Peritoneal dialysis fluid balance monitoring system and method with dynamic weighing and data fusion

Abstract

The present application relates to the technical field of peritoneal dialysis monitoring, and discloses a peritoneal dialysis fluid balance monitoring system and method based on dynamic weighing and data fusion. The system comprises dynamic weighing, multi-source data acquisition, data fusion processing, balance monitoring and early warning output modules. The dynamic weighing module acquires the input and output amounts of dialysis fluid in real time and generates weight data; the multi-source data acquisition module obtains dialysis fluid composition, treatment time and patient physiological parameters; the data fusion processing module fuses the above data to generate dialysis fluid balance characteristic data; the balance monitoring module calculates the net ultrafiltration amount and the fluid balance state according to the above data, and identifies the risk of abnormal leakage or fluid retention; and the early warning output module generates early warning information when the risk is identified. The system does not require manual intervention, can acquire and fuse multi-source data in real time, accurately judges the fluid balance state and timely generates early warning information, and is suitable for hospital and home peritoneal dialysis scenarios.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of peritoneal dialysis monitoring, in particular to a peritoneal dialysis fluid balance monitoring system and method based on dynamic weighing and data fusion. BACKGROUND

[0002] As a common renal replacement therapy for patients with end-stage renal disease, the treatment effect of peritoneal dialysis is closely related to the dialysis fluid balance state. Accurate control of the input, output and net ultrafiltration volume of dialysis fluid is crucial to maintain the patient's fluid balance and ensure treatment safety. In current clinical practice, peritoneal dialysis fluid balance monitoring mainly relies on manual operation and recording. Medical staff need to manually record the initial weight of dialysis fluid, replacement time during dialysis, and manually calculate the weight of discharged fluid after dialysis, and then estimate the net ultrafiltration volume. This manual monitoring method is not only cumbersome to operate, but also prone to inaccurate data due to human recording errors, making it difficult to reflect the patient's dialysis fluid balance state in real time.

[0003] Most existing monitoring methods only focus on the weight data of dialysis fluid, ignoring the influence of dialysis fluid composition changes, treatment time fluctuations and patient physiological parameters (such as blood pressure, heart rate, body weight, etc.) on fluid balance. For example, when the glucose concentration in the dialysis fluid changes, it may affect the ultrafiltration function of the peritoneum, thereby changing the net ultrafiltration volume. Fluctuations in a patient's blood pressure or body weight may also indirectly reflect the body's fluid retention or leakage. Due to the lack of comprehensive analysis of these multi-source data, medical staff have difficulty in fully judging the patient's fluid balance state, and often can only detect abnormalities when obvious clinical symptoms (such as edema, shortness of breath, etc.) occur. At this time, the patient may have already developed more serious complications, delaying the best opportunity for intervention.

[0004] For home peritoneal dialysis patients, the existing monitoring system cannot achieve real-time remote monitoring. Patients need to regularly visit the hospital to provide dialysis data, and medical staff have difficulty in timely grasping the patient's fluid balance changes during home treatment, increasing the risk of home treatment. Once dialysis fluid leaks or fluid retention occurs, the patient may not be able to detect it in time, leading to worsening of the condition, and even causing serious consequences such as infection and cardiovascular events. Currently, although some simple weighing devices are used to assist in monitoring the weight of dialysis fluid, these devices have single functions and do not have data fusion analysis and automatic warning capabilities, which cannot meet the clinical demand for accurate, real-time and comprehensive monitoring of peritoneal dialysis fluid balance. Therefore, developing a peritoneal dialysis fluid balance monitoring system that can integrate multi-source data and achieve real-time monitoring and automatic warning has become a problem that needs to be solved in current clinical practice. SUMMARY

[0005] The present application aims to provide a peritoneal dialysis fluid balance monitoring system and method with dynamic weighing and data fusion to solve the problems raised in the background.

[0006] To achieve the above-mentioned purpose, the present application provides a peritoneal dialysis fluid balance monitoring system with dynamic weighing and data fusion, which comprises:

[0007] A dynamic weighing module is configured to collect the input and output amounts of dialysis fluid in real time during peritoneal dialysis and generate dialysis fluid weight data;

[0008] A multi-source data acquisition module is configured to acquire dialysis fluid composition data, treatment time data and patient physiological parameter data;

[0009] A data fusion processing module is configured to fuse and process the dialysis fluid weight data and the data acquired by the multi-source data acquisition module to generate dialysis fluid balance feature data;

[0010] A balance monitoring module is configured to calculate the net ultrafiltration amount and the fluid balance state based on the dialysis fluid balance feature data and identify the risk of abnormal leakage or fluid retention;

[0011] An early warning output module is configured to generate early warning information when the risk of abnormal leakage or fluid retention is identified.

[0012] Preferably, the data fusion processing module comprises:

[0013] A dialysis fluid balance reference table establishing unit is configured to establish a dialysis fluid balance reference table containing dialysis fluid balance reference data of different treatment stages;

[0014] A feature autoencoder constructing unit is configured to construct a target dialysis fluid balance feature autoencoder based on the dialysis fluid balance reference data and the corresponding real-time dialysis fluid balance encoding;

[0015] A reference data matching unit is configured to match the elementary dialysis fluid balance reference data closest to the current dialysis fluid weight data from the dialysis fluid balance reference table;

[0016] A feature processing unit is configured to perform feature processing on the part of the dialysis fluid balance reference table corresponding to the elementary dialysis fluid balance reference data by using the target dialysis fluid balance feature autoencoder and output the current dialysis fluid balance feature result.

[0017] Preferably, the feature autoencoder constructing unit comprises:

[0018] A reference data matching subunit is configured to match the reference dialysis fluid balance reference data from the dialysis fluid balance reference table;

[0019] The reference table extraction subunit is configured to extract a reference partial dialysis fluid balance reference table corresponding to the reference dialysis fluid balance reference data.

[0020] The pointing determination subunit is configured to perform a pointing determination operation on the dialysis fluid balance matching information in the reference partial dialysis fluid balance reference table, and generate an updated reference partial dialysis fluid balance reference table with matching pointing information.

[0021] The fusion operation subunit is configured to perform a feature fusion operation on the updated reference partial dialysis fluid balance reference table by using the initial dialysis fluid balance feature autoencoder and combining the matching pointing information, and output a representative feature position corresponding to each dialysis fluid balance reference data.

[0022] The parameter adjustment subunit is configured to perform parameter adjustment on the initial dialysis fluid balance feature autoencoder according to the representative feature position and the real-time dialysis fluid balance code, and generate the target dialysis fluid balance feature autoencoder.

[0023] Preferably, the reference table extraction subunit comprises:

[0024] The phase difference degree judgment subunit is configured to determine, from the dialysis fluid balance reference table, dialysis fluid balance reference data with a matching phase difference degree between the reference dialysis fluid balance reference data that is not greater than a preset reference matching phase difference degree.

[0025] The reference table generation subunit is configured to generate the reference partial dialysis fluid balance reference table according to matching information between the reference dialysis fluid balance reference data and the determined dialysis fluid balance reference data.

