Abnormal data filtering method of dialysate intelligent system

Abstract

The invention discloses an abnormal data filtering method of a dialysate intelligent system. The abnormal data filtering method comprises the following steps: S1, collecting patient personalized feature data, environmental factor data and dialysate core parameter data; s2, cleaning and standardizing the collected data, and aligning according to timestamps to form a fusion feature matrix; s3, mapping the fusion feature matrix to a hidden space through training and reconstructing the fusion feature matrix to obtain a reconstruction mode of normal data; and S4, calculating an initial basic threshold value based on the reconstruction error of the personalized reference model, and dynamically adjusting the threshold value according to the real-time physiological state of the patient and environmental factor fluctuation. According to the abnormal data filtering method of the dialysate intelligent system, personalized features of a patient and environmental factors are fused through multi-dimensional data acquisition, a personalized reference model is constructed by using a variational auto-encoder, and a dynamic threshold value is generated, so that the problems of misjudgment and missed judgment caused by neglecting individual differences and environmental interference in a traditional method are effectively solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical data processing, in particular to an abnormal data filtering method of a dialysis solution intelligent system. BACKGROUND

[0002] As an important treatment method for patients with end-stage renal disease, the stable operation of the dialysis solution system and the accuracy of the data are directly related to the treatment effect and patient safety. The dialysis solution system continuously collects a large amount of data during operation, including various physicochemical parameters of the dialysis solution, equipment operation state parameters, and patient-related data, etc. These data are the key basis for monitoring the dialysis process and adjusting the treatment plan.

[0003] However, in the current data processing of the dialysis solution system, there are obvious deficiencies in the abnormal data filtering link. The traditional abnormal data filtering method uses fixed threshold judgment or simple statistical analysis means, and does not fully consider the individual differences of patients. The physical condition, severity of illness, tolerance to dialysis treatment, etc. of different patients are significantly different, and a unified filtering standard is difficult to adapt to personalized dialysis needs, which may easily lead to misjudgment of normal physiological fluctuations of patients as abnormal, or delayed identification of truly abnormal data.

[0004] At the same time, the influence of environmental factors on the data of the dialysis solution system has not been effectively handled. Changes in temperature and humidity in the dialysis room, voltage fluctuations in the power supply system, and performance degradation of the equipment after long-term operation, etc. Environmental and equipment factors may cause deviations in the collected data. Existing methods often ignore the influence of these factors, limiting the accuracy of abnormal data filtering.

[0005] In addition, with the development of dialysis technology, the amount of data has increased dramatically, and the traditional filtering method performs poorly in real-time and adaptability. When the data presents complex nonlinear relationships or dynamic change characteristics, the fixed filtering model cannot be adjusted in time, making it difficult to accurately capture potential abnormal patterns, affecting real-time monitoring and risk warning of the dialysis process. SUMMARY

[0006] The purpose of the present application is to provide an abnormal data filtering method for a dialysis solution intelligent system to solve the problems raised in the background.

[0007] To achieve the above purpose, the present application provides the following technical solution: an abnormal data filtering method for a dialysis solution intelligent system, comprising the following steps:

[0008] S1: Collecting patient individualized feature data, environmental factor data, and dialysis solution core parameter data;

[0009] S2: Cleaning and standardizing the collected data, and aligning the fusion feature matrix according to the time stamp;

[0010] S3: mapping the fusion feature matrix to a latent space and reconstructing to obtain a reconstruction mode of the normal data by training;

[0011] S4: calculating an initial basic threshold value based on the reconstruction error of the personalized benchmark model, and dynamically adjusting the threshold value according to the real-time physiological state of the patient and the fluctuation of environmental factors;

[0012] S5: inputting the fusion features collected in real time into the personalized benchmark model, calculating the reconstruction error and comparing it with the dynamic threshold value, and determining that the data is abnormal and filtering if the threshold value is exceeded;

[0013] S6: periodically updating the model parameters by adding normal data, and optimizing the dynamic threshold value based on the feedback of the artificial review of abnormal data;

[0014] S7: integrating the model and filtering logic into the edge computing unit of the dialysis fluid system to realize real-time abnormal data filtering and alarm.

[0015] Preferably, the patient personalized feature data in step S1 includes fixed features and dynamic features, the fixed features are fixed information related to the patient individual, and the dynamic features are real-time changing physiological indicators in the patient dialysis.

[0016] Preferably, the environmental factor data in step S1 includes dialysis room environmental parameter and dialysis equipment internal running state parameter, and the dialysis fluid core parameter data is a key parameter related to the dialysis fluid. The sampling frequency is set according to actual needs.

[0017] Preferably, in step S2, the sliding window method is used for data cleaning to remove outliers, and the interpolation method based on patient historical data is used to fill in missing values; the standardization processing includes standardization processing for numerical value type data and encoding processing for classification type data.

[0018] Preferably, in step S3, the variational autoencoder model includes an encoder and a decoder, the encoder maps the fusion feature matrix to a latent space of a preset dimension, and the latent space contains latent variables related to patient features, environmental factors and the interaction between the two; the decoder is symmetrical to the encoder, and the output layer uses an activation function that adapts to the data range.

[0019] Preferably, in step S3, the loss function of model training is the sum of reconstruction error and regularization term, and a suitable optimizer is used to stop iteration when the loss of the verification set decreases to a preset range.

[0020] Preferably, the initial basic threshold in step S4 is a preset quantile of the reconstruction error of the patient's historical normal data calculated by the personalized benchmark model; the dynamic adjustment comprises: when the patient's real-time physiological indicators fluctuate by a certain proportion of the basic value, the threshold is adjusted according to the physiological indicator change rate by a preset proportion, wherein the adjustment proportion is positively correlated with the physiological indicator change amplitude; when the environmental parameters fluctuate by a certain range, the threshold is adjusted according to the environmental parameter change value by a preset proportion, wherein the adjustment proportion is positively correlated with the environmental parameter change amplitude, and the threshold is constrained within a preset multiple range of the initial basic threshold.

