immediate intervention is not yet confirmed. The generated intervention plans fully consider individual patient differences and clinical realities, including personalized adjustments based on the patient's historical treatment history, balancing and optimizing multiple treatment goals, and providing actionable recommendations based on clinical realities. All risk predictions and intervention recommendations are output in a structured format, including key information such as risk type, severity, probability of occurrence, time window, recommended measures, expected effects and confidence level, providing medical staff with clear and actionable decision support.

[0169] Step S4.4: Based on the risk prediction report and intervention recommendation plan, generate decision support information for medical staff through interpretable analysis and clinical rule engine, and output graded warning signals and operational guidance suggestions.

[0170] In this step, the system first implements explainable analysis to make the AI's recommendations clinically credible: by tracing the causal chain, a complete derivation process from original observational data to final conclusions is constructed; according to the type and strength of evidence supporting each recommendation, the evidence level is assigned, such as "Level A: supported by multiple randomized controlled trials", "Level B: supported by observational studies" or "Level C: expert consensus or systematic empirical inference"; cases of similar patients and similar situations are retrieved from the historical database to provide reference information and enhance the credibility of the recommendations. Based on interpretable analysis, the system presents information through a multi-level early warning and decision support interface: the hierarchical early warning system designs a four-level early warning mechanism, including blue (early change trends that require attention), yellow (medium risk that requires preparation for intervention), orange (significant risk that requires timely intervention) and red (serious risk that requires immediate intervention). Each level of warning uses different visual and sound prompts; parameter adjustment decision support provides sliding adjustment controls, simulation diagrams of expected parameter adjustment effects and adjustment history records; intervention timing recommendations use timers and time window visualization tools to clearly display the optimal intervention time window, suboptimal intervention time, the attenuation curve of the intervention effect over time, and the predicted curve of risk development in the absence of intervention; the comprehensive decision console integrates all warnings and suggestions, including a patient treatment overview panel, a real-time risk monitoring dashboard, a priority list of intervention suggestions, and interactive controls for one-click execution or modification of suggestions. A specially designed interactive mechanism adapts to the medical workflow, including: context-aware notifications, which adjust the notification method and content based on the medical staff's location and work status; an informed confirmation mechanism, which tracks unconfirmed key information and upgrades the notification method when necessary; a feedback learning loop, which records the medical staff's acceptance, modification, or rejection of system recommendations, the reasons, and the ultimate treatment results; and team collaboration support, which allows the same warning or recommendation to be distributed to medical staff in different roles based on their responsibilities. Finally, a complete decision support record is output, including all generated warning signals and timestamps, the intervention recommendations provided and their basis, the medical staff's response behavior and the actual intervention measures implemented, as well as the post-intervention effect evaluation and system learning points. These records serve as part of the medical documentation and also provide feedback data for the system's continuous learning, forming a closed-loop optimization.

[0171] The detailed process of building and training a deep learning model is as follows:

[0172] Model Training Data Source: The model training set consists of 10,000 blood purification treatment records from three years of clinical data, covering patients of varying ages, genders, and underlying medical conditions. The data was anonymized and divided into training, validation, and test sets in an 8:1:1 ratio.

[0173] Model Architecture Design: A hybrid network architecture is employed, comprising a temporal feature processing module and a static feature processing module. The temporal feature processing module employs a three-layer bidirectional LSTM structure, with each layer containing 128 neurons, to capture the temporal dependencies of parameter changes. The static feature processing module utilizes a three-layer fully connected neural network with 256, 128, and 64 nodes per layer, respectively. ReLU activation functions and batch normalization are used to improve training efficiency. The outputs of these two modules are fused through an attention mechanism to generate the final feature representation.

[0174] Hyperparameter settings: The batch size was set to 64, the initial learning rate was 0.001, and the Adam optimizer was used with a cosine annealing learning rate schedule. To prevent overfitting, the dropout rate was set to 0.3, the L2 regularization coefficient was set to 1e-5, and an early stopping strategy was used (training was stopped after 10 consecutive epochs of no improvement in the validation set loss).

