Hemodialysis quality control and improvement system and hemodialysis quality control method

Abstract

The invention discloses a hemodialysis quality control and improvement system and a hemodialysis quality control method.The hemodialysis quality control and improvement system comprises a multi-source data acquisition layer, a data processing and integration layer, an intelligent quality control evaluation layer, a closed-loop improvement execution layer and a multi-terminal interaction layer, and all the layers achieve two-way communication through an Ethernet bus or a CAN bus; the communication delay is less than or equal to 100ms; hTTPS encryption communication and an Ethernet / CAN bus (delay is less than or equal to 100ms) are adopted, and patient privacy data desensitization is realized during multi-center sharing; the 10GB cache and the 10MB / s supplementary transmission rate of the edge node ensure zero loss of data in an off-network state, and the three-level protection requirement of the Medical Data Safety Guide is met.

Description

Technical Field

[0001] This invention relates to the field of hemodialysis technology, and in particular to a hemodialysis quality control and improvement system and a hemodialysis quality control method. Background Technology

[0002] Hemodialysis is a core alternative treatment for patients with end-stage renal disease, and the quality of treatment directly determines patients' survival time and quality of life. With the number of patients with chronic kidney disease continuing to increase at a rate of 5% per year, the clinical application of hemodialysis technology is becoming increasingly widespread. However, the current quality management system still faces multiple technical bottlenecks and practical challenges.

[0003] (a) The fragmentation problem of data collection and integration The hemodialysis process involves multi-dimensional data, including equipment operation, patient physiology, dialysis water quality, operational procedures, and environmental conditions. Existing systems generally suffer from incomplete data collection dimensions and inconsistent standards. Equipment parameter collection is mostly limited to basic dialysis machine operating data, lacking high-precision real-time capture of key parameters such as blood pump speed and transmembrane pressure. Patient physiological monitoring relies on intermittent blood draws, and the acquisition of core indicators such as blood urea nitrogen and creatinine has a lag of several hours, making it difficult to support dynamic adjustments during treatment. Water quality monitoring is mostly limited to the reverse osmosis outlet, lacking end-to-end monitoring of impurity concentrations and fluid levels in pretreatment pipelines and transfer tanks. Furthermore, various data are scattered across equipment terminals, nursing records, and water quality ledgers, forming "information silos" that cannot achieve unified correlation analysis. At the same time, the communication protocols and data formats used by different devices differ significantly, and the lack of standardized integration solutions results in a data utilization rate of less than 15%.

[0004] (II) Limitations of Quality Assessment and Risk Warning Current quality control assessments largely rely on manual review against the "Quality Control Standards for Hemodialysis" for post-implementation audits. The indicators covered are mostly focused on basic items such as dialysis adequacy (Kt / V value), with insufficient consideration given to influencing factors such as equipment calibration status, operational compliance, and environmental interference. Risk warnings are often triggered by single parameter thresholds, such as judging hypotension risk solely based on blood pressure below 90 / 60 mmHg, without considering the synergistic effects of transmembrane pressure mutations, water quality abnormalities, and operational errors, resulting in a warning accuracy rate of less than 60%. Furthermore, assessment models often use fixed algorithms, failing to dynamically adjust to individual characteristics such as underlying patient diseases and dialysis duration, and lack multi-center cross-sectional comparison mechanisms, making it difficult to identify quality control differences and areas for improvement between institutions.

[0005] (iii) Lack of closed-loop management of improvement measures Current quality improvement efforts largely follow a reactive response model, only tracing the root causes after problems such as inadequate dialysis or patient complications occur. Furthermore, improvement plans rely heavily on the experience of healthcare professionals, lacking data-driven, personalized support. For example, ultrafiltration rate adjustments are often calculated based on average weight changes, without considering real-time parameters such as blood potassium concentration and blood pressure fluctuations for dynamic optimization. After implementation, the effectiveness of improvement measures is only verified through post-treatment indicator testing, lacking process tracking and real-time correction mechanisms. More importantly, the data on improvement results fails to effectively feed back into the evaluation model, leading to recurring similar problems and creating a management gap in the "monitoring-early warning-improvement" process, hindering continuous improvement in quality control.

[0006] (iv) Insufficient multi-terminal collaboration and data security The existing system's interactive terminals have limited functionality; medical staff terminals mostly only support data viewing and lack voice interaction and emergency guide push functions. Patient terminals struggle to obtain personalized treatment progress and quality control scores, resulting in low participation. Furthermore, data transmission largely relies on unencrypted local area network communication, posing privacy risks during multi-center data sharing, and edge nodes lack the ability to resume transmission after network outages, easily leading to the loss of critical data. These issues collectively result in low standardization, high safety risks, and difficulty in continuous improvement of hemodialysis quality control, failing to meet increasingly stringent clinical regulatory requirements and patient treatment needs. Summary of the Invention

[0007] In order to overcome the shortcomings of the prior art, one of the objectives of this invention is to provide a hemodialysis quality control and improvement system and a hemodialysis quality control method.