[0021] Preferably, in step S5, the real-time data acquisition frequency is set according to actual needs, the reconstruction error is calculated by a suitable distance calculation method, and the feature dimensions with error proportion exceeding a preset proportion are recorded synchronously to locate the abnormal source when the abnormality is determined.

[0022] Preferably, in step S6, the model is periodically updated by an incremental training method, the learning rate is reduced by a preset proportion, and the update period is set according to actual conditions; in the abnormal feedback closed loop, the artificial review marks "true abnormality" or "misjudgment", and the misjudgment samples are used to optimize the threshold adjustment module, and the target misjudgment rate is controlled within a preset range.

[0023] Preferably, in step S7, the edge computing unit uses a computing device adapted to system requirements, the real-time processing delay is controlled within a preset range, and the hospital-related system and device control system are connected to trigger the corresponding alarm mechanism and reserve the original data for tracing.

[0024] Compared with the prior art, the beneficial effects of the present application are:

[0025] The abnormal data filtering method of the dialysate intelligent system fuses the patient's personalized features and environmental factors through multi-dimensional data acquisition, uses a variational autoencoder to construct a personalized benchmark model and generate a dynamic threshold, realizes accurate filtering of abnormal data of the dialysate system, effectively solves the misjudgment and omission problems caused by ignoring individual differences and environmental interference in traditional methods, improves the real-time and adaptability of data filtering, provides reliable data support for safety monitoring and scheme adjustment of dialysis treatment, and significantly enhances the accuracy and effectiveness of medical data processing. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 The figure is a flowchart of the method of the present application. DETAILED DESCRIPTION

[0027] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all the other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0028] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all the other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application. Figure 1 The present application provides a technical solution: an abnormal data filtering method of a dialysis solution intelligent system, characterized by comprising the following steps:

[0029] S1: multi-dimensional data acquisition, acquiring patient individualized characteristic data, environmental factor data and dialysis solution core parameter data, and completing the data acquisition work through the sensor network deployed in the dialysis room, the medical monitoring equipment worn by the patient and the acquisition module of the dialysis solution system. Among them, the patient individualized characteristic data is obtained through the hospital information system (HIS) and real-time monitoring equipment, the environmental factor data is collected by multiple types of sensors deployed at different positions in the dialysis room, and the dialysis solution core parameter data is collected by the special detection element of the dialysis solution system. The sensor network is deployed in a combination of wired and wireless modes, the wireless sensor transmits data through low-power Bluetooth or Wi-Fi, ensuring the stability and energy saving of data transmission; the medical monitoring equipment has waterproof and anti-interference functions and meets the safety standards of medical equipment; the special detection element of the dialysis solution system is calibrated regularly, and the calibration period is determined according to the performance of the element. The patient individualized characteristic data in step S1 includes fixed characteristics and dynamic characteristics. The fixed characteristics are fixed information related to the patient individual, and the dynamic characteristics are real-time changing physiological indicators during dialysis. The fixed characteristics can be extracted from the registration information and historical medical records of the patient at the time of admission, and the dynamic characteristics are monitored and transmitted to the data processing center by the wearable device in real time. The wearable device and the data processing center use an encrypted transmission protocol to prevent leakage of patient privacy data; the data processing center has a data buffering function, which can temporarily store data when the network is interrupted, and automatically upload the data after the network is restored. The environmental factor data in step S1 includes dialysis room environmental parameters and internal running state parameters of the dialysis equipment, and the dialysis solution core parameter data is the key parameter related to the dialysis solution. The sampling frequency is set according to actual needs. The sampling points of the environmental factor data are distributed in the corners of the dialysis room and the key components inside the dialysis equipment to ensure that the environment and equipment status can be fully reflected. The sampling frequency can be differentiated according to the change rate of different parameters. For parameters that change rapidly, the sampling frequency can be appropriately increased. The internal running state parameters of the dialysis equipment include motor speed, pump body pressure, pipe flow sensor feedback value, etc.; the sampling frequency of the key parameters is set through preliminary experiments, and is adjusted according to the parameter change rate.

[0030] The fixed features in the patient individualized feature data specifically include patient age, gender, height, weight, underlying disease history, dialysis treatment history, etc.; the dynamic features specifically include heart rate, blood pressure, blood oxygen saturation, body temperature, respiratory rate, and other real-time physiological indicators. The dialysis room environment parameters in the environmental factor data specifically include room temperature, humidity, air pressure, air quality index, illumination intensity, etc.; in addition to the motor speed, pump body pressure, and pipeline flow mentioned above, the internal running state parameters of the dialysis equipment also include equipment running temperature, voltage stability, pipeline pressure difference, filter usage time, etc. The dialysis fluid core parameter data specifically includes dialysis fluid concentration (such as sodium ion, potassium ion, calcium ion concentration), pH value, osmotic pressure, flow rate, temperature, etc. The encryption transmission protocol specifically uses the SSL / TLS protocol, and the buffer storage capacity of the data processing center is designed according to the maximum data acquisition amount of 3 days to ensure that data is not lost in the event of network interruption.

[0031] By integrating various collection devices and systems, a comprehensive data acquisition system is constructed to achieve accurate and real-time collection and transmission of patient individualized feature data, environmental factor data, and dialysis fluid core parameter data. Specifically, in terms of hardware deployment, a sensor network combining wired and wireless is used to cover the dialysis room. Wireless sensors use low-power Bluetooth or Wi-Fi technology to ensure stable data transmission while reducing energy consumption, ensuring that environmental factor data can be effectively captured from different locations in the dialysis room and key components inside the equipment. Patients wear waterproof and anti-interference monitoring equipment that meets medical safety standards to record dynamic physiological indicators in real time, while specialized detection elements of the dialysis fluid system focus on collecting core parameters, and these elements are regularly calibrated based on their performance to ensure data accuracy.

