Hemodialysis fatigue index construction and auxiliary decision generation method and system

Abstract

The application discloses a kind of fatigue index construction and auxiliary decision generation method and system for hemodialysis.The method divides dialysis cycle into four stages before, during, after and interval, retains original irregular sampling time stamp and adds stage, event, quality label;Cross-modal association and whole cycle dependence relationship in event window are established through event hypergraph and time patch graph;Whole cycle state representation is generated by shared-private latent space separation and state space aggregation;Based on historical effective cycle, individual prototype is constructed by exponential moving average, the offset measure of current cycle relative to prototype is calculated, fatigue offset index, recovery delay index, physiological tolerance index and whole cycle fatigue load index are generated;Combined with individualized dynamic threshold, confidence evaluation and priority sorting, auxiliary decision suggestion is output, and closed-loop feedback mechanism is provided.The application realizes individualized fatigue index construction and auxiliary decision support.

Description

Technical Field

[0001] This invention belongs to the field of medical information processing technology, specifically relating to a method and system for constructing and generating multimodal fatigue indicators and supporting decision-making throughout the entire hemodialysis cycle. Background Technology

[0002] Hemodialysis is a core treatment for end-stage renal disease patients to maintain their lives. Dialysis-related fatigue is one of the most common and far-reaching symptoms in these patients, severely reducing their quality of life and closely related to dialysis adequacy, cardiovascular event risk, and long-term survival. However, fatigue is highly subjective, multidimensional, and varies from person to person. How to objectively and accurately quantify dialysis-related fatigue and provide individualized decision support remains a pressing technical challenge.

[0003] Currently, existing technologies mainly suffer from the following four types of solutions and corresponding systemic shortcomings:

[0004] The first type is fatigue assessment schemes based on single-scale assessments. Their shortcomings are that the data source is singular, the time coverage is a single time point, and the output is a single score value, lacking the ability to conduct cross-time period and cross-modal analysis.

[0005] The second type is an evaluation scheme based on fixed thresholds and regular statistical features. Its shortcomings are that the threshold is based on group experience, lacks individualized baselines, cannot handle irregular sampling, and the output is a binary label.

[0006] The third category is based on regular time grids and conventional deep learning models. Its drawbacks are that forced resampling leads to the loss of irregular sampling information, limited time coverage, the use of a unified group model, and the output being a single risk label.

[0007] The fourth category is the closest existing technology, namely the evaluation scheme based on multimodal fusion and full-process deep learning. Its core defects include: information loss due to the use of rule-based time grid alignment; multimodal fusion achieved by simple feature splicing; use of a unified group model; output as a single-dimensional score; time coverage limited to a single dialysis process; and lack of individual prototype modeling, structured index generation, and closed-loop auxiliary decision-making mechanisms.

[0008] In summary, existing technologies have fundamental deficiencies in terms of time coverage, retention of irregular sampling information, multimodal fusion methods, event-driven representation, structured indicator system, individual baseline mechanism, and closed-loop decision support. Summary of the Invention

[0009] The purpose of this invention is to solve the problem that existing technologies cannot perform individualized, multi-dimensional fatigue assessment for hemodialysis patients, and to provide a method and system for constructing and generating multimodal fatigue indicators and assisting decision-making throughout the entire hemodialysis cycle.

[0010] To achieve the above objectives, the present invention adopts the following technical solution: a method for constructing fatigue indicators and generating auxiliary decisions for hemodialysis, comprising the following steps:

[0011] Step S1, Data Access: Acquire multimodal observation data from multiple data sources during a patient's complete dialysis cycle, and attach stage labels, event labels, and data quality labels to each observation data in a way that retains the original timestamps, forming a sequence of observation tuples with time order; wherein, the dialysis cycle is divided into the predialysis stage, the dialysis stage, the postdialysis stage, and the interdialysis stage;

[0012] Step S2, Event Relationship Modeling: Based on the observed tuple sequence, construct an event hypergraph and a time patch graph. Establish cross-modal associations within the event window through the event hypergraph, establish dependencies at different time scales throughout the entire cycle through the time patch graph, and update the feature representation of the observed tuple sequence through message passing.

[0013] Step S3, Fatigue State Characterization and Full-Cycle Aggregation: The feature representation of the updated observation tuple sequence is extracted by the shared-private latent space separation mechanism to extract the shared and private features of each mode. After gating fusion, the input is a state space model with stage switching gating for full-sequence aggregation to generate the stage-level representation and global representation of the current cycle.

[0014] Step S4, Individual Prototype Construction and Offset Quantization: Based on the stage-level and global representations of the patient's historical effective dialysis cycles, an individual prototype of the patient is constructed and updated using an exponential moving average method; the offset measure of the stage-level and global representations of the current cycle relative to the individual prototype is calculated, and the offset measure includes the overall offset measure and the stage offset measure of each stage.

[0015] Step S5: Generation of structured fatigue indices: Based on the offset metric, generate structured fatigue indices including fatigue offset index, recovery delay index, physiological tolerance index, and full-cycle fatigue load index.

[0016] Step S6, Assisted Decision Generation: Based on the structured fatigue index and combined with the individualized thresholds dynamically calculated for each patient, assisted decision suggestions with priority ranking and confidence scores are generated.

[0017] Further, step S1 includes:

[0018] Using the planned start time of the current dialysis cycle as the reference origin, the absolute timestamps of the observation data are converted into relative offset times;

[0019] Based on the different intervals in which the relative offset time is located, the observation data are classified into one of the predialysis stage, dialysis stage, postdialysis stage, and interdialysis stage, and a stage label is attached to each observation data.

[0020] Based on the clinical event information in the nursing records, attach corresponding event labels to the observation data located in the event window;

[0021] Calculate and attach data quality labels based on data integrity, sensor status, and reasonableness of value range.

[0022] Further, in step S2, constructing the event hypergraph includes:

[0023] Each observation data point is used as a node in the hypergraph;

[0024] Construct event hyperedges to group observation nodes from all data sources within the time influence window of the same clinical event into the same event hyperedge;

[0025] Construct a time hyperedge to group consecutive observation nodes with a time interval of less than a preset threshold into the same time hyperedge;

[0026] Construct variable hyperedges to group all nodes belonging to the same data source and the same observed variable within the same dialysis cycle into the same variable hyperedge;

[0027] Perform two-stage aggregated message passing on the event hypergraph, from node to hyperedge and from hyperedge to node, to update node characteristics.

[0028] Further, in step S2, the build time patch graph includes:

[0029] Based on the density of the event-sensitive mask, non-uniform patching is performed on the entire time axis, with smaller patch lengths used in densely populated event regions and larger patch lengths used in sparsely populated event regions.

[0030] The features of all observed nodes within each time patch are weighted and aggregated to form the features of the patch node, where the aggregation weight is determined by the data quality label and the event sensitivity mask.

[0031] Construct a time patch graph whose edge set includes intra-stage adjacency edges connecting adjacent time patches within the same stage, and cross-stage association edges connecting semantically related patches in different stages.

[0032] Perform message passing on the time-patch graph to update patch node characteristics.

[0033] Further, in step S3, the shared-private latent space separation mechanism includes: mapping node features to the shared latent space using shared projection parameters shared by all modalities to obtain a shared representation; mapping node features to the private latent space using private projection parameters independent of each modality to obtain a private representation; applying orthogonal constraints to make the shared representation and private representation of the same node tend to be orthogonal; applying contrast constraints to make the shared representations of nodes of different modalities within the same time window approach each other; and adaptively fusing the shared representation and private representation through a gating mechanism to obtain a fused representation.

[0034] The discretization step size of the state space model is dynamically generated based on the time interval between adjacent observation data. When the stage labels of adjacent observation data are different, a stage switching gating is triggered. The stage switching gating is used to adaptively balance between continuing the state information of the previous stage and resetting to the initial state of the current stage.

[0035] Furthermore, in step S4, the historical effective dialysis cycles need to meet all of the following screening rules: data completeness is higher than the first threshold, there are no records of serious equipment failures, the ratio of actual dialysis duration to prescription duration is higher than the second threshold, and the prescription parameters are within the statistical range of the patient's recent historical prescription parameters.

[0036] The offset metric is calculated by combining the cosine similarity difference and the Euclidean distance difference; the overall offset metric is used to calculate the full-cycle fatigue load index and fatigue offset index; the stage offset metric includes pre-dialysis stage offset, mid-dialysis stage offset, post-dialysis stage offset and inter-dialysis stage offset, which are used to calculate the fatigue offset index, recovery delay index and physiological tolerance index, respectively.

[0037] Furthermore, in step S5, the fatigue deviation index is obtained by weighted summation of pre-dialysis stage deviation, mid-dialysis stage deviation, and overall deviation measure, and is used to quantify the degree of fatigue abnormality caused by the current dialysis cycle process itself.

[0038] The recovery delay index is obtained by weighted summation of post-dialysis stage shift, inter-dialysis stage shift, and the increment of post-dialysis stage shift relative to its historical average level, and is used to quantify the degree of delay in the fatigue recovery process.

[0039] The physiological tolerance index is obtained by mapping the phase shift during dialysis to the [0,1] interval, and is used to quantify the patient's physiological tolerance to the current dialysis treatment;

[0040] The full-cycle fatigue load index is obtained by weighted summation of the overall offset measure of the current cycle, the average overall offset measure within the recent historical window, and the offset trend term, and is used to quantify long-term cumulative fatigue load.

[0041] Further, step S6 includes:

[0042] For each patient, a personalized threshold is dynamically calculated. The personalized threshold is the sum of the moving average and the moving standard deviation of the corresponding structured index over the patient's historical period.

[0043] Calculate the decision confidence level corresponding to the indicator based on the mean of the observation data quality labels related to the structured indicator within the period and the data completeness of the current period;

[0044] Based on the degree of standardization of the structured indicators deviating from their individualized thresholds, decision confidence, and preset clinical importance weights, the priority of each auxiliary decision suggestion to be generated is calculated and sorted from high to low priority.

[0045] Based on individualized thresholds, decision confidence, and priority ranking results, auxiliary decision-making suggestions with priority ranking and confidence scores are generated. The types of auxiliary decision-making suggestions include nursing observation suggestions, recovery management suggestions, exercise suggestions, parameter review prompts, and early warning prompts.

