Personalized planning and management system and method for gynecological postoperative rehabilitation path

Abstract

The invention relates to the technical field of medical information, in particular to a gynecological postoperative rehabilitation path personalized planning and management system and method.The gynecological postoperative rehabilitation path personalized planning and management method comprises the steps that complete-cycle clinical data of a patient receiving a gynecological operation are obtained, and a multi-dimensional time sequence state snapshot with an identifier is generated through recombination according to time and types; disassembling the snapshot into a physiological index group for defining a recovery stage, an action feature group for measuring activity ability and a physical sign fluctuation group for evaluating complication risk; taking the snapshots as indexes, recalling similar cases from the rehabilitation knowledge graph, and extracting rehabilitation task sequences and resource occupation records of the similar cases; and jointly inputting the information into a rehabilitation path planning model, and outputting a personalized rehabilitation task list with a time anchor point and a resource allocation instruction. According to the invention, continuous and multi-dimensional analysis of the rehabilitation state of the patient is realized, and a dynamic and accurate rehabilitation plan is generated according to the individual state and historical similar cases.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, and in particular to a personalized planning and management system and method for postoperative rehabilitation pathways in gynecology. Background Technology

[0002] Current gynecological postoperative rehabilitation management primarily relies on standardized clinical pathway guidelines and physicians' individual experience. Patient clinical data, including preoperative assessments, intraoperative records, and postoperative monitoring information, is scattered across different information systems with inconsistent data formats and recording standards. This data is fragmented in time and isolated in dimensions, lacking effective mechanisms for correlation and integration. This makes it impossible to construct a continuous and complete holistic view reflecting the patient's individualized recovery process, and the timing and content of rehabilitation interventions are difficult to precisely match with the patient's real-time dynamic changes.

[0003] In the rehabilitation assessment and planning stages, existing methods typically rely on monitoring a few core physiological indicators and threshold alarms, or on general subjective functional scoring. These methods fail to systematically isolate and extract specific feature sets serving different rehabilitation management goals from massive amounts of multi-source heterogeneous clinical data. The formulation of rehabilitation tasks and the allocation of medical resources lack data support based on multi-dimensional, fine-grained feature collaborative analysis, resulting in insufficient personalization of rehabilitation pathways and limited forward-looking risk warning capabilities.

[0004] This invention aims to address the problems of fragmented patient data throughout the entire life cycle and the inability to form a time-series panoramic status view in existing technologies, as well as the problem of single-dimensional rehabilitation assessment and the inability to support multi-objective collaborative and precise planning. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a personalized planning and management system and method for postoperative gynecological rehabilitation pathways.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for personalized planning and management of postoperative gynecological rehabilitation pathways, comprising: Acquire all clinical records and monitoring data generated before, during and after gynecological surgery for patients, reorganize the clinical records and monitoring data according to the collection time and data type, and generate a multidimensional time-series snapshot of the patient's status with a unique identifier; The patient's multidimensional time-series snapshots were discretized and decomposed to separate a group of physiological indicators for defining the recovery stage, a group of motor characteristics for measuring activity capacity, and a group of vital sign fluctuations for assessing the risk of complications. Using the patient's multidimensional temporal state snapshot as an index, recall past case databases from the rehabilitation knowledge graph whose similarity to the patient's multidimensional temporal state snapshot exceeds a preset threshold, and extract case rehabilitation task sequences and case rehabilitation resource usage records from the past case database; The patient's multidimensional temporal state snapshot, the case rehabilitation task sequence in the previous case library, and the case rehabilitation resource occupancy record are jointly input into the rehabilitation path planning model, and the output is a rehabilitation task list and resource allocation instructions containing time anchors.

