Blood pressure early warning system for hemodialysis treatment based on wearable monitoring equipment

Abstract

The invention discloses a blood pressure early warning system in hemodialysis treatment based on a wearable monitoring device, comprising: a wearable monitoring device, which comprises an STM32F103C8T6 core board, an MQ-R-12 heart rate and blood pressure module, a DS18B20 temperature sensor, an ESP8266WIFI transmission module, an OLED12864 display module, an LED single indicating lamp, a buzzer, and a Type-c interface; the hemodialysis treatment module adopts a front-end and back-end Web framework of Vue + Django, and a directive tagging blood pressure abnormity prediction algorithm is embedded in the hemodialysis treatment module; real-time, efficient and rapid data transmission between the wearable monitoring device and the hemodialysis treatment module adopts an MQTT internet transmission protocol and an EMQX information server; the medical staff can conveniently operate and record the related treatment information of the patient, monitor and predict the abnormal blood pressure in real time, and visually display the abnormal result, so that the medical staff can be timely reminded to carry out corresponding clinical treatment.

Description

Technical Field

[0001] This invention relates to the field of blood pressure monitoring and early warning technology, and more specifically, to a blood pressure early warning system for hemodialysis treatment based on wearable monitoring devices. Background Technology

[0002] End-stage renal disease (ESRD) is a major chronic disease that threatens human health and consumes significant healthcare resources. Hemodialysis is the primary alternative treatment. However, various complications during hemodialysis can cause immense suffering and even death for patients. Hypotension is a common and serious complication of hemodialysis, occurring in 20%-50% of maintenance hemodialysis (MHD) patients.

[0003] Currently, the traditional method of manually measuring blood pressure at regular intervals in clinical practice presents a series of challenges, including low monitoring frequency, heavy workload, and delayed detection of abnormalities. Furthermore, the diversity of data modalities, the high complexity of disease associations, and the unclear pathogenesis of patients make risk assessment in hemodialysis difficult; difficulties in predicting risks during patient treatment, insufficient effective intervention by medical staff, and unsatisfactory cost-effectiveness lead to challenges in the comprehensive prevention and treatment of hypotension complications in hemodialysis; and the lack of standardized evaluation indicators, the absence of a unified management system, and insufficient demonstration and application by various institutions make the evaluation and management of IDH in hemodialysis difficult.

[0004] Meanwhile, due to the complex etiologies of hypotension during hemodialysis, clinical prediction relies heavily on experience and subjective judgment, resulting in inherent uncertainties and limitations. In medical research, traditional statistical methods such as logistic regression have historically been used to analyze and build predictive models. However, the rapid development of big data and artificial intelligence (AI) technologies in the medical field in recent years has provided new opportunities to improve the treatment of adverse complications in patients with end-stage renal disease. Constructing hypotension prediction models based on machine learning and deep learning methods has become a research hotspot in the field of hemodialysis. Compared to traditional statistical methods, these methods can better uncover potential patterns and regularities in clinical data, providing more reliable support for clinical decision-making.

[0005] However, existing predictive models also have some problems, such as only considering the degree of blood pressure drop while ignoring machine parameter prediction, failing to comprehensively consider factors such as medication use, accompanying clinical symptoms, and whether medical intervention is needed. More importantly, most of these methods directly model data from all patients, ignoring individual differences and heterogeneity among patients, which limits the model's generalization ability and clinical application effectiveness.

[0006] Therefore, clinicians need to identify risk factors for hemodialysis complication IDH, elucidate the hemodialysis risk spectrum, accurately predict hemodialysis risks and implement effective interventions, achieve comprehensive evaluation of hemodialysis, and form an integrated management pathway. To address these issues, we propose a management and early warning system that integrates wearable monitoring devices, a hemodialysis treatment management system, and an embedded blood pressure abnormality prediction algorithm.

[0007] In this invention, the Vue framework refers to the Vue.js front-end framework created by Evan You, which can be downloaded and used from the official Vue.js website. The data and interface presented in Vue can be modified through Django. The Django development system refers to a web framework created by a web development team in Lawrence, Kansas, USA. This framework integrates many commonly used functions, making the development process more efficient and convenient. It can be downloaded and used from the official Django website, and the data and interface presented in Vue can be modified within this development system. The blood pressure abnormality prediction algorithm refers to the K-nearest neighbor clustering model, Hypergraph Neural Network (HGNN), and Transformer deep learning model. Summary of the Invention

[0008] Existing technologies face challenges in the comprehensive prevention and treatment of hypotension complications during hemodialysis. The lack of standardized evaluation indicators, the absence of a unified management system, and insufficient demonstration applications by various units contribute to the difficulty in evaluating and managing IDH during hemodialysis.

[0009] To address the above problems, this invention provides a blood pressure early warning system for hemodialysis treatment based on a wearable monitoring device, the hardware of which includes:

[0010] The wearable monitoring device includes: STM32F103C8T6 core board, MQ-R-12 heart rate and blood pressure module, DS18B20 temperature sensor, ESP8266 WIFI transmission module, OLED12864 display module, LED single indicator light, buzzer, and Type-C interface.

[0011] The STM32F103C8T6 core board is equipped with peripheral interfaces, which can be connected to the MQ-R-12 heart rate and blood pressure module, DS18B20 temperature sensor, ESP8266WIFI module, OLED12864 display, LED single indicator light and buzzer.

[0012] Based on the aforementioned hardware, this system includes the following functional modules in terms of software functionality:

[0013] The system includes a hemodialysis treatment module and a real-time blood pressure monitoring and visualization module. The hemodialysis treatment module comprises: a nurse mobile terminal unit, a doctor PC terminal unit, and a blood pressure abnormality alarm system unit.

[0014] The specific workflow of the above functional modules is as follows:

[0015] S1: The nurse mobile terminal and doctor PC terminal of the hemodialysis treatment module collect routine clinical dialysis data. The data of the dialysis machine, including arterial pressure, venous pressure, conductivity, actual blood flow, average blood flow, dialysate flow, and dialysate temperature, are transmitted to the hemodialysis treatment module via wired transmission.

[0016] S2: The hemodialysis treatment module receives data from the dialysis machine and wearable device, and simultaneously calls data from the cloud database. The algorithm in the blood pressure abnormality alarm system unit generates a prediction result.

[0017] S3: The prediction results will be displayed and alerted on the nurse's mobile terminal and the doctor's PC terminal in the hemodialysis treatment module. At the same time, the prediction results will be transmitted to the blood pressure real-time monitoring visualization module for display and alert.

[0018] In the preferred embodiment, it includes:

[0019] In step S1, routine dialysis data, including the patient's basic information, predialysis assessment, treatment records, and treatment orders, are sent to the cloud database; the MQ-R-12 heart rate and blood pressure module of the wearable device sends blood pressure and heart rate data to the hemodialysis treatment module in real time via the ESP8266WIFI module.

[0020] The blood pressure abnormality alarm system unit in step S2 includes:

[0021] The information entry module, static feature extraction module, dynamic feature processing module, and risk prediction module are implemented as follows:

[0022] M1: The information entry module is implemented by entering multimodal data such as patient basic information, predialysis measurement data, dialysis measurement data, and data generated by wearable monitoring devices and machines during dialysis. Among them, patient basic information, predialysis measurement data, and dialysis measurement data are static data and are denoted as S, while data generated by wearable monitoring devices and machines during dialysis are dynamic data and are denoted as X.