[0026] Preferably, the fusion operation subunit comprises:

[0027] The traversal matching subunit is configured to traverse the dialysis fluid balance reference data in the updated reference partial dialysis fluid balance reference table, and match adjacent dialysis fluid balance reference data of the currently traversed dialysis fluid balance reference data in the table.

[0028] The fusion processing subunit is configured to perform a fusion operation on feature positions of the currently traversed dialysis fluid balance reference data and each adjacent dialysis fluid balance reference data by using the initial dialysis fluid balance feature autoencoder and combining matching pointing information between the currently traversed dialysis fluid balance reference data and each adjacent dialysis fluid balance reference data, and output a current fusion feature position.

[0029] The representative feature determination subunit is configured to determine a representative feature position corresponding to each dialysis fluid balance reference data according to the current fusion feature position.

[0030] Preferably, the fusion processing subunit comprises:

[0031] a direction subunit, configured to analyze a direction of each matching direction information between the current traversed dialysate balance reference data and the adjacent dialysate balance reference data;

[0032] a marking subunit, configured to mark an end position of the direction as elementary matching direction information of the current traversed dialysate balance reference data;

[0033] a fusion data marking subunit, configured to analyze dialysate balance reference data corresponding to an initial position of the direction of the elementary matching direction information, and mark the dialysate balance reference data as fusion-required dialysate balance reference data;

[0034] an information fusion subunit, configured to perform information fusion operation on feature positions of the fusion-required dialysate balance reference data and the current traversed dialysate balance reference data by using the initial dialysate balance feature self-encoder, and output a current fusion feature position.

[0035] Preferably, the information fusion subunit comprises:

[0036] an importance distribution determining subunit, configured to determine an importance parameter distribution corresponding to each fusion-required dialysate balance reference data;

[0037] a feature analysis subunit, configured to perform classification feature analysis processing on the feature positions of the current traversed dialysate balance reference data and each fusion-required dialysate balance reference data according to the importance parameter distribution by using the initial dialysate balance feature self-encoder, and output a classification importance parameter corresponding to each fusion-required dialysate balance reference data;

[0038] a fusion mapping subunit, configured to perform fusion mapping operation on the feature positions and the classification importance parameter corresponding to each fusion-required dialysate balance reference data, and generate fusion mapping information;

[0039] an excitation output subunit, configured to perform fusion excitation output operation on the fusion mapping information corresponding to all fusion-required dialysate balance reference data, and output the current fusion feature position.

[0040] Preferably, the parameter adjusting subunit comprises:

[0041] an identification code generating subunit, configured to perform feature processing on the reference dialysate balance reference data according to the representative feature position, and output an identification dialysate balance code;

[0042] The parameter adjustment operation subunit is configured to analyze a parameter adjustment interval by using the identified dialysis solution balance code and the real-time dialysis solution balance code, and adjust the initial dialysis solution balance feature auto-encoder according to the interval to generate the target dialysis solution balance feature auto-encoder.

[0043] Preferably, the balance monitoring module comprises:

[0044] The state acquisition unit is configured to acquire a balance state type in the dialysis solution balance feature data.

[0045] The judgment unit is configured to judge whether the patient physiological parameter data is lower than a preset physiological threshold value when the balance state type is an abnormal type.

[0046] The duration acquisition unit is configured to acquire a duration of being lower than the threshold value when the patient physiological parameter data is lower than the preset physiological threshold value.

[0047] The first warning unit is configured to generate a first warning instruction when the duration is greater than a preset duration threshold value.

[0048] The movement vector acquisition unit is configured to acquire a movement vector of the dynamic weighing module when the balance state type is not an abnormal type.

[0049] The vector direction acquisition unit is configured to acquire a vector direction of the movement vector when a vector value of the movement vector is greater than a preset movement threshold value.

[0050] The second warning unit is configured to generate a second warning instruction when a detection device corresponding to the vector direction detects the dynamic weighing module.

[0051] Preferably, the present application further comprises a dynamic weighing and data fusion peritoneal dialysis solution balance monitoring method applied to the dynamic weighing and data fusion peritoneal dialysis solution balance monitoring system as described above, and the method comprises:

[0052] S1, real-time collection of input and output amounts of dialysis solution in a peritoneal dialysis process by dynamic weighing, and generation of dialysis solution weight data;

[0053] S2, acquisition of multi-source data including dialysis solution composition data, treatment time data and patient physiological parameter data;

[0054] S3, fusion processing of the dialysis solution weight data and the multi-source data to generate dialysis solution balance feature data;

[0055] S4, calculation of net ultrafiltration and liquid balance state based on the dialysis solution balance feature data, and identification of abnormal leakage or liquid retention risk;

[0056] S5. When an abnormal leakage or fluid retention risk is identified, generate and output a warning message.

[0057] Compared with the prior art, the present application has the following beneficial effects:

[0058] By setting the dynamic weighing module, the input and output amounts of dialysis fluid can be collected in real time and weight data can be generated, thereby eliminating the dependence on traditional manual weighing and recording, avoiding human operation errors, making the acquisition of dialysis fluid weight data more timely and accurate, and reflecting the dynamic changes in fluid volume during dialysis in real time to provide accurate basic data for subsequent fluid balance judgment.

[0059] The addition of the multi-source data acquisition module breaks through the limitation of existing monitoring which only focuses on dialysis fluid weight, and includes dialysis fluid composition data, treatment time data and patient physiological parameter data into the monitoring range, making the monitoring dimension more comprehensive. Changes in dialysis fluid composition may directly affect ultrafiltration effect, fluctuations in treatment time may change dialysis adequacy, and patient physiological parameters can indirectly reflect the body fluid state. The integrated collection of these data enables medical staff to understand the overall situation of the dialysis process from multiple angles, rather than being limited to single data dimension judgment, and provides rich information support for comprehensive evaluation of patient fluid balance state.

[0060] The data fusion processing module fuses and processes multi-source data to generate dialysis fluid balance feature data, realizing effective integration and analysis of different types of data. Different sources of data are interrelated, and separate analysis of a certain type of data cannot fully grasp the true state of fluid balance. Through data fusion, potential relationships between data can be mined, such as combining changes in dialysis fluid glucose concentration with changes in net ultrafiltration volume, and associating changes in patient weight with discharge fluid volume data, to form feature data that better reflects the essence of fluid balance, helping medical staff more clearly identify the trend of fluid balance changes.

[0061] The balance monitoring module calculates the net ultrafiltration volume and fluid balance state based on the balance feature data, and identifies the risk of abnormal leakage or fluid retention, realizing active judgment and risk identification of the dialysis fluid balance state. Instead of relying on the experience-based judgment of medical staff, this module automatically completes the calculation of key indicators and risk assessment through analysis of fused data, can timely discover subtle abnormalities that are easily overlooked in traditional monitoring methods, such as early dialysis fluid leakage or slight fluid retention, making risk identification more timely and accurate, and helping to take intervention measures at the risk germination stage.