[0046] Furthermore, the method also includes a closed-loop feedback step: recording the implementation status of the current period's auxiliary decision-making recommendations and the improvement status of the corresponding structured indicators in the next period, and updating the historical tolerance profile accordingly to adjust the confidence or priority of subsequent similar recommendations.

[0047] This invention also provides a fatigue index construction and auxiliary decision generation system for hemodialysis, used to implement the above-mentioned method, including:

[0048] The data access module is used to acquire multimodal observation data of a patient during a complete dialysis cycle from multiple data sources, and to add stage labels, event labels and data quality labels to the observation data in a way that retains the original timestamps, forming a sequence of observation tuples with time order;

[0049] The event relationship modeling module is used to construct an event hypergraph and a time patch graph based on the observation tuple sequence. It establishes cross-modal associations within the event window through the event hypergraph, establishes dependencies at different time scales throughout the entire cycle through the time patch graph, and updates the feature representation of the observation tuple sequence through message passing.

[0050] The fatigue state characterization and full-cycle aggregation module is used to aggregate the feature representation of the updated observation tuple sequence through a shared-private latent space separation mechanism and gated fusion, and then input it into a state space model with stage switching gating to generate the stage-level representation and global representation of the current cycle.

[0051] The Individual Prototype Construction and Offset Quantization module is used to construct and update the individual prototype of a patient based on the stage-level and global representations of the patient's historical effective dialysis cycles, and to calculate the overall offset metric of the stage-level and global representations of the current cycle relative to the individual prototype and the stage offset metric of each stage.

[0052] The structured fatigue index generation module is used to generate structured fatigue indices, including fatigue offset index, recovery delay index, physiological tolerance index and full-cycle fatigue load index, based on the overall offset metric and the stage offset metric.

[0053] The decision support generation module is used to generate decision support suggestions with priority ranking and confidence scores based on structured fatigue indicators, combined with individualized thresholds, decision confidence and priority ranking rules dynamically calculated for each patient, and to receive closed-loop feedback information.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] 1. Complete preservation of the temporal structure characteristics of the data: This invention preserves the original irregular sampling timestamps, additional stages and event labels, and uses relative offset time transformation instead of forced rule resampling, so that the dense sampling information of the event triggering area and the original temporal structure of multi-source data can be completely preserved, avoiding information dilution and false interpolation.

[0056] 2. Achieving event-driven refined relationship modeling: This invention constructs an event hypergraph containing event hyperedges, time hyperedges, and variable hyperedges, and a non-uniform time patch graph, which explicitly establishes cross-modal associations and long-range dependencies throughout the entire cycle within the event window, fully characterizing the multimodal linkage changes triggered by clinical events.

[0057] 3. Providing a truly individualized fatigue assessment benchmark: The core of this invention lies in establishing a patient's own "individual prototype" (recursively updated through an exponential moving average of historical effective periods), and quantifying fatigue by the "deviation" of the current period's state relative to this prototype, replacing the traditional uniform group scoring. This allows the assessment results to reflect the patient's deviation from their normal state, possessing individualized clinical significance.

[0058] 4. Generate multi-dimensional structured fatigue indicators: This invention outputs four types of indicators: Fatigue Displacement Index (FDI), Recovery Delay Index (RDI), Physiological Tolerance Index (PTI), and Full-Cycle Fatigue Load Index (FBI). These indicators characterize fatigue status from four dimensions: process fatigue, recovery delay, immediate tolerance, and long-term cumulative load, providing a clear basis for targeted intervention.

[0059] 5. Constructing quantifiable, prioritized, and closed-loop feedback-based auxiliary decision-making: This invention employs individualized dynamic thresholds, data-quality-driven confidence adjustment, and clinically important weighted prioritization to generate decision recommendations, avoiding mechanical output from fixed templates. Simultaneously, the closed-loop feedback mechanism allows the decision-making strategy to be continuously optimized based on the patient's response history, forming a complete closed loop from indicators to decisions and from decisions to feedback. Attached Figure Description

[0060] Figure 1 This is a flowchart illustrating the implementation of the fatigue index construction and auxiliary decision generation method provided in this embodiment of the invention.

[0061] Figure 2 This is a schematic diagram of multi-source data access and full-cycle stage division in an embodiment of the present invention;

[0062] Figure 3 This is a schematic diagram of the event hypergraph construction in an embodiment of the present invention;

[0063] Figure 4 This is a schematic diagram of the time patch diagram and cross-stage association in an embodiment of the present invention;

[0064] Figure 5 This is a schematic diagram of the state characterization and aggregation process in an embodiment of the present invention;

[0065] Figure 6 This is a schematic diagram of individual prototype update and offset quantization in an embodiment of the present invention;

[0066] Figure 7 This is a schematic diagram of the structured fatigue index and auxiliary decision output in an embodiment of the present invention;

[0067] Figure 8 This is a block diagram illustrating the implementation principle of the fatigue index construction and auxiliary decision generation system provided in this embodiment of the invention. Detailed Implementation

[0068] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0069] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0070] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0071] I. Specific Implementation of the Method of the Invention

[0072] This invention provides a method for constructing multimodal fatigue indicators and generating auxiliary decisions for the entire hemodialysis cycle. This method does not take a single risk score or classification label as the final goal, but achieves multidimensional quantification and closed-loop management of dialysis-related fatigue through three progressive processing levels: constructing a full-cycle fatigue state representation, forming individualized and structured fatigue indicators, and generating structured auxiliary decisions.

[0073] like Figure 1 As shown, the specific implementation steps of the present invention are as follows:

[0074] Includes the following steps:

[0075] Step S1, Data Access: Acquire multimodal observation data from multiple data sources during a patient's complete dialysis cycle, and attach stage labels, event labels, and data quality labels to each observation data in a way that retains the original timestamps, forming a sequence of observation tuples with time order; wherein, the dialysis cycle is divided into the predialysis stage, the dialysis stage, the postdialysis stage, and the interdialysis stage;

[0076] Step S2, Event Relationship Modeling: Based on the observed tuple sequence, construct an event hypergraph and a time patch graph. Establish cross-modal associations within the event window through the event hypergraph, establish dependencies at different time scales throughout the entire cycle through the time patch graph, and update the feature representation of the observed tuple sequence through message passing.

[0077] Step S3, Fatigue State Characterization and Full-Cycle Aggregation: The feature representation of the updated observation tuple sequence is extracted by the shared-private latent space separation mechanism to extract the shared and private features of each mode. After gating fusion, the input is a state space model with stage switching gating for full-sequence aggregation to generate the stage-level representation and global representation of the current cycle.

[0078] Step S4, Individual Prototype Construction and Offset Quantization: Based on the stage-level and global representations of the patient's historical effective dialysis cycles, an individual prototype of the patient is constructed and updated using an exponential moving average method; the offset measure of the stage-level and global representations of the current cycle relative to the individual prototype is calculated, and the offset measure includes the overall offset measure and the stage offset measure of each stage.

[0079] Step S5: Generation of structured fatigue indices: Based on the offset metric, generate structured fatigue indices including fatigue offset index, recovery delay index, physiological tolerance index, and full-cycle fatigue load index.

[0080] Step S6, Assisted Decision Generation: Based on the structured fatigue index and combined with the individualized thresholds dynamically calculated for each patient, assisted decision suggestions with priority ranking and confidence scores are generated.

[0081] Of the six steps described above, the first three steps serve to provide high-quality state representation input for the fourth step, which is the core inventive step of this invention. The fifth and sixth steps transform the output of the core step into a structured result that is clinically operable.

[0082] 1. Data Access

[0083] 1.1 Multi-source data access

[0084] The data sources involved in this invention include the following seven categories. The first category is basic information from the Hospital Information System (HIS), including patient demographic characteristics (age, gender, weight, height), primary disease diagnosis, comorbidity list, medication records, and dialysis prescription parameters (dialysis mode, target ultrafiltration volume, dialysis duration, blood flow setting, dialysate formulation, etc.). The second category is real-time parameters of the dialysis equipment, including actual blood flow, dialysate flow, transmembrane pressure, venous pressure, arterial pressure, actual ultrafiltration rate, and cumulative ultrafiltration volume. This type of data is continuously recorded by the equipment at a frequency of seconds or minutes during dialysis, and the sampling density is automatically increased when an alarm event occurs. The third category is vital sign monitoring data, including systolic blood pressure, diastolic blood pressure, mean arterial pressure, heart rate, blood oxygen saturation, and body temperature. This type of data is continuously recorded by bedside monitoring equipment during dialysis, with a set of baseline values ​​recorded before and after dialysis. The fourth category is Laboratory Information System (LIS) test indicators, including hemoglobin, albumin, prealbumin, C-reactive protein, creatinine, blood urea nitrogen, potassium, sodium, calcium, phosphorus, parathyroid hormone, etc. This type of data is typically tested by collecting blood samples once before and once after dialysis. The fifth category is fatigue scale and symptom questionnaire data, including fatigue severity scores, fatigue dimension scores, and dialysis-related symptom scores. This type of data is completed by nursing staff with the assistance of patients before and after dialysis, with a sampling frequency of once per dialysis session or once per week. The sixth category is nursing records, including clinical events recorded by nursing staff during dialysis (hypotension, muscle cramps, nausea and vomiting, dizziness, chest tightness, etc.) and their management measures. This type of data is generated by event triggers and has an irregular time distribution. The seventh category is continuous data from wearable devices, including activity levels (steps, minutes of activity), sleep parameters (total sleep duration, deep sleep duration, sleep efficiency, number of awakenings), resting heart rate, etc. This type of data is continuously collected between dialysis sessions, but there may be intermittent data loss due to interruptions in wearing the device.

[0085] 1.2 Four-stage division and relative time conversion

[0086] This invention divides each complete dialysis cycle into four stages. Let the first stage be... The start time of the dialysis session is The end time of dialysis is The four-stage definition is as follows:

[0087] The predialysis phase (Phase I) is defined as ,in The predialysis observation window is typically set to 30 to 60 minutes. This phase covers the data collection window from when the patient arrives at the dialysis center until the start of the dialysis machine. The main data sources are predialysis baseline vital signs, predialysis blood test results, and predialysis fatigue scale scores.