[0023] M2: The static feature processing module uses a hypergraph neural network for feature processing, implemented as follows:

[0024] M2-1: A hypergraph G is a triple, denoted as G=<V,E,W>, where V represents the set of nodes of the hypergraph, E represents the set of hyperedges of the hypergraph, and each hyperedge is a subset of the vertex set;

[0025] M2-2: A KNN model is used for hyperedge construction. The KNN model parameter N is set, where N represents the number of samples clustered together. Basic patient information is entered as input features, and the static data feature attributes of each patient are considered as {s}. (1) ,...,s (m) The KNN model aggregates samples into multiple classes with N samples each. The specific algorithm is as follows:

[0026] M2-2-1: Randomly select a patient as the cluster centroid point μ1∈R;

[0027] M2-2-2: Repeat the following procedure until convergence:

[0028] M2-2-2-1: For each sample μ i ∈R, calculate the distance between the cluster centroids using the following formula:

[0029]

[0030] M2-2-2-2: For each cluster centroid, sort the patients according to their distance from smallest to largest, and then select the N patients with the smallest distance as the nearest neighbors, with each nearest neighbor representing a category;

[0031] M2-2-2-3: The category labels of N patients are {y} 1N ,...,y nN The patient categories are determined by voting. The number of votes for category C is calculated using the following formula:

[0032]

[0033] In the formula, C represents the class with the closest distance among N samples, I(·) represents the indicator function, and the iteration is repeated. c This represents the number of votes N patients received in category C. The higher the number of votes, the closer the N patients are to category C.

[0034] M2-2-2: Until the moving distance between two iterations is less than the preset threshold, the threshold is set to 0.01 in the experiment. Each class C represents a set of patients with similar conditions, that is, a class is a hyperedge.

[0035] After aggregating static data using the KNN algorithm, patients with the same features will be constructed in the same hyperedge, and then all patients and hyperedges will be represented by the correlation matrix H.

[0036] M2-3: The graph structure of the hypergraph G is represented using an incidence matrix H. H is a matrix of dimension |V|×|E|, the specific values ​​of which are determined by h(v,e). h(v,e) represents the specific values ​​of the coordinates (v,e) in the H matrix. h(v,e) is defined as:

[0037]

[0038] For a patient v∈V, its degree is defined as:

[0039]

[0040] For a hyperedge e∈E, its degree is defined as:

[0041]

[0042] For a degree d∈D, the hypergraph association matrix H and patient feature attributes are input into the HGNN layer to obtain the node output features, where a hyper-edge convolutional layer f(S,W,θ) is established, as shown in the following formula:

[0043]

[0044] Among them, S (l) ∈R N×C The hypergraph represents the patient feature attributes at layer l, and S is initialized. (0) =S, where σ() represents the nonlinear activation function, and W is the super-activation function. Figure 3 W, D in the tuple e and D v Let H represent the diagonal matrices for edge degree and patient degree, respectively, with corresponding values ​​d(e) and d(v), where e and v represent the positions of the values ​​in the matrix. T Let S be the transpose of the hypergraph incidence matrix H, and let S be the initial node features for matrix multiplication with the diagonal matrices of edge degree and patient degree. (l) Through the learnable filter matrix θ (l) The HGNN layer extracts higher-order correlations on the hypergraph by transforming the patient-hyperedge-patient transmission method. The new node feature matrix is ​​then multiplied by a trainable weight matrix on the left before being fed into a softmax activation layer to obtain the output, as shown in the following formula:

[0045] O = softmax(W1S) (l+1) )

[0046] Where W1 is the trainable weights and O is the output, which is the final embedding of the hypergraph;

[0047] M3: The dynamic feature processing module is implemented as follows:

[0048] M3-1: Let the input sequence be dynamic data X = [x1, x2, ..., x...]. n ], where x i Representing the i-th element in this sequence, calculate the query vector Q, key vector K, and value vector A, which are obtained by performing a linear transformation on the input sequence:

[0049] Q = XW Q

[0050] K = XW K

[0051] A = XW A

[0052] Among them, W Q W K W A The weight matrix is ​​then learned, and the attention score is calculated using the softmax() function:

[0053]

[0054] Among them, K T d represents the transpose of the key vector matrix. k The dimension of the key vector, divided by Used to scale attention scores;

[0055] Repeat step M3-1 h times, where h is the number of heads, and concatenate the Attention outputs from each head before performing a linear transformation to obtain a dimension of h×d. A The matrix, where d A It is the dimension of the value vector, calculated using the following formula:

[0056] MultiHeadAttention(X)=Concat(head1,head2,...,head h W O

[0057] Among them, head i =Attention(XW i Q XW i K XW i A ), W i Q W i K W i A To obtain the corresponding weight matrix, Concat() concatenates the attention values ​​of each head to get MultiHead, WO To output the projection matrix;

[0058] M3-2: Construct a fully connected feedforward network, the formula is:

[0059] FFN(X) = max(0, MultiHeadW m +b1)W f +b2

[0060] Among them, W m W f Let be the weights, b be the bias, and max() be the maximum value function. The concatenated result MultiHead(Q,K,A) is projected back to the original dimension d using a trainable linear mapping matrix. model This yields the final output FFN of the multi-head attention, where the output is the intermediate encoded information matrix, representing the features of the input time series data.

[0061] M4: The risk prediction module calculates the risk probability results, implemented as follows:

[0062] M4-1: The final embedding matrix O∈R of the hypergraph in the static feature extraction module. N×F Perform a transpose, where N is the number of patients and F is the dimension of the static features, so that rows represent features and columns represent patients. The transposed matrix O T ∈R F×N With dynamic data matrix FFN∈R M×N Maintaining consistency across the feature dimensions, where M is the dimension of the dynamic features, and then horizontally concatenating it with the dynamic data matrix to form a new feature matrix Z∈R. (F+M)×N ;

[0063] M4-2: Z is fed into the decoder of the Transformer model as input, and the output vector of the decoder at each time step is represented as Z. t Where t is the index of the time step, and then a vector Y is obtained by transformation through a linear layer. t The formula for the linear layer is:

[0064] Y t =Z t W y +b y

[0065] Among them, W y It is the weight matrix of the linear layer, b y It is the bias vector;

[0066] M4-3: Output vector Y of the linear layer tThe system is transformed into a probability distribution through a softmax layer, and the final output is the risk probability P of hypotension for each patient. The softmax function is calculated as follows:

[0067]

[0068] in, The original score of the i-th sample in the model output belongs to the c-th class, where C is the total number of classes. In this algorithm, it is 2, which means risky or no risk. The closer the probability value is to 1, the higher the risk of hypotension in the patient during dialysis predicted by the model; the closer the probability value is to 0, the lower the risk of hypotension in the patient during dialysis predicted by the model.

[0069] In the preferred mode, the treatment record includes: dialysis treatment time, dialysis treatment method, preset ultrafiltration volume, total replacement fluid volume, dialysis wave flow rate, and Na+. + Ion concentration, oxygen uptake, and anticoagulation methods.

[0070] In the preferred approach, the hemodialysis treatment module adopts the Vue + Django front-end and back-end web framework, which embeds a blood pressure abnormality prediction algorithm, i.e., a blood pressure abnormality alarm system unit. The joint use of Vue and Django is achieved by using the API interface in Django. In Vue, the HTTP client is used to call the API interface provided by Django, and the data from Django is displayed on the Vue page interface.

[0071] In the preferred method, the patient's basic information includes: age, gender, dialysis duration, and BMI, totaling 4 data points; the pre-dialysis measurement data includes: height measured on the day of dialysis, weight measured on the day of dialysis, percentage of weight loss during dialysis, percentage of weight gain compared to the last dialysis session, amount of weight gain, amount of weight loss during dialysis, pre-dialysis heart rate, pre-dialysis diastolic blood pressure, pre-dialysis systolic blood pressure, pulse pressure, and accurate ultrafiltration rate, totaling 11 data points; the sensor and machine data during dialysis include: arterial pressure (AP), venous pressure (VP), real-time blood flow, average blood flow, dialysate flow rate, temperature, and dialysate conductivity, totaling 7 data points; the data generated by wearable monitoring devices and machines during dialysis include: blood pressure and dialysate flow rate.