[0007] As a further aspect of the present invention, the acquisition of all clinical records and monitoring data generated by patients undergoing gynecological surgery before, during, and after the surgery is further specifically implemented as follows: The hospital information platform interface was used to extract the patient's preoperative laboratory test results, imaging examination reports, gynecological examination records, preoperative nursing assessment forms, and anesthesia visit records. The surgical anesthesia system interface is used to obtain the patient's anesthesia induction and maintenance drug dosage records, fluid balance records, surgical procedure name and surgeon code, intraoperative event markers, and surgical duration during the operation. The system receives real-time data on the patient's dynamic vital signs, wound drainage characteristics and flow rate, visual analog scale (VAS) scores for pain, first flatus and feeding timestamps, and duration and number of steps for first ambulation in the postoperative recovery room and ward via bedside monitoring equipment and nursing handheld terminal. All clinical records and monitoring data are arranged with millisecond precision according to the time of data generation, and multiple indicators generated at the same time are packaged into a data package, ultimately generating a multidimensional time-series snapshot of the patient's status with the patient's hospital number as the index key and the timestamp as the sequence key.

[0008] As a further aspect of the present invention, the patient's multidimensional time-series snapshot is discretized and decomposed to separate a group of physiological indicators for defining the recovery stage, a group of motor characteristics for measuring activity capacity, and a group of vital sign fluctuations for assessing the risk of complications. This is specifically accomplished through the following sub-steps: Extract the peak and trough values ​​of the continuous body temperature curve, the dynamic change rate of white blood cell count, the absolute value of C-reactive protein, the hourly decreasing slope of postoperative drainage volume, and the rolling average value of the visual analog scale score for pain from the patient's multidimensional time-series snapshot, and package the parameter combination into a stage boundary determination feature package. Extract the number of seconds required to transition from a bedridden state to a sitting-up position, the number of seconds required to transition from a sitting-up position to a standing position at the bedside, the number of seconds required to transition from a standing position at the bedside to an indoor walking position, the duration of each walking session, and the self-rating score of fatigue after walking from the patient's multidimensional time-series snapshot, and package the parameters into a quantitative feature package of activity capacity. The extreme values ​​of the frequency domain index of heart rate variability, the amplitude of respiratory rate fluctuation, the cumulative duration of blood oxygen saturation below the threshold, and the frequency of dressing bleeding and exudation replacement are extracted from the patient's multidimensional time-series snapshot. The parameters are combined and packaged into a complication early warning feature package. The stage boundary determination feature package, the activity capacity quantification feature package, and the complication early warning feature package are stored as three types of feature vectors that are independent of each other but related through patient identifiers.

[0009] As a further aspect of the present invention, using the patient's multidimensional temporal state snapshot as an index, a database of past cases from the rehabilitation knowledge graph whose similarity to the patient's multidimensional temporal state snapshot exceeds a preset threshold is retrieved, specifically including: Calculate the Euclidean distance between the feature vector of the latest moment in the patient's multidimensional temporal state snapshot and the initial stage vector of each previous case in the rehabilitation knowledge graph, and select the case set with the smallest distance value as the initial screening case pool; Within the initial screening case pool, the patient's age range, surgical incision type code, and number of previous pregnancies are compared with the values ​​of the corresponding fields in previous cases. Cases with completely identical values ​​are assigned a higher-level similarity label. Retrieve complete rehabilitation path records corresponding to previous cases with higher-level similar tags. The rehabilitation path records include a list of rehabilitation action names performed daily by the case, the single execution duration and frequency of each rehabilitation action, the role category of the person in charge of the execution, and the specific model and time period of the equipment used.

[0010] As a further aspect of the present invention, the patient's multidimensional temporal state snapshot, the case rehabilitation task sequence in the previous case database, and the case rehabilitation resource occupancy record are jointly input into the rehabilitation path planning model, and the model outputs a rehabilitation task list containing time anchors and resource allocation instructions, which are executed in the following order: The patient's multidimensional time-series snapshot and the case rehabilitation task sequence are input into the task anchor point regression module. The task anchor point regression module sets the time anchor points for the patient's first ambulation, first pelvic floor muscle contraction training, and first attempt at voluntary urination, with minutes as the smallest unit, based on the types of tasks performed by the patient at the current time and the corresponding time in the previous cases. The case rehabilitation resource occupancy record, along with the bed location information, family caregiver code, and the hospital's current shift available nurse schedule from the patient's multidimensional time-series snapshot, are input into the resource occupancy prediction module. The resource occupancy prediction module calculates the specific time period during which the patient uses the corridor walker, lower limb pneumatic pump, and electrocardiogram monitor based on the start and end times of equipment use in the case rehabilitation resource occupancy record and the patient's activity ability quantitative feature package. The three time anchors are combined with the specific time period to form a task list with precise execution time. At the same time, the executor code, execution location address and required equipment number of each task are written into the task instruction as auxiliary fields.