[0062] The early warning output module generates early warning information when a risk is identified, which can quickly transmit the risk signal to medical staff or patients. For home peritoneal dialysis patients, the early warning function can help them to timely detect abnormal conditions in the dialysis process and avoid the aggravation of the disease due to the failure to timely detect the risk. At the same time, the timely output of the early warning information also gives medical staff more intervention time, so that they can quickly adjust the treatment plan, reduce the possibility of complications, and improve the safety of treatment. BRIEF DESCRIPTION OF DRAWINGS

[0063] Figure 1 A timing diagram of the dynamic weighing and data fusion peritoneal dialysis fluid balance monitoring system according to the present application;

[0064] Figure 2 A flowchart of the operation of the data fusion processing module;

[0065] Figure 3 A flowchart of the operation of the feature autoencoder construction unit. DETAILED DESCRIPTION

[0066] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0067] Please refer to Figure 1 The present application provides a dynamic weighing and data fusion peritoneal dialysis fluid balance monitoring system, which comprises a dynamic weighing module, a multi-source data acquisition module, a data fusion processing module, a balance monitoring module and an early warning output module.

[0068] The dynamic weighing module monitors the input and output processes of the dialysis fluid in the peritoneal dialysis process in real time through a high-precision sensor, continuously acquires the weight changes and converts them into digital signal form of dialysis fluid weight data; the multi-source data acquisition module integrates multiple sensors and data interfaces for obtaining dialysis fluid composition data such as glucose concentration and electrolyte level, treatment time data including infusion and drainage time points, and patient physiological parameter data such as blood pressure and heart rate; the data fusion processing module receives the dialysis fluid weight data and multi-source data, processes these heterogeneous data using a fusion algorithm, and generates feature data that comprehensively reflects the balance state of the dialysis fluid; the balance monitoring module performs calculation logic based on the dialysis fluid balance feature data, derives the net ultrafiltration volume and evaluates the fluid balance state, detects abnormal leakage or fluid retention risk through pattern recognition technology; the early warning output module triggers an alarm mechanism when a risk is identified, generates visual or audible early warning information and outputs it through a display device or a communication interface.

[0069] Embodiment 1: Refer to Figure 2 The data fusion processing module comprises multiple units working in coordination to realize its functions. The dialysate balance reference table establishing unit accesses the historical treatment database through the system interface, extracts key parameters of different treatment stages as dialysate balance reference data from the database, and these parameters include time series data such as priming rate, drainage efficiency, and ultrafiltration volume change curve; the unit stores these reference data according to treatment stages to form a structured dialysate balance reference table, each record in the table contains a timestamp, a treatment stage identifier, and a corresponding reference value range, and an index mechanism is established to facilitate quick retrieval. The feature autoencoder construction unit adopts a deep learning framework to construct a neural network model, the input layer of which receives dialysate balance reference data and corresponding real-time dialysate balance codes, which are standardized data representations generated by the multi-source data acquisition module; the hidden layer of the autoencoder comprises multiple fully connected layers, which compress the input data into latent space features through forward propagation, and then reconstruct the output through the decoder; the mean square error loss function is used to optimize the network parameters during the training process, and the final generated target dialysate balance feature autoencoder can effectively extract key features from the data. The reference data matching unit receives the dialysate weight data sent by the dynamic weighing module in real time, and calculates the similarity between the data and the records in the dialysate balance reference table, and uses the Euclidean distance-based matching algorithm to find the closest primary dialysate balance reference data; the matching process considers the consistency of the treatment stage and the alignment of the time window to ensure the clinical relevance of the matching result. The feature processing unit calls the trained target dialysate balance feature autoencoder to extract features from the part of the dialysate balance reference table corresponding to the matched primary dialysate balance reference data; the processing process first standardizes the data blocks in the reference table, then inputs the autoencoder for forward propagation calculation, and finally outputs the latent feature representation as the current dialysate balance feature result to the subsequent module.

[0070] The dialysate balance control table establishment unit periodically updates the reference data from the clinical database during operation, and the update strategy is based on a sliding time window mechanism, always retaining typical treatment data in a certain period as reference standards; for each treatment stage data, abnormal value detection and filtering processing are performed to ensure the quality and reliability of the reference data, and the data storage adopts a relational database management system, and multiple indexes are established to support efficient query. The feature self-encoder construction unit uses an adaptive learning rate optimizer during model training, and the training data comes from the dialysate balance reference data and its corresponding encoding in the historical treatment records; the network structure includes an encoder and a decoder, the encoder is composed of multiple fully connected layers and uses a ReLU activation function, the decoder adopts a symmetric structure, and an early stopping strategy is applied during the training process to prevent overfitting, and the final self-encoder model can well capture the nonlinear features in the data. The reference data matching unit uses weighted distance measurement when implementing similarity matching, and different weight coefficients are assigned to data of different dimensions, which are set based on clinical expert experience; during the matching process, data of the same treatment stage will be given priority, and for the time feature, a cyclic processing method is used to deal with periodic patterns, and the matching result will be subjected to confidence evaluation, only the results exceeding the threshold will be adopted. The feature processing unit performs batch normalization on the input data when performing feature extraction to ensure uniform data scale; the forward propagation calculation of the self-encoder uses matrix operation to speed up, and the output features are post-processed including dimension reduction and standardization, and the finally generated feature data contains rich time dynamic information and statistical features. The dialysate balance control table establishment unit also implements data version management function, each update will keep the historical version for backtracking analysis, and at the same time establishes a data quality monitoring mechanism to automatically detect and exclude abnormal reference records; data of special treatment modes will be stored separately and special identifiers will be added in the control table. The feature self-encoder construction unit will perform validation set test after model training is completed, and the cross-validation method is used to evaluate the model generalization ability, and the models with poor performance will be adjusted and retrained; the final model adopts quantization technology to reduce the demand for computing resources, so that it can run efficiently on embedded devices. In addition to the distance-based matching algorithm, the reference data matching unit also integrates a clustering-based method as an auxiliary matching strategy, which improves the robustness of the matching by fusing multiple matching methods; the matching history data will be recorded during the matching process to optimize the efficiency of the subsequent matching process. The feature processing unit supports multiple operating modes, including real-time processing mode and batch processing mode, and selects the appropriate processing mode according to different application scenarios; processing logs will be generated during feature extraction to record feature distribution changes and abnormal situations, providing data support for system optimization.