[0072] In the preferred mode, dynamic data include: arterial pressure, venous pressure, conductivity, actual blood flow, average blood flow, dialysate flow, and dialysate temperature.

[0073] The beneficial effects of this invention are: medical staff can conveniently operate and record relevant patient treatment information; and by constructing a blood pressure early warning system for hemodialysis treatment based on wearable monitoring devices, this invention achieves the technology of predicting the risk of hypotension in advance when patients undergo hemodialysis. It can monitor and predict the occurrence of blood pressure abnormalities in real time and visualize the abnormal results, thereby promptly reminding medical staff to take appropriate clinical measures, solving the problem of not being able to predict the occurrence of hypotension in hemodialysis treatment. Attached Figure Description

[0074] Figure 1 This is a schematic diagram of the modular architecture of the present invention;

[0075] Figure 2 This is an overall flowchart of the present invention;

[0076] Figure 3 This is a schematic diagram of the hypotension prediction algorithm framework in hemodialysis based on artificial intelligence technology of the present invention;

[0077] Figure 4 This is a schematic diagram of the experimental data acquisition method of the present invention;

[0078] Figure 5 This is a schematic diagram of the modeling process for the hypotension event prediction model of the present invention;

[0079] Figure 6 This is a structural diagram of the OLED12846 display module of the present invention;

[0080] Figure 7 This is a structural diagram of the ESP8266 transmission module of the present invention;

[0081] Figure 8 This is a structural diagram of the MQ-R-12 blood oxygen, blood pressure, and heart rate module of the present invention;

[0082] Figure 9 This is a structural diagram of the LED lamp of the present invention;

[0083] Figure 10 This is a structural diagram of the DS18B20 sensor of the present invention;

[0084] Figure 11 This is a structural diagram of the buzzer of the present invention;

[0085] Figure 12 This is a structural diagram of the STM32F103 core board of this invention. Detailed Implementation

[0086] Example 1:

[0087] like Figure 1As shown, a blood pressure early warning system for hemodialysis treatment based on wearable monitoring devices is presented. The system consists of a software system and a wearable monitoring device. The software system includes a hemodialysis treatment module and a real-time blood pressure monitoring and visualization module. The hemodialysis treatment module includes a nurse's mobile terminal unit, a doctor's PC terminal unit, and a blood pressure abnormality alarm system unit. The wearable monitoring device includes heart rate and blood pressure acquisition modules. The hemodialysis treatment module is a crucial part of the entire blood pressure early warning system. It mainly includes the nurse's mobile terminal unit, the doctor's PC terminal unit, and the blood pressure abnormality alarm system unit. Its core function is to provide medical staff with comprehensive treatment management for patients during hemodialysis treatment and to issue alerts for abnormal blood pressure conditions during hemodialysis treatment. The real-time blood pressure monitoring and visualization module, also known as a monitoring BI screen, acquires data from the hemodialysis treatment module. This module uses the acquired data to perform risk assessment and risk level classification, and visualizes the real-time blood pressure monitoring and early warning data, presenting it in a visual monitoring BI screen format. The nurse's mobile terminal unit plays an indispensable role in the hemodialysis treatment module, mainly used for the management, tracking, and recording of the entire dialysis treatment process for patients. The system performs pre-dialysis assessment, dialysis recording, and post-dialysis assessment for patients during hemodialysis treatment. The doctor's PC unit plays a crucial role in inputting and receiving patient-related information, primarily focusing on the diagnosis and treatment throughout the entire dialysis process. Its main functions focus on the diagnosis and treatment of patients throughout the dialysis treatment process, covering professional aspects such as condition assessment and treatment plan formulation, including issuing treatment orders before hemodialysis treatment and issuing temporary medical orders based on the patient's clinical condition during treatment. The blood pressure abnormality alarm system can predict possible blood pressure abnormalities during hemodialysis treatment. By acquiring patient medical information, pre-dialysis assessment, dialysis recording, post-dialysis assessment, and temporary treatment orders from the hemodialysis treatment module, as well as dialysis machine data generated during hemodialysis treatment, it issues alarms for abnormal blood pressure during hemodialysis treatment and displays this information in the real-time blood pressure monitoring visualization module. The wearable monitoring device mainly includes heart rate and blood pressure acquisition modules. This device can process data collected by biosensors to generate alarms and transmit them to the hemodialysis treatment module. The heart rate and blood pressure acquisition module can collect data such as heart rate and blood pressure in real time during the patient's hemodialysis treatment using biosensors.

[0088] Example 2:

[0089] The hemodialysis treatment module plays a crucial role in the blood pressure early warning system, primarily encompassing a mobile unit for nurses, a PC unit for doctors, and a blood pressure abnormality alarm function. The core function of this module is to assist medical staff in comprehensively managing patients' hemodialysis treatment and to issue alarms when abnormal blood pressure occurs. Within the hemodialysis treatment module, the mobile unit for nurses plays a vital role; its main functions include managing, tracking, and recording the entire process of patient dialysis treatment, such as... Figure 2 As shown. Before dialysis treatment, nurses must perform a routine pre-dialysis assessment for the patient (i.e., Figure 2 The process involves pre-dialysis assessment, and relevant information is accurately recorded on the nurse's mobile terminal unit's pre-dialysis assessment interface, including the patient's basic information, pre-dialysis assessment completion details, and pre-dialysis assessment options. The nurse sets the monitoring frequency through the hemodialysis monitoring interface on the nurse's mobile terminal unit and clicks "Start Monitoring." The real-time data monitored by the wearable monitoring device is transmitted to the hemodialysis treatment module, such as the transmission module of the wearable monitoring device in Example 5. The hemodialysis monitoring interface on the nurse's mobile terminal unit displays the patient's real-time monitoring data. The real-time data monitored by the wearable monitoring device transmitted to the hemodialysis treatment module includes the patient's real-time blood pressure, heart rate, and other data. During dialysis treatment, the nurse needs to closely monitor the patient's physical condition, especially when an alarm sounds on the nurse's mobile terminal unit. The nurse must intervene promptly and enter the intervention record into the nurse's mobile terminal unit. The alarm sound on the nurse's mobile terminal unit is an abnormal blood pressure alarm. When the wearable monitoring device detects an abnormal blood pressure, it transmits the abnormal signal to the nurse's mobile terminal unit and sounds an alarm, as shown in the heart rate and blood pressure module and transmission module of the wearable monitoring device in Example 5. This function is stored in the background of the nurse's mobile terminal unit's operating system within the hemodialysis treatment module and is not displayed in the user interface. After a patient's dialysis treatment, the nurse performs a routine post-dialysis assessment and records relevant post-dialysis information on the nurse's mobile terminal unit's post-dialysis recording interface. These records are crucial for doctors to understand the patient's treatment progress. Within the hemodialysis treatment module, the doctor's PC terminal unit plays a key role. It undertakes the important mission of inputting and receiving relevant patient information, a process involving key steps such as data entry and information retrieval. Doctors comprehensively analyze all of the patient's treatment data to provide a basis for developing personalized dialysis treatment plans. Figure 2 As shown. Before dialysis treatment, doctors need to develop a suitable treatment plan for the patient and fill out treatment records based on the routine pre-dialysis assessment information transmitted by the nurse's mobile terminal. These decisions require doctors to comprehensively consider the patient's individual differences and the severity of their condition. The aforementioned treatment records include determining the dialysis treatment time, dialysis treatment method, preset ultrafiltration volume, total replacement fluid volume, dialysis wave flow rate, and Na+. + Ion concentration, oxygen uptake, and anticoagulation methods, etc.