[0011] As a further aspect of the present invention, the method further includes, after outputting the rehabilitation task list containing time anchors and resource allocation instructions, tracking the completion status of each task in the task list in real time: The system receives entry signals from the nurse's handheld terminal as it approaches the patient's bed area via Bluetooth beacon positioning, records the actual time when the nurse performs venous thrombosis prevention procedures, and calculates the deviation from the planned execution time in the task list to obtain the positive delay time or negative advance time. The angle change curve of the patient's lower limb lifting movement is collected by a wearable accelerometer. The number of peaks in the angle change curve is compared with the number of ankle pump repetitions required in the task list to obtain the task completion percentage. When the task completion percentage is lower than a preset completion threshold, or when the positive delay duration exceeds the allowed waiting time, a task replanning event is triggered, and the replanning event code is pushed to the rehabilitation path generation model.

[0012] As a further aspect of the present invention, the method further includes performing an incremental path correction process for the patient state that triggers the replanning event: After receiving the replanning event code, the rehabilitation path generation model automatically freezes all currently unexecuted task entries and uses only the latest five-minute data fragment generated in the patient's multidimensional time-series snapshot and the completed task entries in the task list as input to re-call the previous case library. However, at this time, it only recalls rehabilitation paths of failed cases that are similar to the risk indicator value distribution in the patient's current complication warning feature package. Extract the key nodes that cause task delays or resource conflicts in the recovery path of the failed cases, set the key nodes as avoidance objects, and modify the execution order and executor assignment roles of the remaining tasks in the task list based on the avoidance objects; The revised rehabilitation task instructions are output and updated synchronously with the nurses' handheld terminals via the nursing station's large screen.

[0013] As a further aspect of the present invention, the construction and updating process of the rehabilitation knowledge graph is independent of the invocation process of the rehabilitation path generation model: Complete medical records and nursing records of all gynecological postoperative patients within a specified time window are periodically extracted from the hospital's clinical data warehouse. Natural language segmentation and entity recognition are performed on the medical records and nursing records. The extracted entities include postoperative days, rehabilitation action vocabulary, medical device vocabulary, and personnel role vocabulary. The temporal, causal, and hierarchical relationships between the entities are modeled as directed edges in a knowledge graph, and weight coefficients are assigned to the directed edges. The weight coefficients of the directed edges are determined by the frequency of the co-occurrence of the entities in the medical records and nursing records. The knowledge graph is stored in a graph database instance, and a vectorized index interface is configured for the graph database instance so that the recall step of the similar past case library can be queried in a hybrid mode of graph traversal and vector retrieval.