[0008] One of the objectives of this invention is achieved through the following technical solution: The hemodialysis quality control and improvement system includes a multi-source data acquisition layer, a data processing and integration layer, an intelligent quality control evaluation layer, a closed-loop improvement execution layer, and a multi-terminal interaction layer. Each layer achieves bidirectional communication through an Ethernet bus or a CAN bus, with a communication delay of ≤100ms. The multi-source data acquisition layer includes a device parameter acquisition module, a patient physiological monitoring module, a water quality monitoring module, an operation process recording module, and an environmental sensing module. The device parameter acquisition module establishes a communication connection with the hemodialysis machine, blood pump, and heparin pump via an RS485 interface, and collects parameters in real time such as blood pump speed (range: 0-500mL / min, accuracy: ±1mL / min), transmembrane pressure (range: -50 to 600mmHg, accuracy: ±1mmHg), ultrafiltration rate (range: 0-2000mL / h, accuracy: ±1mL / h), and dialysate flow rate (range: 300-800mL / min, accuracy: ±5mL / min). The data sampling interval is 100ms. The patient physiological monitoring module integrates a near-infrared spectroscopy monitoring unit and a vital signs sensing unit. The near-infrared spectroscopy monitoring unit collects blood spectral data through a probe that is snap-fitted to the arterial and venous circulation catheters. The probe uses a near-infrared light source with a center wavelength of 1100-2500nm, a spectral resolution of ≤10nm, and a sampling frequency of 50Hz. It can simultaneously analyze the concentrations of urea nitrogen (detection range: 1.7-36mmol / L, accuracy: ±0.1mmol / L) and creatinine (detection range: 44-1330μmol / L, accuracy: ±5μmol / L). The vital signs sensing unit collects the patient's systolic blood pressure (range: 60-200mmHg, accuracy: ±2mmHg), heart rate (range: 40-180 beats / min, accuracy: ±1 beat / min), and activated clotting time (range: 60-400s, accuracy: ±5s) through a wristband sensor, with a data update frequency of 1Hz. The water quality monitoring module includes a multi-parameter sensor group deployed in the pretreatment pipeline (after sand filtration and carbon filtration) and the reverse osmosis water supply pipeline. The conductivity sensor has a measurement range of 0-50 μS / cm (accuracy: ±0.1 μS / cm), the total chlorine sensor has a detection limit of ≤0.01 mg / L (accuracy: ±0.005 mg / L), the pressure sensor has a range of 0-1 MPa (accuracy: ±0.01 MPa), and the temperature sensor has a range of 0-100℃ (accuracy: ±0.1℃). Each sensor has an automatic calibration function (calibration cycle: 7 days). The operation process recording module consists of a high-definition camera (1080P resolution, 25fps frame rate) and a fingerprint reader (recognition accuracy ≤50ppm, response time ≤1s). It uses the YOLOv5 image recognition algorithm to determine the compliance of operation steps in real time (such as 12 key operations such as dialyzer installation and pipeline venting), and binds the operator's identity to the operation record. The environmental sensing module uses an integrated sensor to collect data on room temperature (range: 10-40℃, accuracy: ±0.5℃), power frequency noise intensity (range: 30-100dB, accuracy: ±1dB), and air cleanliness (≥0.5μm particle count, unit: particles / L), with a sampling interval of 5 minutes. The data processing and integration layer includes a data standardization module, an error layering correction module, a multimodal fusion module, and edge computing nodes; The data standardization module adopts the HL7FHIR data exchange standard to convert multi-source data such as equipment, physiology, and water quality into JSON format. It establishes an association index through the patient's unique identification code (PID) and treatment session number, and the data timestamp accuracy is down to the millisecond level. The error stratification correction module handles data deviations through a triple correction mechanism: ① Instrument hardware error correction: Autocorrelation analysis is performed on 10 consecutive sampling points. When the autocorrelation coefficient |r|≥0.9, it is determined to be a systematic error, triggering a device self-test command and calling historical calibration parameters for compensation; ② Environmental interference correction: The cross-correlation coefficient between environmental data and physiological data is calculated to generate an interference coupling matrix, and an adaptive Kalman filter algorithm is used to eliminate noise; ③ Individual difference correction: A linear regression model is established based on the patient's treatment data over the past 3 months to correct physiological parameter deviations caused by underlying diseases. The multimodal fusion module employs a convolutional neural network (Attention-CNN) with a fusion attention mechanism. It first extracts temporal features such as device parameters and physiological data through a 1D convolutional layer (kernel size: 3×3, stride: 1), then extracts static features such as water quality and operation through a fully connected layer, and finally dynamically allocates feature weights through an attention weight matrix (dimension: 64×128), ultimately fusing them into a 128-dimensional comprehensive feature vector. The edge computing node uses an ARM Cortex-A9 processor to perform real-time data noise reduction, preliminary identification of abrupt signals (such as a sudden drop in blood pressure exceeding 30 mmHg), and key parameter caching (cache capacity: ≥10GB) locally. It only uploads abnormal data and fused feature vectors to the cloud, with a transmission latency of ≤50ms. The intelligent quality control evaluation layer includes a dynamic quality indicator library, a multi-sub-model AI evaluation system, a risk warning module, and a multi-center comparison module; The dynamic quality indicator library covers 23 core indicators in 4 categories, and is synchronized in real time with the national "Quality Control Standard for Hemodialysis" via API interface. The key indicators and standards are as follows: equipment calibration pass rate ≥98%, dialysis adequacy Kt / V value ≥1.2, coagulation index compliance rate ≥95%, water conductivity (25℃) ≤5μS / cm, and operation compliance rate ≥99%. The indicator weights are determined by the analytic hierarchy process: equipment status (25%), patient physiological indicators (35%), water quality indicators (20%), and operation compliance (20%). The multi-sub-model AI evaluation system includes a basic evaluation model and specific prediction sub-models: ① The basic evaluation model is based on the Attention-DNN architecture, which takes a comprehensive feature vector as input to generate a comprehensive quality score of 0-100 and a blood status evaluation matrix (dimension: 5×5, covering dimensions such as clearance efficiency and circulation stability); ② The infection risk prediction sub-model adopts the matter-element extensional weight theory, takes 8 parameters as input, such as the dialysis fluid bacterial culture results and the patient's white blood cell count, and outputs an infection probability of 0-100%; ③ The cardiovascular complication prediction sub-model integrates active learning and contrastive learning algorithms, extracts patient representations through a 5-layer fully connected encoder (number of nodes: 1024→512→256→128→64, activation function: ReLU), and uses cosine distance to screen high-value positive and negative samples to predict the risk of cardiovascular events. The risk warning module divides the assessment results into three levels of warning: Level 1 warning (score ≥ 90 points, no risk) is only recorded and archived; Level 2 warning (70-89 points, potential risk) triggers local audio and visual prompts and pushes a manual review list to the medical staff terminal; Level 3 warning (< 70 points, high risk) activates an emergency alarm, freezes the dialysis machine's operating permissions, and simultaneously pushes emergency suggestions to the medical staff terminal and the on-duty mobile phone. The multi-center comparison module receives quality control data from peer institutions through the regional cloud platform, performs horizontal comparison using Z-score standardization, and generates institution rankings, indicator difference heatmaps, and differentiated improvement suggestions (such as recommending optimization of the pretreatment process when water quality indicators are lagging behind). The closed-loop improved execution layer includes a personalized solution generation module, a full-process execution tracking module, and an effect backtracking and optimization module; The personalized solution generation module outputs improvement instructions based on quality scores and risk levels: for equipment-related issues, it generates equipment debugging instructions including calibration time, parameter thresholds, and operation steps; for treatment parameter issues, it uses a fractional sliding mode control algorithm to calculate the optimal ultrafiltration rate (sliding mode surface: s=e+λD^αe, α∈[0.1, 0.5]), and dynamically adjusts the dialysate formulation based on changes in patient weight and serum potassium concentration; for anticoagulation protocols, it recommends the heparin dosage adjustment range (step size: 50U / time) based on the rate of change in activated clotting time. The full-process execution tracking module collects the execution status of improvement measures in real time through IoT chips, records parameters of key operations such as ultrafiltration rate adjustment at the 10ms level, and generates operation trajectory curves; it automatically upgrades the warning level for instructions that are not executed on time (timeout ≥ 10min); The effect backtracking optimization module calculates the improvement efficiency based on indicators such as Kt / V value before and after treatment, urea clearance index (URR), and patient blood pressure fluctuation range (efficacy = |improved indicator - standard value| / |improved indicator - standard value| × 100%). When the efficiency is ≥ 70%, the feature vector of the case and the improvement plan are included in the model training set, and the AI ​​evaluation model weights are updated using the gradient descent method (learning rate: 0.001). The multi-terminal interaction layer includes medical terminals, equipment terminals, management terminals and patient terminals, and supports HTTPS encrypted communication; The medical terminal uses a 15.6-inch touch screen and integrates voice interaction (supporting more than 200 commands such as "retrieving PIDXXX quality control report") and a thermal printing module, which can automatically generate weekly equipment inspection reports (including calibration due reminders). The equipment terminal is embedded in the dialysis machine control panel, which displays real-time operating parameters, error status and improvement instructions, and supports manual confirmation and parameter fine-tuning. The management terminal has data visualization functions (line chart, bar chart), which can statistically analyze the quality control compliance rate and the distribution of early warning events for any time period, and generate regional quality control analysis reports. The patient terminal is implemented through a mini-program, which displays treatment progress, historical quality control scores, medical order reminders, and a satisfaction feedback portal.

[0009] The water quality monitoring module also includes a transfer tank status monitoring unit. The transfer tank is made of food-grade 316L stainless steel and has a honeycomb-shaped diversion isolation block (thickness: 5mm, aperture: 2mm) inside. The monitoring unit includes an immersion liquid level sensor (measuring range: 0-2m, accuracy: ±1mm) and a laser particle size sensor (detection range: 0.1-100μm), which respectively collect the liquid level in the tank and the impurity concentration at the drain outlet. When the liquid level is <0.3m, a water replenishment reminder is triggered. When the impurity concentration is >5mg / L or after replacing the sand filter tank or resin tank, a 30-minute pipeline flushing command is automatically generated (flushing flow rate: 800mL / min).

[0010] The probe housing of the near-infrared spectroscopy monitoring unit is made of 316L stainless steel (wall thickness: 2mm). The inner diameter of the blood channel matches that of the dialysis tubing (Φ4-6mm). Coaxial near-infrared emitting and receiving holes (aperture diameter: 3mm) are provided on both sides of the channel. The apertures are covered with quartz glass with a light transmittance of ≥95% (thickness: 1mm). Spectral data is acquired through vertical transmission. The probe and tubing are connected by a quick-connect interface and have an IP67 waterproof rating.

[0011] In the cardiovascular complication prediction sub-model of the intelligent quality control assessment layer, the active learning unit filters samples through a positive and negative sample selection rule: for the normalized patient representation, the angle between the sample and other samples is calculated, and the upper quartile of the angle between samples with the same label is included in the positive sample set, and the lower quartile of the angle between samples with different labels is included in the negative sample set; the contrastive learning unit constructs a loss function through the cosine distance of positive samples (target: ≥0.8) and the cosine distance of negative samples (target: ≤0.2), and updates the encoder parameters once every 50 samples are trained.

[0012] The edge computing node also integrates an edge storage module and a resume transmission function after network outage. In the event of a network outage, it can locally store ≥7 days of original data and retransmit it in the order of timestamps after the network is restored, with a retransmission rate of ≥10MB / s to ensure data integrity.

[0013] The medical terminal is also equipped with an emergency treatment knowledge base module, which contains standardized treatment procedures for 12 common risks such as low blood pressure, coagulation, and abnormal water quality. When a level 3 warning is triggered, a pop-up window will automatically display the corresponding treatment steps (including operation diagrams), and it supports one-click calling of the dialysis room director.