[0090] The clinical pre-dialysis assessment information transmitted from the aforementioned nurse mobile terminal is entered by the nurse on the mobile terminal, and the record is simultaneously written to a cloud database in the background for storage. This cloud database is shared by the nurse mobile terminal, doctor PC terminal, and blood pressure abnormality alarm system unit of the hemodialysis treatment module. The doctor PC terminal, based on the clinical pre-dialysis assessment record information transmitted from the nurse mobile terminal, triggers the retrieval of the pre-dialysis assessment record information stored in the cloud database by the doctor clicking to view the pre-dialysis assessment, and then displays it on the doctor PC terminal's page. During dialysis treatment, the doctor PC terminal can receive real-time patient data from wearable monitoring devices and the dialysis machine to promptly understand the patient's physical condition and make necessary adjustments to the treatment plan. The data received in real-time from the wearable monitoring device by the doctor PC terminal is data measured by the wearable monitoring device, such as blood pressure and heart rate. For example, the heart rate and blood pressure module and transmission module of the wearable monitoring device in Example 5 are transmitted to the hemodialysis treatment module, and the doctor PC terminal can view and display the data on the doctor PC terminal interface by calling the cloud database. The data received in real-time from the dialysis machine by the doctor's PC unit refers to machine data generated during dialysis treatment. This includes: arterial pressure, venous pressure, conductivity, actual blood flow, average blood flow, dialysate flow, and dialysate temperature. The data is transmitted to the hemodialysis treatment module via a wired network using the dialysis machine's built-in data acquisition and transmission module, and stored in the module's cloud database. During dialysis treatment, doctors can issue temporary medical orders for patients on their PC unit, referencing dialysis machine data and wearable monitoring device data, and filling out the corresponding temporary medical order interface. After dialysis treatment, doctors can also view the routine post-dialysis assessment information recorded by the nurse's mobile unit on their PC unit to further understand the patient's treatment effect and provide a reference for subsequent treatment. The aforementioned doctor's PC terminal unit views routine post-dialysis assessment information recorded by the nurses. This is achieved by the doctor clicking to view the post-dialysis assessment, triggering the retrieval of post-dialysis assessment records stored in the nurses' mobile terminal unit in the cloud database, and then displaying them on the doctor's PC terminal unit page. The blood pressure abnormality alarm system unit's function is to acquire patient medical information, pre-dialysis assessment, dialysis records, post-dialysis assessments, and temporary treatment orders from the hemodialysis treatment module, as well as dialysis machine data generated during hemodialysis treatment. It issues warnings in the hemodialysis treatment module and displays them simultaneously in the real-time blood pressure monitoring visualization module. This module can be equipped with any artificial intelligence algorithm.The patient's medical information, pre-dialysis assessment, dialysis records, post-dialysis assessment, and temporary treatment orders in the aforementioned hemodialysis treatment module are data input by the nurse's mobile terminal and the doctor's PC terminal, respectively. The aforementioned medical information, pre-dialysis assessment, dialysis records, and post-dialysis assessment data are patient treatment-related data stored in the cloud database by the nurse's mobile terminal, where medical information refers to the patient's basic information. This module can directly retrieve the patient's treatment-related data from the cloud database. The aforementioned temporary treatment orders are patient treatment-related data stored in the cloud database by the doctor's PC terminal. This module can directly retrieve the patient's treatment-related data from the cloud database. The aforementioned cloud database is shared by the nurse's mobile terminal, the doctor's PC terminal, and the blood pressure abnormality alarm system unit of the hemodialysis treatment module. The dialysis machine data generated by the patient during treatment in the aforementioned hemodialysis treatment module comes from data measured by the dialysis machine, including: arterial pressure, venous pressure, conductivity, actual blood flow, average blood flow, dialysate flow, and dialysate temperature. The aforementioned blood pressure abnormality alarm system unit acquires data measured by the dialysis machine. Data transmission utilizes the dialysis machine's built-in data acquisition and transmission module, connected via a wired network, to transmit machine data to the hemodialysis treatment module and store it in the hemodialysis treatment module's cloud database. The blood pressure abnormality alarm system unit can then directly access real-time dialysis machine data. The aforementioned warning refers to using artificial intelligence technology to predict blood pressure using patient medical information, pre-dialysis assessment, dialysis records, post-dialysis assessments, and temporary treatment orders obtained from the hemodialysis treatment module, as well as dialysis machine data generated during the patient's hemodialysis treatment. The prediction results generated by this module are displayed in the real-time blood pressure monitoring visualization module. The network model of this module can be replaced with any required network model; this module is adaptable to any network model that meets different clinical needs. The warning results generated by this module will be displayed in the real-time blood pressure monitoring visualization module and stored in the cloud database, including: abnormal results, abnormal probabilities, etc. The transmission of the aforementioned warning results to the real-time blood pressure monitoring and visualization module is achieved by connecting the transmission interface of the hemodialysis treatment module to the transmission interface of the real-time blood pressure monitoring and visualization module. This allows data from the hemodialysis treatment module to be transmitted to the real-time blood pressure monitoring and visualization module. Both the transmission interfaces of the hemodialysis treatment module and the real-time blood pressure monitoring and visualization module are located in the background of their respective module systems and are not displayed on the user interface.

[0091] Example 3:

[0092] The real-time blood pressure monitoring visualization module, also known as the monitoring BI dashboard, uses data acquired from the hemodialysis treatment module. This module categorizes the acquired data into risk levels and visualizes real-time blood pressure monitoring alerts, presenting the data as a visual monitoring BI dashboard. The real-time data acquisition for the blood pressure monitoring visualization module comes from the hemodialysis treatment module. This includes patient basic information such as primary disease, age, and dialysis duration; patient treatment records such as weight gain and fluid loss; and abnormal results and probabilities from the alert module. The aforementioned risk level categorization classifies patients into different risk levels: such as low risk, medium risk, high risk, normal, alert, and dangerous, allowing medical staff to intuitively understand the patient's abnormal condition and take appropriate preventative measures. This risk level categorization is based on the abnormal results and probabilities from the blood pressure abnormality alarm system unit within the real-time blood pressure monitoring visualization module. The hypotension early warning monitoring is used by healthcare workers. Risk alert levels are categorized as follows: white (low risk), yellow (medium risk), and red (high risk). These risk alert levels are determined by the abnormal probability of the blood pressure anomaly alarm system unit within the real-time blood pressure monitoring visualization module. The classification rules are: white (low risk) range 0%-40%, yellow (medium risk) range 40%-60%, and red (high risk) range 60%-100%. Basic data display is used by all users, with risk alert levels categorized as follows: green (normal), yellow (alert), and red (dangerous). These risk alert levels are also determined by the abnormal probability of the blood pressure anomaly alarm system unit within the real-time blood pressure monitoring visualization module. The classification rules are: green (normal) range 0%-40%, yellow (alert) range 40%-60%, and red (dangerous) range 60%-100%. The data transmission process for the real-time blood pressure monitoring visualization module to obtain abnormal results and probabilities from the early warning module involves connecting the transmission interface of the hemodialysis treatment module to the transmission interface of the real-time blood pressure monitoring visualization module, enabling data from the hemodialysis treatment module to be transmitted to the real-time blood pressure monitoring visualization module. Other data displayed in the real-time blood pressure monitoring visualization module, such as the patient's primary disease, age, and dialysis duration in basic information, and weight gain and fluid loss data in the patient's treatment records, are all retrieved from the cloud database in the hemodialysis treatment module. The risk grading in this module can be personalized by medical staff, setting different probability ranges for each level based on the actual patient's condition.