[0088] Phase II of dialysis is defined as follows: This stage covers the entire dialysis treatment process and typically lasts 3 to 5 hours. The main data sources are real-time parameters of the dialysis equipment, continuous vital sign monitoring, and nursing event records. This stage has the largest data volume, the highest sampling density, and the most concentrated event-triggered sampling.

[0089] The post-dialysis phase (Phase III) is defined as follows: ,in The observation window after dialysis is typically set to 30 to 120 minutes. This phase covers the immediate recovery period after dialysis. The main data sources are post-dialysis vital signs, post-dialysis blood test indicators, and post-dialysis fatigue scale scores.

[0090] The interdialysis phase (Phase IV) is defined as follows: This phase covers the entire home period between two dialysis sessions and typically lasts 1 to 3 days, with the primary data source being continuous data from wearable devices.

[0091] Starting time of a single dialysis session As a reference origin, the absolute timestamps of all observations Convert to relative offset time The conversion formula is:

[0092] (1)

[0093] in Corresponding to the pre-dialysis stage, The same principle applies to the corresponding stages of dialysis. This transformation standardizes the time coordinate system across different dialysis cycles while preserving the original relative time position of each observation, without performing any forced resampling or grid alignment operations.

[0094] 1.3 Observation Label Attachment

[0095] Each observation record Before proceeding to further processing, four types of labels must be added to form labeled observation tuples:

[0096] (2)

[0097] in For the observation vector ( (The feature dimension of this data source). This is the relative offset time. For stage labels, For data source tags (corresponding to the seven types of data sources mentioned above). For event labels (0 indicates no event, 1 to 10 ... (Each corresponds to a different clinical event type) The data quality label is determined by a comprehensive assessment of data integrity, sensor status, and reasonableness of the value range; 1 indicates the highest quality and 0 indicates that the data is completely unusable.

[0098] The technical advantages of this data access method are: it preserves the temporal structure characteristics of the original irregular sampling, enabling subsequent event relationship modeling and state representation to directly utilize the dense sampling information of the event triggering area; it achieves logical segmentation of the full-cycle data rather than physical truncation through stage labels; and it provides a quantitative basis for subsequent indicator generation and decision confidence adjustment through data quality labels.

[0099] Figure 2 This is a schematic diagram of multi-source data access and the division of the entire lifecycle stages. Figure 2 The figure illustrates the time division of four phases and the sampling distribution characteristics of seven data sources in each phase. The vertical dashed lines in the figure mark the locations of clinical events. The sampling point density in the vicinity of the event is significantly higher than that in the stable area, which intuitively reflects the characteristic of dense sampling triggered by the event.

[0100] 2. Event Relationship Modeling

[0101] After completing data access, this invention does not adopt the conventional approach of simply stitching together all observations after aligning them according to a fixed time grid. Instead, it constructs two complementary relationship structures—an event hypergraph and a time patch graph—to explicitly express cross-modal correlations within the event window and multi-scale time dependencies over the entire period.

[0102] 2.1 Event Sensitive Mask Generation

[0103] Before constructing the relational structure, event-sensitive mask weights are first generated for each observation. For relative time... The observation at the location, its event-sensitive mask weight The calculation method is as follows:

[0104] (3)

[0105] in This represents the total number of clinical events recorded within the current period. For the first The time of occurrence of each event (relative offset time). For the first The impact intensity coefficient of an event (determined by the type of event, for example, the impact intensity of a hypotensive episode is greater than that of a mild muscle spasm). For the first The time-related impact range parameter of an event. This is the sigmoid function. The technical advantage of this formula is that observations closer to the clinical event receive higher mask weights, thus receiving greater attention in subsequent message passing. This enhances the information in the event-triggered region during relational modeling rather than diluting it through averaging.

[0106] 2.2 Event Hypergraph Construction

[0107] Event Hypergraph node set It consists of all tagged observations within the current dialysis cycle. For each observation... First, it is mapped to node features of a uniform dimension through a data source-specific embedding layer:

[0108] (4)

[0109] in and For the first Projection parameters corresponding to different data sources (different projection parameters are used for different data sources to handle differences in feature dimensions). For a unified embedding dimension, Based on relative offset time The consecutive position encoding uses the following formula:

[0110] (5)

[0111] in , This refers to the temperature parameter.

[0112] Hyperedge set of a hypergraph It includes three types of hyperedges. The first type is event hyperedges: for each clinical event... According to the time of its occurrence Centered on, with For time window ( (Assuming the event's influence radius), this incorporates all observation nodes from all data sources within the window into the same event hyperedge. This explicitly establishes cross-modal associations triggered by events. The second type is temporal hyperedges: the entire cycle is naturally grouped according to stage labels and temporal proximity, with time intervals within the same stage less than a threshold. Continuous observations were incorporated into the same time super-edge This type of hyperedge captures temporal locality and intra-stage continuity. The third type is the variable hyperedge: it incorporates all nodes belonging to the same data source and the same observed variable throughout the entire period into the same variable hyperedge. This hyperedge enables information exchange between observations of the same physiological variable at different time points, capturing the long-term evolutionary trend of the variable.

[0113] Message passing on the hypergraph employs a two-phase aggregation mechanism. In the aggregation phase from a node to a hyperedge, for each hyperedge... and the set of nodes it contains The update formula for the hyperedge representation is:

[0114] (6)

[0115] in The projection matrix is ​​the value. The attention coefficient is calculated as follows:

[0116] (7)

[0117] in These are the query and key projection matrices, respectively. During the aggregation phase from hyperedge to node, for the node... and its set of superedges The update formula for the node representation is:

[0118] (8)

[0119] in This represents the attention weights of a node to the representations of its respective hyperedges. This is a feedforward network layer. This two-stage aggregation allows multimodal information within the event window to fully interact through event hyperedges, while maintaining temporal locality and variable consistency through temporal and variable hyperedges.

[0120] Figure 3 This is a schematic diagram of the construction of an event hypergraph. Figure 3 This demonstrates how the event hypergraph is constructed. Using observation nodes as the basic element, three types of hyperedges are presented visually: event hyperedges are represented by dashed ellipses enclosing all nodes within the event window; time hyperedges are represented by rounded rectangles grouping consecutive nodes within the same period; and variable hyperedges are represented by vertical dashed lines grouping the same variable across time nodes. The diagram below shows the two-stage information flow directions and corresponding formulas for node-to-hyperedge aggregation and hyperedge-to-node aggregation.

[0121] 2.3 Time Patch Graph Construction

[0122] Time Patch Diagram The design goal is to capture dependencies across different time scales throughout the entire period. Unlike the node-level modeling of event hypergraphs, time-patch graphs use patch-level modeling, which divides the observation sequence into non-uniform time patches, with each patch serving as a node in the graph.

[0123] The rule for non-uniform patching is: in event-dense regions (event-sensitive masks)... The mean is higher than the threshold (Time period), using a smaller patch length To preserve fine-grained temporal resolution within the event window; and to employ medium patch lengths in stable regions during dialysis phases. During the interdialysis period, due to the sparse data sampling, a larger patch length is used. Let the first... The time range covered by each time patch is: Then the node characteristics of the patch are the weighted aggregation of all observed nodes within that time range after hypergraph message passing:

[0124] (9)

[0125] in For the process Node characteristics after message passing in a hypergraph This is used as a data quality label. The formula assigns greater weight to observations that are of higher quality and more relevant to the events in the patch features.

[0126] Edge set of time patch graph It includes two types of edges: intra-stage temporal adjacency edges connect temporally adjacent patch pairs within the same stage, with their edge weights inversely proportional to the time interval between the two patches; cross-stage association edges connect semantically related patch pairs in different stages (e.g., establishing an association between the pre-dialysis blood pressure baseline patch and the blood pressure fluctuation patch during dialysis), with their edge weights determined by the cosine similarity between patch features. The message passing formula on the patch graph is:

[0127] (10)

[0128] in For patch In the figure The set of neighbors in the middle, For edge weights, The parameters are for graph convolution.

[0129] The core difference between this module and existing technologies is that existing technologies directly stitch all observations into the input sequence model after aligning them at fixed time intervals, thus losing the local dense sampling information of the event window and the real-time correlation across modes. This invention explicitly establishes the cross-modal correlation structure within the event window through an event hypergraph, and achieves long-range relationship modeling of the entire cycle while preserving the high resolution of the event region through a non-uniform time patch graph.

[0130] Figure 4 This is a time patch diagram and a cross-stage association diagram. Figure 4 The diagram shows the results of non-uniform patching across the entire timeline. Event-dense regions (gray fill) have smaller patches to preserve high temporal resolution, stable regions have medium-sized patches, and interdialysis interval patches are the largest. The patch plot below shows intra-stage adjacency edges (solid lines) and cross-stage correlation edges (dashed curves) between patches, illustrating a multi-scale time-dependent modeling approach.

[0131] 2.4 Self-supervised pre-training and reconstruction loss

[0132] To enhance the ability of the event relationship modeling module to represent irregular multimodal data, this invention introduces a self-supervised pre-training mechanism. During the pre-training phase, some observation nodes are randomly occluded (occlusion rate is...). The original features of the occluded nodes are reconstructed using contextual information obtained from hypergraph and patch graph message passing. The reconstruction loss formula is:

[0133] (11)

[0134] in For the set of indices of the occluded observations, These are the reconstructed values. A data quality label is introduced into this loss function. and event-sensitive mask As weights, high-quality and event-related observations receive greater reconstruction optimization during pre-training.

[0135] 3. Fatigue state characterization and full-cycle aggregation

[0136] After event relationship modeling, node-level features have been obtained. and patch-level features The goal of this step is to transform these intermediate representations into a unified, full-cycle state representation for subsequent comparison with individual prototypes. This step is achieved through two sub-mechanisms: shared-private latent space separation and state space aggregation.

[0137] 3.1 Separation of Shared and Private Potential Spaces

[0138] Data from different modalities simultaneously contain shared information reflecting common fatigue characteristics and private information reflecting the state of specific physiological systems. To distinguish between these two types of information, this invention analyzes each data source... The observed features are subjected to shared and private projections.

[0139] The formula for shared projection is:

[0140] (12)

[0141] in and Shared projection parameters for all modes This projection technique aims to share the latent space dimension. Its function is to map features from different data sources onto the same shared space, enabling direct comparison and fusion of common fatigue features across different modalities.