[0175] Model Validation and Evaluation: The model's stability was assessed through 5-fold time-series cross-validation. Key evaluation metrics on an independent test set included: 92.3% risk prediction accuracy, 89.1% precision, 87.6% recall, 88.3% F1 score, and an AUC of 0.94. Regarding time-series prediction capabilities, the model was able to predict clinically significant events an average of 15 minutes in advance, providing medical staff with ample time to intervene.

[0176] To address the class imbalance problem in medical scenarios, this model adopts a combined strategy: using SMOTE (Synthetic Minority Oversampling Technique) for data augmentation on minority class samples; using the Focal Loss loss function to assign higher weights to difficult-to-classify samples; and designing a dynamic sample weighting strategy to adaptively adjust the weights of various types of samples during training.

[0177] In step S5, the data management and traceability system is constructed by using distributed database technology and blockchain verification mechanism to achieve secure storage and retrospective analysis of data throughout the treatment process, including:

[0178] Step S5.1: Based on all data types and traffic characteristics generated in the previous steps, design a highly reliable, high-throughput distributed data storage architecture using sharded storage and time series database technology, and output a data storage solution.

[0179] In this step, the data characteristics generated by the previous steps are first analyzed, including data type, generation frequency, data volume, and access patterns. The data generated during the blood purification process mainly includes: high-frequency sampled physiological parameter time series data (such as blood pressure and blood flow data once per second), medium-frequency sampled equipment operating parameters (such as dialysate temperature and conductivity data once per minute), low-frequency recorded treatment event data (such as operation records and alarm records), and structured patient basic data and treatment plan data. Based on these characteristics, a hierarchical storage architecture is designed: for high-frequency sampled time series data, a dedicated time series database (such as InfluxDB or TimescaleDB) is used for storage, and automatic downsampling and hot and cold data tiering strategies are configured to balance storage efficiency and query performance; for structured patient data and treatment plans, a relational database (such as PostgreSQL) is used for storage to ensure data integrity and consistency; for semi-structured event data and logs, a document-based database (such as MongoDB) is used for storage to provide more flexible query capabilities. To meet the requirements of high availability and scalability, the system implements a data sharding strategy, horizontally sharding data based on time and patient ID, and employs a consistent hashing algorithm to ensure even data distribution. A comprehensive replication strategy is also designed, with at least three copies of critical data stored across different physical nodes. This approach improves system throughput through a read-write splitting mechanism. Given the importance of medical data security, a multi-layered data backup mechanism is implemented, including real-time, incremental, and regular full backups. A disaster recovery process is also designed to ensure data recoverability in extreme situations. Furthermore, a metadata management framework is established to record data sources, processing steps, versions, and lifecycle information, providing support for subsequent data traceability and auditing. Ultimately, this step outputs a complete data storage solution, including database selection, sharding strategy, replica configuration, backup solutions, and expansion planning, providing a technical foundation for subsequent secure data storage and efficient access.

[0180] Step S5.2: Based on the data storage solution, blockchain technology and cryptographic algorithms are used to achieve tamper-proof verification of the data writing process and secure desensitization of patient privacy information, and output an encrypted data set with integrity proof.

[0181] In this step, a data integrity verification mechanism based on blockchain technology was first designed. Specifically, the data generated during the treatment process was organized into data blocks according to chronological order and logical relationships. Each data block contains treatment data within a certain period of time (such as 10 minutes), and the hash value of the data block is calculated using cryptographic hash algorithms such as SHA-256. Each new data block not only contains the hash value of its own data, but also references the hash value of the previous block, forming a chain structure. Once the data is modified, the hash value will change, destroying the chain structure, thereby detecting tampering. The system deploys a blockchain network on multiple physical nodes and adopts consensus mechanisms such as Practical Byzantine Fault Tolerance (PBFT) to ensure that data consistency and integrity can be maintained even if some nodes fail or are attacked. To protect data privacy, a multi-layered security strategy is adopted: first, data desensitization is performed, replacing patient identification information (such as name and ID number) with hash identifiers to retain medically valid data; then, differential privacy technology is implemented to add carefully calibrated noise to statistical query results to prevent inference of individual information through multiple queries; finally, homomorphic encryption technology is applied to allow computational analysis of data in an encrypted state, without the need for decryption to obtain analysis results. The system also establishes a fine-grained access control mechanism, defining different access rights based on roles (such as doctors, nurses, technicians) and scenarios (such as routine queries, emergency rooms, and scientific research), and implements a dynamic authorization strategy to dynamically adjust access rights based on contextual information such as time and location.