[0093] Example 4:

[0094] Figures 6-12This is a system schematic diagram of the various components of a blood pressure early warning system for hemodialysis treatment based on wearable biosensors. This system integrates multiple components, including an STM32F103 core board, heart rate and blood pressure modules, a temperature sensor, a transmission module, a display module, LEDs, and a buzzer, and transmits data to a web server via the MQTT protocol. The device can not only display monitoring data locally but also remotely transmit data via a Wi-Fi module, achieving real-time monitoring and wireless data transmission of human physiological parameters and ambient temperature. It provides users with convenient health monitoring and environmental awareness functions. The following will explain in detail the function of each module and its pins in the schematic diagram, as well as the signal transmission method and process. The STM32F103C8T6 core board is manufactured by STMicroelectronics, a globally renowned semiconductor manufacturer. To more efficiently and conveniently connect the MQ-R-12 blood pressure and heart rate module, the ESP8266 Wi-Fi module, the OLED12864 display, the DS18B20 temperature sensor, and the single LED indicator, this invention expands multiple external interfaces. The functions of the external serial ports used will be described in detail below. Pin 4 (PC15) can be configured to receive signals from a temperature sensor, used to trigger chip algorithms and status monitoring inputs, and has external interrupt and wake-up functions. Pin 7 (PA2) can transmit signals and clocks to the MQ-R-12 blood pressure and heart rate module and the ESP8266WIFI module to enable their operation, and has a reset function; a low level resets the chip. Pin 8 (PA3) can transmit blood pressure and heart rate signals collected by the MQ-R-12 blood pressure and heart rate module to the chip, and is also used for peripheral expansion such as multi-signal synchronous control. Pin 13 (PB0) connects to an external LED, with the chip output signal indicating the device's active status. Pin 14 (PB1) connects to a buzzer, receives signals from the chip, and triggers an alarm in case of an anomaly. Pin 26 (PA9_TX) can be configured with an output, enabling connection to external devices for control and detection, and has special functions when multiplexed. Pin 27 (PA10_RX) is a general-purpose input mode, capable of detecting external levels, such as button status, and has corresponding functions when multiplexed. Pins 34 (PB6) and 35 (PB7) are configured as GPIO outputs, which can be connected to devices such as the OLED12864 to control communication with external devices. Pins 38 (5V) and 40 (3.3V) serve as power input pins, playing a crucial role in ensuring stable power supply to the internal circuitry of the chip. Pin 39 (GND) is used to establish a circuit loop, effectively preventing the impact of poor signal transmission on the chip's performance stability. The MQ-R-12 heart rate and blood pressure module is manufactured by Hanwei Electronics. The RXD and TXD pins of the MQ-R-12 heart rate and blood pressure module are connected to the PA2 and PA3 serial lines of the STM32F103 core board, respectively, thus establishing a data transmission channel. Data interaction with the core board is achieved using the serial port (UART) in an asynchronous communication manner. This connection method ensures that the data collected by the module is accurately transmitted to the STM32F103 core board.The MQ-R-12 heart rate and blood pressure module, as the core data acquisition unit, can accurately sense human blood pressure and heart rate data, and can also output pulse waveforms, vascular microcirculation parameters, and other information in digital signal form for direct use. It works in conjunction with the STM32F103; after data is processed by the STM32F103 core board, it can be transmitted to the OLED12864 display module for intuitive presentation to the user, and can also be remotely monitored via the ESP8266 WIFI module. When data is abnormal, the STM32F103 can trigger an early warning mechanism, whether through a buzzer alarm or flashing indicator lights, to promptly inform the user of any health abnormalities, providing important information for subsequent health management. The DS18B20 temperature sensor is manufactured by Dallas Semiconductor (now part of Maxim Integrated). The DS18B20 temperature sensor's I / O ports are connected to pin PC15 of the STM32F104C8T6 core board as digital signal input and output ports. The DS18B20 temperature sensor can receive commands from the STM32F104C8T6 core board to read the temperature. After receiving the command, the sensor sends the temperature value back via its I / O port. The STM32F104C8T6 core board performs temperature compensation, unit conversion, and other processing upon receiving the data. The processed data can be stored in its internal RAM and displayed on an OLED12864 screen showing the patient's body temperature. It can also be remotely transmitted via the ESP8266WIFI module. The ESP8266WIFI module is a Wi-Fi chip manufactured by Espressif Systems. The ESP8266WIFI module connects to the STM32F103 via a serial port. Its TX and RX pins correspond to relevant pins on the STM32F103 (e.g., receiving data from the STM32F103's PA2 pin), and the two interact using asynchronous communication. This connection method allows for bidirectional data transmission between the two; the STM32F103 can send data to the ESP8266 via serial port, and vice versa. The OLED12864 display is manufactured by Zhongjingyuan Electronics. The OLED12864 receives power by connecting to pin 40 (3.3V) and pin 39 (GND) of the STM32F103 core board via VCC and GND interfaces, respectively. Its clock line (SCL) and signal line (SDA) are connected to pins 35 (PB7) and 34 (PB6) of the core board, respectively. Data transmission with the STM32F103 is achieved via I2C or SPI interface protocols, thus establishing an effective connection between the two. As the system's display module, the OLED12864 primarily visualizes various system data. It can display physiological data such as blood pressure, heart rate, and temperature collected by sensors, as well as system status information such as device connection status and operating mode, making it a key part of the system's feedback mechanism.When the system experiences abnormal data, device connection or disconnection, or other events or status changes, it can promptly display relevant information, allowing users to intuitively understand the system's operational status and the health of the inspector, enabling them to take appropriate measures. The LED is connected to pin 13 of the STM32F103C8T6 core board via PB0, used to control its on / off state. The LED primarily indicates various system states. For example, when the system starts, this LED illuminates, indicating that the system is initializing. When the sensor begins acquiring data, this LED also illuminates, indicating that data acquisition is in progress. The buzzer is manufactured by SEG Electronics. The buzzer is connected to pin 14 of the STM32F103C8T6 core board via PB1, used to receive clock information from the STM32F103C8T6 core board; it only requires a high-level control signal to trigger.

[0095] The above is a detailed introduction to the specific functions and connection methods of each module. The following section will introduce the signal transmission forms and methods of each module.

[0096] After the system powers on, the STM32F103C8T6 core board first performs a self-test, then initializes and configures each peripheral module. The STM32F103C8T6 establishes communication with different peripheral modules through specific pins. The MQ-R-12 heart rate and blood pressure module communicates serially via PA2 (RX) and PA3 (TX) pins; the ESP8266 module also uses PA2 (RX) and PA3 (TX) pins for serial communication; the OLED12864 display uses PA7 (SCL) and PA6 (SDA) pins for I2C communication; the DS18B20 temperature sensor communicates via a single bus through the PC15 pin; and LED indicators are controlled by the PB0 pin. Simultaneously, the STM32F103C8T6's internal RAM / ROM is used to store program code and temporary data. The MQ-R-12 heart rate and blood pressure module and the DS18B20 temperature sensor are powered through their respective power supply pins (VCC and GND or VDD and GND). The physiological parameter signals collected by the MQ-R-12 heart rate and blood pressure module are processed and transmitted to the STM32F103C8T6 in digital form via the RXD (PA2) and TXD (PA3) pins. The DS18B20 temperature sensor converts the detected temperature value into a digital signal and transmits it to the STM32F103C8T6 via the I / O pin (PC15). The STM32F103C8T6 receives and processes the data from the MQ-R-12 heart rate and blood pressure module and the DS18B20 temperature sensor, and then sends the processed data to the OLED12864 display via the I2C interface. The OLED12864 display, powered by VCC (3.3V) and GND pins, receives data and displays it via the SCL (PA7) and SDA (PA6) pins. The processed data can also be sent to the ESP8266 WiFi module via serial port. The ESP8266 module, powered by VCC (3.3V) and GND pins, receives data through the TX (PA2) and RX (PA3) pins, encapsulates the data into WiFi protocol data packets, and finally transmits them through the antenna. During data acquisition or transmission, LED indicators are illuminated under the control of the PB0 pin, providing visual feedback to the user. The STM32F103C8T6 outputs a high level on the PB0 pin to turn on the LED indicators. Furthermore, the STM32F103C8T6 generates a synchronization signal based on sidelink technology, communicating with the ESP8266 module through the PA2 and PA3 pins to achieve synchronous communication between devices. This synchronous communication enables long-distance communication exceeding the CP protection range.