[0032] In terms of data sources and transmission, the fixed feature data of the patient is retrieved from the hospital information system (HIS) admission registration and historical medical records, and the dynamic feature data is monitored by wearable devices in real time and sent to the data processing center through an encrypted transmission protocol. The center has a data buffering function to cope with network interruptions and ensure data is not lost; environmental factor data is collected by multiple types of sensors, and dialysis fluid core parameter data is obtained by specialized detection elements. Both are collected at the actual required sampling frequency (differentially set according to the parameter change rate) and transmitted to the processing center. Through this multi-device collaboration and multi-channel integration approach, comprehensive and reliable multi-dimensional data collection is achieved, laying the foundation for subsequent data processing and analysis.

[0033] S2: Data preprocessing and feature fusion, the collected data is cleaned, standardized, and aligned by timestamp to form a fusion feature matrix, the data preprocessing process is carried out on a dedicated data processing server, and the automatic processing of the data is realized through the preprocessed algorithm. The time stamp alignment uses high-precision clock synchronization technology to ensure the consistency of data from different sources in the time dimension. The dedicated data processing server is configured with multi-core processors and large-capacity memory to meet the demand of parallel processing of multi-channel data. The pre-processing algorithm uses Python and is packaged as a callable function library for easy system integration. The high-precision clock synchronization technology uses Network Time Protocol (NTP) with a synchronization accuracy of milliseconds. In step S2, the sliding window method is used to remove outliers, and the interpolation method based on patient historical data is used to fill in missing values. Standardization processing includes standardization processing of numerical data and encoding processing of categorical data. The size of the sliding window can be adjusted according to the characteristics of the data and the actual application scenario to better adapt to the fluctuation of different data. When the data fluctuates sharply, the window size is automatically reduced to accurately capture abnormal points. The encoding processing of categorical data uses one-hot encoding or label encoding. One-hot encoding is used to process classification data without sequential relationship, and label encoding is used to process classification data with hierarchical relationship.

[0034] The initial size of the window in the sliding window method is set according to the data sampling frequency, usually 5-20 sampling points. The determination basis for detecting sharp data fluctuations is that the change amplitude of the continuous 3 sampling points exceeds 2 times the historical average fluctuation amplitude of the parameter. The interpolation method based on patient historical data specifically uses a combination of linear interpolation and polynomial interpolation. Linear interpolation is used for data with smooth changes, and 3rd order polynomial interpolation is used for data with large fluctuations. Numerical data standardization processing specifically uses Z-score standardization, i.e. (x-μ) / σ, where μ is the mean and σ is the standard deviation. One-hot encoding is suitable for non-sequential categorical data such as dialysis mode type and device model, and label encoding is suitable for hierarchical categorical data such as disease severity, mild / moderate / severe. The Network Time Protocol (NTP) clock synchronization server is calibrated with the national time service center, with a calibration frequency of not less than 2 times a day, ensuring millisecond-level synchronization accuracy.

[0035] Based on the dedicated data processing server and optimized algorithm, the multi-dimensional raw data collected is systematically processed and integrated to form a standardized and unified fusion feature matrix, providing high-quality input for subsequent model construction. Specifically, in terms of hardware support, the dedicated data processing server is configured with multi-core processors and large-capacity memory, with the ability to parallel process multi-channel data, and can efficiently handle massive data processing requirements.

[0036] In the data preprocessing link, the preprocessing algorithm written by Python realizes automatic processing: the sliding window method is used to eliminate abnormal values, the window size can be automatically adjusted according to the characteristics of the data and the application scene, and when the data fluctuation is detected, the window is reduced to accurately identify and eliminate abnormal points; for missing values, interpolation method based on patient historical data is used for filling, combined with historical data law and current data trend to improve the accuracy of filling. At the same time, the data is standardized, the numerical data is standardized to eliminate the difference in dimension, and the classification data is converted by one-hot encoding (no order) or label encoding (with hierarchical relationship) according to whether there is an order relationship, to ensure the uniformity of data format.

[0037] In the feature fusion link, the network time protocol (NTP) is used to realize high-precision clock synchronization, and the data collected from different sources and at different times is aligned according to the time stamp, so that the patient's personalized feature data, environmental factor data and dialysate core parameter data are consistent in the time dimension, and finally integrated to form a fusion feature matrix, providing a structured and complete data basis for the subsequent personalized benchmark model construction.