[0182] Step S5.3: Based on the encrypted data set, through multimodal data indexing and distributed query optimization, build an analysis engine that supports multi-dimensional retrieval of time, parameters, and events, and output an interactive data analysis interface.

[0183] In this step, we first designed an index structure tailored to the characteristics of blood purification data. For time series data, we used a time-partitioned tree index and a time-prefixed R-tree index to accelerate time range queries. For multidimensional numerical parameters, we used KD-tree and R*-tree indexes to support queries based on multiple parameter combinations. For event data, we constructed an inverted index and prefix tree to enable efficient searches based on event type and keywords. A unified query interface has been designed to support multiple query modes: time-based queries, which can retrieve data by absolute time (e.g., "May 1, 2023, 8:00-9:00 AM") or relative time (e.g., "30-60 minutes after dialysis start"); parameter-based queries, which can filter data by single-parameter conditions (e.g., "blood pressure < 90 mmHg") or multi-parameter combinations (e.g., "blood pressure decrease rate > 10 mmHg / min and heart rate increase > 15 beats / min"); event-based queries, which can retrieve relevant records by event type (e.g., "hypotension event") or event characteristics (e.g., "alarm duration > 5 minutes"); and association queries, which support complex queries based on causal relationships or correlations (e.g., "blood pressure changes within 30 minutes after ultrafiltration rate adjustment"). To improve the performance of large-scale data queries, distributed query optimization strategies have been implemented, including query plan optimization, predicate pushdown, parallel execution, and result merging. The system also incorporates a caching mechanism that intelligently caches frequently used query results and intermediate calculation results based on query popularity and patterns, significantly improving the response speed of repeated queries. To support in-depth data analysis, the system integrates a variety of analytical tools: a statistical analysis component that provides basic statistical calculations, hypothesis testing, and correlation analysis; a time series analysis component that supports time series processing such as trend analysis, seasonal decomposition, and anomaly detection; a pattern mining component that can discover recurring parameter change patterns or event sequences from large amounts of data; and a predictive analysis component that predicts future parameter change trends or event probabilities based on historical data. Services are provided to upper-level applications via REST APIs and GraphQL interfaces, supporting flexible query and analysis requirements. Ultimately, this step outputs a fully functional interactive data analysis interface, enabling medical staff to quickly retrieve treatment data, perform multi-dimensional analysis, and identify potential treatment patterns and optimization opportunities.

[0184] Step S5.4: Based on the interactive data analysis interface and complete historical data, through time series reconstruction and parameter correlation analysis, a panoramic review and reproduction of the treatment process is achieved, and a treatment process audit report and quality improvement suggestions are output.

[0185] In this step, the time-series reconstruction of the treatment process is achieved based on the fully preserved historical data. Specifically, all key parameter changes, control commands, alarm events, and manual interventions during a specific treatment process can be reproduced along a timeline, forming a complete spatiotemporal map of the treatment process. The system provides multiple visualization views: Parameter Trend View, which displays the change curves of multiple key parameters with time as the horizontal axis, supports interactive zooming and parameter combination adjustment; Event Marker View, which marks the time and duration of various events (such as alarms, parameter adjustments, and drug interventions) on the timeline and provides pop-up windows with detailed information; Control Response View, which displays the time and content of control commands and the system response, facilitating analysis of control effectiveness; and Risk Score View, which displays the changing trends and key turning points of the risk score during the treatment process. It supports various interactive exploration functions, such as time window dragging, parameter comparison, and event filtering, allowing users to deeply analyze the details of specific time periods or events. In addition to reproducing a single treatment, it also supports cross-treatment comparative analysis, displaying parameter changes and event distributions from different treatments for the same patient or similar treatments for different patients side by side, helping to identify commonalities and differences. Based on these panoramic reproducibility capabilities, a comprehensive treatment audit function is achieved. Treatment process audit reports can be automatically generated, including: basic treatment information (such as date, duration, and main parameter settings), parameter stability assessment (such as the proportion of blood pressure stable time and fluctuation range), abnormal event statistics (such as the number and duration of hypotension events), intervention measure evaluation (such as parameter adjustment and its effect), treatment effect evaluation (such as ultrafiltration target achievement rate and urea clearance rate), and system performance evaluation (such as prediction accuracy and alarm rationality). It can also identify best practices and potential improvement points during the treatment process.