[0071] The units of the data fusion processing module exchange data through a clear interface protocol defined between them, and a unified data format specification is adopted to ensure compatibility; a fault isolation mechanism is implemented within the module, so that an exception in a single unit will not cause the entire module to crash, and the module also has an automatic recovery function. The dialysate balance table establishment unit establishes a secure data connection with the hospital information system, and an encryption protocol is used in the transmission process to protect patient privacy; an access control mechanism is implemented for data storage, and only authorized operations can modify the reference data. The training process of the feature autoencoder construction unit can be performed in the cloud, and the trained model is deployed to the local device through a secure channel; the model update supports hot switching function, which can update the model version without interrupting the service. The reference data matching unit implements multi-threaded parallel processing, can handle multiple matching requests simultaneously, and the optimized matching algorithm supports real-time response; a cache mechanism is established for the matching results, and repeated query requests are directly returned with cached results to improve efficiency. The feature processing unit implements dynamic resource allocation function, automatically adjusts the allocation of computing resources according to the processing load; the feature extraction process supports interruption recovery, and after an unexpected interruption, the process can continue from the checkpoint without starting over. The cooperative work between the units of the module is realized through asynchronous communication by message queue, which reduces the coupling degree between the units; important operations will generate audit logs, recording operation time, operation content and operation result. The dialysate balance table establishment unit provides data quality report function, and generates quality evaluation report of reference data regularly; supports manual review function, and clinical experts can confirm and adjust the automatically generated reference data. The feature autoencoder construction unit provides model interpretability analysis tool, which can visualize the feature importance distribution; supports transfer learning function, which can quickly adapt to new treatment modes by using pre-trained models. The reference data matching unit provides matching result visualization interface, which shows the matching process and matching basis; supports matching strategy configuration, which can adjust the parameter settings of the matching algorithm according to clinical needs. The feature processing unit provides feature visualization tool, which can visually display the extracted feature distribution and trend; supports feature export function, and the extracted features can be exported in standard format for use by other analysis tools.

[0072] Example 2: see Figure 3, the feature autoencoder construction unit completes the construction process of the autoencoder through multiple precisely matched sub-units. The reference data matching sub-unit first performs data retrieval from the established dialysate balance control table, and adopts an algorithm based on density clustering to identify representative reference dialysate balance reference data, which are usually distributed in the core area of the feature space; the statistical characteristics and time sequence characteristics of the data are considered during the clustering process to ensure that the selected reference data can cover different treatment scenarios and physiological states. The control table extraction sub-unit then works, and the phase difference degree judgment sub-unit in it calculates the matching phase difference degree between all data in the control table and the reference data, which is based on the Mahalanobis distance measurement of the multi-dimensional feature space and considers the correlation between dimensions; the preset reference matching phase difference degree threshold is set by clinical experts according to the treatment accuracy requirement, and only those data points with a difference degree within the acceptable range are selected. The control table generation sub-unit then constructs the reference partial dialysate balance control table according to the matching information between the selected data points and the reference data; this local control table not only contains the original data values, but also records the association relationship and similarity weight between the data points.

[0073] The direction determination sub-unit starts to process the generated reference partial dialysate balance control table and performs directional analysis on the dialysate balance matching information in the table; the sub-unit assigns a directional identifier to each matching relationship by analyzing the gradient change trend and correlation direction between data points, and these directional information reveals the influence direction and dependency relationship between data. The fusion operation sub-unit then starts and calls the initial dialysate balance feature autoencoder to process the updated control table with directional information; the sub-unit adopts a method based on graph neural network to perform information propagation and aggregation in the feature space along the direction indicated by the matching directional information, and finally generates a representative feature position coordinate for each dialysate balance reference data through multiple rounds of iteration calculation. The parameter adjustment sub-unit works in the last stage, compares the obtained representative feature position with the real-time dialysate balance code, and calculates the quality index of the feature representation; based on the comparison result, the network parameters of the initial autoencoder, including the weight matrix and the bias vector, are adjusted using the back propagation algorithm, and after multiple iterations of optimization, the target dialysate balance feature autoencoder with improved performance is obtained.

[0074] The reference data matching subunit employs an adaptive clustering algorithm in implementation, which can automatically adjust the clustering radius and density threshold according to the data distribution characteristics; the subunit will periodically recalculate the clustering center to ensure that the reference data can timely reflect the latest treatment mode and patient state changes. The degree of difference judgment subunit uses a dynamic threshold mechanism to adjust the tolerance range of the matching difference degree according to the characteristics of different treatment stages; for key treatment stages such as ultrafiltration period, stricter standards will be adopted, while for stable stages, the requirements will be appropriately relaxed to improve the system robustness. The control table generation subunit adopts a graph structure storage method when constructing the local control table, not only saving the data values themselves, but also recording the topological relationship and similarity measure between data points. The pointing determination subunit implements multiple pointing judgment algorithms, including unidirectional pointing based on time priority, bidirectional pointing based on correlation strength, and pointing relationship based on causal inference; the subunit will generate a confidence score for each matching relationship, and only high-confidence pointing information will be retained for subsequent processing. The fusion operation subunit adopts a feature fusion method enhanced by an attention mechanism, which assigns different fusion weights according to the importance of matching pointing; multiple feature views will be maintained during feature propagation, including time view, frequency view and statistical view, to ensure that the fused feature representation contains multi-angle information. The parameter adjustment subunit implements a multi-objective optimization strategy, considering both the compression quality and reconstruction accuracy of feature representation, to maintain the discriminability of features while ensuring that the decoder can accurately reconstruct the original data.

[0075] The entire feature self-encoder construction unit adopts a pipeline architecture design, and the intermediate results are exchanged between the sub-units through shared memory to reduce data transmission overhead. A state management mechanism is implemented within the unit to record the metadata information of each processing step. The reference data matching sub-unit supports incremental clustering calculation, and only the local area needs to be re-clustered when new data is added without the need to re-calculate everything. The sub-unit also implements an outlier detection function that can identify and exclude abnormal data points that deviate significantly from the normal range. The difference degree judgment sub-unit uses multi-dimensional weighted distance calculation to assign different weight coefficients to feature dimensions of different clinical importance. Distance calculation supports multiple measurement methods, including Euclidean distance, cosine similarity, and dynamic time warping distance, to adapt to the matching needs of different types of feature data. The control table generation sub-unit implements an intelligent memory management function based on data access frequency and importance using a hierarchical storage strategy. For frequently accessed core data, it is retained in the cache, while less frequently used data is stored in external memory. The pointing determination sub-unit provides a pointing relationship visualization tool that can graphically display the pointing network between data points. It supports manual adjustment functions that allow clinical experts to modify the automatically generated pointing relationships based on experience. The fusion operation sub-unit implements distributed computing functions that can distribute large-scale feature fusion tasks to multiple computing nodes for parallel processing. It supports breakpoint resume functions that can resume from the last checkpoint without restarting after interruption. The parameter adjustment sub-unit uses an adaptive learning rate adjustment strategy that dynamically adjusts the learning rate based on the gradient changes during training. It implements various regularization techniques to prevent overfitting, including Dropout and weight decay methods. The components of the unit are designed with loose coupling, allowing individual sub-units to be upgraded or replaced without affecting the overall functionality. The communication interface uses standardized data formats to ensure interoperability between components. The reference data matching sub-unit provides a clustering quality evaluation report that regularly outputs various indicators of the clustering results. It supports multiple clustering algorithm switching, allowing the selection of the most suitable clustering method based on data characteristics. The difference degree judgment sub-unit implements a threshold adaptive adjustment algorithm that can automatically optimize the threshold parameters based on historical matching results. It provides a threshold adjustment suggestion function that recommends the optimal threshold setting based on matching success rate and mismatch rate.