[0038] S3: Construct a personalized benchmark model based on a variational autoencoder, map the fusion feature matrix to the latent space and reconstruct it by training, obtain the reconstruction mode of normal data, and complete the model construction with a deep learning framework. In the model design stage, the number of layers, node number and other key parameters of the network will be determined according to the results of the previous data analysis to ensure that the model can fully learn the characteristics of the data. In the training process, batch training is used to improve the training efficiency and model performance. The deep learning framework can be TensorFlow. When designing the model, the feature dimension that has a greater impact on the model output is determined by drawing a data feature heat map, and then the number of network nodes is adjusted. The batch size of batch training is set according to the memory capacity, and the model parameter verification is performed once after training a certain batch. The variational autoencoder model in step S3 includes an encoder and a decoder. The encoder maps the fusion feature matrix to the latent space with a preset dimension. The latent space contains hidden variables related to patient characteristics, environmental factors and their interaction. The decoder is symmetrical to the encoder. The output layer uses an activation function that adapts to the data range. The connection method and regularization method are used between the layers of the encoder to prevent overfitting. The preset of the latent space dimension is determined based on the analysis of the data feature dimension and the expected model performance, which can better balance the expression ability and computational complexity of the model. The residual connection method is used between the layers of the encoder to alleviate the gradient vanishing problem. The regularization method includes the Dropout layer and the L2 regularization. The latent space dimension is determined through multiple experiments and can be adjusted within a certain range after testing and verification. The loss function of model training in step S3 is the sum of reconstruction error and regularization term. Use a suitable optimizer, iterate until the validation set loss decreases to the preset range, and stop. The loss changes of the training set and the validation set will be monitored in real time during the training process. The training state of the model is judged by analyzing the loss curve, and the training parameters are adjusted in time to ensure that the model can converge to the ideal state. The reconstruction error is calculated by mean square error (MSE), and the regularization term is KL divergence. The Adam optimizer is selected as the optimizer. The initial learning rate is set according to the model training requirements. When the validation set loss does not decrease for several consecutive rounds, the learning rate is automatically decayed. The preset range is determined according to the reconstruction error distribution of historical data.

[0039] The data feature heat map is drawn by calculating the correlation between features by Pearson correlation coefficient, and the feature dimension that has greater influence on the model output is determined as the feature with an absolute correlation coefficient greater than 0.6. The network structure of the encoder is as follows: after the input layer, 2-4 hidden layers are connected, the number of nodes in each hidden layer is 1 / 2-2 / 3 of the previous layer, and the last layer is an implicit space output layer; the decoder is symmetrical to the encoder, and the number of hidden layer nodes gradually recovers from the dimension of the implicit space to the dimension of the input layer. The residual connection mode is specifically adding a jump connection to the input of the second hidden layer after the output of each hidden layer. The dropout rate of the Dropout layer is set to 0.1-0.3, and the L2 regularization coefficient is set to 1e-4-1e-3. The experimental determination process of the implicit space dimension is as follows: the reconstruction error of the model under 10-50 dimensional implicit space is tested respectively, and the dimension with the minimum reconstruction error and meeting the requirement of calculation efficiency is selected as the preset dimension. The initial learning rate of the Adam optimizer ranges from 1e-5 to 1e-3, the number of consecutive non-decreasing rounds of the validation set loss is set to 5-10 rounds, and the learning rate decay ratio is 1 / 2-1 / 10 of the current value each time. The specific determination method of the preset range is as follows: the sum of the mean and 3 times the standard deviation of the reconstruction error of the historical normal data is calculated, which is used as the convergence target of the validation set loss.

[0040] The working principle of constructing a personalized benchmark model based on a variational autoencoder is to use a deep learning framework to perform feature learning and reconstruction on the fusion feature matrix through the cooperative operation of the encoder and the decoder, thereby capturing the pattern of normal data. Specifically, the TensorFlow deep learning framework is selected for model construction. In the design stage, the feature dimensions that have greater influence on the model output are identified by drawing a data feature heat map, and the number of layers and nodes of the network are adjusted to ensure that the model can fully learn the data features.

[0041] The variational autoencoder is composed of an encoder and a decoder. The encoder is responsible for mapping the fusion feature matrix to a preset dimensional implicit space. The implicit space contains implicit variables related to patient characteristics, environmental factors, and their interaction. The dimension is determined through multiple experiments and can be adjusted within a certain range to balance the expression ability and computational complexity of the model. Each layer of the encoder uses a residual connection method to alleviate the gradient vanishing problem, and combines Dropout layer and L2 regularization and other means to prevent overfitting. The decoder is symmetrical to the encoder structure, and the output layer uses an activation function that adapts to the data range to reconstruct the vector of the implicit space into data consistent with the format of the original fusion feature matrix.

[0042] The model training is performed in a batch training manner, the batch size is set according to the memory capacity, and model parameter checking is performed once per certain batch to improve training efficiency and model performance. The loss function of training is the sum of reconstruction error (calculated by mean square error, MSE) and regularization term (KL divergence), the optimizer is Adam optimizer, the initial learning rate is set according to the training requirement, and when the validation set loss does not decrease for continuous multiple rounds, the learning rate is automatically decayed. The loss changes of the training set and the validation set are monitored in real time during the training process, the training parameters are adjusted by analyzing the loss curve, until the validation set loss decreases to the preset range based on the reconstruction error distribution of historical data, so that the model converges to the ideal state, and finally the personalized baseline model which can accurately reflect the normal data pattern is obtained.

[0043] S4: generating a dynamic threshold, calculating an initial basic threshold based on the reconstruction error of the personalized baseline model, and dynamically adjusting the threshold according to the real-time physiological state of the patient and the fluctuation of environmental factors, the generation algorithm of the dynamic threshold is integrated in the subsequent processing module of the model, which can calculate and update the threshold in real time according to the input of various parameters, the subsequent processing module is realized by FPGA chip to realize hardware acceleration, ensuring the real-time performance of threshold calculation; the parameter input interface supports analog signal and digital signal, the analog signal is converted into digital signal by AD converter and then input, the initial basic threshold in step S4 is the preset quantile of the reconstruction error of the patient's historical normal data calculated by the personalized baseline model; dynamic adjustment includes: when the real-time physiological index of the patient fluctuates more than a certain proportion of the basic value, the threshold is adjusted according to the physiological index change rate by a preset proportion, wherein the adjustment proportion is positively correlated with the physiological index change amplitude; when the environmental parameter fluctuates more than a certain range, the threshold is adjusted according to the environmental parameter change value by a preset proportion, wherein the adjustment proportion is positively correlated with the environmental parameter change amplitude, and the threshold is constrained within a preset multiple range of the initial basic threshold, the determination of the preset quantile is obtained by statistical analysis of a large amount of patient historical normal data, which can better reflect the distribution characteristics of normal data. In the dynamic adjustment process, a special parameter control module will be set to ensure the rationality of the adjustment proportion and the effectiveness of the threshold constraint, the preset quantile is set according to the data distribution characteristics, and it can be adjusted appropriately for patients with unstable condition; the parameter control module has upper and lower limits of the adjustment proportion; the preset multiple range can be manually modified on the system setting interface according to clinical requirements.