[0142] The private projection formula is:

[0143] (13)

[0144] in and For the first Private projection parameters specific to the data source class. This is a private latent space dimension. Different data sources use independent private projection parameters, thus preserving the unique physiological state information of each modality.

[0145] To ensure that shared and private spaces truly capture different types of information, the following two constraints are introduced. The orthogonal constraint formula is:

[0146] (14)

[0147] This constraint makes the shared and private representations of the same observation tend to be orthogonal in the vector space, thereby reducing information redundancy and ensuring that the shared space does not absorb mode-specific information that should belong to the private space. The contrast constraint formula is:

[0148] (15)

[0149] in It is a set of positive sample pairs (observation pairs of different modalities within the same time window). The cosine similarity function is used. The temperature parameter is used. This constraint encourages shared representations from different modalities to be close to each other within the same time window, thus ensuring that the shared space truly captures common features across modalities.

[0150] The gating fusion formula is used to fuse shared and private representations into a unified observation representation:

[0151] (16)

[0152] (17)

[0153] in This represents vector concatenation. and For gating parameters, Represents element-wise product. and They are respectively and Unified by linear mapping to The result after dimensioning. Gated vector. Adaptively control the fusion ratio of shared and private information in each dimension.

[0154] 3.2 State Space Aggregation

[0155] The fused observation representations are arranged chronologically to form a full-cycle sequence. (by relative offset time) (arranged in ascending order). This invention employs a state-space model to perform long-range aggregation of the sequence, capturing the evolution trajectory of the fatigue state between four stages. Unlike traditional recurrent networks, the state-space model achieves sequence modeling through the discretization of continuous-time dynamics, making it naturally suitable for handling observation sequences with irregular time intervals.

[0156] The state update formula is:

[0157] (18)

[0158] (19)

[0159] (20)

[0160] in For the first The hidden state at each time step represents the cumulative fatigue evolution state up to that time step; This is the state transition matrix, which describes the autoregressive dynamic characteristics of the hidden states; The current observation information is injected into the hidden state space as the input projection matrix; To output the projection matrix, the hidden states are mapped back to the observation space for subsequent use; The discretization step size is dynamically determined based on the time interval between adjacent observations. The dynamic parameters are generated as follows:

[0161] (twenty one)

[0162] (twenty two)

[0163] in , , , These are trainable parameters. The softplus function ensures that the discretization step size is positive. This design allows the discretization step size to adaptively increase when the time interval between adjacent observations is large (such as when crossing from the dialysis phase to the interdialysis phase), thereby maintaining the continuity and rationality of state transitions.

[0164] The phase transition gating formula is used to handle state transitions between the four phases. When two adjacent observations belong to different phases (i.e., ...), ... Introduce phase switching gating in state updates:

[0165] (twenty three)

[0166] (twenty four)

[0167] in Stage label Trainable embedding vectors, and For gating parameters, For the stage The system can trainable initial state vectors. This gating mechanism enables the system to adaptively balance between continuing the state information of the previous stage and resetting to the initial state of the new stage when a stage switch occurs, thus preserving long-range dependencies across stages while adapting to the differences in state distribution between different stages.

[0168] After state-space aggregation of the entire sequence, the output representation of each time step is obtained. The present invention further aggregates the output representations of each stage to form four stage-level representations and one global representation:

[0169] (25)

[0170] (26)

[0171] in For the first Each cycle in the stage The aggregation representation, For the first A globally concatenated representation of each period. Data quality labels. Weighted participation ensures that high-quality observations contribute more to the stage representation.

[0172] The technical effects of this step are reflected in three aspects: the long-range evolution trajectory of fatigue state between the four stages is captured through the recursive structure of the state-space model; the state transition characteristics between stages are preserved through stage switching gating; and a unified representation basis is provided for subsequent multi-granularity comparison of individual prototypes by simultaneously generating state representations at both the stage level and the global level.

[0173] Figure 5 This is a schematic diagram of the state representation and aggregation process. Figure 5 This demonstrates the branching structure of shared and private projections, the application positions of orthogonal and contrast constraints, the information flow of gating fusion, the recursive links of the state-space model across the four stages, and the triggering position of the stage switching gating (denoted as G) at the Phase boundary. The final output is a global representation composed of the concatenation of the four stage-level representations.

[0174] 4. Individual Prototype Memory and Offset Quantization

[0175] This section is the core creative element of the entire technical solution. Existing technologies use a unified population model to output the relative position within the population distribution as the evaluation result. The fundamental difference of this invention lies in establishing an individual prototype for each patient that reflects their own normal state characteristics, and quantifying the degree of abnormality in fatigue state by the deviation of the current cycle relative to this prototype. The basic principle of this design is that the determination of dialysis-related fatigue should not be based on the population average level, but rather on the patient's own historical normal performance, because the same physiological indicator changes have drastically different clinical significance in different patients.

[0176] 4.1 Screening of historically effective dialysis cycles

[0177] Not all of a patient's historical dialysis cycles are suitable for building an individual prototype. This invention defines the following four screening rules, and only historical cycles that meet all four rules are included in prototype construction.

[0178] Rule 1 (Data Integrity Rule): Data integrity for this period The threshold must be reached Data integrity is defined as the data quality label within that period. The ratio of the number of observations to the total number of observations in that period:

[0179] (27)

[0180] Rule 2 (Rule for Normal Equipment Operation): There are no records of serious dialysis equipment malfunctions during this cycle, such as dialyzer clotting leading to premature termination, or equipment malfunction alarms lasting longer than the preset limit.

[0181] Rule 3 (Dialysis Completion Rule): The ratio of the actual dialysis duration to the prescribed dialysis duration for the cycle shall not be less than the threshold. This means that dialysis was not terminated prematurely due to uncontrollable reasons.

[0182] Rule 4 (Regular Range Rule for Prescription Parameters): The prescription parameters (blood flow, dialysis fluid flow, target ultrafiltration volume, etc.) for this cycle are within the range of the mean of the patient's recent historical prescription parameters plus or minus 2 standard deviations, in order to exclude non-representative cycles caused by temporary prescription adjustments.

[0183] Let the set of historical valid periods that satisfy the above four rules be . .

[0184] 4.2 Individual Prototype Construction and Update

[0185] For the set of historical effective periods Each cycle The stage-level representation is obtained according to the method in Section 3. and global representation .

[0186] This invention uses the Exponential Moving Average (EMA) method to recursively update individual prototypes. Assume that after processing the... The individual prototype after one effective cycle is Then the individual prototype update formula is:

[0187] Stage-based prototype update:

[0188] (28)

[0189] Global prototype update:

[0190] (29)

[0191] in This is the momentum coefficient, controlling the decay rate of historical information. When... When the value approaches 1, the prototype updates slowly, retaining more early historical information, making it suitable for patients with relatively stable conditions; when... When smaller, the prototype adapts more quickly to long-term drift in patient condition, making it suitable for patients with rapidly changing conditions. For the first effective cycle for each patient, the prototype is initialized with the state representation for that cycle, i.e. .

[0192] The technical advantages of the EMA update method are reflected in three aspects: there is no need to store and reprocess the original data of all historical periods, only the current prototype vector needs to be maintained to achieve incremental updates, and the computation and storage overhead is constant; the prototype representation naturally reflects the patient's recent normal state characteristics, while retaining a certain historical memory through the momentum coefficient; when the patient's condition changes over a long period of time, the prototype can gradually adapt to the new normal, rather than being rigidly fixed in the early state.

[0193] 4.3 Offset Quantization

[0194] In obtaining the current cycle (the... After representing the state of each cycle and the individual prototype of the patient, the present invention calculates the offset metric at both the overall and phased levels.

[0195] The formula for overall offset measurement is:

[0196] (30)

[0197] The definitions of each item are as follows:

[0198] (31)

[0199] (32)

[0200] Formula (31) is a cosine difference measure, reflecting the degree of deviation in direction between the current periodic global representation and the prototype global representation, with a value range of [value range missing]. The larger the value, the greater the directional deviation. Formula (32) is the normalized Euclidean distance metric, reflecting the degree of deviation in scale between the current periodic global representation and the prototype global representation, where the denominator is... This is a normalization factor to eliminate the influence of vector dimension on the distance value. and For the preset weighting coefficients, satisfy They control the contribution ratio of orientation offset and scale offset to the overall metric, respectively.

[0201] The formula for measuring phased offset is:

[0202] (33)

[0203] The formula has the same structure as the overall offset metric, but it is calculated independently in the representation spaces of the four stages, thus obtaining the degree of offset of the current cycle relative to the prototype in each stage. The clinical significance of the staged offset metric lies in: predialysis stage offset... Reflects the degree of abnormality of a patient's baseline status relative to their normal baseline when entering dialysis; stage shift during dialysis. Reflects the degree of abnormality in physiological responses during dialysis relative to their normal tolerance pattern; postdialysis phase shift. Reflects the degree of abnormality in the acute recovery phase relative to its normal recovery pattern; interdialysis interval shift. It reflects the degree of abnormality of the home recovery period relative to its normal interval pattern.

[0204] The fundamental reason for using individual prototype shift rather than population risk scores is that population models can only determine a patient's relative position within the population distribution, but the same relative position has completely different clinical significance for different patients. For example, a slight shift in a patient with consistently good tolerance may indicate a problem, while the same index value may be considered normal for a patient who has been in a poor condition for a long time. The individual prototype shift mechanism accurately captures the core information of how much a patient deviates from their normal state, which is the basis for the individualized clinical significance of structured fatigue indices.

[0205] Figure 6 This is a schematic diagram of individual prototype update and offset quantization. Figure 6 This is a visual representation of the core inventiveness of the present invention. The top section shows the state representation sequence formed after state aggregation of historical effective cycles, which is then recursively updated using EMA to form an individual prototype. The middle section shows the offset quantification process between the current cycle state representation and the prototype. The bottom section shows the correspondence between the overall offset and the phased offsets (four dimensions: pre-dialysis, during dialysis, post-dialysis, and interdialysis), as well as the clinical significance of each dimension. The bottom section compares the fundamental differences between the individual prototype offset method of the present invention and the existing group-based unified scoring method.

[0206] 5. Generation of Structured Fatigue Indices

[0207] Based on the overall offset metric and the phased offset metric obtained in Section 4, this invention generates the following four types of structured fatigue indices. These indices are all directly derived from the offset of the current cycle relative to the individual prototype, rather than arbitrarily set artificial scores.

