Hemodialysis room whole-process information management system based on active learning

Abstract

The invention provides a hemodialysis room whole-process information management system based on active learning, and relates to the technical field of data processing, and the system comprises a feature construction module which is used for carrying out the structural processing of original dialysis information data, and extracting a plurality of feature factors related to the current physical state of a patient; the risk identification module is used for constructing a training sample group, performing supervised training by introducing a label, and generating a risk judgment model; the early warning generation module is used for generating a risk probability result and generating early warning information associated with the current patient identifier when the risk probability result reaches a preset risk threshold value; the scheduling module is used for dynamically adjusting the treatment scheduling sequence of the current patient according to the early warning information and recording the actual intervention effect after the treatment scheduling sequence is adjusted; the feedback learning module is used for endowing each feature factor with a weight value according to an actual intervention feedback result and feeding back the weight value to the risk identification module; according to the invention, autonomy and accuracy of information management are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a full-process information management system for a hemodialysis room based on active learning. Background Art

[0002] Existing technologies often utilize static, rule-driven process control systems for managing the entire hemodialysis process, including registration, evaluation, pre-dialysis examinations, treatment records, and post-dialysis follow-up, based on traditional hospital information systems (HIS) or electronic medical records (EMR). Data collection primarily relies on manual input, with some systems incorporating barcodes or RFID to assist in identifying errors. While these systems can cover basic operational processes, they lack the ability to dynamically adapt information, often failing to automatically adjust processes based on patient conditions. Furthermore, they lack the ability to predict and proactively intervene in special circumstances.

[0003] In the hemodialysis centers of some county-level hospitals, most patients are elderly patients with chronic kidney disease, accompanied by multiple complications, and their conditions are complex and changing. During the pre-dialysis assessment phase, traditional systems often use fixed tables or static indicator thresholds to classify patients, failing to identify individuals with potential risks. For example, a patient's heart rate was slightly elevated before dialysis but did not exceed the set threshold. The system did not issue an alarm, and medical staff continued treatment based on this information, ultimately causing intraoperative hypotension. Such systems lack active learning capabilities and fail to continuously optimize judgment criteria based on historical data. This leads to delayed identification of clinical risks, an inability to provide timely warnings and personalized treatment recommendations, and poses potential safety risks. Summary of the Invention

[0004] The purpose of the present invention is to provide a full-process information management system for hemodialysis rooms based on active learning, aiming to solve the problems mentioned in the background technology.

[0005] In order to solve the above technical problems, the technical solutions of the present invention are as follows:

[0006] A full-process information management system for hemodialysis rooms based on active learning, the system comprising:

[0007] The feature construction module is used to perform structured processing on the original dialysis information data to generate a standard dialysis information dataset, and extract multiple feature factors related to the patient's current physical condition from it, including heart rate variability, blood pressure fluctuation trend, and weight residual, to generate a feature dataset;

[0008] The risk identification module is used to compare the feature dataset with the standard dialysis information dataset, construct a training sample group, and introduce labels for supervised training to generate a risk judgment model for identifying specific high-risk states;

[0009] An early warning generation module is used to input the feature data set into the risk judgment model to generate a risk probability result. When the result reaches a preset risk threshold, an early warning message is generated that is associated with the current patient identifier. The early warning message includes the patient identifier, risk level, and processing time requirement.

[0010] The scheduling module is used to dynamically adjust the treatment schedule of the current patient based on the early warning information, and record the actual intervention effect after the adjustment of the treatment schedule to form the actual intervention feedback results;

[0011] The feedback learning module is used to calculate the correlation between the improvement degree of post-dialysis physiological monitoring indicators reflected in the actual intervention feedback results and the characteristic factors to determine the contribution of each characteristic factor, assign a corresponding weight value to each characteristic factor, generate a characteristic weighted data set, and feed it back to the risk identification module to execute the next round of training.

[0012] Preferably, the feature construction module includes:

[0013] The feature preprocessing submodule is used to standardize the format and unit of the numerical fields in the original dialysis information data, and to fill in missing values. It also analyzes and identifies the medical record fields contained therein, and automatically extracts information with physician risk event markers, intraoperative intervention records, or complication codes as high-risk event annotation fields to generate a standard dialysis information dataset.

[0014] A feature extraction submodule is used to extract multiple feature factors related to the patient's current physical state from the standard dialysis information data set, wherein the feature factors include heart rate change rate, blood pressure fluctuation trend and weight residual;

[0015] The feature generation submodule is used to convert the feature factor set into a feature vector representation of uniform dimension to generate a feature data set.

[0016] Preferably, the risk identification module includes:

[0017] The sample construction submodule is used to compare the feature dataset with the historical high-risk annotated data in the standard dialysis information dataset and generate a training sample group containing positive and negative samples and corresponding labels;

[0018] The model training submodule is used to automatically generate multiple risk judgment model structure templates based on the training sample group, input the training sample group into each model structure template, perform iterative training and error optimization, and output the performance evaluation indicators of each model structure template;

[0019] The model publishing submodule is used to calculate the performance score based on the performance evaluation indicators, and store the risk judgment model with the highest ranking and the performance score reaching the set score threshold as a callable version.

[0020] Preferably, the shift scheduling module includes:

[0021] The task management submodule is used to extract the patient identification, risk level and processing time requirements of the current patient based on the early warning information, and generate a queue of tasks to be scheduled;

[0022] The resource matching submodule is used to obtain the current idle status, emergency capacity status, expected idle time and historical dialysis success rate of each dialysis unit and quantify them into resource scoring factors;

[0023] The resource matching submodule is further configured to calculate a matching priority score based on the risk level and processing time requirements of each patient in the waiting task queue, combined with a resource scoring factor, and sort the patients from high to low according to the matching priority score, match the optimal dialysis resources, and form a resource matching plan;

[0024] The execution record submodule is used to perform scheduling adjustments according to the resource matching plan, record the scheduling execution time, intervention conditions and actual effects during the dialysis process, and generate actual intervention feedback results.

[0025] Preferably, the feedback learning module includes:

[0026] The result analysis submodule is used to analyze the improvement degree of post-dialysis physiological monitoring indicators based on the actual intervention feedback results and generate an improvement indicator data set;

[0027] The factor contribution calculation submodule is used to match the improvement index dataset with each characteristic factor in the characteristic dataset, and calculate the contribution of each characteristic factor to the improvement index to generate a characteristic contribution dataset;

[0028] The weight generation submodule is used to assign corresponding weight values ​​to each feature factor according to the feature contribution dataset, generate a feature weighted dataset, and output the feature weighted dataset to the risk identification module for the next round of model training.

[0029] Preferably, the sample construction submodule includes:

[0030] A labeling and screening unit is used to extract historical dialysis records with labeled high-risk events from a standard dialysis information dataset to form a labeled sample dataset;

[0031] A feature label fusion unit is used to pair each set of feature vectors in the feature data set with the corresponding label in the labeled sample data set to construct a training sample group, wherein the training sample group includes feature-label combinations of positive samples and negative samples;

[0032] The sample integration unit is used to divide the training sample group into training set and test set, and record the sample source, label type and construction timestamp.

[0033] Preferably, the model training submodule includes:

[0034] A structure initialization unit is used to automatically generate multiple risk judgment model structure templates based on the characteristic dimensions and label distribution characteristics of the training sample group;

[0035] The error optimization unit is used to input the training sample group into each model structure template, calculate the prediction error of each iteration, and adjust the parameters according to the error feedback results until the training error is lower than the preset error threshold;

[0036] The performance evaluation unit is used to receive multiple risk judgment models that have completed training and calculate their precision, recall, F1-score and AUC indicators on the test set.

[0037] Preferably, the resource matching submodule includes:

[0038] The resource status quantification unit is used to obtain the idle time period, current load rate, equipment status fluctuation and historical dialysis success rate of each dialysis unit, and assign corresponding resource scoring factor values ​​to form a resource status matrix;

[0039] The priority score calculation unit is used to assign the patient's risk level and processing time requirement as risk score factors and time limit score factors respectively according to the queue of tasks to be scheduled, and perform weighted fusion with the resource status matrix to calculate the matching priority score value between each resource and each task;

[0040] The matching screening unit is used to sort resources according to the priority score value and screen the resources with free time within the expected treatment time period in each dialysis unit. For each high-priority patient, the dialysis unit with the highest matching priority score and the available time that does not conflict with the treatment time requirements is selected as the recommended scheduling result; if there is a conflict of the same score or time overlap, a secondary screening is performed based on the historical scheduling success rate of the patient and the resource, and finally a resource matching plan is formed.

[0041] Preferably, the factor contribution calculation submodule includes:

[0042] A data matching unit is used to receive the improvement index data set and the feature data set, and to match each improvement index with its corresponding feature factor one by one to form a matching data set;

[0043] The contribution calculation unit is used to calculate the fit between each characteristic factor and the corresponding improvement index through the least squares regression method based on the matching data set, and use the fit as the characteristic contribution value;

[0044] The contribution integration unit is used to merge the feature contribution values ​​of each feature factor into a data set to generate a feature contribution data set.