[0014] As a further aspect of the present invention, the construction steps of the rehabilitation pathway planning model are specifically completed in the following manner: Historical rehabilitation datasets covering the entire year are extracted from the hospital information platform. The preoperative basic index vector, intraoperative anesthesia record vector, postoperative daily physiological index peak sequence, daily activity ability quantitative feature package sequence, complication warning feature package sequence, daily rehabilitation task name and execution time, and functional recovery score at discharge are aligned for each patient in the historical rehabilitation dataset to generate training sample units identified by the patient's hospital number. The training sample units are truncated according to the number of days after the patient's surgery. The length of the input window is fixed as a complete state snapshot sequence of seven days after the surgery. The patient's functional recovery score at the end of the seventh day is used as the prediction target label. The truncated training sample units are input into a task temporal prediction network based on a long short-term memory network. The task temporal prediction network extracts temporal dependency features from the input state snapshot sequence to generate a hidden state vector sequence. The hidden layer state vector sequence is input into a fully connected layer for dimensionality transformation, and then a linear regression output layer is used to generate the probability distribution of the rehabilitation task categories to be executed on that day and the corresponding execution time offset. Calculate the cross-entropy loss function value between the probability distribution of the rehabilitation task category and the actual executed task category in the historical rehabilitation dataset. Simultaneously, calculate the mean squared error loss function value between the execution time offset and the actual execution time offset. The two loss function values ​​are weighted and summed to obtain the total loss value. The weight parameters of the task encoding network and the fully connected layer are updated through the backpropagation algorithm until the total loss value converges. The converged model is then saved as a rehabilitation path planning model instance.

[0015] As a further aspect of the present invention, the present invention also includes a personalized planning and management system for post-gynecological postoperative rehabilitation pathways. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the personalized planning and management method for post-gynecological postoperative rehabilitation pathways described above.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: The patient's full-cycle clinical records and monitoring data are reorganized according to the collection time and data type to generate a multi-dimensional time-series snapshot with a unique identifier. This method constructs a unified structured data carrier, realizing the alignment and fusion of multi-source heterogeneous data from preoperative, intraoperative, and postoperative periods on a timeline. It eliminates the isolation of data in time and space, forming a complete, continuous, and machine-readable digital archive of the patient's individual recovery process, providing a high-quality data foundation with strict temporal correlation for subsequent analysis.

[0017] The multidimensional time-series snapshots are discretized and decomposed into physiological indicator groups, motor characteristic groups, and vital sign fluctuation groups. This decomposition allows for targeted screening and classification of the raw data based on different clinical goals. The physiological indicator group quantifies the recovery stage, the motor characteristic group represents the level of functional activity, and the vital sign fluctuation group identifies risk signals. This process transforms comprehensive state information into multiple feature subsets with clearly defined objectives and dimensions, enabling subsequent computational models to process quantitative information from different dimensions of rehabilitation concern in parallel.

[0018] Based on structured temporal state snapshots and decoupled feature groups, similar historical cases are retrieved from the rehabilitation knowledge graph, and their task sequences and resource records are extracted. The planning model then generates a task list and resource instructions with time anchors. This process ensures that the generation of the rehabilitation path directly relies on a multi-dimensional analysis of the patient's overall condition and matching with historically effective patterns. The output rehabilitation plan demonstrates high adaptability to the individual's real-time condition and past successful cases in terms of task content, execution timing, and resource allocation. Attached Figure Description

[0019] Figure 1 This is a flowchart of the personalized planning and management method for gynecological postoperative rehabilitation pathways described in this invention; Figure 2 A flowchart for generating a multidimensional temporal state snapshot of a patient; Figure 3 Flowchart for recalling similar past cases; Figure 4 Heat map of factors influencing postoperative recovery failure in gynecological surgery; Figure 5 Line graph showing the predictive performance of postoperative gynecological rehabilitation. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0021] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0022] See Figure 1 The system acquires all clinical records and monitoring data generated before, during, and after gynecological surgery. These records and data are then reorganized according to collection time and data type to generate a multidimensional temporal snapshot of the patient's state, each with a unique identifier. This snapshot is then discretized to separate physiological indicators for defining the recovery phase, motor characteristics for measuring activity levels, and vital sign fluctuations for assessing complication risk. Using this snapshot as an index, a database of past cases with similarity exceeding a preset threshold is retrieved from the rehabilitation knowledge graph. Case rehabilitation task sequences and resource usage records are extracted from this database. Finally, the patient's multidimensional temporal snapshot, the case rehabilitation task sequences from the database, and the resource usage records are input into a rehabilitation pathway planning model, which outputs a rehabilitation task list and resource allocation instructions with time anchors.