[0097] Example 5:

[0098] A blood pressure early warning system for hemodialysis treatment based on a wearable monitoring device includes the following hardware: a wearable monitoring device consisting of an STM32F103C8T6 core board, an MQ-R-12 heart rate and blood pressure module, a DS18B20 temperature sensor, an ESP8266 WIFI transmission module, an OLED12864 display module, a single LED indicator, a buzzer, and a Type-C interface; the STM32F103C8T6 core board is equipped with a peripheral interface, through which the MQ-R-12 heart rate and blood pressure module, DS18B20 temperature sensor, ESP8266 WIFI module, OLED12864 display, single LED indicator, and buzzer are connected;

[0099] Based on the aforementioned hardware, this system includes the following functional modules in terms of software functionality:

[0100] The system includes a hemodialysis treatment module and a real-time blood pressure monitoring and visualization module. The hemodialysis treatment module comprises: a nurse mobile terminal unit, a doctor PC terminal unit, and a blood pressure abnormality alarm system unit.

[0101] The specific workflow of the above functional modules is as follows:

[0102] S1: The nurse mobile terminal and doctor PC terminal of the hemodialysis treatment module collect routine clinical dialysis data, including patient basic information, predialysis assessment, treatment records, and treatment orders, and send them to the cloud database; at the same time, the dialysis machine data, including arterial pressure, venous pressure, conductivity, actual blood flow, average blood flow, dialysate flow, and dialysate temperature, are transmitted to the hemodialysis treatment module via wired transmission.

[0103] The MQ-R-12 heart rate and blood pressure module of the wearable device transmits blood pressure and heart rate data to the hemodialysis treatment module in real time via the ESP8266 WIFI module;

[0104] S2: The hemodialysis treatment module receives data from the dialysis machine and wearable device, and can also call up data from the cloud database. The algorithm in the blood pressure abnormality alarm system unit generates a prediction result.

[0105] S3: The prediction results will be displayed and alerted on the nurse's mobile terminal and the doctor's PC terminal in the hemodialysis treatment module. At the same time, the prediction results will be transmitted to the real-time blood pressure monitoring and visualization module for display and alert.

[0106] The blood pressure abnormality alarm system unit in the above workflow includes:

[0107] The specific implementations of the information entry module, static feature extraction module, and dynamic risk prediction module are as follows:

[0108] M1: The information entry module is implemented by entering multimodal data such as patient basic information, pre-dialysis measurement data, dialysis measurement data, and data generated by wearable monitoring devices and machines during dialysis. Among them, patient basic information, pre-dialysis measurement data, and dialysis measurement data are static data and are denoted as S, while data generated by wearable monitoring devices and machines during dialysis are dynamic data and are denoted as X.

[0109] M2: The static feature processing module uses a hypergraph neural network for feature processing, implemented as follows:

[0110] M2-1: A hypergraph G is a triple, denoted as G=<V,E,W>, where V represents the set of nodes of the hypergraph, E represents the set of hyperedges of the hypergraph, and each hyperedge is a subset of the set of vertices;

[0111] M2-2: A KNN model is used for hyperedge construction. The KNN model parameter N is set, where N represents the number of samples clustered together. Basic patient information is entered as input features, and the static data feature attributes of each patient are considered as {s}. (1) ,...,s (m) The KNN model aggregates samples into multiple classes with N samples each. The specific algorithm is as follows:

[0112] M2-2-1: Randomly select a patient as the cluster centroid point μ1∈R;

[0113] M2-2-2: Repeat the following procedure until convergence:

[0114] M2-2-2-1: For each sample μ i ∈R, calculate the distance between the cluster centroids using the following formula:

[0115]

[0116] M2-2-2-2: For each cluster centroid, sort the patients according to their distance from smallest to largest, and then select the N patients with the smallest distance as the nearest neighbors, with each nearest neighbor representing a category;

[0117] M2-2-2-3: The category labels of N patients are {y} 1N ,...,y nN The patient categories are determined by voting. The number of votes for category C is calculated using the following formula:

[0118]

[0119] In the formula, C represents the class with the closest distance among N samples, I(·) represents the indicator function, and the iteration is repeated. cThis represents the number of votes N patients received in category C. The higher the number of votes, the closer the N patients are to category C.

[0120] M2-2-2: Until the moving distance between two iterations is less than the preset threshold, the threshold is set to 0.01 in the experiment. Each class C represents a set of patients with similar conditions, that is, a class is a hyperedge.

[0121] After aggregating static data using the KNN algorithm, patients with the same features will be constructed in the same hyperedge, and then all patients and hyperedges will be represented by the correlation matrix H.

[0122] M2-3: The graph structure of the hypergraph G is represented using an incidence matrix H. H is a matrix of dimension |V|×|E|, the specific values ​​of which are determined by h(v,e). h(v,e) represents the specific values ​​of the coordinates (v,e) in the H matrix. h(v,e) is defined as:

[0123]

[0124] For a patient v∈V, its degree is defined as:

[0125]

[0126] For a hyperedge e∈E, its degree is defined as:

[0127]

[0128] For a degree d∈D, the hypergraph association matrix H and patient feature attributes are input into the HGNN layer to obtain the node output features, where a hyper-edge convolutional layer f(S,W,θ) is established, as shown in the following formula:

[0129]

[0130] Among them, S (l) ∈R N×C The hypergraph represents the patient feature attributes at layer l, and S is initialized. (0) =S, where σ() represents the nonlinear activation function, and W is the super-activation function. Figure 3 W, D in the tuple e and D v Let H represent the diagonal matrices for edge degree and patient degree, respectively, with corresponding values ​​d(e) and d(v), where e and v represent the positions of the values ​​in the matrix. T Let S be the transpose of the hypergraph incidence matrix H, and let S be the initial node features for matrix multiplication with the diagonal matrices of edge degree and patient degree. (l) Through the learnable filter matrix θ (l)The HGNN layer extracts higher-order correlations on the hypergraph by transforming the patient-hyperedge-patient transmission method. The new node feature matrix is ​​then multiplied by a trainable weight matrix on the left before being fed into a softmax activation layer to obtain the output, as shown in the following formula:

[0131] O = softmax(W1S) (l+1) )

[0132] Where W1 is the trainable weights and O is the output, which is the final embedding of the hypergraph;

[0133] M3: The dynamic feature processing module is implemented as follows:

[0134] M3-1: Let the input sequence be dynamic data X = [x1, x2, ..., x...]. n ], where x i Representing the i-th element in this sequence, calculate the query vector Q, key vector K, and value vector A, which are obtained by performing a linear transformation on the input sequence:

[0135] Q = XW Q

[0136] K = XW K

[0137] A = XW A

[0138] Among them, W Q W K W A The weight matrix is ​​then learned, and the attention score is calculated using the softmax() function:

[0139]

[0140] Among them, K T d represents the transpose of the key vector matrix. k The dimension of the key vector, divided by Used to scale attention scores;

[0141] Repeat step M3-1 h times, where h is the number of heads, and concatenate the Attention outputs from each head before performing a linear transformation to obtain a dimension of h×d. A The matrix, where d A It is the dimension of the value vector, calculated using the following formula:

[0142] MultiHeadAttention(X)=Concat(head1,head2,...,head h W O

[0143] Among them, head i =Attention(XW i Q XW i K XW i A ), W i Q W i K W i A To obtain the corresponding weight matrix, Concat() concatenates the attention values ​​of each head to get MultiHead, W O To output the projection matrix;

[0144] M3-2: Construct a fully connected feedforward network, the formula is:

[0145] FFN(X) = max(0, MultiHeadW m +b1)W f +b2

[0146] Among them, W m W f Let be the weights, b be the bias, and max() be the maximum value function. The concatenated result MultiHead(Q,K,A) is projected back to the original dimension d using a trainable linear mapping matrix. model This yields the final output FFN of the multi-head attention, where the output is the intermediate encoded information matrix, representing the features of the input time series data.