[0044] The FPGA chip is specifically an Xilinx series industrial chip, the hardware acceleration logic of which is written in Verilog HDL language to realize parallel processing of threshold calculation. The conversion accuracy of the AD converter is 12-16 bits, and the sampling rate is not less than 1 kHz, ensuring the accuracy and timeliness of analog signal conversion. The preset quantile calculation of the initial basic threshold value uses the empirical distribution function method to statistically analyze the historical normal data of at least 100 patients (not less than 30 dialysis data per patient). The basic value of the patient's real-time physiological index fluctuation is the average value of the corresponding index in the patient's last three dialysis; the normal range of environmental parameter fluctuation is the daily fluctuation range of the dialysis room environmental parameters, such as room temperature 18-26℃, humidity 40%-60%, which is determined by statistical analysis of environmental monitoring data in the last three months. The upper and lower limits of the adjustment ratio in the parameter control module are determined by clinical experiments to ensure that the adjusted threshold value can identify abnormalities without being overly sensitive; manual modification of the preset multiple range requires authorization by the system administrator, and the modification record is automatically archived.

[0045] is the reconstruction error based on the personalized benchmark model, combined with the patient's real-time physiological state and environmental factor fluctuation, to construct a threshold system that can be self-adaptively adjusted to accurately determine whether the dialysate data is abnormal. Specifically, the generation of dynamic threshold relies on the special algorithm integrated in the model subsequent processing module, which is realized by FPGA chip to ensure the real-time performance of threshold calculation, and the parameter input interface supports analog and digital signals. After the analog signal is converted into digital signal by the AD converter, it is inputted to provide support for threshold calculation of multiple types of parameters.

[0046] The determination of the initial basic threshold value is based on the patient's historical normal data, the reconstruction error of which is calculated by the personalized benchmark model, and then the preset quantile is selected as the initial value according to the data distribution characteristics. For patients with unstable conditions, the quantile can be adjusted appropriately to better reflect the distribution characteristics of normal data.

[0047] The dynamic adjustment mechanism optimizes the threshold value according to the patient's real-time physiological indicators and environmental parameter fluctuations: when the patient's real-time physiological indicators fluctuate beyond a certain proportion of the basic value, the threshold value will be adjusted according to the preset proportion based on the physiological indicator change rate, and the adjustment ratio is positively correlated with the physiological indicator change amplitude; when the environmental parameter fluctuates beyond a certain range, the threshold value will be adjusted according to the preset proportion based on the environmental parameter change value, and the adjustment ratio is also positively correlated with the environmental parameter change amplitude. At the same time, a special parameter control module is set, which has an upper and lower limit of the adjustment ratio, and the threshold value is constrained within the preset multiple range of the initial basic threshold value. This range can be manually modified on the system settings interface according to clinical needs to ensure the rationality of the adjustment ratio and the effectiveness of the threshold value constraint. Finally, a threshold value that can dynamically adapt to changes in patient status and environment is formed to provide accurate standards for abnormal data determination.

[0048] S5: real-time data processing and anomaly determination, input the fusion features collected in real time into the personalized benchmark model, calculate the reconstruction error and compare it with the dynamic threshold, if the threshold is exceeded, it is determined to be abnormal data and filtered, real-time data processing uses edge computing technology to complete preliminary processing near the data collection end, reducing data transmission and processing delay, improving the timeliness of anomaly determination, the edge computing unit uses an industrial-grade embedded computer with anti-electromagnetic interference capability, suitable for the working environment of the dialysis room; preliminary processing includes data compression and feature extraction, controlling the appropriate compression rate, the real-time data collection frequency in step S5 is set according to actual needs, the reconstruction error uses a suitable distance calculation method, when determining the anomaly, record the feature dimensions whose error proportion exceeds the preset proportion to locate the abnormal source, the selection of the distance calculation method will consider the calculation efficiency and accuracy, for scenes with high real-time requirements, a method with less calculation amount can be selected. The abnormal source positioning information is stored in real time and can be displayed through a visual interface, which is convenient for medical staff to view and analyze, the distance calculation method selects a method with less calculation amount when real-time requirements are high, and a method with high precision when accuracy requirements are high; the preset proportion is set according to the anomaly determination requirements, that is, when the error proportion of a certain feature dimension exceeds the proportion, it is marked as a suspected abnormal source; the visual interface uses a touch display screen, supporting data curve zooming, abnormal point marking and other functions.

[0049] The industrial-grade embedded computer specifically uses an ARM architecture or x86 architecture processor with a main frequency not less than 1.5 GHz, at least 2 Ethernet interfaces and 4 USB interfaces, a working temperature range of -20°C-70°C, and a protection level not less than IP30. Data compression uses LZ77 or DEFLATE compression algorithm, with a compression rate controlled at 30%-70%, ensuring that the compressed data can completely retain the feature information. In the distance method of reconstruction error calculation, the method with less calculation amount is specifically Manhattan distance (L1 norm), and the method with high precision is specifically Euclidean distance (L2 norm). The preset proportion of abnormal source positioning is set to 20%-40%, which is determined by analyzing the contribution of each feature dimension in historical abnormal data. The size of the touch display screen of the visual interface is 10-15 inches, with a resolution not less than 1920x1080, the data curve refresh frequency is consistent with the real-time data collection frequency, and the abnormal point marking uses a red flashing icon, which also displays the abnormal occurrence time, duration and error value.