[0076] The control table generation subunit supports multiple data storage format outputs, including relational databases, NoSQL databases, and file storage, etc.; realizes data compression function, reduces storage space occupation while maintaining information integrity. The pointing determination subunit provides pointing relationship verification tools, which can evaluate the significance level of the pointing relationship through statistical test methods; supports the export function of the pointing relationship, and the generated relationship network can be imported into professional analysis software for further research. The fusion operation subunit realizes the fusion process monitoring interface, which can display the state and progress of feature propagation in real time; provides fusion quality evaluation indicators, including feature consistency and information retention, etc. The parameter adjustment subunit realizes the visualization function of the training process, which can display the loss function change and model performance curve in real time; supports multiple optimization algorithm selection, including Adam, SGD and RMSprop, etc. The whole feature autoencoder construction unit adopts modular design, and each subunit can be tested and verified independently; the code implementation pays attention to readability and maintainability, and contains detailed comments and development documents. The unit running state is monitored in real time, and the resource usage and service quality indicators are recorded and analyzed continuously; the abnormal situation will trigger the early warning mechanism, and the system maintenance personnel will be notified in time for processing. The finally realized feature autoencoder construction unit can generate high-quality target dialysis fluid balance feature autoencoder, which can effectively extract the deep features of the dialysis fluid balance data.

[0077] In embodiment 3, the fusion operation subunit realizes efficient feature fusion processing through multiple precisely matched components. The traversal matching subunit first performs systematic data traversal on the updated reference dialysis fluid balance control table, and processes each dialysis fluid balance reference data record in chronological order according to the data entry time; in the traversal process, a matching mechanism based on k-nearest neighbor algorithm is used to find several adjacent dialysis fluid balance reference data closest to the current dialysis fluid balance reference data in space or feature space, and the calculation of proximity degree considers both the multidimensional feature distance and the time proximity of the data points. The fusion processing subunit then starts working, calling the initial dialysis fluid balance feature autoencoder to process the current traversed dialysis fluid balance reference data and its adjacent data, which fully utilizes the matching pointing information between data to guide the direction and strength of feature fusion; for each current data point, the subunit will perform targeted fusion operation on the feature representation of each adjacent data point, and the fusion weight is dynamically determined according to the strength and direction of the matching pointing information, finally outputting the fused current fusion feature position. The representative feature determination subunit integrates all the current fusion feature positions in the last stage, and uses a density peak-based clustering algorithm to identify representative position points in the feature space, which can summarize the core feature distribution pattern of the entire data set.

[0078] The orientation analysis subunit inside the fusion processing subunit is responsible for analyzing the directional features in the matching pointing information. The subunit determines the position orientation of each matching pointing information by calculating the gradient vectors and correlation directions between data points. The time sequence relationship and feature dependent direction of data points are considered in the analysis process, and a complete orientation description containing the direction angle and intensity value is generated for each pointing relationship. The labeling subunit performs the labeling work of elementary matching pointing information based on the orientation analysis results. The subunit uses the end position attribution judgment algorithm to label those matching relationships pointing to the current traversal data point as elementary matching pointing information. Confidence evaluation is performed during the labeling process, and only those matching information with clear pointing relationship and intensity reaching the threshold requirement are retained. The data fusion marking subunit then works, determines the source data points that need to participate in the fusion operation by tracing the starting position of the elementary matching pointing information in reverse, and marks these data points as dialysis fluid balance reference data that need to be fused. The marking process follows the transitivity of data dependency relationship, ensuring that all related data points are included in the fusion range. The information fusion subunit finally performs the substantive feature fusion operation. The subunit uses the fusion algorithm based on the attention mechanism to assign appropriate fusion weights to each dialysis fluid balance reference data that needs to be fused. The weight value is dynamically calculated according to the data importance and relationship strength. The fusion process is carried out in the feature space, and the enriched current fusion feature position is generated through weighted combination and nonlinear transformation.

[0079] The traversal matching subunit adopts an optimized traversal strategy in the implementation process, giving priority to processing data regions with significant feature changes and increasing the sampling density in high dynamic change intervals. The matching algorithm supports adaptive k value adjustment, automatically determining the optimal number of neighboring points according to the data distribution density. The fusion processing subunit implements a multi-granularity fusion mechanism, supporting feature fusion operations at different levels. Effective fusion can be performed from the original data layer to the abstract feature layer. The subunit also implements real-time fusion quality monitoring function, dynamically evaluating the effect of each fusion operation and adjusting the fusion parameters accordingly. The representative feature determination subunit uses an incremental clustering algorithm, which can quickly update the representative feature position when new data arrives without having to recalculate all data. The subunit also provides a representative feature visualization function, which projects high-dimensional representative features to two-dimensional space through dimensionality reduction techniques for intuitive observation.

[0080] The orientation analysis subunit implements multiple orientation determination algorithms, including direction detection based on gradient descent, direction inference based on correlation analysis, and direction determination based on causal discovery algorithms; for complex multi-directional cases, the subunit uses a main direction extraction method to select the most important direction as the final orientation. The labeling subunit implements a fuzzy labeling mechanism, and for boundary case direction determination, a probabilistic label is used instead of a hard decision, retaining a certain degree of uncertainty for subsequent processing steps to make comprehensive judgments. The data labeling subunit supports multi-level data labeling, and according to different fusion requirements, different ranges of data points can be labeled; the subunit also implements a label propagation mechanism, when a data point is labeled, its label state is transmitted to adjacent data points with strong correlation. The information fusion subunit uses a neural network-based fusion function to learn the optimal fusion method through training; the fusion function can be represented as:

[0081]

[0082] wherein: represents the fused feature output, represents the total number of data points to be fused, is the fusion strength adjustment factor, represents the importance score of the ith data point to be fused, is the feature representation of the ith data point to be fused. The fusion strength adjustment factor is dynamically adjusted according to data quality and application scenarios, and the importance score is determined by the feature quality of the data point itself and the correlation strength with the current data.

[0083] The entire fusion operation subunit adopts a pipeline architecture design, but a buffer mechanism is provided between each processing stage to allow asynchronous execution to some extent to improve overall processing efficiency. The traversal matching subunit implements intelligent cache management, optimizes the cache for frequently accessed data regions, and reduces repeated calculation overhead; the subunit also supports parallel traversal processing, which can handle different regions simultaneously using multiple threads for large-scale data tables. The fusion processing subunit implements fusion strategy configuration functions, allowing different fusion algorithms and parameter settings to be selected according to different application requirements; the subunit also provides fusion process recording functions, which record the parameters and results of each fusion operation in detail for subsequent analysis and optimization.

[0084] The orientation analysis subunit provides an orientation relationship visualization tool to graphically display the pointing network and direction distribution between data points; it supports orientation relationship verification functions to evaluate the significance level of the orientation relationship through statistical hypothesis testing methods. The labeling subunit implements a labeling quality evaluation mechanism to score the confidence of automatically generated labeling results, and labeling results with low confidence will be specially marked for manual review. The data-to-be-fused labeling subunit supports labeling propagation control, which can adjust the strength and range of labeling propagation to prevent feature dilution caused by excessive labeling. The information fusion subunit implements multiple fusion modes, including conservative fusion mode, aggressive fusion mode and adaptive fusion mode, and selects the most suitable fusion strategy according to different scenarios. The representative feature determination subunit uses multi-dimensional evaluation indicators to evaluate the quality of representative features, including feature representativeness index, feature discriminability index and feature stability coefficient; the subunit also implements a dynamic updating mechanism for representative features, which can automatically adjust the position of representative features when the data distribution changes. The running state of the entire fusion operation subunit is comprehensively monitored, including processing progress, resource usage, fusion quality indicators, etc., which are all recorded and analyzed in real time; the unit internally implements a fault detection and recovery mechanism, and the failure of a single component will not cause the entire unit to fail, and it can automatically recover from failure.