[0147] M4: The risk prediction module calculates the risk probability results, implemented as follows:

[0148] M4-1: The final embedding matrix O∈R of the hypergraph in the static feature extraction module. N×F Perform a transpose, where N is the number of patients and F is the dimension of the static features, so that rows represent features and columns represent patients. The transposed matrix O T ∈R F×N With dynamic data matrix FFN∈R M×N Maintaining consistency across the feature dimensions, where M is the dimension of the dynamic features, and then horizontally concatenating it with the dynamic data matrix to form a new feature matrix Z∈R. (F+M)×N ;

[0149] M4-2: Z is fed into the decoder of the Transformer model as input, and the output vector of the decoder at each time step is represented as Z. t Where t is the index of the time step, and then a vector Y is obtained by transformation through a linear layer. tThe formula for the linear layer is:

[0150] Y t =Z t W y +b y

[0151] Among them, W y It is the weight matrix of the linear layer, b y It is the bias vector;

[0152] M4-3: Output vector Y of the linear layer t The system is transformed into a probability distribution through a softmax layer, and the final output is the risk probability P of hypotension for each patient. The softmax function is calculated as follows:

[0153]

[0154] in, The original score of the i-th sample in the model output belongs to the c-th class, where C is the total number of classes. In this algorithm, it is 2, which means risky or no risk. The closer the probability value is to 1, the higher the risk of hypotension in the patient during dialysis predicted by the model; the closer the probability value is to 0, the lower the risk of hypotension in the patient during dialysis predicted by the model.

[0155] Treatment records include: dialysis treatment time, dialysis treatment method, preset ultrafiltration volume, total replacement fluid volume, dialysis wave flow rate, and Na+. + Ion concentration, oxygen uptake, and anticoagulation methods.

[0156] The hemodialysis treatment module uses the Vue + Django front-end and back-end web framework. It embeds a blood pressure abnormality prediction algorithm, namely the blood pressure abnormality alarm system unit, which is a directional labeling algorithm. The joint use of Vue and Django is achieved by using the API interface in Django. In Vue, the HTTP client is used to call the API interface provided by Django to display the data of Django on the Vue page.

[0157] The basic patient information entered includes: age, gender, dialysis duration, and BMI, totaling 4 data points; the pre-dialysis measurement data includes: height measured on the day of dialysis, weight measured on the day of dialysis, percentage of weight loss during dialysis, percentage of weight gain compared to the last dialysis session, amount of weight gain, amount of weight loss during dialysis, pre-dialysis heart rate, pre-dialysis diastolic blood pressure, pre-dialysis systolic blood pressure, pulse pressure, and accurate ultrafiltration rate, totaling 11 data points; the sensor and machine data during dialysis include: arterial pressure (AP), venous pressure (VP), real-time blood flow, average blood flow, dialysate flow rate, temperature, and dialysate conductivity, totaling 7 data points; the data generated by wearable monitoring devices and machines during dialysis include: blood pressure and dialysate flow rate. The real-time blood pressure monitoring and visualization module acquires patient data from the hemodialysis treatment module, including basic information, treatment records, and abnormal results and probabilities from the early warning module. It then classifies these data into risk levels and displays them visually. Risk levels are based on abnormal probability, with different grading standards for different user groups. Medical staff can personalize the risk grading probability range according to the patient's condition. Real-time, efficient, and rapid data transmission between the wearable monitoring device and the hemodialysis treatment module utilizes the MQTT (Message Queuing Telemetry Transport) internet transmission protocol and the EMQX information server.

[0158] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A blood pressure early warning system for hemodialysis treatment based on wearable monitoring devices, characterized in that, The hardware includes: The wearable monitoring device includes: STM32F103C8T6 core board, MQ-R-12 heart rate and blood pressure module, DS18B20 temperature sensor, ESP8266 WIFI transmission module, OLED12864 display module, LED single indicator light, buzzer, and Type-C interface. The STM32F103C8T6 core board is equipped with peripheral interfaces, which can be connected to the MQ-R-12 heart rate and blood pressure module, DS18B20 temperature sensor, ESP8266WIFI module, OLED12864 display, LED single indicator light and buzzer. Based on the aforementioned hardware, this system includes the following functional modules in terms of software functionality: The system includes a hemodialysis treatment module and a real-time blood pressure monitoring and visualization module. The hemodialysis treatment module comprises: a nurse mobile terminal unit, a doctor PC terminal unit, and a blood pressure abnormality alarm system unit. The specific workflow of the above functional modules is as follows: S1: The nurse mobile terminal and doctor PC terminal of the hemodialysis treatment module collect routine clinical dialysis data. The data of the dialysis machine, including arterial pressure, venous pressure, conductivity, actual blood flow, average blood flow, dialysate flow, and dialysate temperature, are transmitted to the hemodialysis treatment module via wired transmission. S2: The hemodialysis treatment module receives data from the dialysis machine and wearable device, and simultaneously calls data from the cloud database. The algorithm in the blood pressure abnormality alarm system unit generates a prediction result. S3: The prediction results will be displayed and alerted on the nurse's mobile terminal and the doctor's PC terminal in the hemodialysis treatment module. At the same time, the prediction results will be transmitted to the blood pressure real-time monitoring visualization module for display and alert.