[0208] 5.1 Fatigue Deviation Index (FDI)

[0209] Indicator Definition: FDI reflects the overall state deviation of the current dialysis cycle in the two stages of predialysis and dialysis, and is used to quantify the degree of fatigue abnormality caused by the dialysis process itself in this cycle.

[0210] The formula for calculating the indicator is:

[0211] (34)

[0212] in For the weighting coefficients, satisfying . This is a pre-dialysis stage offset measure that reflects the pre-contribution of baseline anomalies to fatigue. It is a measure of phase deviation during dialysis, reflecting deviations from physiological tolerance during treatment. This is a measure of overall offset, reflecting the comprehensive deviation over the entire period. A larger value (such as 0.5) is usually chosen because the dialysis phase is the main source of fatigue.

[0213] Clinical significance: The higher the FDI value, the greater the deviation of the current dialysis cycle process characteristics from the patient's normal pattern, suggesting that the fatigue load caused by this dialysis is higher than its usual level, and it may be necessary to pay attention to the adjustment of dialysis parameters in subsequent treatment.

[0214] 5.2 Recovery Delay Index (RDI)

[0215] RDI Definition: The RDI reflects the degree of deviation of the two phases of immediate recovery after dialysis and home recovery between dialysis periods relative to the individual prototype. It is used to quantify the degree of delay or inadequacy in the fatigue recovery process.

[0216] The formula for calculating the indicator is:

[0217] (35)

[0218] in For the weighting coefficients, satisfying . This is a measure of post-dialysis stage shift. This is a measure of interdialysis shift. This is the moving average of the post-dialysis stage offset measure over the patient's historical effective cycles. The third term in the formula... To recover abnormal increments, a positive contribution is only made when the post-dialysis offset exceeds the historical average level, which is used to capture sudden recovery delays.

[0219] Clinical significance: The higher the RDI value, the slower or worse the patient's recovery process in the current cycle is compared to their normal recovery pattern, suggesting possible factors hindering recovery (such as malnutrition, sleep disorders, infection, etc.) that need attention.

[0220] 5.3 Physiological Tolerance Index (PTI)

[0221] Indicator Definition: PTI focuses on the mid-dialysis phase, quantifying the deviation of a patient's current cycle of physiological tolerance to dialysis treatment relative to their individual normal tolerance pattern.

[0222] The formula for calculating the indicator is:

[0223] (36)

[0224] in This is the scaling factor, which controls the sensitivity of PTI to changes in offset. The median of the phase shift measurement during dialysis throughout the patient's historical effective cycles was used as an individualized reference benchmark. When the phase shift during dialysis... equal to the historical median hour, This indicates that the patient's tolerance is at a normal level; when Significantly greater than When the PTI approaches 0, it indicates a significant decrease in tolerance; when much smaller When the PTI approaches 1, it indicates good tolerance.

[0225] Clinical significance: PTI reflects a patient's tolerance to dialysis treatment in a normalized manner. A PTI that is consistently below 0.5 may indicate that the dialysis prescription parameters may need to be reviewed (such as reducing the ultrafiltration rate, adjusting the dialysis duration, and checking vascular access function).

[0226] 5.4 Fatigue Burden Index (FBI)

[0227] Indicator definition: The FBI integrates offset information from all four phases and the cumulative trend of recent historical cycles to quantify the cumulative fatigue load of patients over a time span.

[0228] The formula for calculating the indicator is:

[0229] (37)

[0230] in This refers to the set of valid periods within the recent historical window (e.g., the last five valid periods). For the first The overall offset measure for each historical period. The slope of the linear trend for the offset metric is determined by... Each cycle The results were obtained by linear regression in chronological order. Positive values ​​indicate an upward trend in the offset, meaning that fatigue is accumulating and worsening, while negative values ​​indicate a downward trend in the offset, meaning that the fatigue condition is improving. For the weighting coefficients, satisfying .

[0231] Clinical significance: The FBI not only reflects the fatigue state of the current cycle but also integrates recent historical averages and trends, enabling it to capture cumulative fatigue loads that cannot be reflected by single-cycle indicators. A persistently elevated FBI suggests that the patient may be in a phase of cumulative and worsening fatigue, requiring comprehensive intervention.

[0232] The core advantages of the above four types of indicators over the traditional single fatigue score are as follows: FDI focuses on fatigue induced by the dialysis process, RDI focuses on abnormalities in the recovery process, PTI focuses on immediate tolerance, and FBI focuses on long-term cumulative trends. These four indicators characterize the fatigue state from different dimensions and time scales, enabling subsequent decision support to adopt targeted intervention strategies based on specific abnormalities, rather than being helpless in the face of an indivisible comprehensive score.

[0233] 6. Decision-making support

[0234] The decision support generation module of this invention transforms structured fatigue indicators into clinically operable decision support outputs. Instead of simply mapping fixed text templates to indicators, this module generates structured decisions through three calculation steps: dynamic threshold calculation, decision confidence adjustment, and intervention priority ranking, combined with the patient's clinical context and historical tolerance records.

[0235] 6.1 Input Information

[0236] The input for decision support generation consists of five parts: the first part is the structured fatigue index, i.e. The first part contains the current values ​​of four indicators; the second part contains the patient's clinical context, including structured information obtained from the HIS such as current dialysis prescription parameters, a list of comorbidities, and current medication regimens; the third part contains a contraindication rule base, which includes intervention contraindications for specific comorbidities or drugs (e.g., patients with severe heart failure should not be advised to increase exercise intensity, and patients with bleeding risk should not be advised to increase anticoagulation dose); the fourth part contains a historical tolerability profile, recording the patient's response to various recommendations in previous cycles (e.g., whether the corresponding indicator improved in the next cycle after a certain type of recommendation was implemented); and the fifth part contains a set of data quality labels for the current cycle, used to assess the reliability of the data on which the indicators are calculated.

[0237] 6.2 Dynamic Threshold Calculation

[0238] This invention does not use a globally fixed threshold to determine whether an indicator is abnormal; instead, it dynamically calculates an individualized threshold for each patient. The dynamic threshold calculation formula is:

[0239] (38)

[0240] in , This is the moving average of this indicator over the patient's historical effective treatment periods. The corresponding moving standard deviation (based on the most recent) (Calculation of each effective period) This is the sensitivity coefficient (adjustable according to clinical needs; a smaller value indicates greater sensitivity, and a larger value indicates greater conservatism). When a certain indicator exceeds its corresponding dynamic threshold, that dimension is considered abnormal. The technical advantage of this formula is that for patients with poor but stable baseline conditions, the threshold automatically adjusts to a higher level, avoiding persistent false alarms; for patients with good baseline conditions, the threshold is at a lower level, enabling sensitive detection of minor abnormalities. It should be noted that the PTI indicator uses a lower limit threshold for judgment, i.e., when the PTI is lower than... It was believed at the time that tolerance was abnormally decreased.

[0241] 6.3 Adjustment of Decision Confidence

[0242] Since the calculation of the structured fatigue index depends on the quality of multi-source data, the reliability of the index value decreases accordingly when the data quality is low. The decision confidence adjustment formula is as follows:

[0243] (39)

[0244] in The mean of the observed data quality labels within the phases that are primarily related to the indicator (for example, if FDI is primarily related to Phase I and Phase II, then the mean of the data quality of these two phases is taken; if RDI is primarily related to Phase III and Phase IV, then the mean of the data quality of the latter two phases is taken). This represents the data completeness for the current period. When both data quality and completeness are at a high level, the confidence level approaches 1; when data quality declines, the confidence level decreases accordingly. A confidence level below a preset threshold... Decision recommendations triggered by certain indicators will be marked as low-confidence recommendations, prompting clinicians to make a comprehensive judgment based on other information.

[0245] 6.4 Intervention Prioritization

[0246] When multiple indicators show abnormalities simultaneously, it is necessary to prioritize the various intervention recommendations. The formula for prioritizing interventions is:

[0247] (40)

[0248] in This is the current value of the indicator. For dynamic thresholds, This represents the historical standard deviation (used to indicate the degree of standardized deviation). For decision confidence, Clinical importance weights are assigned (based on domain knowledge; for example, abnormal PTIs involve immediate safety, so their clinical importance weight is higher than that of RDIs). Intervention recommendations are based on... Output sorted from highest to lowest. For the PTI metric, since it uses a lower threshold, the priority calculation is adjusted accordingly. .

[0249] 6.5 Decision Output Types and Generation Logic

[0250] Based on the above dynamic threshold judgment, confidence adjustment and priority ranking results, the auxiliary decision output of the present invention includes the following five types of structured suggestions.

[0251] The first category is nursing observation recommendations: triggered when the PTI falls below the lower limit of the dynamic threshold. Recommendations include increasing the frequency of vital sign monitoring during dialysis and closely monitoring for early signs of hypotension. The specific observation frequency is proportional to the degree of PTI deviation. The second category is recovery management recommendations: triggered when the RDI exceeds the upper limit of the dynamic threshold. Recommendations include extending the in-hospital observation time after dialysis and assessing interdialysis nutritional intake and sleep quality. The intensity of the intervention is determined by both the degree of RDI deviation and the FBI trend. The third category is exercise recommendations: based on the current values ​​of FDI and RDI and referring to exercise restriction entries in the contraindication rule base, appropriate activity level recommendations are generated. When FDI is high and RDI is normal, it is recommended to moderately reduce exercise intensity; when FDI is normal and RDI is high, it is recommended to increase light recovery activities. The fourth category is parameter review prompts: triggered when the PTI remains below the dynamic threshold and the FBI shows an upward trend. This prompts clinicians to review dialysis prescription parameters (including ultrafiltration rate, dialysis duration, blood flow, etc.) and provides specific parameters to focus on. The fifth category is early warning alerts: when the FBI exceeds the upper limit dynamic threshold and the trend slope... When the value is positive, a comprehensive early warning is triggered, indicating that the patient is in a phase of cumulative and deteriorating fatigue, requiring multi-dimensional collaborative intervention.

[0252] Each of the above recommendations includes the following structured attributes: the triggering indicator and its current value, the corresponding dynamic threshold, the standardized value of the deviation, the confidence score, and the priority ranking. These recommendations are not mechanical outputs from fixed templates, but rather structured results driven by the degree of indicator anomaly, individualized dynamic thresholds, data quality, and historical tolerance profiles.