[0014] The hemodialysis quality control method using the system described in any one of claims 1-6 includes the following steps: S1. Synchronous Acquisition and Preprocessing of Multi-Source Data: S11. Equipment parameter acquisition: The parameters such as blood pump speed and transmembrane pressure are acquired from the dialysis machine, blood pump and other equipment via RS485 interface. The sampling frequency is 10Hz and abnormal values ​​that exceed the range (such as transmembrane pressure > 600mmHg) are automatically filtered. S12. Patient physiological data acquisition: The near-infrared spectral probe acquires blood spectral data in real time (50Hz), and the wristband sensor acquires vital signs (1Hz), which are then linked to the patient's electronic medical record via PID; S13. Water quality data acquisition: Sensors at each node of the pipeline synchronously collect indicators such as conductivity and total chlorine (2Hz), and automatically perform calibration every 7 days (using standard solutions: conductivity 10μS / cm, total chlorine 0.1mg / L). S14. Operation and Environmental Data Acquisition: The camera captures the operation process (25fps), the fingerprint reader records the operator's information, and the environmental sensor collects data such as room temperature (1Hz). S2. Data Standardization and Error Correction: S21. Standardization Processing: All data is converted to JSON format using the HL7FHIR standard, and PID, treatment session, and timestamp index are added; S22. Stratification error correction: ① Instrument error: Calculate the autocorrelation coefficient of 10 consecutive sampling points, and trigger self-check and compensation when |r|≥0.9; ② Environmental interference: Calculate the cross-correlation matrix between environmental and physiological data, and reduce noise through adaptive Kalman filtering; ③ Individual differences: Call the patient's historical model to correct physiological parameters; S3. Multimodal data fusion and feature extraction: S31. Preliminary feature extraction: For time-series data, a 1D convolutional layer (3×3 kernel) is used to extract local features, and for static data, a fully connected layer is used to extract global features; S32. Attention Fusion: Weights are assigned to the two types of features using an attention weight matrix (64×128), and after noise reduction by fractional integral (α=0.3), a 128-dimensional comprehensive feature vector is generated. S33. Edge preprocessing: Edge computing nodes locally identify parameter mutations (such as a sudden drop in blood pressure >30mmHg) and only upload abnormal data and feature vectors to the cloud; S4. Intelligent Quality Assessment and Risk Warning: S41. Comprehensive score calculation: Input the feature vector into the Attention-DNN model, and output a comprehensive score of 0-100 points by combining the index weights; S42. Specific Risk Prediction: The infection risk sub-model takes 8 parameters as input and outputs the infection probability, while the cardiovascular risk sub-model outputs the cardiovascular event risk through active-contrast learning. S43. Warning Judgment and Output: Trigger the corresponding level of warning based on the score and risk probability. Level 3 warning will simultaneously freeze the device operation permissions. S44. Multi-center comparison: Regularly upload data to the regional platform and receive horizontal comparison results and suggestions for differentiation; S5. Generation and Implementation of Personalized Improvement Plans: S51. Solution Formulation: Prioritize based on scores; for scores <70, automatically generate instructions for equipment debugging, parameter adjustment, etc.; for scores 70-89, generate manual review suggestions. S52. Command Push: The solution is pushed to the corresponding terminal through the multi-terminal interaction layer, and the operation guide is displayed simultaneously on the medical terminal; S53. Execution Tracking: Real-time collection of execution status, recording of key parameters at the 10ms level, and automatic escalation of warnings for instructions not executed within the timeout period; S6. Effect Retrospective and Model Optimization: S61. Effectiveness Assessment: After treatment, the improvement rate of indicators such as Kt / V value and URR is calculated. An effective rate of ≥70% is judged as an effective case. S62. Model Update: Incorporate the feature vectors, evaluation results, and improvement plans of valid cases into the training set, and update the AI ​​model parameters using gradient descent (learning rate 0.001); S63. Indicator Database Synchronization: Automatically synchronizes national quality control standards monthly, updating indicator standards and weights.

[0015] The calculation process of the comprehensive score in step S41 is as follows: the judgment matrix is ​​constructed using the analytic hierarchy process, and the index weights are determined after passing the consistency test (CR < 0.1). The score of each index = (measured value / standard value) × 100 × index weight. The comprehensive score = Σ score of each index, where the score of Kt / V is calculated in segments: ≥1.4 gets full marks, 1.2-1.39 is scored proportionally, and <1.2 gets 0 marks.

[0016] The process of generating the ultrafiltration rate adjustment command in step S51 includes: ① Input parameters: real-time patient weight, pre-treatment weight, serum potassium concentration, and blood pressure change rate; ② Algorithm calculation: constructing a sliding surface s=e+0.5D^0.3e through a fractional sliding mode control algorithm, where e is the deviation between the actual ultrafiltration rate and the target value; ③ Parameter output: outputting the optimal ultrafiltration rate according to the control law u=-ksign(s), and dynamically calibrating the sliding surface parameters every 30s based on the patient's blood pressure response.

[0017] Step S3 also includes a data amplification step: the fused original features are amplified using a single-feature randomization method. When the number of positive samples (with complications) is less than the number of negative samples, the remaining features are fixed, and key features such as infection risk and blood pressure are randomized within ±5% until the number of positive and negative samples is balanced. The amplified data is labeled "amplified" to distinguish it from the original data. The input parameters of the infection risk prediction sub-model in step S42 include dialysate bacterial colony count (CFU / mL), endotoxin content (EU / mL), and patient white blood cell count (×10). 9 The correlation was calculated using the matter-element extensional weight theory, with a correlation of ≥0.6 indicating a high risk of infection. The parameters included: C-reactive protein (mg / L), dialysis duration (months), history of underlying diseases (present / absent), puncture site condition (normal / red and swollen), and history of infection in the past 3 months (present / absent).

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: (i) Achieving precise integration of multi-source data across all dimensions Comprehensive data acquisition: The device parameter acquisition module enables 100ms-level acquisition of 12 device parameters, including blood pump speed (accuracy ±1mL / min) and transmembrane pressure (accuracy ±1mmHg). Combined with the near-infrared spectroscopy monitoring unit (1100-2500nm wavelength, accuracy ±0.1mmol / L), it enables real-time detection of urea nitrogen and creatinine. With the addition of a multi-parameter water quality sensor (total chlorine detection limit ≤0.01mg / L) and operation image recognition (YOLOv5 algorithm), a comprehensive data acquisition network covering 23 indicators in 5 categories is constructed, improving data integrity to over 99%.

[0019] Standardized processing is efficient and reliable: The HL7FHIR standard is used to convert multi-source data into JSON format, and a unique association index is established between PID and treatment session; the innovative error stratification correction mechanism reduces data deviation caused by instrument error, environmental interference and individual differences through autocorrelation analysis, adaptive Kalman filtering and individual regression model respectively, and the data accuracy is improved to over 98%.

[0020] Edge-cloud collaborative processing: Edge nodes based on ARM Cortex-A9 processors achieve local noise reduction and mutation identification, only uploading abnormal data and fused feature vectors, with transmission latency controlled within 50ms. It also has 7 days of local caching and network outage resume capability to ensure a balance between data integrity and real-time performance.

[0021] (II) Constructing an intelligent and precise quality control and evaluation system Dynamic indicator database adapted to regulatory requirements: Establish a dynamic indicator database that is synchronized with national quality control standards in real time, covering 23 indicators in 4 categories: equipment, physiology, water quality, and operation. The weights are determined by the analytic hierarchy process (patient physiological indicators account for 35%, equipment status accounts for 25%) to ensure the authority and scientific nature of the assessment standards.

[0022] Multi-sub-model improves assessment accuracy: By integrating the Attention-DNN basic model with specific prediction sub-models for infection and cardiovascular complications, feature weights are allocated through the attention mechanism. Combined with the matter-element extensional weight theory and active contrastive learning algorithm, the overall quality score error is ≤3 points, the accuracy of infection risk prediction is ≥92%, and the early warning of cardiovascular events is ≥10 minutes, which is more than 50% better than existing technologies.

[0023] Multi-center comparison expands the perspective for improvement: By standardizing Z-score, the quality control data of institutions at the same level in the region can be compared horizontally, generating heat maps of indicator differences and personalized improvement suggestions to help institutions quickly identify shortcomings. For example, when water quality indicators are lagging behind, the system can automatically recommend optimization plans for the replacement cycle of sand filter tanks.