2. The blood pressure early warning system for hemodialysis treatment based on wearable monitoring devices according to claim 1, characterized in that, include: In step S1, routine dialysis data, including the patient's basic information, predialysis assessment, treatment records, and treatment orders, are sent to the cloud database; the MQ-R-12 heart rate and blood pressure module of the wearable device sends blood pressure and heart rate data to the hemodialysis treatment module in real time via the ESP8266WIFI module. The blood pressure abnormality alarm system unit in step S2 includes: The information entry module, static feature extraction module, dynamic feature processing module, and risk prediction module are implemented as follows: M1: The information entry module is implemented by entering multimodal data such as patient basic information, predialysis measurement data, dialysis measurement data, and data generated by wearable monitoring devices and machines during dialysis. Among them, patient basic information, predialysis measurement data, and dialysis measurement data are static data and are denoted as S, while data generated by wearable monitoring devices and machines during dialysis are dynamic data and are denoted as X. M2: The static feature processing module uses a hypergraph neural network for feature processing, implemented as follows: M2-1: The hypergraph G is a triple, denoted as G =<V,E,W> , where V represents the set of nodes of the hypergraph and E represents the set of hyperedges of the hypergraph, and each hyperedge is a subset of the set of vertices; M2-2: A KNN model is used for hyperedge construction. The KNN model parameter N is set, where N represents the number of samples clustered together. Basic patient information is entered as input features, and the static data feature attributes of each patient are considered as {s}. (1) ,...,s (m) The KNN model aggregates samples into multiple classes with N samples each. The specific algorithm is as follows: M2-2-1: Randomly select a patient as the cluster centroid point μ1∈R; M2-2-2: Repeat the following procedure until convergence: M2-2-2-1: For each sample μ i ∈R, calculate the distance between the cluster centroids using the following formula: M2-2-2-2: For each cluster centroid, sort the patients according to their distance from smallest to largest, and then select the N patients with the smallest distance as the nearest neighbors, with each nearest neighbor representing a category; M2-2-2-3: The category labels of N patients are {y} 1N ,...,y nN The patient categories are determined by voting. The number of votes for category C is calculated using the following formula: In the formula, C represents the class with the closest distance among N samples, I(·) represents the indicator function, and the iteration is repeated. c This represents the number of votes N patients received in category C. The higher the number of votes, the closer the N patients are to category C. M2-2-2: Until the moving distance between two iterations is less than the preset threshold, the threshold is set to 0.01 in the experiment. Each class C represents a set of patients with similar conditions, that is, a class is a hyperedge. After aggregating static data using the KNN algorithm, patients with the same features will be constructed in the same hyperedge, and then all patients and hyperedges will be represented by the correlation matrix H. M2-3: The graph structure of the hypergraph G is represented by the incidence matrix H. H is a matrix of dimension |V|×|E|, the specific values ​​of which are determined by h(v,e). h(v,e) represents the specific values ​​of the coordinates (v,e) in the H matrix, and h(v,e) is defined as: For a patient v∈V, its degree is defined as: For a hyperedge e∈E, its degree is defined as: For a degree d∈D, the hypergraph association matrix H and patient feature attributes are input into the HGNN layer to obtain the node output features, where a hyper-edge convolutional layer f(S,W,θ) is established, as shown in the following formula: Among them, S (l) ∈R N×C The hypergraph represents the patient feature attributes at layer l, and S is initialized. (0) =S, σ() represents the nonlinear activation function, W is the W in the hypergraph triple, D e and D v Let H represent the diagonal matrices for edge degree and patient degree, respectively, with corresponding values ​​d(e) and d(v), where e and v represent the positions of the values ​​in the matrix. T Let S be the transpose of the hypergraph incidence matrix H, and let S be the initial node features for matrix multiplication with the diagonal matrices of edge degree and patient degree. (l) Through the learnable filter matrix θ (l) The HGNN layer extracts higher-order correlations on the hypergraph by transforming the patient-hyperedge-patient transmission method. The new node feature matrix is ​​then left-multiplied by a trainable weight matrix before being fed into a softmax activation layer to obtain the output, as shown in the following formula: O=softmax(W1S (l+1) ) Where W1 is the trainable weights and O is the output, which is the final embedding of the hypergraph; M3: The dynamic feature processing module is implemented as follows: M3-1: Let the input sequence be dynamic data X = [x1, x2, ..., x...]. n ], where x i Representing the i-th element in this sequence, calculate the query vector Q, key vector K, and value vector A, which are obtained by performing a linear transformation on the input sequence: Q=XW Q K=XW K A=XW A Among them, W Q 、W K 、W A The weight matrix is ​​then learned, and the attention score is calculated using the softmax() function: Among them, K T d represents the transpose of the key vector matrix. k The dimension of the key vector, divided by Used to scale attention scores; Repeat step M3-1 h times, where h is the number of heads, and concatenate the Attention outputs from each head before performing a linear transformation to obtain a dimension of h×d. A The matrix, where d A It is the dimension of the value vector, calculated using the following formula: MultiHeadAttention(X)=Concat(head1,head2,...,head h )W O Among them, head i =Attention(XW i Q XW i K XW i A ), W i Q 、W i K 、W i A To obtain the corresponding weight matrix, Concat() concatenates the attention values ​​of each head to get MultiHead, W O To output the projection matrix; M3-2: Construct a fully connected feedforward network, the formula is: FFN(X)=max(0,MultiHeadW m +b1)W f +b2 Among them, W m 、W f Let be the weights, b be the bias, and max() be the maximum value function. The concatenated result MultiHead(Q,K,A) is projected back to the original dimension d using a trainable linear mapping matrix. model This yields the final output FFN of the multi-head attention, where the output is the intermediate encoded information matrix, representing the features of the input time series data. M4: The risk prediction module calculates the risk probability results, implemented as follows: M4-1: The final embedding matrix O∈R of the hypergraph in the static feature extraction module. N×F Perform a transpose, where N is the number of patients and F is the dimension of the static features, so that rows represent features and columns represent patients. The transposed matrix O T ∈R F×N With dynamic data matrix FFN∈R M×N Maintaining consistency across the feature dimensions, where M is the dimension of the dynamic features, and then horizontally concatenating it with the dynamic data matrix to form a new feature matrix Z∈R. (F+M)×N ; M4-2: Z is fed into the decoder of the Transformer model as input, and the output vector of the decoder at each time step is represented as Z. t Where t is the index of the time step, and then a vector Y is obtained by transformation through a linear layer. t The formula for the linear layer is: Y t =Z t W y +b y Among them, W y It is the weight matrix of the linear layer, b y It is the bias vector; M4-3: Output vector Y of the linear layer t The system is transformed into a probability distribution through a softmax layer, and the final output is the risk probability P of hypotension for each patient. The softmax function is calculated as follows: in, The original score of the i-th sample in the model output belongs to the c-th class, where C is the total number of classes. In this algorithm, it is 2, which means risky or no risk. The closer the probability value is to 1, the higher the risk of hypotension in the patient during dialysis predicted by the model; the closer the probability value is to 0, the lower the risk of hypotension in the patient during dialysis predicted by the model.

3. The blood pressure early warning system for hemodialysis treatment based on wearable monitoring devices according to claim 2, characterized in that, Treatment records include: dialysis treatment time, dialysis treatment method, preset ultrafiltration volume, total replacement fluid volume, dialysis wave flow rate, and Na+. + Ion concentration, oxygen uptake, and anticoagulation methods.

4. The blood pressure early warning system for hemodialysis treatment based on wearable monitoring devices according to claim 2, characterized in that, The hemodialysis treatment module uses the Vue + Django front-end and back-end web framework. It embeds a blood pressure abnormality prediction algorithm, namely the blood pressure abnormality alarm system unit, which is a directional labeling algorithm. The joint use of Vue and Django is achieved by using the API interface in Django. In Vue, the HTTP client is used to call the API interface provided by Django to display the data of Django on the Vue page.

5. The blood pressure early warning system for hemodialysis treatment based on wearable monitoring devices according to claim 2, characterized in that, The basic patient information entered includes: age, gender, dialysis duration, and BMI, totaling 4 data points; the pre-dialysis measurement data includes: height measured on the day of dialysis, weight measured on the day of dialysis, percentage of weight loss during dialysis, percentage of weight gain compared to the last dialysis session, amount of weight gain, amount of weight loss during dialysis, pre-dialysis heart rate, pre-dialysis diastolic blood pressure, pre-dialysis systolic blood pressure, pulse pressure, and accurate ultrafiltration rate, totaling 11 data points; the sensor and machine data during dialysis include: arterial pressure (AP), venous pressure (VP), real-time blood flow, average blood flow, dialysate flow rate, temperature, and dialysate conductivity, totaling 7 data points; the data generated by wearable monitoring devices and machines during dialysis include: blood pressure and dialysate flow rate.

6. The blood pressure early warning system for hemodialysis treatment based on wearable monitoring devices according to claim 2, characterized in that, Dynamic data include: arterial pressure, venous pressure, conductivity, actual blood flow, average blood flow, dialysate flow, and dialysate temperature.

Images