[0050] The working principle of real-time data processing and abnormality determination is to perform efficient processing on the real-time collected data through edge computing technology, and to realize accurate determination of abnormal data and positioning of abnormal sources according to the comparison between reconstruction error and dynamic threshold. Specifically, real-time data processing uses edge computing technology, and deploys an industrial-grade embedded computer as an edge computing unit near the data collection end. The unit has anti-electromagnetic interference capability and can adapt to the working environment of the dialysis room. The collected fusion feature data is first processed in the edge computing unit, including data compression and feature extraction. By controlling the appropriate compression rate, the data transmission amount and processing delay are reduced, and the overall processing efficiency is improved.

[0051] In the abnormality determination process, the real-time collected fusion features are input into the personalized reference model, and the reconstruction error is calculated by the model. The reconstruction error is calculated by a suitable distance calculation method. When the real-time requirement is high, a method with small calculation amount is selected, and when the accuracy requirement is high, a method with high precision is selected. The reconstruction error is compared with the dynamic threshold. If the reconstruction error exceeds the dynamic threshold, it is determined as abnormal data and filtered.

[0052] At the same time, abnormal source positioning is performed simultaneously during abnormality determination. According to a preset proportion, the proportion is set according to the abnormality determination requirement. When the error of a certain feature dimension accounts for more than the proportion of the total error, it is marked as a suspected abnormal source. The abnormal source positioning information is stored in real time and can be displayed through a visual interface in the form of a touch display screen. The interface supports data curve zooming, abnormal point marking and other functions, which facilitates medical staff to view and analyze, so as to quickly master the abnormal situation. In this way, real-time monitoring, accurate abnormality determination and efficient abnormal source tracing of the dialysate system data are realized.

[0053] S6: Model iterative optimization, periodically update model parameters by adding normal data, and optimize dynamic threshold based on artificial review of abnormal data feedback. Model iterative optimization is automatically performed in the background without human intervention, and an optimization report is generated to record the changes in model parameters and optimization effect. Artificial review is completed by professional medical staff on a dedicated review platform, and the review results are fed back to the model optimization module. The background automatic optimization is performed during the system idle period to avoid affecting normal dialysis work; the optimization report includes the comparison of model performance before and after parameter adjustment; the dedicated review platform has data labeling tools that support box selection of abnormal data segments and adding text descriptions. In step S6, the model is periodically updated using incremental training, the learning rate is reduced by a preset ratio, and the update cycle is set according to actual conditions; in the abnormal feedback closed loop, artificial review marks "true abnormality" or "misjudgment", misjudgment samples are used for threshold adjustment module, the target misjudgment rate is controlled within a preset range, incremental training adds a certain proportion of the original training set each time, and the bottom network parameters are frozen and the top parameters are fine-tuned; the learning rate preset ratio is set according to the model training state, i.e. the learning rate is reduced by a certain percentage each time; the update cycle is adjusted according to the data accumulation speed and clinical needs; the preset range is set according to clinical requirements and monitored by periodically calculating the proportion of misjudgment sample quantity to total sample quantity.

[0054] The proportion of incremental training data to the original training set is 10%-30%, the frozen bottom layer network parameters are the first 1-2 layers of the encoder and decoder, and the fine-tuned top layer parameters are the last 1-2 layers and the hidden space mapping parameters. The learning rate preset ratio is reduced to 1 / 2-1 / 5 of the current value each time, and the update cycle is initially set to once a week, which can be triggered for early update according to the data accumulation speed (e.g. every 50 new data). The annotation tool of the dedicated review platform supports adding text descriptions including abnormal type (e.g. equipment failure / patient physiological fluctuation / environmental interference), handling suggestions, etc., and the annotation results are stored in association with the original data. The preset range of the target misjudgment rate is 1%-5%, and the statistical period is every 7 days. When the range is exceeded for 2 consecutive statistical periods, the system automatically sends an optimization reminder.

[0055] By periodically introducing new data to update model parameters and combining artificial review of abnormal data feedback, the model performance and dynamic threshold are continuously optimized to adapt to the changes in dialysis data. Specifically, model iterative optimization is automatically performed in the background without human intervention, and is carried out during the system idle period to avoid affecting normal dialysis work.

[0056] In terms of model periodic updating, incremental training is used, with each new data amount being a certain proportion of the original training set. The bottom layer network parameters are frozen during training, and only the top layer parameters are fine-tuned to efficiently utilize new data information without damaging the basic features learned by the model. The learning rate is reduced by a preset proportion, which is set according to the model training state. The update cycle is flexibly adjusted according to the data accumulation speed and clinical needs to ensure that the model can timely absorb effective information from new data.

[0057] At the same time, an abnormal feedback closed loop is constructed, and artificial review is completed by professional medical staff on a special review platform. The platform has data labeling tools, supports box selection of abnormal data segments and adds text explanations. After the medical staff label "true abnormality" or "misjudgment", the results are fed back to the model optimization module. Misjudgment samples are used to optimize the threshold adjustment module to make the dynamic threshold more accurate. The whole process generates an optimization report, including the comparison of model performance before and after parameter adjustment, and the target misjudgment rate is controlled within the preset range according to clinical requirements. The proportion of misjudgment sample quantity to total sample quantity is monitored regularly to realize continuous optimization of the model and threshold, and improve the accuracy and adaptability of abnormal data filtering.