[0024] (III) Forming a data-driven closed-loop improvement mechanism Personalized solutions are generated precisely: Improvement instructions are automatically generated based on quality scores and risk levels. For high-risk cases with scores <70, the optimal ultrafiltration rate is calculated using a fractional sliding mode control algorithm (α∈[0.1,0.5]). Combined with the change rate of activation clotting time, the heparin dosage is recommended (step size 50U / time), improving the suitability of the solution to over 90%.

[0025] End-to-end tracking ensures effective implementation: IoT chips enable 10ms-level parameter recording and trajectory visualization of improvement measures, automatically escalating warnings for instructions not executed within 10 minutes, increasing the execution rate to 99%; the effect backtracking module ensures the clinical value of improvement measures by calculating the Kt / V value and the urea clearance index (URR) effectiveness rate (≥70% is considered effective).

[0026] Model iteration enables continuous optimization: Effective improvement cases are included in the training set, and the AI ​​model parameters are updated using gradient descent (learning rate 0.001), which reduces the recurrence rate of similar problems by more than 75%, forming a virtuous cycle of "data collection - intelligent evaluation - precise improvement - model optimization".

[0027] (iv) Strengthen multi-terminal collaboration and data security. Multi-terminal functionality is enhanced: the medical terminal supports 200+ voice commands and emergency guide pop-ups, the device terminal enables parameter fine-tuning and self-check reminders, the management terminal generates visual quality control reports, and the patient terminal accesses treatment progress and satisfaction feedback through a mini-program, improving multi-role collaboration efficiency by 60%.

[0028] Data transmission is secure and reliable: HTTPS encrypted communication and Ethernet / CAN bus (latency ≤100ms) are used to desensitize patient privacy data when sharing data across multiple centers; the 10GB cache and 10MB / s retransmission rate of the edge nodes ensure zero data loss in the event of network outage, meeting the Level 3 protection requirements of the "Guidelines for Medical Data Security".

[0029] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0030] Figure 1 This is a flowchart of the data acquisition and preprocessing stage in this embodiment; Figure 2 This is a flowchart of the intelligent assessment and risk warning stage in this embodiment. Detailed Implementation

[0031] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0032] It should be noted that when a component is described as "fixed to" another component, it can be directly on the other component or may have a component in between. When a component is considered "connected to" another component, it can be directly connected to the other component or may have a component in between. When a component is considered "set on" another component, it can be directly set on the other component or may have a component in between. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0033] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items. I. Specific Implementation Methods This embodiment takes the clinical application scenario of a tertiary hospital hemodialysis center (equipped with 20 hemodialysis machines) as an example to explain in detail the composition, deployment and workflow of the hemodialysis quality control and improvement system. The hardware equipment and software algorithms used all meet industrial-grade standards and clinical specifications.

[0035] (I) System Hardware Deployment and Module Configuration 1. Hardware configuration of multi-source data acquisition layer Equipment parameter acquisition module: A distributed data acquisition unit (model: ModbusRTU485-ETH08) based on an RS485 bus is used to establish communication connections with 20 Fresenius 4008S hemodialysis machines, B. Braun Diapact blood pumps, and heparin pumps via dedicated shielded cables. The data acquisition unit's sampling frequency is set to 10Hz, and it can simultaneously capture parameters such as blood pump speed (0-500mL / min, accuracy ±1mL / min), transmembrane pressure (-50 to 600mmHg, accuracy ±1mmHg), ultrafiltration rate (0-2000mL / h, accuracy ±1mL / h), and dialysate flow rate (300-800mL / min, accuracy ±5mL / min). After preprocessing, the data is transmitted to the edge nodes via an Ethernet bus.

[0036] Patient physiological monitoring module: The near-infrared spectroscopy monitoring unit uses a MicroNIR1700 miniature spectrometer. The probe is made of 316L stainless steel housing (2mm wall thickness) and is fixed to the dialysis arterial / venous tubing (compatible with Φ5mm tubing) by quick-connect clips. The emission and receiving apertures are coaxially set (aperture diameter 3mm) and covered with quartz glass with 96% light transmittance. The center wavelength of the light source is 1100-2500nm, the spectral resolution is 5nm, the sampling frequency is 50Hz, and it can analyze the concentration of urea nitrogen (1.7-36mmol / L, ±0.1mmol / L) and creatinine (44-1330μmol / L, ±5μmol / L) in real time. The vital signs sensing unit uses Mindray PM-10 wristband sensor to collect systolic blood pressure (60-200mmHg, ±2mmHg), heart rate (40-180 beats / min, ±1 beat / min) and activated clotting time (60-400s, ±5s). The data update frequency is 1Hz and it is transmitted to the edge node via Bluetooth 5.0.

[0037] Water quality monitoring module: Multi-parameter sensor groups are deployed at the outlet of the pretreatment sand filter tank, the outlet of the carbon filter tank, and the reverse osmosis water supply pipeline: conductivity sensor (model: E+HCondumaxCLS15) with a measurement range of 0-50μS / cm (±0.1μS / cm), total chlorine sensor (model: HACHCL17) with a detection limit of 0.01mg / L (±0.005mg / L), pressure sensor (model: SMCISE80-02-NL) with a range of 0-1MPa (±0.01MPa), and temperature sensor (model: Pt1000) with a range of 0-100℃ (±0.1℃). The transfer tank (made of 316L stainless steel, with a volume of 500L) has a built-in honeycomb-shaped diversion isolation block (5mm thick, 2mm aperture), and is equipped with an immersion liquid level sensor (model: UHC-5100, 0-2m, ±1mm) and a laser particle size sensor (model: BettersizeBT-9300S, 0.1-100μm). All sensors are equipped with a weekly automatic calibration mechanism (using a 10μS / cm conductivity standard solution and a 0.1mg / L total chlorine standard solution).

[0038] Operation process recording module: A Hikvision DS-2CD3T47WD-L high-definition camera (1080P, 25fps) is deployed on the ceiling of the dialysis unit, with the lens covering the dialysis machine operation panel and tubing connection area, and powered by a PoE switch; the nurse station is equipped with a fingerprint reader (model: FPC1020, recognition accuracy 30ppm, response time 0.8s), and operators are required to sign in with their fingerprints before treatment and at key operation points (such as dialyzer installation and tubing venting), and the image data and identity information are uploaded synchronously via Ethernet.

[0039] Environmental sensing module: The SensirionSHT40 integrated sensor is deployed in the corner of the dialysis unit (1.5m above the ground) to collect data on room temperature (10-40℃, ±0.5℃), power frequency noise intensity (30-100dB, ±1dB) and the number of particles ≥0.5μm (unit: particles / L). The sampling interval is 5min, and the data is transmitted to the data acquisition unit via the I2C interface.

[0040] 2. Data Processing and Integration Layer Configuration Edge computing nodes: Utilize TIAM335x development boards (ARM Cortex-A9 processor, 1GHz clock speed, 2GB memory, 16GB storage), pre-installed with a Linux system and edge computing engine (using the MQTT protocol). Local implementation includes data denoising (adaptive Kalman filtering), abrupt signal identification (e.g., triggering a flag when blood pressure drops >30mmHg), and key parameter caching (16GB cache capacity, supporting 7 days of local data storage). Only abnormal data and fused feature vectors are uploaded to the cloud server via Gigabit Ethernet, with a transmission latency ≤50ms. A resume function is enabled in case of network outage (15MB / s retransmission rate after network recovery).

[0041] Cloud-based data processing servers: A distributed computing cluster is built using two Huawei FusionServerPro2288HV5 servers (dual Intel Xeon Gold 6248 processors, 128GB RAM, 4TB SSD storage). The data standardization module, developed based on the HL7 FHIRR4 standard, converts multi-source data into JSON format and establishes an association index using the patient's unique identifier (PID) and treatment session number (format: YYYYMMDD-XXX, where XXX is the treatment sequence number for the day), with timestamp accuracy down to the millisecond level. The error stratification and correction module implements a triple correction logic using Python: autocorrelation analysis uses the NumPy library to calculate the Pearson coefficients for 10 consecutive sampling points; environmental interference correction uses a Kalman filter model built with the scikit-learn library; and individual difference correction uses a linear regression equation based on the patient's data from the past three months. The multimodal fusion module uses the PyTorch framework to implement the Attention-CNN model. The 1D convolutional layer has 8 3×3 convolutional kernels, the fully connected layer has 256 nodes, the attention weight matrix has a dimension of 64×128, and the final output is a 128-dimensional feature vector.