[0186] Figure 2 The figure shows the overall structure of the blood purification intelligent monitoring system provided by the present invention. Figure 2 As shown, the system includes the following modules:

[0187] The dual-sensor monitoring module 101 is used to obtain the core parameter requirements of the blood purification equipment and build a monitoring subsystem consisting of two independent sensor networks. By parallelly collecting and processing key physiological parameters during the blood purification process, a preliminarily verified parameter data stream is obtained;

[0188] The controller module 102 is configured to obtain the preliminarily verified parameter data stream, control the actuator of the blood purification system through an adaptive task space non-singular terminal super-warp sliding mode control algorithm, and obtain a control instruction;

[0189] Anomaly detection module 103 is used to obtain the parameter data stream and the control instruction that have been preliminarily verified, and to implement redundant detection and cross-validation of parameter anomalies through a dual-channel comparative analysis algorithm to obtain anomaly detection results;

[0190] Intelligent diagnosis module 104 is used to obtain the preliminarily verified parameter data stream, the control instructions and the abnormality detection results, analyze the data characteristics of the patient's treatment process through the trained deep learning model, predict potential risks, and obtain intelligent intervention suggestions;

[0191] The data management module 105 is used to obtain the parameter data stream that has been preliminarily verified, the control instructions, the abnormality detection results and the intelligent intervention suggestions, and build a data management and traceability system through distributed database technology and blockchain verification mechanism to achieve secure storage and retrospective analysis of data throughout the treatment process.

[0192] Figure 3 FIG. 1 shows a schematic diagram of the structure of the dual sensing monitoring subsystem provided by an embodiment of the present invention. Figure 2 As shown in Figure 1, the dual-sensor monitoring subsystem includes a primary sensor network, an auxiliary sensor network, and a data verification processing unit. The primary and auxiliary sensor networks use different technical principles to collect the same functional parameters. The data verification processing unit compares and verifies the data from the two sensor networks for plausibility, outputting a validation parameter data stream with confidence markers.

[0193] Figure 4 FIG. 1 shows a flow chart of a dual exception determination and redundant processing mechanism provided by an embodiment of the present invention. Figure 4 As shown in the figure, the mechanism first establishes a parameter abnormality pattern library based on clinical experience and historical cases, then uses the verification parameter data stream and control instructions to perform real-time parameter deviation analysis through a multi-dimensional anomaly detection algorithm, confirms the authenticity of the detected abnormal event candidates through cross-validation, and finally activates the corresponding level of emergency response measures according to the severity of the anomaly.

[0194] Figure 5 FIG1 shows a flow chart of artificial intelligence-assisted diagnosis and intervention provided by an embodiment of the present invention. Figure 5 As shown in the figure, the system converts historical data and current treatment status into clinical decision support information through four steps: feature extraction, model training, risk prediction and decision support, providing medical staff with intelligent intervention recommendations.

[0195] The above description is only a preferred embodiment of the present invention and is 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 in the scope of protection of the present invention.