[0253] 6.6 Closed-loop feedback mechanism

[0254] This invention establishes a closed-loop feedback mechanism, incorporating the implementation status and subsequent responses of decision support measures into a historical record for updating decisions in the next cycle. Specifically, when a suggestion is adopted and implemented by clinicians in the current cycle, the system records the direction of change of the corresponding indicator in the next dialysis cycle: if the indicator improves in the next cycle (e.g., FDI decreases, PTI increases), the positive response count for that type of suggestion is incremented by one in the historical tolerability record, enhancing the confidence level when generating similar suggestions in the future; if the indicator does not improve or worsens in the next cycle, the negative response count is incremented by one, and the priority of similar suggestions in the future will be appropriately reduced. This feedback mechanism enables the decision support system to continuously optimize suggestion strategies based on the patient's individual historical response patterns, forming a complete closed loop from indicator to decision, from decision to feedback, and from feedback to the next cycle's indicator.

[0255] Figure 7 This is a schematic diagram of structured fatigue indicators and auxiliary decision-making outputs. Figure 7 It demonstrates the computational relationship from offset measurement to four types of structured indicators (FDI, RDI, PTI, FBI), the decision generation process from indicators to dynamic threshold judgment, confidence adjustment, and priority ranking, five types of auxiliary decision outputs, and the complete path from decision output to closed-loop feedback. Each suggestion is accompanied by structured information such as trigger indicator, current value, dynamic threshold, confidence level, and priority ranking.

[0256] II. Specific Implementation of the System of the Invention

[0257] This invention also provides a multimodal fatigue index construction and auxiliary decision generation system for the entire hemodialysis cycle, used to implement the above-mentioned method. For example... Figure 8 As shown, the system contains six functional modules:

[0258] The data access module is used to acquire multimodal observation data of a patient during a complete dialysis cycle from multiple data sources, and to add stage labels, event labels and data quality labels to the observation data in a way that retains the original timestamps, forming a sequence of observation tuples with time order;

[0259] The event relationship modeling module is used to construct an event hypergraph and a time patch graph based on the observation tuple sequence. It establishes cross-modal associations within the event window through the event hypergraph, establishes dependencies at different time scales throughout the entire cycle through the time patch graph, and updates the feature representation of the observation tuple sequence through message passing.

[0260] The fatigue state characterization and full-cycle aggregation module is used to aggregate the feature representation of the updated observation tuple sequence through a shared-private latent space separation mechanism and gated fusion, and then input it into a state space model with stage switching gating to generate the stage-level representation and global representation of the current cycle.

[0261] The Individual Prototype Construction and Offset Quantization module is used to construct and update the individual prototype of a patient based on the stage-level and global representations of the patient's historical effective dialysis cycles, and to calculate the overall offset metric of the stage-level and global representations of the current cycle relative to the individual prototype and the stage offset metric of each stage.

[0262] The structured fatigue index generation module is used to generate structured fatigue indices, including fatigue offset index, recovery delay index, physiological tolerance index and full-cycle fatigue load index, based on the overall offset metric and the stage offset metric.

[0263] The decision support generation module is used to generate decision support suggestions with priority ranking and confidence scores based on structured fatigue indicators, combined with individualized thresholds, decision confidence and priority ranking rules dynamically calculated for each patient, and to receive closed-loop feedback information.

[0264] Figure 8 The system's six functional modules and their data flow connections are illustrated. Seven types of data sources are listed at the bottom, starting with the data access module. The main processing chain sequentially passes through six modules: event relationship modeling, fatigue state characterization and full-cycle aggregation, individual prototype construction and offset quantification, structured fatigue index generation, and auxiliary decision generation. Offset measurement is completed in the "Individual Prototype Construction and Offset Quantification Module," while the "Structured Fatigue Index Generation Module" generates indicators such as FDI, RDI, PTI, and FBI based solely on the offset measurement. The auxiliary decision generation module uses a dotted feedback path to send the implementation status of suggestions and subsequent indicator responses back to the Individual Prototype Construction and Offset Quantification Module to support subsequent periodic updates.

[0265] 1. Data Access Module

[0266] Inputs include: patient basic information and prescription parameters from the HIS system, real-time operating parameters from the dialysis equipment, vital sign data from the bedside monitoring equipment, laboratory test results from the LIS system, electronically collected data from fatigue scales and symptom questionnaires, clinical event records entered by nursing staff, continuous monitoring data between dialysis intervals from wearable devices, and dialysis scheduling information from the HIS system.

[0267] Processing Logic: This module corresponds to method step S1 and performs two types of operations. The first type of operation is data access: Raw observation data is obtained from the aforementioned data sources through a standardized data interface, retaining the original absolute timestamps without forced resampling or grid alignment, and marking missing fields, device status, and value range rationality. The second type of operation is also data access: Using the start time of a single dialysis session as the reference origin, the absolute timestamp is converted to a relative offset time; based on the relative offset time, four stages are divided: pre-dialysis, during dialysis, post-dialysis, and inter-dialysis; event labels are added to the observations within the event window based on clinical events in the nursing records; and data quality labels are formed by comprehensively considering data integrity, sensor status, and value range rationality.

[0268] Output: A sequence of observation tuples in time order, each observation tuple including an observation vector, relative offset time, stage label, data source label, event label, and data quality label.

[0269] Connection relationship: The output of this module is fed into the event relationship modeling module.

[0270] 2 Event Relationship Modeling Module

[0271] Input: A sequence of labeled observation tuples from the data access module.

[0272] Processing logic: This module corresponds to method step S2 and performs four operations. The first operation is event-sensitive mask generation: according to the clinical event occurrence time, event impact intensity and time impact range, calculate the event-sensitive mask weight of each observation according to formula (3). The second operation is node feature initialization: according to formula (4) and formula (5), through the specific embedding layer and continuous position encoding of the data source, map each observation to a node feature of a unified dimension. The third operation is event hypergraph construction and message passing: construct an event hypergraph containing three types of hyperedges: event hyperedges, time hyperedges and variable hyperedges, and perform two-stage aggregation message passing of node-hyperedge-node according to formula (6) to formula (8). The fourth operation is time patch graph construction and message passing: divide the observation sequence into time patches according to the non-uniform patch partitioning rule, calculate the patch node features according to formula (9), construct a patch graph containing adjacent edges within the stage and cross-stage related edges, and perform patch-level message passing according to formula (10); in the model training stage, self-supervised reconstruction pre-training can be further performed according to formula (11).

[0273] Output: Node-level features after message passing through both the hypergraph and patch graph. and patch-level features .

[0274] Connection relationship: The output of this module is connected to the state representation and the full-cycle aggregation module.

[0275] 3 Fatigue State Characterization and Full-Cycle Aggregation Module

[0276] Input: Node-level features from the event relationship modeling module .

[0277] Processing Logic: This module performs three operations. The first operation is the separation of shared and private latent spaces: according to formulas (12) to (15), the node features of each observation are subjected to shared and private projections, and the information separation of the two spaces is ensured by orthogonal constraints and contrast constraints. The second operation is gated fusion: according to formulas (16) and (17), the shared representation and the private representation are fused into a unified observation representation. The third operation is state space aggregation: after sorting the fused representations by time, the entire sequence is aggregated using the state space model according to formulas (18) to (24), where a stage switching gating mechanism is introduced at the stage switching point. After aggregation, four stage-level representations and one global concatenation representation are calculated according to formulas (25) and (26).

[0278] Output: Stage-level representation of the current period and global representation .

[0279] Connection relationship: The output of this module is fed into the individual prototype construction and offset quantization module.

[0280] 4. Individual Prototype Construction and Offset Quantization Module

[0281] Input: The current cycle stage-level representation and global representation from the fatigue state representation and full-cycle aggregation module, and the historical individual prototype vectors stored for this patient.

[0282] Processing Logic: This module corresponds to step S4 of the method and performs three operations. The first operation is historical effective cycle screening: according to formula (27) and rules two to four, determine the historical effective dialysis cycles that can be used for prototype construction or updating. The second operation is individual prototype construction and updating: based on the stage-level and global representations of historical effective cycles, construct and update the patient's stage-level prototype and global prototype using EMA according to formulas (28) and (29). The third operation is offset quantification: calculate the overall offset metric and four sub-stage offset metrics of the current cycle relative to the individual prototype according to formulas (30) to (33). This module maintains a persistent individual prototype library and stores the moving average and moving standard deviation of historical offset metrics.

[0283] Output: The patient's current individual prototype vector, overall offset metric, four phase offset metrics, and statistical information on historical offset metrics.

[0284] Connection relationship: The output of this module is fed into the structured fatigue index generation module; at the same time, it receives closed-loop feedback information from the auxiliary decision generation module.

[0285] 5. Structured Fatigue Index Generation Module

[0286] Input: Overall offset metric, four phase offset metrics, and historical offset statistics from the individual prototype construction and offset quantization module.

[0287] Processing logic: This module corresponds to step S5 of the method. Instead of repeatedly calculating the offset metric, it directly generates the structured fatigue index based on the overall offset metric and the stage offset metric: calculate the fatigue offset index FDI according to formula (34), calculate the recovery delay index RDI according to formula (35), calculate the physiological tolerance index PTI according to formula (36), and calculate the full-cycle fatigue load index FBI according to formula (37).

[0288] Output: A set of structured fatigue indices, and corresponding offset measures to explain the source of each index.

[0289] Connection relationship: The output of this module is fed into the auxiliary decision generation module.

[0290] 6. Decision Support Generation Module

[0291] Inputs: Structured fatigue indicators from the structured fatigue indicator generation module, patient clinical context from the HIS system, a pre-configured contraindication rule base, historical tolerance profiles from the individual prototype construction and offset quantification module, and a set of data quality labels for the current period.

[0292] Processing Logic: This module performs four operations. The first operation is dynamic threshold calculation: calculates individualized dynamic thresholds for each indicator according to formula (38). The second operation is confidence adjustment: calculates the decision confidence of each indicator based on data quality according to formula (39). The third operation is priority ranking: prioritizes the intervention suggestions corresponding to all abnormal indicators according to formula (40). The fourth operation is structured suggestion generation: generates structured auxiliary decision outputs including nursing observation suggestions, recovery management suggestions, exercise suggestions, parameter review prompts, and early warning prompts based on the abnormality type, deviation degree, confidence level, and contraindication rules of each indicator. Each suggestion is accompanied by trigger indicators, confidence level, and priority ranking.