[0042] 3. Intelligent quality control evaluation layer configuration Dynamic quality indicator library: Built on a MySQL database, it covers 23 indicators in 4 categories and is synchronized in real time with the National Health Commission's "Quality Control Standards for Hemodialysis" via API interface (automatically updated monthly). Key indicators and standards include: equipment calibration pass rate ≥98%, Kt / V value ≥1.2, coagulation index compliance rate ≥95%, water conductivity (25℃) ≤5μS / cm, and operation compliance rate ≥99%. The indicator weights are determined using the analytic hierarchy process (consistency test CR=0.07<0.1): equipment status (25%), patient physiological indicators (35%), water quality indicators (20%), and operation compliance (20%).

[0043] The multi-sub-model AI evaluation system is developed using the TensorFlow framework and deployed on a cloud GPU server (NVIDIA Tesla V100 graphics card). The basic evaluation model is an Attention-DNN architecture, taking a 128-dimensional feature vector as input, which is processed through three fully connected layers (512→256→128 nodes) and an attention mechanism to output a comprehensive score of 0-100. The infection risk prediction sub-model is based on the matter-element extensional weight theory, taking eight parameters such as the number of bacterial colonies in the dialysis fluid as input, and calculating the correlation degree through the matter-element matrix (≥0.6 indicates high risk). The cardiovascular complication prediction sub-model integrates active learning and contrastive learning, with a five-layer fully connected network encoder (ReLU activation function). Positive samples are selected using the upper quartile rule, and the contrastive learning loss function aims for a cosine distance of ≥0.8 for positive samples and ≤0.2 for negative samples, updating parameters every 50 samples.

[0044] Risk warning and multi-center comparison module: The warning module is developed in Java and integrated into a cloud application server. It triggers three levels of warnings based on scores and risk probabilities: Level 1 warnings are only logged; Level 2 warnings are pushed to medical staff terminals via TCP protocol and trigger audible and visual alerts (buzzer frequency 1Hz, indicator light yellow); Level 3 warnings freeze dialysis machine operation permissions (sending commands to the dialysis machine's RS485 interface) and simultaneously push emergency advice to the on-duty mobile phone via SMS. The multi-center comparison module uses Z-score standardization to process data from five hospitals of the same level within the region, generates a heatmap of indicator differences using ECharts, and constructs a differentiated suggestion rule base based on expert experience and evidence-based medicine (e.g., recommending increased reverse osmosis membrane flushing frequency when water conductivity exceeds the standard).

[0045] 4. Closed-loop improvement of execution layer configuration Personalized solution generation module: Based on rule engine and algorithm model, it generates improvement instructions: Equipment debugging instructions include calibration time (accurate to minute), target parameter threshold and operation steps (e.g., dialysate flow calibration needs to be verified using HDM99XP detector); Ultrafiltration rate adjustment adopts fractional sliding mode control algorithm (sliding surface s=e+0.5D^0.3e, α=0.3), calculates the optimal value through MATLAB, and sets the step size to 5mL / h; Heparin dosage adjustment is based on the activated clotting time change rate (ΔACT / Δt), with a recommended step size of 50U / time and an upper limit of no more than 500U / h.

[0046] Execution tracking and effect feedback module: Execution tracking collects operation status in real time through an IoT chip (model: NB-IoTBC28), records ultrafiltration rate adjustments with 10ms-level parameter recording, and generates an operation trajectory curve in Excel format; if an instruction is not executed within 10 minutes, the warning level is automatically upgraded. Effect feedback module calculates the improvement effectiveness rate: Effectiveness rate = |Improved indicator - Standard value| / |Improved indicator - Standard value| × 100%, ≥70% is considered a valid case, and the AI ​​model weights are updated using gradient descent (learning rate 0.001), with one model iteration performed every 100 valid cases.

[0047] 5. Multi-terminal interaction layer configuration Medical terminal: It adopts a 15.6-inch capacitive touch screen (model: industrial grade T156XW03-V5), pre-installed with Windows 10 IoT system, and integrates a voice interaction module (supporting iFlytek API, 200+ commands) and a thermal printer (model: Zebra ZD420). It can automatically generate weekly equipment inspection reports (including calibration due reminders), and has a built-in emergency handling knowledge base with 12 types of risk handling procedures (including operation diagrams). It automatically displays a pop-up window when a level 3 warning is issued.

[0048] Equipment terminal: Embedded dialysis machine control panel (customized UI interface), using a 7-inch LCD screen to display real-time operating parameters, error status and improvement instructions, and supports confirmation of execution or fine-tuning of ultrafiltration rate parameters within ±5% via physical buttons.

[0049] Management terminal: Deployed in the quality control department, using a 27-inch 4K monitor, running a data visualization platform (developed based on Tableau), which can statistically analyze the quality control compliance rate and the distribution of early warning events on a daily / weekly / monthly basis, and generate regional quality control analysis reports in PDF format.

[0050] Patient terminal: Implemented via WeChat mini program, it displays treatment progress (remaining time, ultrafiltration completion rate), historical quality control scores (last 3 treatments), medical advice reminders (such as dietary restrictions), and a 1-5 point satisfaction feedback entry. Data transmission is encrypted using HTTPS.

[0051] (II) System Workflow Taking a 4-hour hemodialysis treatment of a patient with end-stage renal disease (PID: 20240510001) as an example, the system workflow is as follows: 1. Treatment preparation stage (T-30min): The nurse signs in using the fingerprint reader, and the system automatically starts the operation process recording module; the water quality monitoring module completes self-test and calibration, showing conductivity of 3.2μS / cm and total chlorine of 0.003mg / L, which meet the standards; the equipment parameter acquisition module initializes, establishes communication with the dialysis machine, and confirms the blood pump speed setting of 200mL / min and the dialysate flow rate of 500mL / min.

[0052] 2. Data Acquisition and Preprocessing Stage (T0-T2 40min): The near-infrared spectral probe collects blood spectral data in real time, and analyzes the initial concentration of urea nitrogen (28.5 mmol / L) and creatinine (980 μmol / L); the wristband sensor displays systolic blood pressure (135 mmHg), heart rate (72 beats / min), and activated clotting time (120 s).

[0053] The water quality sensor updates data every 2 seconds, the liquid level in the transfer tank is maintained at 1.2m, and the impurity concentration at the sewage outlet is 0.8mg / L; the environmental sensor shows a room temperature of 24℃, a noise level of 52dB, and a particle count of 350 particles / L.

[0054] The edge computing node performs local data noise reduction and identifies a sudden drop in systolic blood pressure to 95 mmHg (a decrease of 40 mmHg) at T35min. It marks this as an anomaly and uploads it to the cloud.

[0055] 3. Intelligent assessment and early warning phase (T0-T240min, real-time): The cloud-based data processing server corrects the errors in the standardized data and then fuses it into a 128-dimensional feature vector using the Attention-CNN model. The basic evaluation model outputs a T35min comprehensive score of 82, triggering a level-two warning.

[0056] The cardiovascular complication prediction sub-model takes parameters such as blood pressure fluctuations as input and outputs a risk probability of 65%. The multi-center comparison module shows that the incidence of hypotension in this type of patient in our institution (8%) is higher than the regional average (5%).

[0057] 4. Improvement Solution Generation and Implementation Phase (T36-T38min): The treatment parameter generation module determined that the problem was related to treatment parameters. It then used a fractional sliding mode control algorithm to calculate the optimal ultrafiltration rate, reducing it from 500 mL / h to 420 mL / h, and sent instructions to the medical staff terminal and the equipment terminal.

[0058] After the nurse confirms the instruction, the device terminal automatically adjusts the parameters, executes the tracking module to record the adjustment process (10ms-level sampling), and generates an ultrafiltration rate trajectory curve.

[0059] 5. Effect Retrospective and Model Optimization Phase (T240min+): The system calculated that after treatment, blood urea nitrogen was 7.2 mmol / L, creatinine was 156 μmol / L, Kt / V value was 1.4, and systolic blood pressure rose to 120 mmHg after improvement. The effective rate was 85%, and it was judged as an effective case.

[0060] Case data is incorporated into the training set, and the AI ​​evaluation model updates the weights using gradient descent to optimize the low blood pressure risk identification threshold; treatment reports are updated synchronously across multiple terminals, and patient satisfaction surveys are pushed to patient terminals.

[0061] II. Experimental Records (I) Experimental Objective The performance of the system of the present invention in terms of data acquisition accuracy, quality control assessment accuracy, early warning response speed and closed-loop improvement effectiveness is verified, and its advantages are compared with those of existing traditional quality control methods.