[0045] Preferably, the weight generation submodule includes:

[0046] The weight update unit is used to assign an initial weight value to each feature factor based on the feature contribution data set, and update the current round weight based on the difference between its feature contribution value and the weight value of the previous round;

[0047] A normalization processing unit is used to normalize the current round weight values ​​of all feature factors to ensure that the sum of all weight values ​​is 1;

[0048] The weighted data generating unit is used to perform a numerical product operation on each set of feature vectors in the feature data set and the corresponding normalized weight to generate a feature weighted data set.

[0049] The above solution of the present invention includes at least the following beneficial effects:

[0050] This invention achieves intelligent, closed-loop management of the patient dialysis process by constructing a full-process information management system for hemodialysis rooms based on an active learning mechanism. Unlike existing process control systems that rely on static rules and manual input, this system continuously acquires and structures multi-dimensional raw dialysis data, including recent vital signs, historical treatment data, and basic pathology information. Based on this, it constructs a dynamic feature dataset reflecting the patient's status, providing accurate support for subsequent risk identification and scheduling interventions.

[0051] By introducing a risk identification module, the system compares and analyzes the current patient's characteristic data with historical data samples, constructing a supervised learning model to identify high-risk conditions. This proactively generates risk probability results in the pre-transmission phase. This, in conjunction with the early warning generation module, outputs early warning information in real time, including patient identification, risk level, and processing time requirements, ensuring that medical actions are proactive and accurate. Compared to traditional fixed threshold triggering mechanisms, this method can still issue risk warnings in a timely manner even when multiple dimensions change simultaneously but a single indicator remains within its limit, effectively avoiding misjudgments.

[0052] The system also features a scheduling and feedback learning module, which dynamically adjusts treatment sequences based on early warning information and records intervention effectiveness. Improved metrics then statistically analyze the contribution of characteristic factors, generating a weighted feature dataset for optimizing the next round of models. This closed-loop mechanism empowers the system with continuous learning and strategic evolution capabilities, significantly improving model adaptability and judgment accuracy. This is particularly applicable to chronic dialysis patients with fluctuating conditions and complex comorbidities. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is an architecture diagram of a full-process information management system for a hemodialysis room based on active learning provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0054] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0055] like Figure 1 As shown, an embodiment of the present invention proposes a full-process information management system for hemodialysis rooms based on active learning, the system comprising:

[0056] The information collection module is used to obtain original dialysis information data, including recent vital signs data, previous dialysis treatment data and basic pathology data;

[0057] The feature construction module is used to perform structured processing on the original dialysis information data to generate a standard dialysis information dataset, and extract multiple feature factors related to the patient's current physical condition from it, including heart rate variability, blood pressure fluctuation trend, and weight residual, to generate a feature dataset;

[0058] The risk identification module is used to compare the feature dataset with the standard dialysis information dataset, construct a training sample group, and introduce labels for supervised training to generate a risk judgment model for identifying specific high-risk states;

[0059] An early warning generation module is used to input the feature data set into the risk judgment model to generate a risk probability result. When the result reaches a preset risk threshold, an early warning message is generated that is associated with the current patient identifier. The early warning message includes the patient identifier, risk level, and processing time requirement.

[0060] The scheduling module is used to dynamically adjust the treatment schedule of the current patient based on the early warning information, and record the actual intervention effect after the adjustment of the treatment schedule to form the actual intervention feedback results;

[0061] The feedback learning module is used to calculate the correlation between the improvement degree of post-dialysis physiological monitoring indicators reflected in the actual intervention feedback results and the characteristic factors to determine the contribution of each characteristic factor, assign a corresponding weight value to each characteristic factor, generate a characteristic weighted data set, and feed it back to the risk identification module to execute the next round of training.

[0062] In an embodiment of the present invention, by constructing a complete system including an information collection module, a feature construction module, a risk identification module, an early warning generation module, a shift scheduling module, and a feedback learning module, closed-loop management of the entire process information of patients in the hemodialysis room can be achieved. The information collection module can collect key physiological data of patients during the dialysis process in real time. The original dialysis information data obtained covers recent vital signs data, previous dialysis treatment data, and basic pathology data, ensuring the comprehensiveness of the data. The feature construction module can standardize and structure the data, screen out characteristic factors with clinical relevance, such as heart rate variability, blood pressure fluctuation trend, and weight residual, and form a feature data set.

[0063] Based on the feature data set, the risk identification module constructs a risk judgment model through training, which can judge the risk level of different patient conditions. When the risk probability of a patient exceeds the threshold, the warning generation module will promptly output the warning information associated with the patient's identification, including its risk level and processing time requirements. The scheduling module dynamically adjusts the treatment scheduling order based on these warning information and records the actual effect after the intervention, thereby improving the efficiency of resource allocation. Finally, the feedback learning module quantitatively adjusts the importance of the characteristic factors through the degree of improvement of physiological monitoring indicators after dialysis, realizes weight update, and then feeds back the intervention results to the risk judgment model for the next round of training, forming a self-optimization closed loop.

[0064] The implementation of this system can significantly improve the accuracy of predictions and the timeliness of interventions for high-risk patients. For example, in actual operation, the system can proactively identify patients at high risk of hypotension and prioritize their placement in dialysis units equipped with emergency medical equipment. The results of these interventions are also recorded and used for model optimization, making the overall dialysis process safer, more efficient, and more intelligent.

[0065] The information collection module is used to obtain and summarize the original dialysis information data of dialysis patients, including:

[0066] This module realizes the unified collection of data from multiple sources by connecting to information sources such as the hospital information system (HIS), electronic medical record system (EMR) and dialysis equipment monitoring terminal. The collected data include the patient's recent vital signs data (such as heart rate, blood pressure, body temperature, respiratory rate, etc.), previous dialysis treatment data (such as dialysis date, dialysis method, dialysis duration, concurrent event records, etc.), and the patient's basic pathological data (such as diagnosis code, past medical history, combined medication information, etc.). To ensure the continuity and accuracy of data collection, the information collection module supports communication with ICU monitors, blood pressure monitors, blood filtration equipment, etc., and regularly reads the latest indicators through the Internet of Things interface. It also supports the use of RFID or barcode tags to automatically identify the patient's identity to prevent information mismatch.

[0067] The module also has a standardized data structure interface, which uniformly parses data formats from different sources (such as JSON structure, XML format, or database fields) during the collection process and outputs them in a unified structure for use in subsequent modules. For example, when vital signs data is transmitted via the serial port protocol, the system automatically extracts the numerical fields and timestamps and maps them to structured fields such as "HR", "SBP", and "DBP". The data refresh frequency of this module can be set according to the dialysis scenario, usually every 5 to 15 minutes, to ensure that clinical decisions are based on the latest status.

[0068] The warning generation module is used to receive the high-risk probability results output by the risk identification module and, in combination with the risk determination rules set by the system, generate structured warning information associated with the current patient identifier, specifically including:

[0069] First, the system inputs the patient's feature dataset into the previously generated risk assessment model. Based on the sample discrimination boundaries learned during training, the model calculates the probability of a high-risk event occurring in the patient's current state. This probability, a value between 0 and 1, reflects the system's risk assessment of the patient's condition.

[0070] The system then compares this probability value with a pre-set risk threshold. This threshold can be set empirically based on clinical trial data, historical intervention records, and other factors, such as 0.65. If the patient's risk probability result is equal to or higher than this threshold, the system deems the patient to be significantly high-risk and triggers an immediate alert.

[0071] To enhance the clinical applicability of these assessments, the system supports configuring different risk thresholds and comprehensively determines whether to trigger an alert based on factors such as the rate of risk change and historical trends. For example, even if a patient's risk value remains within the current threshold, but it shows a continuous upward trend compared to previous predictions, the system may determine that the patient is in a "needs attention" state and generate a low-level alert.

[0072] The structure of early warning information includes three categories of content: (1) patient unique identification information, such as medical record number, dialysis bed number, etc.; (2) the current risk level, which is usually divided into three levels: "high risk", "medium risk", and "needs attention"; (3) processing time requirements, that is, the recommended latest scheduling processing time, such as "immediate processing" and "processing within 30 minutes".

[0073] This module can also include a summary of the key features currently influencing the prediction, listing the names of the features that contribute most to the current prediction and their corresponding values. For example, in one case, the system generated an alert with the following additional note: "The current heart rate variability is abnormal, and the weight residual is continuing to increase. It is suspected that volume control is poor. Immediate scheduling adjustments are recommended."