[0293] Output: A structured list of decision support suggestions. Each suggestion includes suggestion type, suggestion content, triggering indicator and current value, dynamic threshold, confidence score, and priority ranking.

[0294] Connection relationship: The output of this module is provided to the clinical decision support terminal for display; at the same time, the implementation status of the recommendations and the changes in indicators in subsequent cycles are fed back to the individual prototype construction and offset quantification module to achieve closed-loop update.

[0295] The modules are connected via standardized data interfaces, which define data formats, field types, timestamp standards, and quality label coding specifications. Data transfer between modules uses a message flow approach, with each message carrying a patient identifier, cycle identifier, and processing stage identifier to ensure correct data routing and traceability between modules.

[0296] Example: Construction of fatigue indicators and auxiliary decision generation for a single patient's complete dialysis cycle

[0297] 1. Implementation Scenarios

[0298] This example uses a maintenance hemodialysis patient (hereinafter referred to as Patient A) as the case study. Patient A is a 62-year-old male with diabetic nephropathy as his primary disease, complicated by hypertension and mild heart failure. He undergoes hemodialysis three times a week, with each prescribed dialysis session lasting 4 hours, a prescribed blood flow rate of 250 mL / min, and a prescribed ultrafiltration volume of 3.0 L. Patient A has accumulated data from 24 historical dialysis cycles in the system, of which 20 cycles were included in the valid cycle set after four screening rules. This example demonstrates the complete treatment process for patient A's 25th dialysis cycle (the current cycle).

[0299] 2. Four-stage data access

[0300] The current dialysis schedule information shows that the scheduled start time for dialysis is 08:00 (set as the reference origin). The planned end time is 12:00 (i.e.) According to the four-stage division rule, set... , The time ranges for the four stages are as follows:

[0301] Phase I (Predialysis Stage): Relative Time Minutes, corresponding to 07:15 to 08:00. Data collected during this phase includes: pre-dialysis blood test results collected at 07:20 (14 indicators including hemoglobin, albumin, creatinine, etc., data source labels). Quality label ); Predialysis vital signs record at 07:30 (systolic blood pressure 138 mmHg, heart rate 76 bpm, etc., 6 items); , ); At 07:35, the patient completed the predialysis fatigue scale (FSS score 4.8). , ).

[0302] Phase II (Dialysis Stage): Relative Time Minutes. The data accessed during this phase includes: real-time parameters recorded by the dialysis equipment every minute (a total of 240 time points, with 7 parameters at each time point). , ); routine data recorded every 5 minutes for vital signs monitoring (a total of 48 time points, , ).exist A hypotensive event occurred at 1 minute (systolic blood pressure dropped to 85 mmHg), and the nursing staff recorded the event. (corresponding to the type of hypotension attack), vital sign monitoring before and after the event is automatically encrypted to record once every 30 seconds, lasting for 10 minutes (generating an additional 20 high-density sampling points). , ).exist A mild muscle spasm occurred at 1 minute. ).

[0303] Phase III (Post-dialysis stage): Relative time Minutes. Data accessed during this phase included: post-dialysis vital signs records at 12:05 (minutes). , ); Blood test results after dialysis at 12:10 ( , ); At 12:20, the patient completed the Postdialysis Fatigue Scale (FSS score 6.2). , ).

[0304] Phase IV (interdialysis phase): Relative time Minutes (i.e., from 12:00 noon on the same day's dialysis to 07:15 the following day before the next dialysis, approximately 42 hours). Data collected during this phase includes: activity data recorded by wearable devices (an hourly summary value, totaling approximately 40 time points). , Three time points are missing due to interrupted use; sleep parameters (one sleep report each for the current night and the following night). , ).

[0305] Following the above data integration, approximately 420 tagged observation records were generated in the current period, each with an appended relative offset time. Stage tags Data source tags Event Tags and data quality labels .

[0306] 3. Event Window Markers and Event Sensitivity Masks

[0307] for A low blood pressure event lasting several minutes was analyzed, and an impact intensity coefficient was set. Time-related influence range parameter Minutes, radius of event impact minutes. For Minutes of muscle spasm events, set , minute, Minutes. According to formula (3), located in the center of the hypotension event window ( The mask weight obtained from observations at 10 minutes from the event center is approximately 0.92, the mask weight of observations at 10 minutes from the event center drops to approximately 0.62, and the mask weight of observations in the stable region far from all events is approximately 0.50.

[0308] In the event hypergraph, the event hyperedge corresponding to the low blood pressure event. connect The observation nodes across all data sources within a minute range (including observations from dialysis equipment parameters, vital signs, and nursing records) total approximately 35 nodes. The event hyperedge corresponding to the muscle spasm event. connect Approximately 18 nodes within a minute range.

[0309] 4. State Representation and Individual Prototype Invocation

[0310] After event hypergraph message passing (assuming...) After message passing through the layer and time patch graph, 420 observations are transformed into a fused representation sequence. Let the embedding dimension be... Shared space dimension Private space dimension , integration dimension .

[0311] The result of non-uniform patch partitioning is: event-dense regions ( and (minutes) The patch length is 1 minute, resulting in approximately 16 fine-grained patches; the stable region during dialysis uses... Minutes, approximately 20 patches in total; dialysis intervals are used Each time period has approximately 40 patches, with a total of about 82 time patches throughout the entire lifecycle.

[0312] After state space aggregation, four stage-level representation vectors (each 128-dimensional) and one global concatenated representation vector (512-dimensional) are obtained.

[0313] The individual prototype of patient A is retrieved from the individual prototype construction and offset quantization module. This prototype is recursively updated using EMA based on the first 20 effective cycles and stores four stage-level prototype vectors and one global prototype vector.

[0314] 5. Offset Quantification and Index Calculation

[0315] Let the weighting coefficients be set. , The offset metric is calculated according to formulas (30) to (33). Assume the following illustrative values ​​are obtained from the calculation:

[0316] Overall offset The phased offsets are as follows: (The baseline before dialysis is close to that of the prototype). (Significant deviation during dialysis due to hypotension and muscle spasm). (Recovery after dialysis deviates from the normal pattern). (The interdialysis interval has basically returned to normal, but with a slight deviation).

[0317] According to formula (34), let Calculate FDI:

[0318]

[0319] According to formula (35), let Patient A's historical postdialysis shift moving average Calculate RDI:

[0320]

[0321] According to formula (36), let Patient A's median deviation in his / her history of dialysis Calculate PTI:

[0322]

[0323] According to formula (37), let Patient A's most recent 5 effective cycles The values ​​are 0.22, 0.25, 0.28, 0.31, and 0.35 respectively, with a moving average of 0.282 and a linear trend slope of [missing value]. Calculate the FBI:

[0324]

[0325] 6. Decision-making support

[0326] According to formula (38), let the sensitivity coefficient be... Patient A's historical FDI moving average is 0.28 and moving standard deviation is 0.06. Therefore, the dynamic threshold for FDI is... The current FDI = 0.404 > 0.37, therefore the FDI is considered abnormal.

[0327] Patient A's historical PTI moving average is 0.52 and moving standard deviation is 0.08. Therefore, the dynamic threshold for the lower limit of PTI is: The current PTI = 0.147 < 0.40, therefore the PTI is considered abnormal.

[0328] Patient A's historical RDI moving average is 0.20 and moving standard deviation is 0.04. Therefore, the dynamic threshold for RDI is... The current RDI = 0.296 > 0.26, therefore the RDI is considered abnormal.

[0329] According to formula (39), FDI is mainly related to Phase I and Phase II, and the average data quality of these two phases is: Current cycle completeness ,set up Then the confidence level of FDI The confidence level of PTI, calculated in the same way, is also 1.0. RDI is mainly related to Phase III and Phase IV, and the average data quality of these two phases is approximately 0.87. .

[0330] According to formula (40), let the clinical importance weight of PTI be set. FDI is 1.0, RDI is 1.0. Priority is calculated as follows:

[0331]

[0332]

[0333]

[0334] Sorted from highest to lowest priority, the following structured output suggestions are provided:

[0335] Priority 1 (Abnormal PTI, Priority=4.74): Nursing observation recommendations—Given that the current physiological tolerance index (PTI) during dialysis is significantly lower than the individual's dynamic threshold, it is recommended to increase the frequency of vital sign monitoring to once every 10 minutes during the next dialysis session, and closely monitor for signs of hypotension. Simultaneous parameter review prompts—Clinicians are advised to review the ultrafiltration rate setting and consider whether it needs to be reduced to 80% to 90% of the current level.

[0336] Priority 2 (Abnormal RDI, Priority=0.90): Recovery Management Recommendations – Given that the Recovery Delay Index (RDI) exceeds the individual dynamic threshold, it is recommended to extend the in-hospital observation time after the next dialysis to 90 minutes and assess the adequacy of interdialysis nutrition intake and sleep quality.

[0337] Priority 3 (Abnormal FDI, Priority=0.57): Exercise Recommendation - Given the high Fatigue Deviation Index (FDI) and low PTI, it is recommended to reduce the exercise intensity by one level during the interdialysis period and replace moderate-intensity activities with light restorative activities (walking, light stretching). Adjustments can be made again after the indicators recover in the next cycle.

[0338] 7. Confidence Adjustment When Data Quality Deteriorates

[0339] Suppose in a modified scenario, patient A's wearable device experiences prolonged detachment during the current dialysis interval, causing the mean data quality of Phase IV to drop from 0.82 to 0.35. Simultaneously, the limited number of observations during this phase results in a decrease in cycle integrity to 0.65. In this case, the confidence level of the RDI is adjusted as follows:

[0340]

[0341] Because the confidence level is lower than the preset threshold The recovery management recommendations corresponding to RDIs will be marked as "low confidence recommendations," with a note stating, "The quality of the interdialysis interval data on which this recommendation is based is insufficient; it is recommended to combine it with clinical observation for comprehensive judgment." Furthermore, the priority of RDIs is also reduced in the ranking calculation due to the confidence factor.