[0062] (II) Experimental Subjects and Environment Experimental subjects: 20 dialysis machines (8 from brand A, 6 from brand B, and 6 from brand C) were selected from a tertiary hospital hemodialysis center, and 100 patients with end-stage renal disease (58 males and 42 females, aged 35-78 years, with dialysis duration of 3-60 months) were included. Among them, 42 patients had diabetic nephropathy, 38 patients had hypertensive nephropathy, and 20 patients had other causes.

[0063] Experimental environment: The dialysis center is laid out according to standard (40m×20m), with Ethernet and CAN bus deployed (communication latency ≤80ms), cloud server access bandwidth of 1000Mbps, and regional cloud platform connected to 5 hospitals of the same level (each providing historical data of 50 patients).

[0064] Control group setup: A randomized controlled trial was conducted, with 50 patients included in the experimental group (using the system of this invention) and 50 patients included in the control group (using traditional quality control methods: manual parameter recording, offline calculation of Kt / V, and empirical adjustment of treatment plan).

[0065] (III) Experimental Design and Methods 1. Data acquisition accuracy verification experiment (duration: 1 month) Equipment parameter accuracy: The HDM99XP hemodialysis machine quality tester (JJF1353-2012 standard) was used as the gold standard. The blood pump speed, transmembrane pressure and dialysate flow rate of 20 dialysis machines were tested 100 times per machine. The error rate of the values ​​collected by this system was compared with that of the gold standard.

[0066] Accuracy of patient physiological data: Venous blood was collected from each patient before and after treatment, and the concentrations of urea nitrogen and creatinine were detected by Hitachi 7600 biochemical analyzer. The accuracy of the near-infrared spectroscopy detection value and the biochemical detection value of this system were compared. The Mindray BSM-6000 monitor was used as the gold standard to compare the detection errors of blood pressure and heart rate by wristband sensors.

[0067] Water quality data accuracy: The conductivity and total chlorine concentration at each node of the pipeline were detected using a Hach HQ40d water quality analyzer, and the deviation of the sensor values ​​detected by this system was compared.

[0068] Statistical methods: Calculate the absolute error (|system value - gold standard value|), relative error (absolute error / gold standard value × 100%), and accuracy (number of tests with relative error ≤ 5% / total number of tests × 100%).

[0069] 2. AI Evaluation Model Performance Validation Experiment (Duration: 3 months) Data preparation: 300 treatment data (including equipment, physiological, water quality, operation and environmental data) of 100 patients were collected and divided into training set (210 times) and test set (90 times) in a 7:3 ratio. Among them, there were 60 positive samples (complications such as hypotension / infection) and 240 negative samples.

[0070] Model training: The basic assessment model was iterated 1000 times, the infection risk sub-model adopted 5-fold cross-validation, and the cardiovascular complication prediction sub-model was supplemented by training with 100 high-value samples selected through active learning.

[0071] Performance metrics: Calculate the overall scoring error (|model score - manual review score|), infection risk prediction accuracy (true positive + true negative / total number of samples × 100%), and cardiovascular event early warning lead time (system warning time - clinical diagnosis time).

[0072] Comparison benchmark: Traditional manual assessment (double-blind review by two senior nurses) was used as the scoring benchmark, and existing commercial quality control systems (which only support single-parameter early warning) were used as the early warning performance comparison.

[0073] 3. Early warning response time verification experiment (cycle: 1 month) Simulated scenario design: Three typical risk scenarios were simulated manually, with each scenario repeated 20 times: ① Equipment failure (humanly adjusting the transmembrane pressure of the dialysis machine to 650 mmHg); ② Abnormal water quality (injecting a trace amount of chlorine solution into the water supply line to raise the total chlorine to 0.15 mg / L); ③ Patient hypotension (induced by adjusting the dialysate temperature to 38°C).

[0074] Detection metrics: Record the time from scenario triggering to the medical terminal receiving the warning signal (response time), and compare the response latency of this system with that of existing commercial systems.

[0075] 4. Validation experiment on the effectiveness of the closed-loop improvement mechanism (duration: 3 months) Intervention measures: The experimental group adopted the personalized improvement plan generated by this system, while the control group adopted the traditional empirical plan (such as setting the ultrafiltration rate at 5% of body weight / h).

[0076] Monitoring indicators: The Kt / V target achievement rate (≥1.2), complication rate (hypotension, infection, coagulation), equipment calibration qualification rate, and water quality compliance rate of the two groups of patients were recorded weekly; the improvement effectiveness rate and recurrence rate of similar problems in the experimental group were calculated.

[0077] Statistical methods: SPSS 26.0 was used for data analysis. Quantitative data were expressed as mean ± standard deviation (s), and t-tests were used for comparisons between groups. Categorical data were expressed as [n(%)], and χ² tests were used for comparisons between groups. 2 The test was performed, and P < 0.05 was considered statistically significant.

[0078] 5. Multi-center comparison functional verification experiment (duration: 1 month) Data exchange: Our organization uploaded 300 treatment data to the regional cloud platform and received concurrent data (including equipment, water quality, quality control scores, etc.) from 5 hospitals of the same level.

[0079] Verification metrics: Check the integrity of data transmission (missing rate ≤0.1%), the accuracy of Z-score standardization (consistency with platform calculation results ≥99%), and the rationality of differentiated suggestions (blindly reviewed by 3 quality control experts, with a consistency rate ≥85% to be considered qualified).

[0080] (iv) Experimental Results and Analysis 1. Data acquisition accuracy verification results

[0081] Analysis: The relative error of all detection items in this system is ≤2%, and the accuracy rate is ≥97.5%. Among them, the accuracy rate of equipment parameters and core water quality indicators exceeds 99%, which is significantly higher than that of traditional manual recording (accuracy rate 85%-90%), meeting the high-precision collection needs of clinical practice.

[0082] 2. Performance Validation Results of the AI ​​Evaluation Model

[0083] Analysis: The overall scoring error of this system is significantly lower than that of existing commercial systems, the accuracy of infection risk prediction is improved by 20.8 percentage points, and the advance warning time for cardiovascular diseases is extended by 9.4 minutes, which can intervene in potential complications in advance; the consistency of indicator weights has passed the test, and the evaluation results are scientific and reliable.

[0084] 3. Verification results of early warning response time

[0085] Analysis: The response time of this system in all three scenarios is ≤2s, which is 5-9s shorter than that of existing commercial systems. This is thanks to edge-cloud collaborative processing and high-speed bus transmission, which saves time for clinical emergency response.

[0086] 4. Results of the effectiveness verification of the closed-loop improvement mechanism

[0087] Analysis: All quality control indicators in the experimental group were significantly better than those in the control group. The Kt / V compliance rate increased by 22.8 percentage points, the complication rate decreased by 12 percentage points, the improvement effectiveness exceeded 85%, and the recurrence rate of similar problems was only 5%, proving that the closed-loop improvement mechanism can achieve continuous improvement in quality control level.

[0088] 5. Verification results of multi-center comparison function Data transmission integrity: The missing rate of 300 treatment data was 0.03%, which meets the requirements for clinical data traceability; Standardization accuracy: The Z-score calculation results of this system and the regional platform are 99.6% consistent, with no statistical difference (P>0.05). Reasonableness of recommendations: The blind evaluation of 20 differentiated recommendations by 3 experts showed a 90% consensus rate. Among them, after the implementation of the recommendation to "optimize the sand filter tank replacement cycle to 15 days", the water conductivity compliance rate of our organization increased from 98.0% to 99.8%.

[0089] The above embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of protection of the present invention. Any non-substantial changes and substitutions made by those skilled in the art based on the present invention shall fall within the scope of protection claimed by the present invention.