[0074] Once generated, warning information is pushed through an interface to the medical scheduling system, the nurse's PDA terminal, or the doctor's workstation. It can also be written to the patient's dialysis electronic medical record, creating a traceable warning log. This module significantly enhances the information system's responsiveness to high-risk patients, particularly in scenarios where traditional systems often delay identification due to a single indicator failing to cross a threshold. This module provides a more sensitive and intelligent basis for judgment.

[0075] In a preferred embodiment of the present invention, the feature construction module includes:

[0076] The feature preprocessing submodule is used to standardize the format and unit of the numerical fields in the original dialysis information data, and to fill in missing values. It also analyzes and identifies the medical record fields contained therein, and automatically extracts information with physician risk event markers, intraoperative intervention records, or complication codes as high-risk event annotation fields to generate a standard dialysis information dataset.

[0077] A feature extraction submodule is used to extract multiple feature factors related to the patient's current physical state from the standard dialysis information data set, wherein the feature factors include heart rate change rate, blood pressure fluctuation trend and weight residual;

[0078] The feature generation submodule is used to convert the feature factor set into a feature vector representation of uniform dimension to generate a feature data set.

[0079] In an embodiment of the present invention, by setting up three submodules in the feature construction module, the system can efficiently extract feature data sets that are critical for risk prediction from the original dialysis information data and provide a structured data format that can be used for subsequent model training.

[0080] The feature preprocessing submodule not only completes the format unification, unit conversion and missing field completion of the original dialysis information data from different sources, but also further enhances the semantic recognition capability of the system. In particular, when processing text or structured fields containing medical records, intraoperative monitoring documents and dialysis abnormality registration, the system can automatically identify entries containing doctors' subjective risk markers (such as "intraoperative hypotension", "dialysis interruption", etc.), complication diagnosis codes (such as ICD codes corresponding to high-risk conditions) or nursing abnormal operation records, and set them as "high-risk event annotation fields". This makes the final generated standard dialysis information data set not only structurally complete, but also has traceable label information, which can provide a clear basis for the subsequent construction of supervised learning sample sets.

[0081] In the feature extraction submodule, the system combines clinical experience and statistical analysis methods to extract multiple dynamic factors that are strongly correlated with the patient's current physical condition from the standard dialysis information data set. For example, it calculates the heart rate change rate, blood pressure fluctuation trend, and weight residual through multiple dialysis records. These factors not only reflect the patient's current health fluctuations, but also have a certain predictive value for future dialysis risks. Unlike traditional systems that only collect static indicators, this invention calculates trend factors through multi-period historical data, improving the time sensitivity of the risk judgment dimension.

[0082] The feature generation submodule integrates the extracted multiple feature factors into a uniformly structured feature vector, ensuring consistent data dimensions for all subsequent machine learning modules. Feature vector data undergoes normalization and outlier smoothing before generation, ensuring that the system maintains model convergence and generalization capabilities even when individual data scales vary.

[0083] For example, after deploying this system in a dialysis center, by identifying the "premature termination of dialysis during surgery" event in the patient's medical record as a risk label, and combining it with the weight residual change trend calculated from historical data, 100 sets of training data containing positive and negative samples were successfully constructed, effectively supporting the first round of risk judgment capability improvement of the model. This shows that this module plays an important role in improving data value density and ensuring model availability. Compared with the "one-way, unstructured" operation mechanism of traditional systems in clinical data preprocessing, the present invention provides a data preparation method with a clear structure, rich content, and friendly to subsequent modules, which has significant technological advancement and application promotion value.

[0084] The feature preprocessing submodule is used to unify, clean, and annotate the numerical fields and event record fields in the original dialysis information data to generate a structured standard dialysis information dataset that can be processed later. Specifically, it includes:

[0085] First, this submodule standardizes the format of various numerical fields in the raw dialysis information data. Because different devices or information sources (such as blood pressure monitors, dialysis recording systems, and EMR platforms) may have inconsistent numerical units and data formats, the system uses built-in field mapping tables and unit conversion logic to standardize fields such as heart rate (standardized to "beats per minute"), weight (standardized to "kilograms"), and blood pressure (standardized to "mmHg"), ensuring that all samples have consistent semantics and dimensions under the same fields.

[0086] Secondly, the system identifies and fills missing values. When a key field is missing (e.g., pre-dialysis systolic blood pressure is blank), the system interpolates the time series mean based on the patient's historical records. If historical records are unavailable, the system backfills with the median of the same risk group to maintain data continuity.

[0087] More importantly, this submodule also introduces the "high-risk event labeling and identification" function. The system automatically identifies medical events that represent potential high-risk states by parsing information such as structured forms (such as dialysis abnormality registration forms), semi-structured text (such as nursing records), and diagnostic coding fields (such as ICD-10, CCS codes). For example, the system can identify events such as "premature termination of dialysis during surgery", "multiple hypotension", and "blood circuit obstruction", and convert these records into Boolean or multi-level labeling labels, which are bound to the corresponding original data entries to form training data with supervision attributes. The system can also extract risk labels and structure them based on manually labeled fields by doctors.

[0088] The final output standard dialysis information dataset not only includes basic physiological indicators after cleaning, but also comes with high-risk event annotation fields that can be used for supervised learning. It is the core foundation for the subsequent construction of training sample groups and risk identification models.

[0089] The feature extraction submodule is used to extract multiple feature factors related to the patient's current physical condition from the standard dialysis information dataset, including:

[0090] This submodule first performs trend analysis on time series data to identify dynamic changes in the patient's physiological state. For example, by calculating the range and direction of systolic blood pressure changes over three consecutive dialysis cycles, a blood pressure fluctuation trend factor can be constructed. By calculating the average deviation of weight fluctuations over the past seven days, a weight residual factor can be derived to characterize the patient's degree of volume overload imbalance.

[0091] For the heart rate variability factor, the system compares the patient's average heart rate before, during, and after dialysis. It then combines the magnitude and duration of the fluctuation to determine the patient's heart rate fluctuation, reflecting the level of sympathetic tone. Furthermore, this submodule supports the extraction of custom feature rules. For example, by combining a history of low-sodium dialysis use with recent differences in blood sodium levels, it can automatically generate a potential risk indicator for electrolyte imbalance.

[0092] All extracted characteristic factors will be attached to the patient data structure as multidimensional vector fields. Each factor has a clear name, calculation source and physiological meaning to ensure its traceability and clinical interpretability.

[0093] The feature generation submodule is used to convert the feature factor set into a feature vector representation of uniform dimension and generate a feature data set, which specifically includes:

[0094] This submodule receives the feature factor results output by the previous submodule and constructs a numerical vector according to the preset feature field order. To avoid dimensionality inconsistencies during model training, the system dimensionalizes all features. That is, regardless of whether a patient has certain feature items, the system uses padding values ​​to ensure that the vector length is constant.

[0095] During the construction process, the system also normalizes feature values ​​across different dimensions. For example, for the heart rate variability field, the system linearly maps it within a certain range, ensuring that the value fluctuates within the interval [0, 1], thereby maintaining a consistent scale with indicators such as weight residuals. For discrete features (such as dialysis mode classification), the system uses one-hot encoding to convert them into 0-1 vectors and incorporates them into the feature set.

[0096] Ultimately, the module outputs a feature vector corresponding to each patient, which serves as input for the risk assessment model. This vector structure supports batch generation and caching, allowing for subsequent rounds of training and model feedback, ensuring high system performance and feature consistency.

[0097] In a preferred embodiment of the present invention, the risk identification module includes:

[0098] The sample construction submodule is used to compare the feature dataset with the historical high-risk annotated data in the standard dialysis information dataset and generate a training sample group containing positive and negative samples and corresponding labels;

[0099] The model training submodule is used to automatically generate multiple risk judgment model structure templates based on the training sample group, input the training sample group into each model structure template, perform iterative training and error optimization, and output the performance evaluation indicators of each model structure template;

[0100] The model publishing submodule is used to calculate the performance score based on the performance evaluation indicators, and store the risk judgment model with the highest ranking and the performance score reaching the set score threshold as a callable version.

[0101] In this embodiment of the present invention, the risk identification module comprises a sample construction submodule, a model training submodule, and a model publishing submodule, offering advantages such as a clear structure and intelligent processing. By comparing a feature dataset with historically high-risk annotated data in a standard dialysis information dataset, the sample construction submodule can quickly identify clearly labeled event data, such as dialysis records of hypotension and potassium disturbances, thereby constructing a training sample set containing positive and negative samples and corresponding labels.

[0102] The model training submodule performs structural modeling on the training sample set, automatically generating multiple risk assessment model structure templates. These are then fed into the training sample set for iterative training and error optimization, generating a batch of risk assessment models. After training, each model is evaluated on its accuracy, recall, F1-score, and AUC values ​​on the test set, with these metrics output as performance evaluation indicators. Subsequently, the model publishing submodule scores all models according to predefined metrics, such as using a weighted calculation to comprehensively score four metrics. The model with the highest score that meets the predefined threshold is then marked as the current best model and used as the currently available version of the system for subsequent risk prediction.