[0342] 8. Closed-loop feedback update

[0343] Assume that the priority 1 recommendation (reducing the ultrafiltration rate) is adopted and implemented by the clinician in the next dialysis cycle (cycle 26), with the ultrafiltration rate decreasing from 750 mL / h to 650 mL / h. After cycle 26, the system observes that the PTI recovers from 0.147 to 0.45, and the FDI decreases from 0.404 to 0.31. This positive response is recorded in patient A's historical tolerability profile: "Recommendation to reduce ultrafiltration rate → PTI improvement 0.303, FDI improvement 0.094". In subsequent cycles, if similar abnormal patterns of PTI and FDI reappear, the confidence level of such recommendations will be enhanced.

[0344] This embodiment also provides a computer device, including: at least one processor, at least one memory, and computer program instructions stored in the memory, which implement the above-described method when executed by the processor.

[0345] This embodiment also provides a computer-readable storage medium storing computer program instructions that, when executed by a processor, implement the above-described method.

[0346] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0347] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0348] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0349] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0350] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for constructing fatigue indicators and generating auxiliary decisions for hemodialysis, characterized in that, Includes the following steps: Step S1, Data Access: Acquire multimodal observation data from multiple data sources during a patient's complete dialysis cycle, and attach stage labels, event labels, and data quality labels to each observation data in a way that retains the original timestamps, forming a sequence of observation tuples with time order; wherein, the dialysis cycle is divided into the predialysis stage, the dialysis stage, the postdialysis stage, and the interdialysis stage; Step S2, Event Relationship Modeling: Based on the observed tuple sequence, construct an event hypergraph and a time patch graph. Establish cross-modal associations within the event window through the event hypergraph, establish dependencies at different time scales throughout the entire cycle through the time patch graph, and update the feature representation of the observed tuple sequence through message passing. Step S3, Fatigue State Characterization and Full-Cycle Aggregation: The feature representation of the updated observation tuple sequence is extracted by the shared-private latent space separation mechanism to extract the shared and private features of each mode. After gating fusion, the input is a state space model with stage switching gating for full-sequence aggregation to generate the stage-level representation and global representation of the current cycle. Step S4, Individual Prototype Construction and Offset Quantization: Based on the stage-level and global representations of the patient's historical effective dialysis cycles, an individual prototype of the patient is constructed and updated using an exponential moving average method; the offset measure of the stage-level and global representations of the current cycle relative to the individual prototype is calculated, and the offset measure includes the overall offset measure and the stage offset measure of each stage. Step S5: Generation of structured fatigue indices: Based on the offset metric, generate structured fatigue indices including fatigue offset index, recovery delay index, physiological tolerance index, and full-cycle fatigue load index. Step S6, Assisted Decision Generation: Based on the structured fatigue index and combined with the individualized thresholds dynamically calculated for each patient, assisted decision suggestions with priority ranking and confidence scores are generated.

2. The method for constructing fatigue indicators and generating auxiliary decisions for hemodialysis according to claim 1, characterized in that, Step S1 includes: Using the planned start time of the current dialysis cycle as the reference origin, the absolute timestamps of the observation data are converted into relative offset times; Based on the different intervals in which the relative offset time is located, the observation data are classified into one of the predialysis stage, dialysis stage, postdialysis stage, and interdialysis stage, and a stage label is attached to each observation data. Based on the clinical event information in the nursing records, attach corresponding event labels to the observation data located in the event window; Calculate and attach data quality labels based on data integrity, sensor status, and reasonableness of value range.

3. The method for constructing fatigue indicators and generating auxiliary decisions for hemodialysis according to claim 1, characterized in that, In step S2, constructing the event hypergraph includes: Each observation data point is used as a node in the hypergraph; Construct event hyperedges to group observation nodes from all data sources within the time influence window of the same clinical event into the same event hyperedge; Construct a time hyperedge to group consecutive observation nodes with a time interval of less than a preset threshold into the same time hyperedge; Construct variable hyperedges to group all nodes belonging to the same data source and the same observed variable within the same dialysis cycle into the same variable hyperedge; Perform two-stage aggregated message passing on the event hypergraph, from node to hyperedge and from hyperedge to node, to update node characteristics.

4. The method for constructing fatigue indicators and generating auxiliary decisions for hemodialysis according to claim 1, characterized in that, In step S2, the build time patch graph includes: Based on the density of the event-sensitive mask, non-uniform patching is performed on the entire time axis, with smaller patch lengths used in event-dense regions and larger patch lengths used in event-sparse regions. The features of all observed nodes within each time patch are weighted and aggregated to form the features of the patch node, where the aggregation weight is determined by the data quality label and the event sensitivity mask. Construct a time patch graph whose edge set includes intra-stage adjacency edges connecting adjacent time patches within the same stage, and cross-stage association edges connecting semantically related patches in different stages. Perform message passing on the time-patch graph to update patch node characteristics.

5. The method for constructing fatigue indicators and generating auxiliary decisions for hemodialysis according to claim 1, characterized in that, In step S3, the shared-private latent space separation mechanism includes: mapping node features to the shared latent space using shared projection parameters shared by all modalities to obtain a shared representation; mapping node features to the private latent space using private projection parameters independent of each modality to obtain a private representation; applying orthogonal constraints to make the shared representation and private representation of the same node tend to be orthogonal; applying contrast constraints to make the shared representations of nodes of different modalities within the same time window approach each other; and adaptively fusing the shared representation and private representation through a gating mechanism to obtain a fused representation. The discretization step size of the state space model is dynamically generated based on the time interval between adjacent observation data. When the stage labels of adjacent observation data are different, a stage switching gating is triggered. The stage switching gating is used to adaptively balance between continuing the state information of the previous stage and resetting to the initial state of the current stage.

6. The method for constructing fatigue indicators and generating auxiliary decisions for hemodialysis according to claim 1, characterized in that, In step S4, the historical effective dialysis cycles need to meet all of the following screening rules: data completeness is higher than the first threshold, there are no records of serious equipment failures, the ratio of actual dialysis duration to prescription duration is higher than the second threshold, and the prescription parameters are within the statistical range of the patient's recent historical prescription parameters. The offset metric is calculated by combining the cosine similarity difference and the Euclidean distance difference; the overall offset metric is used to calculate the full-cycle fatigue load index and fatigue offset index; the stage offset metric includes pre-dialysis stage offset, mid-dialysis stage offset, post-dialysis stage offset and inter-dialysis stage offset, which are used to calculate the fatigue offset index, recovery delay index and physiological tolerance index, respectively.

7. The method for constructing fatigue indicators and generating auxiliary decisions for hemodialysis according to claim 1, characterized in that, In step S5, the fatigue deviation index is obtained by weighted summation of pre-dialysis stage deviation, mid-dialysis stage deviation and overall deviation measure, and is used to quantify the degree of fatigue abnormality caused by the current dialysis cycle process itself. The recovery delay index is obtained by weighted summation of post-dialysis stage shift, inter-dialysis stage shift, and the increment of post-dialysis stage shift relative to its historical average level, and is used to quantify the degree of delay in the fatigue recovery process. The physiological tolerance index is obtained by mapping the phase shift during dialysis to the [0,1] interval, and is used to quantify the patient's physiological tolerance to the current dialysis treatment; The full-cycle fatigue load index is obtained by weighted summation of the overall offset measure of the current cycle, the average overall offset measure within the recent historical window, and the offset trend term, and is used to quantify long-term cumulative fatigue load.

8. The method for constructing fatigue indicators and generating auxiliary decisions for hemodialysis according to claim 1, characterized in that, Step S6 includes: For each patient, a personalized threshold is dynamically calculated. The personalized threshold is the sum of the moving average and the moving standard deviation of the corresponding structured index over the patient's historical period. Calculate the decision confidence level corresponding to the indicator based on the mean of the observation data quality labels related to the structured indicator within the period and the data completeness of the current period; Based on the degree of standardization of the structured indicators deviating from their individualized thresholds, decision confidence, and preset clinical importance weights, the priority of each auxiliary decision suggestion to be generated is calculated and sorted from high to low priority. Based on individualized thresholds, decision confidence, and priority ranking results, auxiliary decision-making suggestions with priority ranking and confidence scores are generated. The types of auxiliary decision-making suggestions include nursing observation suggestions, recovery management suggestions, exercise suggestions, parameter review prompts, and early warning prompts.

9. The method for constructing fatigue indicators and generating auxiliary decisions for hemodialysis according to claim 1, characterized in that, It also includes a closed-loop feedback step: recording the implementation status of the decision support recommendations in the current cycle and the improvement status of the corresponding structured indicators in the next cycle, and updating the historical tolerance profile accordingly to adjust the confidence level or priority of subsequent similar recommendations.

10. A fatigue index construction and auxiliary decision generation system for hemodialysis, used to implement the method as described in any one of claims 1-9, characterized in that, include: The data access module is used to acquire multimodal observation data of a patient during a complete dialysis cycle from multiple data sources, and to add stage labels, event labels and data quality labels to the observation data in a way that retains the original timestamps, forming a sequence of observation tuples with time order; The event relationship modeling module is used to construct an event hypergraph and a time patch graph based on the observation tuple sequence. It establishes cross-modal associations within the event window through the event hypergraph, establishes dependencies at different time scales throughout the entire cycle through the time patch graph, and updates the feature representation of the observation tuple sequence through message passing. The fatigue state characterization and full-cycle aggregation module is used to aggregate the feature representation of the updated observation tuple sequence through a shared-private latent space separation mechanism and gated fusion, and then input it into a state space model with stage switching gating to generate the stage-level representation and global representation of the current cycle. The Individual Prototype Construction and Offset Quantization module is used to construct and update the individual prototype of a patient based on the stage-level and global representations of the patient's historical effective dialysis cycles, and to calculate the overall offset metric of the stage-level and global representations of the current cycle relative to the individual prototype and the stage offset metric of each stage. The structured fatigue index generation module is used to generate structured fatigue indices, including fatigue offset index, recovery delay index, physiological tolerance index and full-cycle fatigue load index, based on the overall offset metric and the stage offset metric. The decision support generation module is used to generate decision support suggestions with priority ranking and confidence scores based on structured fatigue indicators, combined with individualized thresholds, decision confidence and priority ranking rules dynamically calculated for each patient, and to receive closed-loop feedback information.

Images