Claims

1. A hemodialysis quality control and improvement system, characterized in that, It includes a multi-source data acquisition layer, a data processing and integration layer, an intelligent quality control and evaluation layer, a closed-loop improvement execution layer, and a multi-terminal interaction layer. Each layer achieves bidirectional communication through an Ethernet bus or a CAN bus, with a communication latency of ≤100ms. The multi-source data acquisition layer includes a device parameter acquisition module, a patient physiological monitoring module, a water quality monitoring module, an operation process recording module, and an environmental sensing module. The device parameter acquisition module establishes a communication connection with the hemodialysis machine, blood pump, and heparin pump via an RS485 interface, and collects parameters in real time such as blood pump speed (range: 0-500mL / min, accuracy: ±1mL / min), transmembrane pressure (range: -50 to 600mmHg, accuracy: ±1mmHg), ultrafiltration rate (range: 0-2000mL / h, accuracy: ±1mL / h), and dialysate flow rate (range: 300-800mL / min, accuracy: ±5mL / min). The data sampling interval is 100ms. The patient physiological monitoring module integrates a near-infrared spectroscopy monitoring unit and a vital signs sensing unit. The near-infrared spectroscopy monitoring unit collects blood spectral data through a probe that is snap-fitted to the arterial and venous circulation catheters. The probe uses a near-infrared light source with a center wavelength of 1100-2500nm, a spectral resolution of ≤10nm, and a sampling frequency of 50Hz. It can simultaneously analyze the concentrations of urea nitrogen (detection range: 1.7-36mmol / L, accuracy: ±0.1mmol / L) and creatinine (detection range: 44-1330μmol / L, accuracy: ±5μmol / L). The vital signs sensing unit collects the patient's systolic blood pressure (range: 60-200mmHg, accuracy: ±2mmHg), heart rate (range: 40-180 beats / min, accuracy: ±1 beat / min), and activated clotting time (range: 60-400s, accuracy: ±5s) through a wristband sensor, with a data update frequency of 1Hz. The water quality monitoring module includes a multi-parameter sensor group deployed in the pretreatment pipeline (after sand filtration and carbon filtration) and the reverse osmosis water supply pipeline. The conductivity sensor has a measurement range of 0-50 μS / cm (accuracy: ±0.1 μS / cm), the total chlorine sensor has a detection limit of ≤0.01 mg / L (accuracy: ±0.005 mg / L), the pressure sensor has a range of 0-1 MPa (accuracy: ±0.01 MPa), and the temperature sensor has a range of 0-100℃ (accuracy: ±0.1℃). Each sensor has an automatic calibration function (calibration cycle: 7 days). The operation process recording module consists of a high-definition camera (1080P resolution, 25fps frame rate) and a fingerprint reader (recognition accuracy ≤50ppm, response time ≤1s). It uses the YOLOv5 image recognition algorithm to determine the compliance of operation steps in real time (such as 12 key operations such as dialyzer installation and pipeline venting), and binds the operator's identity to the operation record. The environmental sensing module uses an integrated sensor to collect data on room temperature (range: 10-40℃, accuracy: ±0.5℃), power frequency noise intensity (range: 30-100dB, accuracy: ±1dB), and air cleanliness (≥0.5μm particle count, unit: particles / L), with a sampling interval of 5 minutes. The data processing and integration layer includes a data standardization module, an error layering correction module, a multimodal fusion module, and edge computing nodes; The data standardization module adopts the HL7FHIR data exchange standard to convert multi-source data such as equipment, physiology, and water quality into JSON format. It establishes an association index through the patient's unique identification code (PID) and treatment session number, and the data timestamp accuracy is down to the millisecond level. The error stratification correction module handles data deviation through a triple correction mechanism: ① Instrument hardware error correction: Autocorrelation analysis is performed on 10 consecutive sampling points. When the autocorrelation coefficient |r|≥0.9, it is determined to be a systematic error, triggering the device self-test command and calling historical calibration parameters for compensation; ② Environmental interference correction: The cross-correlation coefficient between environmental data and physiological data is calculated to generate an interference coupling matrix, and an adaptive Kalman filter algorithm is used to eliminate noise. ③ Individual difference correction: A linear regression model was established based on the patient's treatment data over the past 3 months to correct for physiological parameter deviations caused by underlying diseases; The multimodal fusion module employs a convolutional neural network (Attention-CNN) with a fusion attention mechanism. It first extracts temporal features such as device parameters and physiological data through a 1D convolutional layer (kernel size: 3×3, stride: 1), then extracts static features such as water quality and operation through a fully connected layer, and finally dynamically allocates feature weights through an attention weight matrix (dimension: 64×128), ultimately fusing them into a 128-dimensional comprehensive feature vector. The edge computing node uses an ARM Cortex-A9 processor to perform real-time data noise reduction, preliminary identification of abrupt signals (such as a sudden drop in blood pressure exceeding 30 mmHg), and key parameter caching (cache capacity: ≥10GB) locally. It only uploads abnormal data and fused feature vectors to the cloud, with a transmission latency of ≤50ms. The intelligent quality control evaluation layer includes a dynamic quality indicator library, a multi-sub-model AI evaluation system, a risk warning module, and a multi-center comparison module; The dynamic quality indicator library covers 23 core indicators in 4 categories, and is synchronized in real time with the national "Quality Control Standard for Hemodialysis" via API interface. The key indicators and standards are as follows: equipment calibration pass rate ≥98%, dialysis adequacy Kt / V value ≥1.2, coagulation index compliance rate ≥95%, water conductivity (25℃) ≤5μS / cm, and operation compliance rate ≥99%. The indicator weights are determined by the analytic hierarchy process: equipment status (25%), patient physiological indicators (35%), water quality indicators (20%), and operation compliance (20%). The multi-sub-model AI evaluation system includes a basic evaluation model and specific prediction sub-models: ① The basic evaluation model is based on the Attention-DNN architecture, which takes a comprehensive feature vector as input to generate a comprehensive quality score of 0-100 and a blood status evaluation matrix (dimension: 5×5, covering dimensions such as clearance efficiency and circulation stability); ② The infection risk prediction sub-model adopts the matter-element extensional weight theory, takes 8 parameters as input, such as the dialysis fluid bacterial culture results and the patient's white blood cell count, and outputs an infection probability of 0-100%; ③ The cardiovascular complication prediction sub-model integrates active learning and contrastive learning algorithms, extracts patient representations through a 5-layer fully connected encoder (number of nodes: 1024→512→256→128→64, activation function: ReLU), and uses cosine distance to screen high-value positive and negative samples to predict the risk of cardiovascular events. The risk warning module divides the assessment results into three levels of warning: Level 1 warning (score ≥ 90 points, no risk) is only recorded and archived; A Level 2 warning (70-89 points, potential risk) triggers a local audio-visual alert and pushes a list for manual review to the medical staff terminal; A Level 3 warning (<70 points, high risk) will trigger an emergency alarm, freeze the dialysis machine's operating permissions, and simultaneously push emergency advice to the medical staff terminal and the on-duty mobile phone. The multi-center comparison module receives quality control data from peer institutions through the regional cloud platform, performs horizontal comparison using Z-score standardization, and generates institution rankings, indicator difference heatmaps, and differentiated improvement suggestions (such as recommending optimization of the pretreatment process when water quality indicators are lagging behind). The closed-loop improved execution layer includes a personalized solution generation module, a full-process execution tracking module, and an effect backtracking and optimization module; The personalized solution generation module outputs improvement instructions based on quality scores and risk levels: for equipment-related issues, it generates equipment debugging instructions that include calibration time, parameter thresholds, and operation steps. For treatment parameters, a fractional sliding mode control algorithm is used to calculate the optimal ultrafiltration rate (sliding surface: s=e+λD^αe, α∈[0.1,0.5]), and the dialysate formulation is dynamically adjusted in conjunction with changes in patient weight and serum potassium concentration; for anticoagulation protocols, the heparin dosage adjustment range is recommended based on the rate of change in activated clotting time (step size: 50U / time). The full-process execution tracking module collects the execution status of improvement measures in real time through IoT chips, records parameters of key operations such as ultrafiltration rate adjustment at the 10ms level, and generates operation trajectory curves; it automatically upgrades the warning level for instructions that are not executed on time (timeout ≥ 10min); The effect backtracking optimization module calculates the improvement efficiency based on indicators such as Kt / V value before and after treatment, urea clearance index (URR), and patient blood pressure fluctuation range (efficacy = |improved indicator - standard value| / |improved indicator - standard value| × 100%). When the efficiency is ≥ 70%, the feature vector of the case and the improvement plan are included in the model training set, and the AI ​​evaluation model weights are updated using the gradient descent method (learning rate: 0.001). The multi-terminal interaction layer includes medical terminals, equipment terminals, management terminals and patient terminals, and supports HTTPS encrypted communication; The medical terminal uses a 15.6-inch touch screen and integrates voice interaction (supporting more than 200 commands such as "retrieving PIDXXX quality control report") and a thermal printing module, which can automatically generate weekly equipment inspection reports (including calibration due reminders). The equipment terminal is embedded in the dialysis machine control panel, which displays real-time operating parameters, error status and improvement instructions, and supports manual confirmation and parameter fine-tuning. The management terminal has data visualization functions (line chart, bar chart), which can statistically analyze the quality control compliance rate and the distribution of early warning events for any time period, and generate regional quality control analysis reports. The patient terminal is implemented through a mini-program, which displays treatment progress, historical quality control scores, medical order reminders, and a satisfaction feedback portal.