[0023] See Figure 2In one embodiment of the present invention, when implementing the personalized planning and management method for postoperative gynecological rehabilitation pathways, the data acquisition and feature decomposition process is performed according to the following procedure: Preoperative laboratory test results, imaging reports, gynecological examination records, preoperative nursing assessment forms, and anesthesia visit records are extracted through the hospital information platform interface. These records constitute the patient's baseline medical file. Records of anesthesia induction and maintenance drug dosages, fluid balance records, surgical procedure name and surgeon code, intraoperative event markers, and surgical duration are obtained through the surgical anesthesia system interface to accurately characterize the specific details of the surgical intervention. The system continuously monitors the postoperative recovery process by receiving real-time dynamic vital signs, wound drainage characteristics and flow rates, self-reported pain visual analog scale (VAS) scores, first flatus and feeding timestamps, and duration and number of steps for the first ambulation from bedside monitoring devices and nursing handheld terminals. All clinical records and monitoring data are arranged in a time series with millisecond precision according to the time of data generation, and multiple indicators generated at the same millisecond time are packaged into an independent data package. Finally, a multi-dimensional time-series snapshot of the patient's status is generated with the patient's hospital number as the global index key and the timestamp as the internal sequence key.

[0024] In some embodiments, the patient's multidimensional temporal state snapshot is discretized and decomposed. From the patient's multidimensional temporal state snapshot, continuous peak and trough values ​​of body temperature curve, dynamic change rate of white blood cell count, absolute value of C-reactive protein, hourly decreasing slope of postoperative drainage volume, and rolling average of visual analog scale (VAS) pain scores are extracted and packaged into a stage boundary determination feature package. From the patient's multidimensional temporal state snapshot, the time required to transition from a bedridden state to a sitting-up position, the time required to transition from a sitting-up position to a bedside standing position, the time required to transition from a bedside standing position to an indoor walking position, the duration of each walking session, and the self-rating value of fatigue after walking are extracted and packaged into an activity capacity quantification feature package. From the patient's multidimensional temporal state snapshot, the extreme values ​​of heart rate variability frequency domain indicators, the amplitude of respiratory rate fluctuations, the cumulative duration of blood oxygen saturation below a preset threshold, and the frequency of dressing bleeding and exudation replacement are extracted and packaged into a complication warning feature package.

[0025] It can be understood that the stage boundary determination feature package, the activity capacity quantification feature package, and the complication early warning feature package are stored as three independent but linked feature vectors through patient identifiers. In specific implementation, further processing of the activity capacity quantification feature package can introduce a quantitative scoring function to integrate multiple action transition time parameters into a comprehensive activity capacity score. This scoring function is expressed as: in: This represents the normalized activity ability score. Indicates the first The number of seconds recorded for each action transition process. These are preset weighting coefficients for corresponding movement transitions, reflecting the relative importance of the movement in the overall rehabilitation process. This indicates the total number of action transitions in the current action sequence.

[0026] See Figure 3 In one embodiment of the present invention, the process of recalling a database of past cases similar to the patient's current state from the rehabilitation knowledge graph, based on the personalized planning and management method for post-gynecological rehabilitation pathways, is as follows: The Euclidean distance between the feature vector of the latest moment in the patient's multidimensional temporal state snapshot and the initial stage vector of each past case in the rehabilitation knowledge graph is calculated. This distance value represents the overall state difference between the current patient and historical cases at the start of the disease course. All calculated Euclidean distances are sorted in ascending order, and the top K cases with the smallest distance values ​​are selected to form a set. This set is defined as the initial screening case pool, where K is a preset positive integer.

[0027] In some embodiments, more refined similarity screening is performed within the initial screening case pool, comparing the patient's age range, surgical incision type code, and number of previous pregnancies with the values ​​of the corresponding fields in each previous case in the initial screening case pool. The age range refers to mapping the patient's age to a predefined discrete range such as "20-30 years old" or "31-40 years old"; the surgical incision type code is a standardized code used in the hospital information system to identify the surgical approach, for example, "01" represents laparoscopic surgery and "02" represents open surgery; the number of previous pregnancies is recorded as an integer. For cases where the patient's age range, surgical incision type code, and number of previous pregnancies are completely consistent with the values ​​of the corresponding fields in historical cases, the system assigns a higher-level similarity label.