Claims

1. A blood purification intelligent monitoring method, characterized in that: The method comprises: Obtain the core parameter requirements of the blood purification equipment and build a monitoring subsystem consisting of two independent sensor networks. By parallelly collecting and processing key physiological parameters during the blood purification process, a preliminarily verified parameter data stream is obtained. Obtaining the parameter data stream that has been preliminarily verified, controlling the actuator of the blood purification system through an adaptive task space non-singular terminal super-twisted sliding mode control algorithm to obtain a control instruction; Obtaining the preliminarily verified parameter data stream and the control instruction, and implementing redundant detection and cross-validation of parameter anomalies through a dual-channel comparative analysis algorithm to obtain an anomaly detection result; Obtaining the preliminarily verified parameter data stream, the control instructions, and the abnormality detection results, analyzing the data characteristics of the patient's treatment process through the trained deep learning model, predicting potential risks, and obtaining intelligent intervention recommendations; The parameter data stream that has undergone preliminary verification, the control instructions, the abnormality detection results and the intelligent intervention suggestions are obtained, and a data management and traceability system is constructed through distributed database technology and blockchain verification mechanism to achieve secure storage and retrospective analysis of data throughout the entire treatment process.

2. The method according to claim 1, characterized in that The proposed system includes two independent sensor networks for monitoring subsystems. By concurrently collecting and processing key physiological parameters during the blood purification process, a preliminarily validated parameter data stream is generated, including: Based on the monitoring requirements of key blood purification parameters, a main sensor network containing the key physiological parameters is constructed through sensor deployment to obtain the main sensor data stream; Constructing an auxiliary sensor network for collecting the key physiological parameters, wherein the auxiliary sensor network is constructed using a different principle from the main sensor network and generates an auxiliary sensor data stream through an independent power supply and signal processing unit; The main sensor data stream and the auxiliary sensor data stream are obtained, and data consistency verification is performed through a difference comparison algorithm and rationality verification to obtain the parameter data stream that has undergone preliminary verification.

3. The method according to claim 2, characterized in that The key physiological parameters in the blood purification process are collected and processed in parallel to obtain a preliminarily verified parameter data stream, including: Acquiring the key physiological parameters collected by the primary sensor network and the auxiliary sensor network, and simultaneously collecting the key physiological parameters through two sets of independent sensors to generate dual-channel raw data; wherein the key physiological parameters include blood pressure, blood flow, dialysate temperature, conductivity, and ultrafiltration rate data; Based on the dual-channel raw data, removing noise and interference in the sensor signal through digital filtering and noise reduction processing to generate pre-processed dual-channel raw data; Based on the pre-processed dual-channel raw data, the consistency of the two sets of sensor data is compared by a statistical test method, and a data confidence index is calculated to obtain the parameter data stream that has been preliminarily verified and is marked with confidence.

4. The method according to claim 1, wherein The actuators of the blood purification system are controlled by the adaptive task space non-singular terminal super-twisted sliding mode control algorithm to obtain control instructions, including: Obtaining the parameter data stream that has been preliminarily verified, and constructing a task space model including key parameters of the blood purification process through a multi-parameter state space equation to obtain a dynamic task space model; Based on the dynamic task space model, singular states that may cause control instability are identified through matrix condition number analysis and eigenvalue decomposition, and a preprocessed non-singular control space mapping is obtained; Based on the preprocessed non-singular control space mapping, a sliding mode structure is constructed and a second-order derivative term is introduced to design a control law with fast convergence and strong robustness to obtain the control instruction.

5. The method according to claim 4, characterized in that The control instructions are obtained by constructing a sliding mode structure and introducing a second-order derivative term to design a control law with fast convergence and strong robustness, including: Based on the preprocessed non-singular control space mapping, a sliding surface is constructed by defining a deviation function between the system state and the desired state to generate an initial control law; Based on the initial control law, a super-twisted controller is designed by introducing high-order derivative terms with fractional exponents to eliminate the chattering phenomenon of the control signal and generate a smooth control signal; Based on the smooth control signal and system operation status feedback, parameters are optimized and adjusted online through Lyapunov stability analysis to obtain the control instruction.