[0103] This structure enables the system to continuously maintain optimal performance and enable rapid updates. For example, during a system update cycle, a number of models with weak generalization capabilities were eliminated through the model scoring mechanism, ensuring that the AUC of the final model being called always remained above 0.93, effectively avoiding the risk of misjudgment caused by model degradation.

[0104] In a preferred embodiment of the present invention, the shift scheduling module includes:

[0105] The task management submodule is used to extract the patient identification, risk level and processing time requirements of the current patient based on the early warning information, and generate a queue of tasks to be scheduled;

[0106] The resource matching submodule is used to obtain the current idle status, emergency capacity status, expected idle time and historical dialysis success rate of each dialysis unit and quantify them into resource scoring factors;

[0107] The resource matching submodule is further configured to calculate a matching priority score based on the risk level and processing time requirements of each patient in the waiting task queue, combined with a resource scoring factor, and sort the patients from high to low according to the matching priority score, match the optimal dialysis resources, and form a resource matching plan;

[0108] The execution record submodule is used to perform scheduling adjustments according to the resource matching plan, record the scheduling execution time, intervention conditions and actual effects during the dialysis process, and generate actual intervention feedback results.

[0109] In this embodiment of the present invention, the scheduling module implements risk-prioritized allocation of dialysis treatment resources through the collaborative work of the task management submodule, resource matching submodule, and execution recording submodule. The task management submodule receives warning information output by the warning generation module and extracts information such as the current patient's identification, risk level, and processing time requirement to construct a queue of tasks to be scheduled in real time. This queue is dynamically updated and reflects the urgency of patient risk management in the form of time windows.

[0110] The resource matching submodule obtains key resource information such as each dialysis unit's current availability, emergency response capacity, expected idle time, and historical dialysis success rate, quantifies this resource data, and converts it into a unified scoring factor. The system calculates a priority matching score for each "patient-resource" pair by weightedly integrating the risk level of the patient task queue with the resource scoring factor. The system then sorts patients by this score, prioritizing the dialysis unit with the highest score for treatment resource allocation, thereby forming a resource matching plan.

[0111] The execution record submodule is used to record the entire execution process of the resource matching plan, including the adjusted treatment start time, whether any abnormal events occurred during the intervention, and the patient's actual treatment results. Based on this data, the system can generate actual intervention feedback results and provide input data for the feedback learning module.

[0112] Through the implementation of this module, the system not only improved resource utilization but also optimized scheduling strategies based on patient risk. For example, during a peak treatment period, the system dynamically rescheduled the original plan based on the time requirements and resource availability of multiple high-risk patients, prioritizing three high-risk patients to dialysis units with emergency care capabilities, significantly reducing clinical processing delays for the emergency.

[0113] The task management submodule is used to identify high-risk patients and their urgency of treatment based on early warning information, and to generate a queue of tasks to be scheduled, including:

[0114] This submodule receives structured warning information from the warning generation module, including key parameters such as the patient identification code, current risk level, and recommended treatment time window. The system associates the patient's identification code with the visit number, using a unique index to link the patient to the corresponding dialysis task. Warning information is automatically categorized and sorted by risk level (e.g., high risk, medium risk, or requiring attention), and a task priority field is established.

[0115] The system then assigns different time weights to patient tasks based on the processing time requirement field, such as: processing within 15 minutes, scheduling within 1 hour, processing within the same day, etc. It also generates an absolute timestamp based on the current clock and calculates the remaining time for each task. This remaining time is used as one of the important factors in task scheduling priority.

[0116] The task management submodule supports real-time task queue updates, dynamically adjusting the queue's ranking and status label upon patient risk status updates, manual intervention completion, or system reassessment. Task statuses typically include: assigned, pending, manually suspended, and processed. The task list is then pushed to the scheduling module for subsequent resource matching and scheduling.

[0117] For example, when the system was deployed in a county-level hospital, a patient identified as "high-risk and requiring urgent treatment" was assigned the labels "scheduling priority 1" and "remaining processing time 10 minutes" in the system. The system automatically placed the patient at the top of the task queue, replacing the originally planned patient scheduling order, significantly improving emergency treatment efficiency.

[0118] The execution record submodule is used to record the actual execution process of the scheduling task and generate actual intervention feedback results, including:

[0119] After the shift adjustment is executed, this submodule automatically records the actual start time of the dialysis task and compares it with the original scheduled time to generate the scheduling delay time field. Simultaneously, this module communicates with the dialysis workstation to collect records of key events that occur during treatment, such as dialysis interruptions, hypotensive episodes, and emergency treatments, and categorizes them as "abnormal treatment events."

[0120] Furthermore, the module supports the tagging and review of interventions. After dialysis, the system reads data from physiological monitoring terminals, such as blood pressure recovery curves, net weight loss, and dialysis completion indicators, to determine the effectiveness of the intervention. If physiological indicators stabilize after the intervention, the system automatically records the task as "intervention successful"; if they remain abnormal, it records them as "intervention failed" or "effect unclear."

[0121] All recorded information is stored in the system log database and simultaneously pushed to the feedback learning module as input for the next round of active learning. The system also supports the automatic generation of intervention history archives by patient dimension for reference in subsequent treatment decisions.

[0122] In an actual operation, this sub-module recorded that a high-risk patient received early intervention due to the forward scheduling, and the residual value of his intraoperative weight returned to a safe range. The system automatically marked "intervention effective" and used this case for subsequent factor contribution analysis, effectively supporting the iterative optimization of the model.

[0123] In a preferred embodiment of the present invention, the feedback learning module includes:

[0124] The result analysis submodule is used to analyze the improvement degree of post-dialysis physiological monitoring indicators based on the actual intervention feedback results and generate an improvement indicator data set;

[0125] The factor contribution calculation submodule is used to match the improvement index dataset with each characteristic factor in the characteristic dataset, and calculate the contribution of each characteristic factor to the improvement index to generate a characteristic contribution dataset;

[0126] The weight generation submodule is used to assign corresponding weight values ​​to each feature factor according to the feature contribution dataset, generate a feature weighted dataset, and output the feature weighted dataset to the risk identification module for the next round of model training.

[0127] In an embodiment of the present invention, the feedback learning module includes a result analysis submodule, a factor contribution calculation submodule, and a weight generation submodule, which constitute the core closed loop of the system's self-learning and self-adjustment. The result analysis submodule can receive the actual intervention feedback results returned by the scheduling module and analyze the changes in multiple physiological monitoring indicators of patients after dialysis, such as weight change rate, systolic blood pressure stability, and heart rate range fluctuation range. The system calculates the improvement range by comparing the differences in key indicator values ​​before and after treatment, and then generates an improvement indicator data set.

[0128] The Factor Contribution Calculation submodule maps the improvement indicator dataset to the previously constructed feature dataset, analyzing each feature factor's ability to consistently predict specific risks. Using least squares regression, the system fits the correlation between each feature factor and the improvement indicator, deriving a contribution value for each factor. If a feature consistently demonstrates a high degree of fit in the majority of successful intervention cases, its contribution value will increase significantly.

[0129] The weight generation submodule assigns new weights to each feature factor based on the feature contribution dataset. The weights are adjusted based on the difference between the previous round's weight and the current contribution. After adjustment, they are normalized to maintain a stable total weight of 1, preventing any one feature from overly interfering with the model. Finally, the system feeds the weighted feature dataset back to the risk identification module for the next round of model training.

[0130] This feedback mechanism effectively improves the model's generalization across different patient populations. For example, after one optimization cycle, the system identified a significant increase in the contribution of "weight residual" in predicting short-term hypotension and automatically increased its weight, resulting in a 0.07 increase in the AUC of subsequent models on similar cases.

[0131] The result analysis submodule is used to analyze the post-dialysis physiological monitoring indicators reflected in the actual intervention feedback results to form an improvement indicator data set, which specifically includes:

[0132] This submodule primarily processes data from two time points: before and after the intervention. By interfacing with the execution record submodule, the system extracts pre-dialysis risk characteristic indicator values ​​and corresponding post-dialysis physiological monitoring results, such as mean heart rate, systolic blood pressure change, net weight loss, and mean blood flow rate. While ensuring consistent patient identification, the system calculates the difference between the "before" and "after" values ​​of the same indicator or analyzes the change trend to determine whether the intervention has led to an improvement in the indicator.

[0133] The degree of improvement can be described in a variety of ways. For example, for weight residuals, the fluctuation range of the residuals for three consecutive dialysis sessions before and after the intervention can be compared; for blood pressure fluctuations, the improvement in stability can be reflected by the change in standard deviation. If the value of a particular indicator moves toward the pre-set safety interval and shows a trend of convergence, the indicator is judged to have "effectively improved."