[0028] Assigning higher-level similarity labels is understandable; it's a classification operation designed to further differentiate from the initial case pool those cases that perfectly match the current patient's demographic and surgical baseline characteristics. The complete rehabilitation pathway records corresponding to previous cases with higher-level similarity labels are retrieved from the graph database. These records are structured data objects whose fields include a list of rehabilitation actions performed daily during postoperative recovery, the duration and frequency of each action, the role category of the personnel responsible for performing each action, and the specific model and time period of the equipment used during the rehabilitation process.

[0029] In practice, calculating the Euclidean distance is a core step in case recall. Its calculation formula is defined as: in: This represents the Euclidean distance between the current patient feature vector and the feature vector of a historical case. This represents the feature vector of the latest moment in the patient's multidimensional temporal state snapshot. This represents the initial stage feature vector of a past case in the rehabilitation knowledge graph. The index represents the total dimension of the feature vector. The vector represents the first Each dimension has a component. The calculation process traverses the initial vectors of all registered past cases in the knowledge graph to obtain a set of distance values. Optionally, a preset distance threshold is used for initial screening; if the Euclidean distance of a case in the rehabilitation knowledge graph is... If the value is less than the preset threshold, the case will be included in the initial screening case pool; otherwise, it will be excluded. When comparing the consistency of field values, if a patient's feature information is missing, the system will consider it as not matching any corresponding field value in any historical case, and that historical case will not be assigned a higher-level similarity label.

[0030] In one embodiment of the present invention, the output process of the rehabilitation pathway planning model and the model construction method work together according to the following process: The patient's multidimensional temporal state snapshot and the case rehabilitation task sequence are input into the task anchor point regression module. The task anchor point regression module sets the time anchor points for the patient's first ambulation, first pelvic floor muscle contraction training, and first attempt at voluntary urination, with minutes as the smallest unit, based on the types of tasks performed by the patient at the corresponding time in previous cases. The case rehabilitation resource occupancy record and the bed location information, family caregiver code, and the hospital's current shift available nurse schedule in the patient's multidimensional temporal state snapshot are input into the resource occupancy prediction module. The resource occupancy prediction module calculates the specific time period for the patient to use the corridor walker, lower limb pneumatic pump, and electrocardiogram monitor based on the start and end times of equipment use in the case rehabilitation resource occupancy record and the patient's activity ability quantitative feature package. The three time anchor points and specific time periods are combined to form a task list with precise execution times. At the same time, the executor code, execution location address, and required equipment number of each task are written into the task instruction as auxiliary fields.

[0031] In some embodiments, the construction steps of the rehabilitation pathway planning model are completed in the following manner: Historical rehabilitation datasets covering patients undergoing gynecological surgery throughout the year are extracted from the hospital information platform. The preoperative baseline index vector, intraoperative anesthesia record vector, postoperative daily physiological index peak sequence, daily activity capacity quantitative feature package sequence, complication warning feature package sequence, daily rehabilitation task name and execution time, and discharge functional recovery score of each patient in the historical rehabilitation dataset are aligned to generate training sample units identified by the patient's hospital number. The training sample units are truncated according to the number of days after surgery, and the input window length is fixed as a complete state snapshot sequence for seven days post-surgery, using the patient's functional recovery score at the end of the seventh day as the prediction target label. The truncated training sample units are input into a task temporal prediction network constructed based on a long short-term memory network. The task temporal prediction network extracts temporal dependency features from the input state snapshot sequence to generate a hidden layer state vector sequence. The hidden layer state vector sequence is input into a fully connected layer for dimensionality transformation, and then a linear regression output layer is used to generate the probability distribution of the rehabilitation task category to be executed that day and the corresponding execution time offset.