[0058] S7: System deployment, integrate the model and filtering logic to the edge computing unit of the dialysate system, realize real-time abnormal data filtering and alarm, comprehensive testing will be carried out before system deployment, including function test, performance test and compatibility test, to ensure the stable and reliable operation of the system. The installation location of the edge computing unit is selected near the dialysate system and data acquisition equipment to reduce data transmission distance and delay. Function test contains filtering test of multiple groups of simulated abnormal data to ensure a certain accuracy. Performance test is carried out under different loads to verify the system response time. Compatibility test is carried out for mainstream brands of dialysate systems and hospital information systems to ensure data interface compatibility. In step S7, the edge computing unit uses a computing device that meets the system requirements, the real-time processing delay is controlled within the preset range, and it is connected with the hospital related system and equipment control system to trigger the corresponding alarm mechanism and keep the original data for traceability. The computing device that meets the system requirements is selected and tested strictly to meet the system requirements for computing power, power consumption and stability. The connection with the hospital related system and equipment control system uses standardized interface protocol to ensure smooth and compatible data transmission. The alarm mechanism can be realized through various ways such as sound and light alarm, SMS notification, system pop-up window, etc. The original data is kept by using distributed storage technology to ensure data security and traceability. The configuration of the computing device meets the system requirements for computing power, memory and storage capacity. The preset range of real-time processing delay is set according to the real-time requirement. The interface protocol uses medical health information exchange standard and equipment control protocol. The distributed storage uses multi-copy mechanism and the storage period meets the medical data preservation requirement. The volume and light mode of the sound and light alarm adapt to the alarm requirements of the dialysis room, and the SMS notification is sent to the mobile phones of designated medical staff.

[0059] The simulation abnormal data of the function test covers device faults (such as pump body abnormalities, pipeline blockage), patient physiological abnormalities, such as sudden blood pressure rise / fall, environmental interference, such as sudden temperature change, electromagnetic interference, and more than 20 common abnormal types, each type of test is not less than 50 times, and the accuracy rate needs to reach more than 95%. The different loads of the performance test are specifically the data of 1-20 dialysis machines processed at the same time, and the system response time requirement is not more than 500 ms. The mainstream brand dialysate system of the compatibility test includes Fresenius, B Braun, Becton Dickinson, etc., and the hospital information system includes HIS, LIS, PACS, etc. The interface protocol is specifically HL7 FHIRv4.0.1 and Modbus RTU / TCP. The installation position of the edge computing unit is not more than 5 meters away from the dialysate system and not more than 3 meters away from the data acquisition device, and is fixed with a shockproof support. The volume range of the audible and visual alarm is 60-100 dB, the light is a red LED light, and the flicker frequency is 1-2 times per second; the system pop-up window is displayed on the nurse station monitoring terminal and the doctor workstation at the same time, containing abnormal device number, abnormal time, abnormal type and other information. The distributed storage uses a 3-node cluster architecture, each node has a storage capacity of not less than 1 TB, the data backup strategy is daily incremental backup + weekly full backup, the storage period is 5 years, which meets the requirements of the Medical Quality Management Method for medical data preservation.

[0060] The model and filtering logic are integrated into the edge computing unit of the dialysate system, and through comprehensive testing, adaptive device selection, system docking and data storage management, the real-time abnormal data filtering and alarm function is realized. Specifically, before deployment, three major tests of function, performance and compatibility are carried out: function test verifies the filtering accuracy of the system through multiple groups of simulation abnormal data to ensure that the preset requirements are met; performance test runs under different loads to verify whether the system response time meets the real-time processing requirements; compatibility test for mainstream brand dialysate systems and hospital information systems ensures data interface compatibility and lays a foundation for stable system operation.

[0061] The edge computing unit selects computing devices that meet system requirements, and its configuration is strictly selected and tested to meet the requirements of computing power, memory, storage capacity, power consumption and stability, and the real-time processing delay is controlled within the preset range according to the real-time requirements. The unit is docked with hospital-related systems and device control systems using standardized interface protocols, medical health information exchange standards and device control protocols to ensure smooth and compatible data transmission.

[0062] When the system is running, the edge computing unit processes data in real time and filters out abnormalities. Once an abnormality is determined, multiple alarm mechanisms are triggered: the volume and light mode of the audible and visual alarm adapt to the dialysis room environment, and can attract the attention of medical staff in a timely manner; an SMS notification is sent to the mobile phones of designated medical staff, and a pop-up window is displayed on the relevant terminal, achieving multi-channel alarm. In addition, the raw data is saved through distributed storage technology, and the safety is ensured by the multi-copy mechanism. The storage period meets the medical data preservation requirements, providing reliable data support for subsequent traceability analysis, and finally building a complete system integrating real-time processing, abnormal alarm, and data traceability.

[0063] Through multi-dimensional data acquisition, processing and fusion, model construction, threshold generation, real-time determination, iterative optimization and system deployment, accurate and efficient filtering of abnormal data in the dialysate system data is achieved.

[0064] First, multi-dimensional data acquisition integrates sensor networks, medical monitoring equipment and dialysate system acquisition modules to build a comprehensive data acquisition system. Combined with encryption transmission and data buffering mechanism, it ensures the comprehensive and reliable collection of patient's personalized characteristics, environmental factors and dialysate core parameters, laying a foundation for subsequent processing.

[0065] Next, data preprocessing and feature fusion rely on special servers and optimization algorithms to clean and standardize the raw data. Through high-precision clock synchronization, time stamp alignment is achieved to form a standardized fusion feature matrix, eliminating data differences and noise interference and improving data quality.