[0134] This submodule can also jointly assess multiple indicators, generating a set of improvement indicator labels and grouping them into an "Improvement Indicator Dataset." Each data item in the dataset is labeled with its source task number, intervention time point, improvement direction, and improvement range, ensuring that the data is traceable, calculable, and verifiable.

[0135] The improvement index dataset generated by this submodule will be input into the factor contribution calculation submodule and the weight generation submodule to promote model updates and feature weight adjustments, and build an active learning system with a real clinical feedback loop.

[0136] The weight generation submodule is used to generate a feature-weighted dataset for model iteration based on the actual impact of the feature factors on the improvement indicators. Specifically, it includes:

[0137] This submodule receives the feature contribution dataset from the factor contribution calculation submodule, where each record represents the explanatory power of a feature factor in improving the indicator across multiple intervention samples. The system uses this contribution value as an initial adjustment reference and compares it with the currently used feature factor weights.

[0138] The system then updates the weight based on the difference between the contribution and the current weight. If a factor's contribution is higher than its current weight, the system increases its weight by a certain amount; if its contribution is lower, it decreases it. To prevent certain factors from being eliminated due to low contribution, the system sets a minimum weight threshold to ensure that all factors maintain a certain level of participation.

[0139] All adjusted weights are normalized so that the sum of the weights of all feature factors equals 1, maintaining data stability and consistency with model calculations. After normalization, the system multiplies the feature vector of each sample in the feature dataset by the updated weight to generate a new feature-weighted dataset.

[0140] The weighted feature dataset will serve as input for the next round of risk assessment models, enabling them to prioritize features that have a real impact on improvement. For example, during one round of model optimization, the system detected that "blood pressure fluctuation trends" significantly contributed to improvements in weight residuals. The weight was increased from 0.15 to 0.27, significantly improving the model's ability to discriminate among high-risk groups.

[0141] In a preferred embodiment of the present invention, the sample construction submodule includes:

[0142] A labeling and screening unit is used to extract historical dialysis records with labeled high-risk events from a standard dialysis information dataset to form a labeled sample dataset;

[0143] A feature label fusion unit is used to pair each set of feature vectors in the feature data set with the corresponding label in the labeled sample data set to construct a training sample group, wherein the training sample group includes feature-label combinations of positive samples and negative samples;

[0144] The sample integration unit is used to divide the training sample group into training set and test set, and record the sample source, label type and construction timestamp.

[0145] In an embodiment of the present invention, the sample construction submodule consists of a labeling and screening unit, a feature label fusion unit, and a sample integration unit, which can efficiently complete the labeling and structuring preparation of dialysis sample data. The labeling and screening unit automatically extracts historical dialysis records that have been labeled as "high-risk events" by doctors from the standard dialysis information dataset. The extracted events include types such as hypotension during dialysis, emergency interruptions, and drastic weight fluctuations, thereby forming a labeled sample dataset. The system identifies these event data through label matching rules to ensure the accuracy of the samples.

[0146] The feature label fusion unit maps the labeled samples to the feature vectors in the feature dataset, constructing a training sample set. Each piece of data in the training sample consists of a feature vector and a label pair. The labels include binary labels for positive samples (indicating a high-risk event) and negative samples (indicating a normal dialysis process). The system supports setting strategies for different label ratios to ensure a balanced classification of the training sample set.

[0147] The sample integration unit divides the generated training sample group into training set and test set. The division ratio can be set to 8:2 or 7:3, and records metadata such as sample source information, label type and construction time to facilitate tracking and verification operations during model training.

[0148] Through this sample construction mechanism, the system can efficiently generate high-quality training datasets that meet clinical labeling requirements without manual intervention. For example, during one model update, the system completed the screening and matching of 5,000 sets of samples within one hour, achieving an accuracy rate of over 98%. This significantly reduced the cost of manual sample preparation and significantly accelerated the risk assessment model construction cycle.

[0149] In a preferred embodiment of the present invention, the model training submodule includes:

[0150] A structure initialization unit is used to automatically generate multiple risk judgment model structure templates based on the characteristic dimensions and label distribution characteristics of the training sample group;

[0151] The error optimization unit is used to input the training sample group into each model structure template, calculate the prediction error of each iteration, and adjust the parameters according to the error feedback results until the training error is lower than the preset error threshold;

[0152] The performance evaluation unit is used to receive multiple risk judgment models that have completed training and calculate their precision, recall, F1-score and AUC indicators on the test set.

[0153] In this embodiment of the present invention, the model training submodule includes a structure initialization unit, an error optimization unit, and a performance evaluation unit, forming a multi-structure training framework for dialysis risk discrimination. The structure initialization unit automatically sets a variety of structure templates, including shallow networks, decision tree models, and integrated logic models, based on the number of feature dimensions, label distribution, and fluctuation range of historical samples in the training sample group. Each structure template is assigned an independent parameter space during initialization.

[0154] The error optimization unit sequentially feeds training samples into each structural template for iterative training. After each iteration, the prediction error for that structure is calculated. Using an error feedback mechanism, the unit dynamically adjusts internal weights and biases until the error falls below a preset threshold or the maximum number of training rounds is reached. This process ensures that each model type is fitted using its optimal parameters, avoiding model bias caused by fixed structures.

[0155] The performance evaluation unit calculates multiple key performance indicators for each trained model structure on the test set, including prediction accuracy, recall rate, F1-score, and AUC. The introduction of these indicators enables the system to evaluate the model's stability and generalization ability in risk event identification from multiple dimensions, providing a sufficient basis for the model release phase.

[0156] Through the implementation of this module, in a comparative experiment, the system tested four types of structural templates under the same training data, and finally selected a model with an AUC value of 0.945 and a recall rate of 0.91 as the current callable version. The overall discrimination performance is about 15% better than the traditional static logic model, effectively improving the accuracy of the system in the abnormal dialysis prediction task.

[0157] The structure initialization unit is used to automatically generate multiple risk judgment model structure templates based on the feature dimensions and label distribution characteristics of the training sample group, specifically including:

[0158] This unit first analyzes the number of feature dimensions and the distribution of sample labels in the input training sample set. For example, if the number of features is 25, the label ratio is 1:1.8 for positive samples and 1:1 for negative samples. Based on this information, the system automatically determines the appropriate model structure type to use, such as a shallow feedforward neural network, logistic regression, or ensemble tree model.

[0159] For each structural template, the system sets a set of initial parameter combinations, including the number of hidden layer nodes, activation function type, regularization factor, etc. For example, for a 25-dimensional input, the system can configure a neural network structure with two hidden layers, with the first layer set to 64 nodes and the second layer to 32 nodes, and the ReLU function can be used as the activation function.

[0160] In addition, to enhance the diversity and robustness of model structures, the unit also supports structure combination generation, which involves building multiple model structures and training them in parallel to form a candidate model pool. After initialization, each structure template is saved in the model repository for subsequent error optimization and performance screening.

[0161] In actual deployment, such as a model optimization task in a hemodialysis center, this unit can automatically construct five types of model templates, covering two categories: shallow neural networks and integrated decision models, providing a basis for the system's subsequent balance between accuracy and efficiency.

[0162] The error optimization unit is used to train the model structure output by the structure initialization unit and adjust the parameters according to the error feedback results, including:

[0163] This unit receives a set of training samples and multiple initialized model structures, training them one by one. During each round of training, the model's predictions are compared with the actual labels to generate a prediction error, such as the number of cases predicted as high risk but actually low risk. This error value serves as feedback for parameter correction.

[0164] The parameter adjustment process is accomplished using a gradient descent-like method. This involves the system calculating the direction and magnitude of adjustments to the model's internal parameters (such as weights and biases) based on the error value, and then updating the model structure in the next round of training. This adjustment process is repeated until the training error falls below a preset threshold or the maximum number of training rounds is reached.

[0165] To prevent the model from falling into local optimality, the unit also supports the use of strategies such as dynamic learning rate and early stopping mechanism. For example, if the error stops decreasing in several consecutive rounds of training, the system can terminate training early to save computing resources.

[0166] Furthermore, the unit can record the error changes after each round of training as an error convergence curve for subsequent evaluation of model training efficiency. This mechanism ensures that each model achieves optimal performance within its parameter space, providing an accurate basis for performance scoring and screening.

[0167] In a preferred embodiment of the present invention, the calculation formula for the performance score is:

[0168]

[0169] Among them, Score is the comprehensive performance score of the model, and its value range is [0,1];

[0170] is the accuracy index;

[0171] is the recall rate indicator;

[0172] is the F1-score indicator;

[0173] TP is the number of samples that are high risk and are actually high risk, that is, the number of true positive cases;

[0174] FP is the number of samples that are high risk but actually low risk, that is, the number of false positives;

[0175] FN is the number of samples with low risk but actually high risk, that is, the number of false negatives;

[0176] A is the area under the curve of the model, that is, the AUC indicator;

[0177] w1, w2, w3, and w4 are weighted coefficients of the four indicators, and the weights can be set according to the scenario.