[0085] The data exchange between components uses a unified format of intermediate representation to ensure smooth data processing flow; important operation steps generate detailed operation logs, including timestamp, input data features, processing parameters and output result summary, etc. The traversal matching subunit provides matching result statistical analysis functions, and regularly outputs matching success rate and matching quality reports; it supports matching algorithm parameter optimization suggestions, and recommends the optimal parameter settings according to historical matching effect data. The fusion processing subunit implements fusion effect backtracking analysis functions, which can compare the differences in output results under different fusion strategies to provide data support for fusion algorithm optimization. The orientation analysis subunit supports multiple orientation representation format outputs, including vector form, angle form and matrix form, etc., to meet the needs of different downstream processing components. The labeling subunit provides a labeling consistency checking tool to detect contradictions and inconsistencies in the labeling results and provide solutions. The data-to-be-fused labeling subunit implements labeling propagation visualization functions, which can graphically display the propagation process and range of labels in the data network.

[0086] In Example 4, the information fusion subunit and the parameter adjustment subunit work together to achieve deep processing of dialysate balance data. The importance distribution determination subunit first performs multi-dimensional evaluation on each dialysate balance reference data to be fused, considering factors such as clinical relevance, time freshness, and measurement accuracy. For example, when processing a set of ultrafiltration volume data from different treatment stages, the subunit assigns higher importance parameters to recent data and gives additional weight to monitoring data from high-risk patients. The evaluation results are output in the form of importance parameter distribution, forming a quantitative evaluation system with weight coefficients. The feature analysis subunit then receives these importance parameter distribution data and uses the initial dialysate balance feature autoencoder to perform deep feature analysis on the current dialysate balance reference data and each data to be fused. The analysis process uses a multi-layer neural network architecture to extract deep feature representations through forward propagation calculations, and combines importance parameters to adjust the feature extraction process. Finally, the classification importance parameters corresponding to each data to be fused are output, reflecting the discriminability and contribution value of each data point in the feature space.

[0087] The fusion mapping subunit starts processing the weighted feature data. This unit uses feature transformation techniques to combine the feature position of each data point to be fused with its classification importance parameters. The mapping process maintains the topological structure of the feature space while enhancing the representation strength of important features. Through nonlinear transformation, enriched fusion mapping information is generated, which contains the comprehensive expression of original features and importance weights. The excitation output subunit finally performs information aggregation operations. This unit uses an aggregation function based on the attention mechanism to weight and combine the fusion mapping information corresponding to all data to be fused. During the combination process, the contribution of each data point is dynamically adjusted according to the classification importance parameters. The final output of the current fusion feature position not only retains the detailed information of the original data but also highlights the dominant role of important features.

[0088] The recognition code generation subunit in the parameter adjustment subunit starts working. This unit encodes the reference dialysate balance reference data based on the representative feature position. The encoding process uses feature extraction and dimension reduction techniques to convert high-dimensional representative feature positions into compact recognition dialysate balance codes, which can summarize the core feature patterns of the reference data. The parameter adjustment operation subunit then starts, calculates the parameter interval that needs to be adjusted by comparing the difference between the recognition dialysate balance code and the real-time dialysate balance code. The calculation process uses the gradient descent algorithm to analyze the direction and magnitude of the coding difference, determines the direction and step size of the network parameter update, and finally generates the adjusted target dialysate balance feature autoencoder. Referring to Table 1, the importance parameter distribution of the dialysate balance reference data to be fused in a specific example is shown:

[0089] Table 1: Fusion dialysis fluid balance reference data importance parameter distribution table

[0090]

[0091] The importance distribution determination subunit adopts a multi-factor weighting algorithm in the calculation process, and each influence factor is standardized and given different weight coefficients; the time freshness factor is calculated based on the interval between the data acquisition time and the current time, and the recent data obtains a higher score; the measurement accuracy factor considers the accuracy level of the data acquisition equipment and the measurement environment conditions; the clinical relevance factor is dynamically adjusted according to the importance of the treatment stage represented by the data point and the patient state. The feature analysis subunit will first standardize each data point when processing the data in the table, and then adjust the attention distribution of feature extraction according to the importance parameters; for data points with high importance parameters, this subunit will extract more fine-grained feature representations, while reducing the depth of feature extraction for low-importance data, which not only improves the processing efficiency but also ensures the complete retention of important information.

[0092] The fusion mapping subunit adopts feature space transformation technology to synthesize the original feature vector of each data point with the importance parameter; the synthesis process keeps the metric properties of the feature space unchanged, but adjusts the density distribution of the feature space, making the feature representation around the important data points more dense. The excitation output subunit uses a softmax-based attention allocation mechanism to automatically calculate the contribution weight of each data point in the final fusion result according to the importance parameter; this mechanism ensures that high-importance data has a greater impact on the final feature representation, while allowing low-importance data to participate in fusion to maintain the diversity of features. The recognition code generation subunit adopts an autoencoder structure to compress the representative feature position into a low-dimensional code when generating the recognition code; the encoding process retains the discriminative information in the original features while eliminating redundant and noisy components. The parameter adjustment operation subunit compares the differences between the recognition code and the real-time code, and uses the backpropagation algorithm to calculate the update amount of network parameters; the update process uses an adaptive learning rate strategy to dynamically adjust the update step size according to the importance of the parameters, and uses a smaller step size for fine adjustment of key parameters and a larger step size for fast convergence of secondary parameters.

[0093] Throughout the entire processing process, strict data consistency is maintained between the sub-units, and all intermediate results adopt a unified data format and interface specification; the data processing flow adopts a pipeline architecture but allows parallel execution to improve overall processing efficiency. The importance distribution determination sub-unit implements an importance parameter caching mechanism, reusing previously calculated importance parameter distributions for similar data patterns to reduce repeated calculation overhead. The feature analysis sub-unit supports multiple feature extraction modes and can select the most suitable feature analysis method according to the data type and application scenario. The fusion mapping sub-unit provides mapping quality monitoring functions to assess the information retention level and feature distortion during the fusion mapping process in real time. The excitation output sub-unit implements an output result verification mechanism to verify the quality level of the fused features through reconstruction error.

[0094] The identification code generation sub-unit in the parameter adjustment sub-unit supports adaptive adjustment of code length and automatically selects the optimal code dimension based on data complexity. The parameter adjustment operation sub-unit implements parameter update trajectory recording to save historical records of previous parameter adjustments for traceability analysis and model version management. The entire implementation process adopts an incremental learning strategy, allowing the model to continuously learn and optimize without downtime, gradually improving the accuracy and efficiency of feature fusion and parameter adjustment. The final generated target dialysis fluid balance feature autoencoder can better capture key features in the data, providing more accurate feature representations for subsequent balance state monitoring. The entire implementation fully considers the real-time requirements of clinical applications and the limitations of computing resources in its design, ensuring processing accuracy while meeting the performance requirements of practical applications.