[0066] On this basis, a personalized benchmark model based on variational autoencoder is constructed. The encoder and decoder work together through a deep learning framework. Through reasonable network design and training strategy, the reconstruction mode of normal data is captured to provide accurate benchmark reference for abnormality determination. When generating dynamic threshold, the initial basic threshold is determined based on model reconstruction error, and adaptive adjustment is made in combination with the real-time physiological state of the patient and the fluctuation of environmental factors. Through the parameter control module, the threshold is ensured to be reasonable, so that the abnormality determination standard can dynamically adapt to the actual situation. Real-time data processing and abnormality determination use edge computing technology to complete preliminary processing near the data acquisition end. Abnormal data filtering is achieved through comparison of reconstruction error and dynamic threshold, and the abnormal source is located synchronously, improving the timeliness and accuracy of determination.

[0067] Model iterative optimization uses incremental training and manual review feedback to continuously optimize model parameters and dynamic thresholds without affecting normal dialysis, improving the adaptability of the system to data changes. Finally, the system deployment integrates the model and filtering logic into the edge computing unit, which is tested comprehensively to ensure stable operation. Through multi-channel alarm and distributed storage, real-time response and data traceability of abnormal data are achieved.

[0068] The method effectively improves the accuracy, real-time performance and adaptability of the abnormal data filtering through close connection and technical innovation of each link, reduces the misjudgment and omission, and provides a strong guarantee for the safety of dialysis treatment and data reliability.

[0069] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely divergences of the principles and specific embodiments of the application and that numerous modifications, changes, substitutions, and alterations can be made thereto without departing from the spirit and scope of the application as defined by the appended claims and their equivalents.

Claims

1. A method for filtering abnormal data in a smart dialysis fluid system, characterized in that, Includes the following steps: S1: Collect patient-specific characteristic data, environmental factor data, and core parameter data of dialysis fluid; S2: Clean and standardize the collected data, and align it according to timestamps to form a fusion feature matrix; S3: By training, the fused feature matrix is ​​mapped to the latent space and reconstructed to obtain the reconstruction pattern of normal data; S4: Calculate the initial baseline threshold based on the reconstruction error of the personalized benchmark model, and dynamically adjust the threshold according to the patient's real-time physiological state and environmental fluctuations. S5: Input the real-time collected fusion features into the personalized benchmark model, calculate the reconstruction error and compare it with the dynamic threshold. If the error exceeds the threshold, it is judged as abnormal data and filtered. S6: Regularly update model parameters by adding normal data, and optimize dynamic thresholds based on feedback from abnormal data reviewed by humans; S7: Integrate the model and filtration logic into the edge computing unit of the dialysis fluid system for real-time abnormal data filtering and alarm.

2. The abnormal data filtering method for a dialysis fluid intelligent system according to claim 1, characterized in that: The patient-specific characteristic data in step S1 includes fixed characteristics and dynamic characteristics. The fixed characteristics are fixed information related to the individual patient, and the dynamic characteristics are physiological indicators that change in real time during dialysis.

3. The abnormal data filtering method for a dialysis fluid intelligent system according to claim 1, characterized in that: The environmental factor data in step S1 includes dialysis room environmental parameters and dialysis equipment internal operating status parameters. The core parameters of the dialysis fluid are the dialysis fluid parameters, and the sampling frequency is set.

4. The abnormal data filtering method for a dialysis fluid intelligent system according to claim 2, characterized in that: In step S2, the data cleaning uses a sliding window method to remove outliers and fills in missing values ​​using interpolation based on patient historical data; the standardization process includes standardizing numerical data and encoding categorical data.

5. The abnormal data filtering method for a dialysis fluid intelligent system according to claim 3, characterized in that: In step S3, a variational autoencoder model is used, which includes an encoder and a decoder. The encoder maps the fused feature matrix to a latent space of a preset dimension. The latent space contains latent variables related to patient features, environmental factors, and their interactions. The decoder is symmetrical to the encoder, and the output layer uses an activation function adapted to the data range.

6. The abnormal data filtering method for a dialysis fluid intelligent system according to claim 5, characterized in that: In step S3, the loss function for model training is the sum of the reconstruction error and the regularization term. The optimizer iterates until the loss on the validation set decreases to a preset range and then stops.

7. The abnormal data filtering method for a dialysis fluid intelligent system according to claim 4, characterized in that: The initial baseline threshold in step S4 is the preset quantile of the reconstruction error calculated by the personalized benchmark model based on the patient's historical normal data. The dynamic adjustment includes: when the patient's real-time physiological indicators fluctuate beyond the baseline value, the threshold is adjusted according to the rate of change of the physiological indicators by a preset ratio, wherein the adjustment ratio is exactly matched with the magnitude of change of the physiological indicators; when the environmental parameters fluctuate beyond the range, the threshold is adjusted according to the change value of the environmental parameters by a preset ratio, wherein the adjustment ratio is exactly matched with the magnitude of change of the environmental parameters, and the threshold is constrained within a preset multiple range of the initial baseline threshold.

8. The abnormal data filtering method for a dialysis fluid intelligent system according to claim 6, characterized in that: In step S5, the real-time data acquisition frequency is set according to actual needs, the reconstruction error is calculated using an appropriate distance method, and the feature dimension whose error ratio exceeds a preset proportion is recorded synchronously during anomaly determination to locate the anomaly source.

9. The abnormal data filtering method for a dialysis fluid intelligent system according to claim 8, characterized in that: In step S6, the model is updated periodically using incremental training, with the learning rate decreasing by a preset ratio and the update cycle set according to the actual situation. In the anomaly feedback loop, manual review and labeling are performed to indicate "true anomaly" or "false positive". False positive samples are used to optimize the threshold adjustment module, and the target false positive rate is controlled within a preset range.

10. The abnormal data filtering method for a dialysis fluid intelligent system according to claim 8, characterized in that: In step S7, the edge computing unit uses computing devices adapted to the system requirements, controls the real-time processing latency within a preset range, and interfaces with the hospital's relevant systems and equipment control systems to trigger corresponding alarm mechanisms and retain the original data for traceability.

Images