[0178] In an embodiment of the present invention, the system sets a performance scoring mechanism to comprehensively evaluate the actual performance of multiple candidate risk judgment models, so as to screen the optimal model and store it as a callable version.

[0179] Specifically, after training multiple risk assessment models with different structures, the system will conduct a standardized evaluation of each model. This evaluation not only considers the model's prediction accuracy on the test data, but also focuses on its recall ability for high-risk events, precision performance, overall balance, and the model's ability to distinguish borderline samples. In this embodiment, the performance dimensions of the system evaluation include but are not limited to the following four items:

[0180] First, the accuracy of the model in identifying high-risk events, that is, the proportion of events predicted as high-risk that are actually high-risk (corresponding to the accuracy indicator);

[0181] Secondly, the model's coverage of all true high-risk patients, that is, the proportion of all actual high-risk samples that are correctly identified (corresponding to recall rate);

[0182] Thirdly, the balance between precision and recall, that is, whether the model can avoid excessive false positives while improving recognition capabilities (corresponding to the comprehensive evaluation index);

[0183] Finally, the discrimination ability of the model over all predicted probability intervals is expressed by plotting the receiver operating curve and calculating the area under it.

[0184] The system weights these multiple evaluation results according to pre-set weight coefficients to generate a unified performance score. A higher score indicates a better balance between precision, recall, and generalization, making the model more suitable for deployment in real-world clinical early warning scenarios.

[0185] Through this scoring mechanism, the present invention can effectively distinguish the performance differences between different structural models on real high-risk samples. For example, in a comparison, when two models have similar overall accuracy, one model, due to its higher recall and better area under the curve, has a higher final performance score and is selected as the target model by the system. This screening mechanism overcomes the shortcomings of traditional model selection based on a single metric (such as accuracy), significantly improving actual deployment results.

[0186] Performance scoring is used to comprehensively evaluate the effectiveness of multiple risk assessment models, involving multiple model metrics such as precision, recall, F1-score, and AUC. When weighting multiple metrics, setting appropriate weighting coefficients is key to ensuring the objectivity and clinical applicability of the scoring.

[0187] In the implementation of the present invention, the weight coefficients in the performance score are used to adjust the degree of influence of each model performance indicator on the final score result. The setting strategies include:

[0188] Experience-based configuration: Based on actual clinical needs, the model's risk identification ability (i.e., recall rate) and overall balance (i.e., F1-score) are prioritized. Therefore, the values ​​of w2 and w3 are usually set to be greater than w1 and w4. For example: w1 = 0.2, w2 = 0.35, w3 = 0.35, w4 = 0.1.

[0189] Data-driven parameter tuning: By using the model scoring results of historical samples, we fine-tune the weights in actual deployment and select the weight combination that best distinguishes excellent models.

[0190] Weight normalization principle: The sum of all weight coefficients is kept constant to 1 to maintain the linear interpretability of the performance score value.

[0191] This strategy enables the system to balance model accuracy with sensitivity in identifying high-risk patients, enhancing the clinical guidance of the model screening mechanism. Observations from actual deployments have shown that different weight combinations are sensitive to the impact of score ranking, and after optimization, model screening stability can be improved by approximately 8%.

[0192] In a preferred embodiment of the present invention, the resource matching submodule includes:

[0193] The resource status quantification unit is used to obtain the idle time period, current load rate, equipment status fluctuation and historical dialysis success rate of each dialysis unit, and assign corresponding resource scoring factor values ​​to form a resource status matrix;

[0194] The priority score calculation unit is used to assign the patient's risk level and processing time requirement as risk score factors and time limit score factors respectively according to the queue of tasks to be scheduled, and perform weighted fusion with the resource status matrix to calculate the matching priority score value between each resource and each task;

[0195]

[0196] Among them, P mn is the matching priority score between the mth patient and the nth dialysis unit;

[0197] R m is the risk scoring factor, i.e., the risk level of the mth patient, such as 0 to 3;

[0198] is the time limit scoring factor, t m is the remaining allowed waiting time for the mth patient, t max To preset the maximum waiting time, a larger value of the time limit scoring factor indicates a higher urgency;

[0199] is the current load rate of the nth dialysis unit, i.e., current usage time / total operating time;

[0200] S n is the historical dialysis success rate of the nth dialysis unit;

[0201] is the fluctuation of the equipment status of the nth dialysis unit, that is, the standard deviation / mean of the failure rate in the past 7 days;

[0202] α, β, γ, θ, and δ are weight coefficients;

[0203] The matching screening unit is used to sort resources according to the priority score value and screen the resources with free time within the expected treatment time period in each dialysis unit. For each high-priority patient, the dialysis unit with the highest matching priority score and the available time that does not conflict with the treatment time requirements is selected as the recommended scheduling result; if there is a conflict of the same score or time overlap, a secondary screening is performed based on the historical scheduling success rate of the patient and the resource, and finally a resource matching plan is formed.

[0204] In this embodiment of the present invention, the resource matching submodule includes a resource status quantification unit, a priority scoring calculation unit, and a matching screening unit, forming a refined allocation mechanism for dialysis resource scheduling tasks. The resource status quantification unit collects multidimensional status information for each dialysis unit, including idle time periods, emergency equipment availability, equipment abnormality rate, and dialysis success rate over the past week. It then generates corresponding resource scoring factors, which are further organized into a resource status matrix.

[0205] The priority score calculation unit receives the risk level and processing time requirements of patients in the waiting-for-scheduling task queue and maps them into risk scoring factors and time limit scoring factors, respectively. During the scoring calculation, a matching priority score is calculated for each patient-dialysis unit pair.

[0206] The matching screening unit sorts the patients according to the above-mentioned scoring values ​​and preferentially matches the dialysis unit with the highest score to the patients with higher priority. If a resource scoring conflict occurs, the historical resource scheduling success rate is further compared to make a decision.

[0207] This module enhances the system's intelligent scheduling capabilities for dialysis scheduling, particularly during periods of resource constraints. In a peak scheduling simulation, the system successfully allocated seven high-risk patients to four dialysis units with different emergency equipment based on a priority scoring mechanism, reducing average processing latency by 32%, demonstrating the robustness and effectiveness of scheduling results.

[0208] In the formula for matching priority scores, this formula supports the priority allocation of dialysis units with "low load, high success rate, and low failure rate" to "high-risk + high-urgency" patients. At the same time, the device status fluctuation factor V is introduced to improve the overall reliability of the system. This formula is particularly suitable for scheduling optimization in resource-constrained scenarios.

[0209] Specifically, the system extracts each patient's risk level and remaining waiting time based on their pre-scheduled information. This is then combined with the patient's dialysis resource availability, equipment performance, emergency response capabilities, and historical dialysis success rates to conduct a comprehensive assessment. Each factor is quantified into a scoring factor, which the system then integrates into an overall priority score based on configured weights.

[0210] Patients with higher risk levels and shorter waiting times receive higher scores. Furthermore, patients with greater resource availability, better device status fluctuations, and higher success rates are more likely to be matched first. The system sorts all scores from high to low, prioritizing the allocation of limited resources based on risk.

[0211] Matching priority scores is not only used to prioritize patients, but also to resolve resource conflicts. For example, if two high-priority patients compete for the same dialysis bed, the system compares their scores. If they have the same score, the system further considers the resource's historical treatment success rate for different patients to assist in decision-making.

[0212] This scoring mechanism effectively avoids unfair scheduling issues caused by "first-come, first-served" or "manual experience," particularly during peak hours when patient volume exceeds available resources. This optimizes both clinical safety and operational efficiency. In one hospital deployment, the system helped identify three high-risk elderly patients in advance and dynamically adjusted their scheduling, enabling early intervention and effectively preventing intraoperative hypotension events.

[0213] In summary, this scoring mechanism achieves automatic optimization of complex multi-objective scheduling problems in a quantitative, multi-factor, and traceable manner, which is a significant technological advancement compared with the traditional static scheduling mechanism.

[0214] Among them, the historical scheduling success rate is used to reflect whether a patient and a dialysis unit have good scheduling execution stability in historical scheduling tasks. It is an important auxiliary indicator for the system to perform differentiated screening when the priority score values ​​are the same or there are resource conflicts.

[0215] The historical scheduling success rate is a continuous value between 0 and 1. It is calculated by recording whether the patient's dialysis treatment was completed as planned after being assigned to the dialysis unit within a set evaluation period (e.g., the last 30 days, the last 10 dialysis treatments). If the patient starts and completes treatment as planned after being assigned to a dialysis unit, it is counted as a success. If there is a resource change, a scheduling delay timeout, or a forced unit change due to equipment failure or personnel conflict, it is counted as a failure.