[0095] In Example 5, the balance monitoring module achieves comprehensive monitoring and early warning functions through multiple precisely coordinated units. The state acquisition unit first receives dialysis fluid balance feature data from the data fusion processing module and determines the current balance state type by analyzing the state identification field in the data packet; for example, when processing data for a dialysis treatment, the unit identifies the pattern features of abnormal increase in ultrafiltration volume and incomplete drainage from the feature data and classifies it as the "poor drainage" state in the abnormal type, while recording the timestamp and confidence score of state identification. The judgment unit immediately starts the patient physiological parameter verification process when the state type is identified as abnormal; this unit accesses the real-time physiological parameter data stream and compares it with the pre-set physiological threshold parameter table; for example, when the system detects dialysis fluid balance abnormalities in a patient, the judgment unit continuously monitors the patient's blood pressure, heart rate, and respiratory rate data, and finds that the systolic blood pressure is consistently below the pre-set threshold of 90 mmHg, and the heart rate is abnormally fast, which confirms the abnormalities in both physiological parameters and dialysis fluid balance.

[0096] The duration acquisition unit starts timing work in the case of physiological parameter anomaly, which uses high-precision timestamp recording system to track the time length of the parameter below the threshold; in the above patient example, the unit starts timing from the first time the blood pressure is below 90 mmHg, and updates the duration record every second, while monitoring the duration pattern of other related parameters to form a complete timing anomaly record. The first warning unit automatically generates a warning instruction when the duration exceeds the preset 5-minute threshold, which contains structured information such as anomaly type, severity, duration and recommended treatment measures; the warning instruction is transmitted to the warning output module through the system internal bus, while triggering the local alarm device and remote notification system to ensure that medical staff can learn about the risk situation in time.

[0097] The movement vector acquisition unit starts work when the balance state type is judged to be normal, which collects motion data of the dynamic weighing module through the built-in acceleration sensor and position sensor; taking the device movement in a treatment process as an example, the unit detects that the weighing module has an unexpected horizontal displacement, and calculates the value and direction of the movement vector by integrating the acceleration data, and immediately triggers further processing procedures when the movement exceeds the preset safety threshold of 50 mm. The vector direction acquisition unit then analyzes the spatial orientation of the movement vector, which uses a nine-axis sensor data fusion algorithm to accurately calculate the angular coordinates of the movement direction; in the above device movement example, the system detects that the movement vector is mainly directed to the patient side, with a vector value of 65 mm, which is significantly higher than the safety threshold, at which time the unit converts the movement direction into azimuth angle data in the device coordinate system. The second warning unit starts the environment detection verification process after the vector direction analysis is completed, which uses the visual detection device integrated by the scheduling system to monitor the direction indicated by the movement vector in real time; when the camera captures the displacement image of the dynamic weighing module in the specified direction, the second warning unit immediately generates a device movement warning instruction, which contains detailed information such as movement distance, direction, possible cause and recommended disposal scheme.

[0098] The state acquisition unit adopts a multi-modal state recognition algorithm in the implementation process, combines a machine learning classifier and a rule engine to determine the state type, maintains a state type knowledge base containing feature descriptions and recognition rules of various typical abnormal patterns, and performs real-time matching analysis on the input balance feature data every second. The judgment unit implements physiological parameter correlation analysis function, not only checks whether a single parameter is over the threshold value, but also analyzes the correlation mode between multiple parameters; the unit adopts a dynamic threshold adjustment mechanism, adjusts the threshold parameter according to the patient's basic physiological data and historical records, and improves the accuracy of monitoring. The continuous duration acquisition unit adopts a distributed time recording architecture, synchronously records time information at multiple nodes of the system, and prevents time record loss caused by single point failure; the unit also implements time record verification function, and verifies the accuracy of continuous time calculation by cross-comparing timestamp data of different sensors. The first warning unit supports a multi-level warning mechanism, generates warning instructions of different levels according to the severity and duration of the abnormality; the warning content is generated by using a standardized template to ensure the accuracy and integrity of the information expression, and multiple language outputs are supported to meet the needs of medical staff in different regions.

[0099] The movement vector acquisition unit implements motion pattern recognition function, which can distinguish between intentional device adjustment and accidental movement; the unit adopts an adaptive filtering algorithm to eliminate the influence of environmental vibration and accidental interference on movement detection, and improves the reliability of movement vector measurement. The vector direction acquisition unit integrates multiple direction sensing technologies, including magnetometer, gyroscope and visual positioning data, improves the accuracy and stability of direction measurement through sensor fusion; the unit also implements direction coordinate conversion function, which can convert the direction data in the device coordinate system into clinical position description which is easier for medical staff to understand. The second warning unit implements intelligent device scheduling function, which can automatically select the most suitable detection device for verification according to the characteristics of the movement vector; the unit supports the cooperative work of multiple detection devices, including fixed cameras, mobile inspection devices and infrared sensors, etc., forming a multi-angle verification network. The whole balance monitoring module adopts real-time stream processing architecture, which can process the monitoring data of multiple patients at the same time; the module realizes dynamic resource allocation mechanism, which allocates computing resources according to the urgency and complexity of monitoring tasks. The state acquisition unit provides state recognition transparency report function, which can explain the basis and confidence of each state judgment, and facilitate medical staff to understand the system decision-making process. The judgment unit implements threshold adaptive learning function, which can automatically optimize the threshold parameter according to the historical monitoring data, and continuously improve the monitoring sensitivity. The continuous duration acquisition unit provides timeline visualization tool, which can graphically display the time distribution and duration mode of parameter abnormality.

[0100] The first early warning unit realizes the early warning effect tracking function, records the response and processing result of each early warning instruction, and forms the early warning closed loop management. The mobile vector acquisition unit provides the mobile mode analysis report, and statistically calculates the frequency, amplitude and regularity of the equipment movement, to provide data support for the equipment placement optimization. The vector direction acquisition unit realizes the direction data quality monitoring, and evaluates the accuracy and reliability of the direction measurement in real time. The second early warning unit provides the equipment linkage test function, and regularly verifies the working state and coordination performance of the detection equipment.

[0101] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications, changes, variations, substitutions, and equivalents will occur to those of ordinary skill in the art to which the application relates and are encompassed within the spirit and scope of the application as defined by the following claims.