6. The method according to claim 1, characterized in that The dual-channel comparative analysis algorithm is used to achieve redundant detection and cross-validation of parameter anomalies, and obtain anomaly detection results, including: Based on clinical experience and historical case analysis data sets, through data mining and expert knowledge extraction, a parameter abnormality pattern library that may occur during blood purification is established, and a structured abnormal pattern feature set is obtained; Based on the preliminarily verified parameter data stream, the control instruction, and the structured abnormal pattern feature set, a real-time parameter deviation analysis is performed using a multi-dimensional anomaly detection algorithm to obtain an abnormal event candidate set; Based on the abnormal event candidate set, the abnormal event is reconfirmed through cross comparison and temporal continuity analysis of different sensor network data to obtain the abnormal event detection result.

7. The method according to claim 1, characterized in that The redundant detection and cross-validation of parameter anomalies are achieved by the dual-channel comparative analysis algorithm to obtain anomaly detection results, including: Obtaining the preliminarily verified parameter data stream and the control instruction, performing sensor redundancy verification by calculating data consistency indexes of the primary sensor network and the auxiliary sensor network, and generating a sensor state assessment; Obtaining historical parameter change patterns, combining them with the sensor status assessment, and performing temporal continuity analysis by calculating the autocorrelation function and entropy value of the parameter changes to distinguish between persistent anomalies and occasional interference; Based on the temporal continuity analysis results and clinical knowledge rules, the severity of the abnormal event is determined by a risk level assessment algorithm to obtain the abnormality detection result; Based on the abnormality detection result, the corresponding level of response measures are activated according to the preset emergency processing process to obtain the emergency processing instructions.

8. The method according to claim 1, characterized in that The deep learning model analyzes data characteristics during patient treatment, predicts potential risks, and obtains intelligent intervention recommendations, including: Based on the preliminarily verified parameter data stream, the control instruction, and the abnormality detection result, extracting key features of the patient's treatment process through a time-frequency domain feature extraction algorithm and physiological signal processing technology to obtain a standardized feature vector; Obtaining historical treatment results, combining them with the standardized feature vectors, and constructing and training a deep neural network including a temporal feature processing module and a static feature processing module through a hybrid network architecture design and transfer learning method to obtain a trained deep learning model; The current treatment status data is obtained, combined with the trained deep learning model, and through forward reasoning and sensitivity analysis, possible risk conditions in the future are predicted to obtain the intelligent intervention recommendation.

9. The method according to claim 1, characterized in that The deep learning model analyzes data characteristics during patient treatment, predicts potential risks, and obtains intelligent intervention recommendations, including: Obtain the current treatment status features and continuously extract features from the most recent observation window using sliding window technology to generate a real-time feature set; Based on the real-time feature set, a trained deep learning model is used to predict the changing trends of key physiological parameters in the next 10-30 minutes and generate a multidimensional risk score. Based on the multidimensional risk score, a complete causal reasoning chain is constructed through causal chain tracing and evidence level assessment to generate the intelligent intervention suggestion including parameter adjustment suggestion, intervention timing suggestion and graded intervention plan.

10. A blood purification intelligent monitoring system, characterized in that: The system comprises: The dual-sensor monitoring module is used to obtain the core parameter requirements of the blood purification equipment. It builds a monitoring subsystem consisting of two independent sensor networks. By parallelly collecting and processing key physiological parameters during the blood purification process, it obtains a preliminarily verified parameter data stream; A controller module is used to obtain the parameter data stream that has been preliminarily verified, and control the actuator of the blood purification system through an adaptive task space non-singular terminal super-distortion sliding mode control algorithm to obtain a control instruction; An anomaly detection module is used to obtain the parameter data stream that has been preliminarily verified and the control instruction, and to implement redundant detection and cross-validation of parameter anomalies through a dual-channel comparative analysis algorithm to obtain an anomaly detection result; An intelligent diagnosis module is used to obtain the preliminarily verified parameter data stream, the control instructions, and the abnormality detection results, analyze the data characteristics of the patient's treatment process through the trained deep learning model, predict potential risks, and obtain intelligent intervention recommendations; The data management module is used to obtain the parameter data stream that has been preliminarily verified, the control instructions, the abnormality detection results and the intelligent intervention suggestions, and to build a data management and traceability system through distributed database technology and blockchain verification mechanism to achieve secure storage and retrospective analysis of data throughout the treatment process.

Images