[0216] Ultimately, the historical scheduling success rate is equal to the number of successful scheduling divided by the total number of scheduling attempts, that is: the number of successful scheduling executions / the total number of scheduling attempts.

[0217] This metric not only considers resource availability but also reflects the resource's suitability for scheduling a specific patient and its interoperability with system execution. For example, if Patient A has a 90% success rate with dialysis unit X over the past 10 scheduling attempts, but only a 60% success rate with unit Y, then if the priority scores are the same, the system will prioritize unit X to reduce the risk of delays caused by resource switching.

[0218] It should be noted that the success rate is a historical correlation indicator calculated separately for each "patient-resource" combination, and different patients may have different success rate results for the same dialysis unit.

[0219] The matching priority score comprehensively considers patient risk, waiting time, and the current status of dialysis resources, and is the core indicator for multi-objective scheduling. The weight coefficient in the weighting process reflects the relative importance of different factors.

[0220] The weight coefficients in the matching priority score are set according to the following:

[0221] Patient factor weighting priority: The system assigns higher weights to risk level scoring factors and processing time limit scoring factors to reflect the "patient-centric" scheduling principle. For example: α = 0.3, β = 0.25.

[0222] Resource status weight balance: Assign medium weights to factors such as resource load rate, success rate, and device volatility, usually setting each between 0.1-0.15.

[0223] Adaptive policy adjustment: allows hospital administrators to dynamically adjust weight parameters based on the current operating stage (such as holidays and emergencies), such as increasing the weight of equipment stability during equipment maintenance.

[0224] Normalization: The sum of all weights is 1, which facilitates a uniform interpretation of matching score fluctuations in the range of 0-1.

[0225] By reasonably setting the weight coefficient, the system can give priority to high-risk + short-term patients during peak periods, and give priority to matching dialysis units with strong stability when resources are tight, thereby achieving dual optimization of risk control and resource utilization.

[0226] In a preferred embodiment of the present invention, the factor contribution calculation submodule includes:

[0227] A data matching unit is used to receive the improvement index data set and the feature data set, and to match each improvement index with its corresponding feature factor one by one to form a matching data set;

[0228] The contribution calculation unit is used to calculate the fit between each characteristic factor and the corresponding improvement index through the least squares regression method based on the matching data set, and use the fit as the characteristic contribution value;

[0229] The contribution integration unit is used to merge the feature contribution values ​​of each feature factor into a data set to generate a feature contribution data set.

[0230] In this embodiment of the present invention, the factor contribution calculation submodule includes a data matching unit, a contribution calculation unit, and a contribution integration unit, establishing a feature evaluation mechanism for model optimization feedback. The data matching unit receives the improvement indicator dataset and the feature dataset. By matching the ID with the timestamp, it matches each physiological indicator improvement result with the factors in its preceding feature vector to form a matching dataset, ensuring the accuracy of data association.

[0231] The contribution calculation unit uses the least squares regression method to fit each characteristic factor to its corresponding improvement indicator. The coefficient of determination of the fitting result is used as the contribution value of the characteristic. The system monitors the fitting residual and average fluctuation of the factor across all matching samples. If a factor has a stable predictive effect across multiple samples, it will be assigned a higher contribution.

[0232] The contribution integration unit is responsible for integrating the contribution results of all feature factors to form a feature contribution dataset with a unified structure for use in subsequent weight generation operations.

[0233] This mechanism allows the system to scientifically quantify each characteristic factor after intervention and identify which input parameters have a positive impact on actual treatment improvement. For example, after three rounds of model training, the system identified that the "blood pressure fluctuation trend" significantly enhanced the goodness of fit for improving weight residuals. In subsequent optimization, the contribution of this factor was increased from 0.23 to 0.41, significantly strengthening the clinical relevance of the model's predictions.

[0234] The data matching unit is used to match the improvement index dataset with each characteristic factor in the characteristic dataset to construct a matching dataset that can be used for subsequent fitting analysis, specifically including:

[0235] The unit first receives the improvement indicator dataset generated by the outcome analysis submodule. This dataset contains the magnitude of improvement in multiple physiological monitoring indicators after dialysis, such as decreased blood pressure variability and convergence of weight residuals. The system then extracts the characteristic factor values ​​corresponding to each improvement outcome from the original feature dataset, ensuring that both are records of the same patient and the same intervention task.

[0236] To achieve effective matching, the system aligns each set of data based on task number, patient ID, and timestamp. After a successful match, the system generates several matching data pairs, each consisting of a "feature factor value - improvement indicator value."

[0237] This matching dataset provides a direct data foundation for subsequent analysis of the impact of each factor on improved results. It also serves as an important prerequisite for factor contribution analysis and weight updates in the system. The matching process strictly controls field consistency and sample synchronization to avoid analytical bias caused by data drift or misalignment.

[0238] The contribution calculation unit is used to evaluate the influence of characteristic factors on the improvement index and generate characteristic contribution values, including:

[0239] After receiving the matching dataset, the unit first performs a relationship analysis on each "characteristic factor value - improvement indicator value" data pair. Without explicitly expressing the relationship using regression formulas, the system builds a statistical correlation model to measure the consistency between fluctuations in the characteristic factor and changes in the improvement indicator.

[0240] During the analysis, the system uses factor values ​​as input and improvement results as output, simulates the fitting trend between the two, and calculates the fitting error. If the value of a characteristic factor shows a stable, positive, or negative correspondence with the improvement indicator in multiple intervention groups, that is, the change in the characteristic factor can effectively "explain" the occurrence of improvement, the system will assign a higher contribution value to the characteristic factor.

[0241] The system also considers the impact of sample size on the confidence level of contribution. If a feature has a small number of matching samples, its contribution will be appropriately adjusted to prevent overfitting. In addition, all contribution values ​​will be normalized within a unified scale to ensure comparability in subsequent weight generation operations.

[0242] In an actual case, "weight residual" showed a high correlation with the systolic blood pressure improvement index in 30 improvement task samples, and its corresponding fitting error continued to be lower than other factors. Based on this, the system calculated that its contribution value was significantly higher than similar features, providing an effective basis for subsequent weight adjustments.

[0243] In a preferred embodiment of the present invention, the weight generation submodule includes:

[0244] The weight update unit is used to assign an initial weight value to each feature factor based on the feature contribution data set, and update the current round weight based on the difference between its feature contribution value and the weight value of the previous round;

[0245] A normalization processing unit is used to normalize the current round weight values ​​of all feature factors to ensure that the sum of all weight values ​​is 1;

[0246] The weighted data generating unit is used to perform a numerical product operation on each set of feature vectors in the feature data set and the corresponding normalized weight to generate a feature weighted data set.

[0247] In this embodiment of the present invention, the weight generation submodule, consisting of a weight update unit, a normalization processing unit, and a weighted data generation unit, implements dynamic weighting control of feature factors during model training. The weight update unit assigns initial weights to each feature factor based on the aforementioned feature contribution dataset and updates the feature weight values ​​for the current round based on changes in contribution after each round of training.

[0248] The normalization processing unit normalizes all updated weight values ​​to ensure that the sum of the weights of all feature factors is equal to 1, preventing the weight of a certain factor from being abnormally amplified or ignored, thereby maintaining model stability.

[0249] The weighted data generation unit performs numerical product calculations on the normalized weight values ​​and the original feature vectors one by one to generate a feature weighted data set, which is output to the risk identification module as input for a new round of model training.

[0250] This mechanism enables the system to self-adjust with each round of feedback, strengthening high-value factors and weakening ineffective, redundant factors. In actual deployments, this module has accelerated model convergence by an average of 26% and maintained a stable error fluctuation range, effectively improving the efficiency and stability of model training.

[0251] The weight updating unit is used to assign a weight value corresponding to the current round to each feature factor based on the feature contribution dataset, and adjust the difference with the weight of the previous round to generate an updated feature factor weight set, which specifically includes:

[0252] This unit receives the feature contribution dataset output by the Contribution Calculation submodule. This dataset records the contribution of each feature factor in the current training round. The value range is usually between zero and one, indicating the closeness of the correlation between the feature factor and the improvement indicator. The system extracts the weight value of the feature factor in the previous round from the historical records and compares it with the current contribution value.

[0253] To reflect the system's responsiveness to new, high-performing features, the system uses an adjustment rate parameter as a learning intensity control factor. When a feature's contribution value significantly exceeds its original weight, the system adjusts its weight closer to its contribution by a specified amount, generating a new weight for the current round. Conversely, when the contribution value falls below its historical weight, the system appropriately adjusts the feature's weight downward to prevent the model from focusing on ineffective factors.

[0254] This update strategy prevents the system from over-reliance on historical structures while integrating current clinical feedback to achieve dynamic weight iteration. To maintain numerical stability, the system also sets minimum and maximum weight boundaries to prevent individual feature factors from being mistakenly judged as invalid due to low weights, or from causing model oscillation due to excessive updates.