2. The system according to claim 1, characterized in that, The water quality monitoring module also includes a transfer tank status monitoring unit. The transfer tank is made of food-grade 316L stainless steel and has a honeycomb-shaped diversion isolation block (thickness: 5mm, aperture: 2mm) inside. The monitoring unit includes an immersion liquid level sensor (measuring range: 0-2m, accuracy: ±1mm) and a laser particle size sensor (detection range: 0.1-100μm), which respectively collect the liquid level in the tank and the impurity concentration at the drain outlet. When the liquid level is <0.3m, a water replenishment reminder is triggered. When the impurity concentration is >5mg / L or after replacing the sand filter tank or resin tank, a 30-minute pipeline flushing command is automatically generated (flushing flow rate: 800mL / min).

3. The system according to claim 1, characterized in that, The probe housing of the near-infrared spectroscopy monitoring unit is made of 316L stainless steel (wall thickness: 2mm). The inner diameter of the blood channel matches that of the dialysis tubing (Φ4-6mm). Coaxial near-infrared emitting and receiving holes (aperture diameter: 3mm) are provided on both sides of the channel. The apertures are covered with quartz glass with a light transmittance of ≥95% (thickness: 1mm). Spectral data is acquired through vertical transmission. The probe and tubing are connected by a quick-connect interface and have an IP67 waterproof rating.

4. The system according to claim 1, characterized in that, In the cardiovascular complication prediction sub-model of the intelligent quality control assessment layer, the active learning unit filters samples through a positive and negative sample selection rule: for the normalized patient representation, the angle between the sample and other samples is calculated, and the upper quartile of the angle between samples with the same label is included in the positive sample set, and the lower quartile of the angle between samples with different labels is included in the negative sample set; the contrastive learning unit constructs a loss function through the cosine distance of positive samples (target: ≥0.8) and the cosine distance of negative samples (target: ≤0.2), and updates the encoder parameters once every 50 samples are trained.

5. The system according to claim 1, characterized in that, The edge computing node also integrates an edge storage module and a resume transmission function after network outage. In the event of a network outage, it can locally store ≥7 days of original data and retransmit it in the order of timestamps after the network is restored, with a retransmission rate of ≥10MB / s to ensure data integrity.

6. The system according to claim 1, characterized in that, The medical terminal is also equipped with an emergency treatment knowledge base module, which has built-in standardized treatment procedures for 12 common risks such as low blood pressure, coagulation, and abnormal water quality. When a level 3 warning is triggered, a pop-up window will automatically display the corresponding treatment steps (including operation diagrams), and it supports one-click calling of the dialysis room director.

7. A quality control method for hemodialysis, characterized in that, Applying the system according to any one of claims 1-6 includes the following steps: S1. Synchronous Acquisition and Preprocessing of Multi-Source Data: S11. Equipment parameter acquisition: The parameters such as blood pump speed and transmembrane pressure are acquired from the dialysis machine, blood pump and other equipment via RS485 interface. The sampling frequency is 10Hz and abnormal values ​​that exceed the range (such as transmembrane pressure > 600mmHg) are automatically filtered. S12. Patient physiological data acquisition: The near-infrared spectral probe acquires blood spectral data in real time (50Hz), and the wristband sensor acquires vital signs (1Hz), which are then linked to the patient's electronic medical record via PID; S13. Water quality data acquisition: Sensors at each node of the pipeline synchronously collect indicators such as conductivity and total chlorine (2Hz), and automatically perform calibration every 7 days (using standard solutions: conductivity 10μS / cm, total chlorine 0.1mg / L). S14. Operation and Environmental Data Acquisition: The camera captures the operation process (25fps), the fingerprint reader records the operator's information, and the environmental sensor collects data such as room temperature (1Hz). S2. Data Standardization and Error Correction: S21. Standardization Processing: All data is converted to JSON format using the HL7FHIR standard, and PID, treatment session, and timestamp index are added; S22. Stratification error correction: ① Instrument error: Calculate the autocorrelation coefficient of 10 consecutive sampling points, and trigger self-check and compensation when |r|≥0.9; ② Environmental interference: Calculate the cross-correlation matrix between environmental and physiological data, and reduce noise through adaptive Kalman filtering; ③ Individual differences: Call the patient's historical model to correct physiological parameters; S3. Multimodal data fusion and feature extraction: S31. Preliminary feature extraction: For time-series data, a 1D convolutional layer (3×3 kernel) is used to extract local features, and for static data, a fully connected layer is used to extract global features; S32. Attention Fusion: Weights are assigned to the two types of features using an attention weight matrix (64×128), and after noise reduction by fractional integral (α=0.3), a 128-dimensional comprehensive feature vector is generated. S33. Edge preprocessing: Edge computing nodes locally identify parameter mutations (such as a sudden drop in blood pressure >30mmHg) and only upload abnormal data and feature vectors to the cloud; S4. Intelligent Quality Assessment and Risk Warning: S41. Comprehensive score calculation: Input the feature vector into the Attention-DNN model, and output a comprehensive score of 0-100 points by combining the index weights; S42. Specific Risk Prediction: The infection risk sub-model takes 8 parameters as input and outputs the infection probability, while the cardiovascular risk sub-model outputs the cardiovascular event risk through active-contrast learning. S43. Warning Judgment and Output: Trigger the corresponding level of warning based on the score and risk probability. Level 3 warning will simultaneously freeze the device operation permissions. S44. Multi-center comparison: Regularly upload data to the regional platform and receive horizontal comparison results and suggestions for differentiation; S5. Generation and Implementation of Personalized Improvement Plans: S51. Solution Formulation: Prioritize based on scores; for scores <70, automatically generate instructions for equipment debugging, parameter adjustment, etc.; for scores 70-89, generate manual review suggestions. S52. Command Push: The solution is pushed to the corresponding terminal through the multi-terminal interaction layer, and the operation guide is displayed simultaneously on the medical terminal; S53. Execution Tracking: Real-time collection of execution status, recording of key parameters at the 10ms level, and automatic escalation of warnings for instructions not executed within the timeout period; S6. Effect Retrospective and Model Optimization: S61. Effectiveness Assessment: After treatment, the improvement rate of indicators such as Kt / V value and URR is calculated. An effective rate of ≥70% is judged as an effective case. S62. Model Update: Incorporate the feature vectors, evaluation results, and improvement plans of valid cases into the training set, and update the AI ​​model parameters using gradient descent (learning rate 0.001); S63. Indicator Database Synchronization: Automatically synchronizes national quality control standards monthly, updating indicator standards and weights.

8. The method according to claim 7, characterized in that, The calculation process of the comprehensive score in step S41 is as follows: the judgment matrix is ​​constructed using the analytic hierarchy process, and the index weights are determined after passing the consistency test (CR < 0.1). The score of each index = (measured value / standard value) × 100 × index weight. The comprehensive score = Σ score of each index, where the score of Kt / V is calculated in segments: ≥1.4 gets full marks, 1.2-1.39 is scored proportionally, and <1.2 gets 0 marks.

9. The method according to claim 7, characterized in that, The process of generating the ultrafiltration rate adjustment command in step S51 includes: ① Input parameters: real-time patient weight, pre-treatment weight, serum potassium concentration, and blood pressure change rate; ② Algorithm calculation: constructing a sliding surface s=e+0.5D^0.3e through a fractional sliding mode control algorithm, where e is the deviation between the actual ultrafiltration rate and the target value; ③ Parameter output: outputting the optimal ultrafiltration rate according to the control law u=-ksign(s), and dynamically calibrating the sliding surface parameters every 30s based on the patient's blood pressure response.

10. The method according to claim 7, characterized in that, Step S3 also includes a data amplification step: the fused original features are amplified using a single-feature randomization method. When the number of positive samples (with complications) is less than the number of negative samples, the remaining features are fixed, and key features such as infection risk and blood pressure are randomized within ±5% until the number of positive and negative samples is balanced. The amplified data is labeled "amplified" to distinguish it from the original data. The input parameters of the infection risk prediction sub-model in step S42 include dialysate bacterial colony count (CFU / mL), endotoxin content (EU / mL), and patient white blood cell count (×10). 9 The correlation was calculated using the matter-element extensional weight theory, with a correlation of ≥0.6 indicating a high risk of infection. The parameters included: C-reactive protein (mg / L), dialysis duration (months), history of underlying diseases (present / absent), puncture site condition (normal / red and swollen), and history of infection in the past 3 months (present / absent).

Images