Claims

1. A peritoneal dialysis fluid balance monitoring system based on dynamic weighing and data fusion, characterized in that, include: The dynamic weighing module is used to collect the input and output of dialysate during peritoneal dialysis in real time and generate dialysate weight data. The multi-source data acquisition module is used to acquire data on dialysate composition, treatment time, and patient physiological parameters. The data fusion processing module is used to fuse the dialysate weight data and the data collected by the multi-source data acquisition module to generate dialysate balance characteristic data. The balance monitoring module is used to calculate the net ultrafiltration volume and fluid balance state based on the dialysate balance characteristic data, and to identify the risk of abnormal leakage or fluid retention. The early warning output module is used to generate early warning information when abnormal leakage or liquid retention risk is detected; The data fusion processing module includes: The dialysate balance reference table creation unit is used to create a dialysate balance reference table containing dialysate balance reference data for different treatment stages. The feature autoencoder construction unit is used to construct a target dialysate balance feature autoencoder based on the dialysate balance reference data and the corresponding real-time dialysate balance code. The reference data matching unit is used to match the primary dialysate balance reference data that is closest to the current dialysate weight data from the dialysate balance comparison table. The feature processing unit is used to perform feature processing on a partial dialysate balance comparison table corresponding to the primary dialysate balance reference data using the target dialysate balance feature autoencoder, and output the current dialysate balance feature result.

2. The peritoneal dialysis fluid balance monitoring system based on dynamic weighing and data fusion as described in claim 1, characterized in that, The feature autoencoder construction unit includes: The reference data matching subunit is used to match reference dialysate balance reference data from the dialysate balance comparison table; The reference table extraction subunit is used to extract the reference portion of the dialysate balance reference table corresponding to the reference dialysate balance reference data; The pointing determination subunit is used to perform a pointing determination operation on the dialysate balance matching information in the reference partial dialysate balance comparison table, and generate an updated reference partial dialysate balance comparison table with matching pointing information. The fusion operation subunit is used to perform feature fusion operation on the updated reference part of the dialysate balance comparison table by using the initial dialysate balance feature autoencoder and combining the matching pointing information, and output the representative feature position corresponding to each dialysate balance reference data. The parameter adjustment subunit is used to adjust the parameters of the initial dialysate balance feature autoencoder based on the representative feature position and the real-time dialysate balance code, and generate the target dialysate balance feature autoencoder.

3. The peritoneal dialysis fluid balance monitoring system based on dynamic weighing and data fusion as described in claim 2, characterized in that, The lookup table extraction subunit includes: The difference degree judgment subunit is used to determine, from the dialysate balance comparison table, the dialysate balance reference data whose matching difference degree with the reference dialysate balance reference data is not greater than the preset reference matching difference degree; The reference table generation subunit is used to generate the reference partial dialysate balance reference table based on the matching information between the reference dialysate balance reference data and the determined dialysate balance reference data.

4. The peritoneal dialysis fluid balance monitoring system based on dynamic weighing and data fusion as described in claim 2, characterized in that, The fusion operation subunit includes: The matching sub-unit is used to traverse the dialysate balance reference data in the updated reference portion dialysate balance comparison table and match the adjacent dialysate balance reference data of the currently traversed dialysate balance reference data in the reference portion dialysate balance comparison table. The fusion processing subunit is used to utilize the initial dialysate balance feature autoencoder, combined with the matching pointing information between the currently traversed dialysate balance reference data and each adjacent dialysate balance reference data, to perform a fusion operation on the feature positions of the currently traversed dialysate balance reference data and each adjacent dialysate balance reference data, and output the current fused feature position. The representative feature determination subunit is used to determine the representative feature position corresponding to each dialysate balance reference data based on the current fusion feature position.

5. The peritoneal dialysis fluid balance monitoring system based on dynamic weighing and data fusion as described in claim 4, characterized in that, The fusion processing subunit includes: The orientation analysis subunit is used to analyze the position orientation of each matching pointing information between the currently traversed dialysate balance reference data and the adjacent dialysate balance reference data; The marking subunit is used to mark the matching pointing information of the end position of the position orientation belonging to the currently traversed dialysate balance reference data as primary matching pointing information; The data fusion tagging subunit is used to analyze the dialysate balance reference data corresponding to the initial position of the position orientation of the elementary matching pointing information, and to tag it as the dialysate balance reference data to be fused. The information fusion subunit is used to perform information fusion operations on the feature positions of the dialysate balance reference data to be fused and the currently traversed dialysate balance reference data using the initial dialysate balance feature autoencoder, and output the current fused feature position.

6. The peritoneal dialysis fluid balance monitoring system based on dynamic weighing and data fusion as described in claim 5, characterized in that, The information fusion subunit includes: The importance distribution determination subunit is used to determine the distribution of importance parameters corresponding to each reference data point for fusion of dialysate balance; The feature analysis subunit is used to perform classification feature analysis processing on the feature positions of the currently traversed dialysate balance reference data and each dialysate balance reference data to be merged according to the distribution of their importance parameters using the initial dialysate balance feature autoencoder, and output the classification importance parameters corresponding to each dialysate balance reference data to be merged. The fusion mapping subunit is used to perform fusion mapping operations on the feature positions and classification importance parameters corresponding to each dialysis fluid balance reference data to be fused, and generate fusion mapping information; The excitation output subunit is used to perform fusion excitation output operation on the fusion mapping information corresponding to all fusion reference data of the dialysate to be fused, and output the current fusion feature position.

7. The peritoneal dialysis fluid balance monitoring system based on dynamic weighing and data fusion as described in claim 2, characterized in that, The parameter adjustment subunit includes: The identification coding generation subunit is used to perform feature processing on the reference dialysate balance reference data based on the representative feature positions, and output the identification dialysate balance code; The parameter adjustment operation subunit is used to analyze the parameter adjustment range by using the identified dialysate balance code and the real-time dialysate balance code, and adjust the parameters of the initial dialysate balance feature autoencoder according to the range to generate the target dialysate balance feature autoencoder.

8. The peritoneal dialysis fluid balance monitoring system based on dynamic weighing and data fusion as described in claim 1, characterized in that, The balance monitoring module includes: A state acquisition unit is used to acquire the equilibrium state type in the dialysate equilibrium characteristic data; The judgment unit is used to determine whether the patient's physiological parameter data is lower than a preset physiological threshold when the balance state type is abnormal. The duration acquisition unit is used to acquire the duration below the preset physiological threshold when the patient's physiological parameter data is below the preset physiological threshold. The first early warning unit is used to generate a first early warning instruction when the duration exceeds a preset duration threshold. The movement vector acquisition unit is used to acquire the movement vector of the dynamic weighing module when the balance state type is not an abnormal type. A vector direction acquisition unit is used to acquire the vector direction of the moving vector when the vector value of the moving vector is greater than a preset moving threshold. The second early warning unit is used to generate a second early warning command when the detection device corresponding to the vector direction detects the dynamic weighing module.

9. A method for monitoring peritoneal dialysis fluid balance using dynamic weighing and data fusion, applied to the peritoneal dialysis fluid balance monitoring system using dynamic weighing and data fusion as described in any one of claims 1 to 8, characterized in that, The method includes: S1. Real-time collection of dialysate input and output during peritoneal dialysis through dynamic weighing, and generation of dialysate weight data; S2. Acquire multi-source data, including dialysate composition data, treatment time data, and patient physiological parameter data; S3. The dialysate weight data and the multi-source data are fused to generate dialysate balance characteristic data. S4. Based on the dialysate balance characteristic data, calculate the net ultrafiltration volume and fluid balance state, and identify the risk of abnormal leakage or fluid retention. S5. When an abnormal leak or liquid retention risk is detected, generate and output an early warning message.

Images