[0255] For example, in a certain dialysis data round, the "weight residual" factor performed well in the improvement prediction of 30 samples, and its contribution was significantly improved. Based on this, the system adjusted its original weight from 0.12 to 0.21, improving the model's ability to identify capacity control risks.

[0256] The normalization processing unit is used to perform numerical normalization on the current round weight values ​​of all feature factors to ensure that the sum of all weights is one and to maintain the relative stability of the model input proportional relationship. Specifically, it includes:

[0257] After the weight update unit completes the adjustment of the individual weight values ​​of all feature factors, this unit performs a traversal scan of all updated weight values ​​and calculates their sum. The system then divides the weight value of each feature factor by this sum to obtain the new "normalized weight value."

[0258] This normalization process maintains the sum of the weights of all feature factors at unity, thus avoiding training anomalies caused by imbalanced weight ratios in subsequent models. This normalization process does not change the relative weight relationships between feature factors; it only performs numerical mapping.

[0259] In addition, to avoid calculation errors caused by division by zero or too small a value, the system checks whether the total weight is greater than the minimum tolerance value before performing the normalization operation. If it is lower than the threshold, the default weight reset logic is triggered, which redistributes all weights into an average value to ensure stable operation of the model.

[0260] In an actual deployment scenario, the system processed feature data from multiple high-risk patients. After completing the weight update, it normalized the weights originally distributed in the range of 0.05 to 0.30 and uniformly mapped them to a new standardized interval to ensure the consistency and numerical controllability of the subsequent feature weighted calculation process.

[0261] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A full-process information management system for hemodialysis rooms based on active learning, characterized in that: The system comprises: The feature construction module is used to perform structured processing on the original dialysis information data to generate a standard dialysis information dataset, and extract multiple feature factors related to the patient's current physical condition from it, including heart rate variability, blood pressure fluctuation trend, and weight residual, to generate a feature dataset; The risk identification module is used to compare the feature dataset with the standard dialysis information dataset, construct a training sample group, and introduce labels for supervised training to generate a risk judgment model for identifying specific high-risk states; An early warning generation module is used to input the feature data set into the risk judgment model to generate a risk probability result. When the result reaches a preset risk threshold, an early warning message is generated that is associated with the current patient identifier. The early warning message includes the patient identifier, risk level, and processing time requirement. The scheduling module is used to dynamically adjust the treatment schedule of the current patient based on the early warning information, and record the actual intervention effect after the adjustment of the treatment schedule to form the actual intervention feedback results; The feedback learning module is used to calculate the correlation between the improvement degree of post-dialysis physiological monitoring indicators reflected in the actual intervention feedback results and the characteristic factors to determine the contribution of each characteristic factor, assign a corresponding weight value to each characteristic factor, generate a characteristic weighted data set, and feed it back to the risk identification module to execute the next round of training.

2. The active learning-based hemodialysis room full-process information management system according to claim 1, characterized in that: The feature construction module includes: The feature preprocessing submodule is used to standardize the format and unit of the numerical fields in the original dialysis information data, and to fill in missing values. It also analyzes and identifies the medical record fields contained therein, and automatically extracts information with physician risk event markers, intraoperative intervention records, or complication codes as high-risk event annotation fields to generate a standard dialysis information dataset. A feature extraction submodule is used to extract multiple feature factors related to the patient's current physical state from the standard dialysis information data set, wherein the feature factors include heart rate change rate, blood pressure fluctuation trend and weight residual; The feature generation submodule is used to convert the feature factor set into a feature vector representation of uniform dimension to generate a feature data set.

3. The active learning-based hemodialysis room full-process information management system according to claim 2, characterized in that: The risk identification module includes: The sample construction submodule is used to compare the feature dataset with the historical high-risk annotated data in the standard dialysis information dataset and generate a training sample group containing positive and negative samples and corresponding labels; The model training submodule is used to automatically generate multiple risk judgment model structure templates based on the training sample group, input the training sample group into each model structure template, perform iterative training and error optimization, and output the performance evaluation indicators of each model structure template; The model publishing submodule is used to calculate the performance score based on the performance evaluation indicators, and store the risk judgment model with the highest ranking and the performance score reaching the set score threshold as a callable version.

4. The active learning-based hemodialysis room full-process information management system according to claim 3, characterized in that: The shift scheduling module includes: The task management submodule is used to extract the patient identification, risk level and processing time requirements of the current patient based on the early warning information, and generate a queue of tasks to be scheduled; The resource matching submodule is used to obtain the current idle status, emergency capacity status, expected idle time and historical dialysis success rate of each dialysis unit and quantify them into resource scoring factors; The resource matching submodule is further configured to calculate a matching priority score based on the risk level and processing time requirements of each patient in the waiting task queue, combined with a resource scoring factor, and sort the patients from high to low according to the matching priority score, match the optimal dialysis resources, and form a resource matching plan; The execution record submodule is used to perform scheduling adjustments according to the resource matching plan, record the scheduling execution time, intervention conditions and actual effects during the dialysis process, and generate actual intervention feedback results.

5. The active learning-based hemodialysis room full-process information management system according to claim 4, characterized in that: The feedback learning module includes: The result analysis submodule is used to analyze the improvement degree of post-dialysis physiological monitoring indicators based on the actual intervention feedback results and generate an improvement indicator data set; The factor contribution calculation submodule is used to match the improvement index dataset with each characteristic factor in the characteristic dataset, and calculate the contribution of each characteristic factor to the improvement index to generate a characteristic contribution dataset; The weight generation submodule is used to assign corresponding weight values ​​to each feature factor according to the feature contribution dataset, generate a feature weighted dataset, and output the feature weighted dataset to the risk identification module for the next round of model training.

6. The active learning-based hemodialysis room full-process information management system according to claim 3, characterized in that: The sample construction submodule includes: A labeling and screening unit is used to extract historical dialysis records with labeled high-risk events from a standard dialysis information dataset to form a labeled sample dataset; A feature label fusion unit is used to pair each set of feature vectors in the feature data set with the corresponding label in the labeled sample data set to construct a training sample group, wherein the training sample group includes feature-label combinations of positive samples and negative samples; The sample integration unit is used to divide the training sample group into training set and test set, and record the sample source, label type and construction timestamp.

7. The active learning-based hemodialysis room full-process information management system according to claim 6, characterized in that: The model training submodule includes: A structure initialization unit is used to automatically generate multiple risk judgment model structure templates based on the characteristic dimensions and label distribution characteristics of the training sample group; The error optimization unit is used to input the training sample group into each model structure template, calculate the prediction error of each iteration, and adjust the parameters according to the error feedback results until the training error is lower than the preset error threshold; The performance evaluation unit is used to receive multiple risk judgment models that have completed training and calculate their precision, recall, F1-score and AUC indicators on the test set.

8. The active learning-based hemodialysis room full-process information management system according to claim 4, characterized in that: The resource matching submodule includes: The resource status quantification unit is used to obtain the idle time period, current load rate, equipment status fluctuation and historical dialysis success rate of each dialysis unit, and assign corresponding resource scoring factor values ​​to form a resource status matrix; The priority score calculation unit is used to assign the patient's risk level and processing time requirement as risk score factors and time limit score factors respectively according to the queue of tasks to be scheduled, and perform weighted fusion with the resource status matrix to calculate the matching priority score value between each resource and each task; The matching screening unit is used to sort resources according to the priority score value and screen the resources with free time within the expected treatment time period in each dialysis unit. For each high-priority patient, the dialysis unit with the highest matching priority score and the available time that does not conflict with the treatment time requirements is selected as the recommended scheduling result; if there is a conflict of the same score or time overlap, a secondary screening is performed based on the historical scheduling success rate of the patient and the resource, and finally a resource matching plan is formed.

9. The active learning-based hemodialysis room full-process information management system according to claim 5, characterized in that: The factor contribution calculation submodule includes: A data matching unit is used to receive the improvement index data set and the feature data set, and to match each improvement index with its corresponding feature factor one by one to form a matching data set; The contribution calculation unit is used to calculate the fit between each characteristic factor and the corresponding improvement index through the least squares regression method based on the matching data set, and use the fit as the characteristic contribution value; The contribution integration unit is used to merge the feature contribution values ​​of each feature factor into a data set to generate a feature contribution data set.

10. The active learning-based hemodialysis room full-process information management system according to claim 9, characterized in that: The weight generation submodule includes: The weight update unit is used to assign an initial weight value to each feature factor based on the feature contribution data set, and update the current round weight based on the difference between its feature contribution value and the weight value of the previous round; A normalization processing unit is used to normalize the current round weight values ​​of all feature factors to ensure that the sum of all weight values ​​is 1; The weighted data generating unit is used to perform a numerical product operation on each set of feature vectors in the feature data set and the corresponding normalized weight to generate a feature weighted data set.

Images