[0032] Understandably, the model training process involves calculating the loss function and updating parameters. It calculates the cross-entropy loss function value between the probability distribution of rehabilitation task categories and the actual task categories executed in the historical rehabilitation dataset, and simultaneously calculates the mean squared error loss function value between the execution time offset and the actual execution time offset. The weighted sum of these two loss function values ​​is then used as the total loss value. The weight parameters of the task temporal prediction network and the fully connected layer are updated using the backpropagation algorithm until the total loss value converges. The converged model is then saved as a rehabilitation path planning model instance. In specific implementation, the total loss value... The calculation is defined by the following formula: in: This represents the total loss value during the training process. Represents the cross-entropy loss function. This represents the probability distribution of rehabilitation task categories predicted by the model. Labels indicating actual rehabilitation task categories, This represents the mean squared error loss function. This represents the offset of the execution time predicted by the model. This represents the actual execution time offset. and It is a preset weighting coefficient used to balance the contributions of the two losses.

[0033] Optionally, when calculating time anchor points, the task anchor point regression module will refer to the average postoperative time of the corresponding task in the rehabilitation task sequence of previous cases, and make fine adjustments based on the deviation between the latest physiological indicators and the average value in the patient's multidimensional time-series snapshot. Optionally, when estimating equipment usage time periods, the resource occupancy prediction module will not only rely on the case rehabilitation resource occupancy records, but also check the skill tags and equipment operation qualifications of nurses in the hospital's current shift available nurse schedule to ensure that the assigned executor has the corresponding capabilities.

[0034] In one embodiment of the present invention, the task completion status tracking and path correction process is initiated after outputting a rehabilitation task list containing time anchors and resource allocation instructions, and tracks the completion status of each task in the task list in real time; it receives the entry signal of the nurse's handheld terminal approaching the patient's bed area through a Bluetooth beacon positioning system, records the actual time when the nurse performs venous thrombosis prevention operations, and calculates the deviation with the planned execution time in the task list to obtain the positive delay time or negative advance time; it collects the angle change curve of the patient's lower limb lifting movement through a wearable accelerometer, compares the number of peaks of the angle change curve with the number of ankle pump repetitions required in the task list to obtain the task completion percentage value; when the task completion percentage value is lower than the preset completion threshold, or the positive delay time exceeds the allowed waiting time, a task replanning event is triggered, and the replanning event code is pushed to the rehabilitation path planning model.

[0035] In some embodiments, for the patient state that triggers a replanning event, an incremental path correction process is performed. After receiving the replanning event code, the rehabilitation path planning model automatically freezes all currently unexecuted task entries. It only uses the latest five-minute data fragment generated in the patient's multidimensional time-series state snapshot and the completed task entries in the task list as input to re-call the previous case library. However, at this time, it only recalls rehabilitation paths of failed cases that are similar to the risk indicator value distribution in the patient's current complication warning feature package. It extracts the key nodes in the failed case rehabilitation paths that cause task delays or resource conflicts, sets the key nodes as avoidance objects, and modifies the execution order and executor assignment roles of the remaining tasks in the task list based on the avoidance objects. It outputs the corrected rehabilitation task instructions and updates them synchronously with the nurse's handheld terminal through the nursing station's large screen.

[0036] As is understandable, deviation calculation and completion assessment are the core of status tracking, and the system maintains a status record for each task. See Table 1 for an example structure of a task execution status tracking table.

[0037] Table 1: Patient Task Performance Status Tracking Table In practical implementation, the judgment logic for triggering the task replanning event is implemented by a comprehensive conditional function, which is formally defined as: in: It is a Boolean function that triggers a replanning event when its value is true; This represents the percentage of task completion calculated in real time. This indicates a pre-set completion threshold; This represents the calculated positive delay duration; Indicates the allowed waiting time; symbol This represents a logical "OR" operation. This formula ensures that the system initiates a correction process if either the task completion quality is insufficient or the execution delay is excessive.

[0038] See Figure 4 This is a heatmap showing the influencing factors of postoperative gynecological rehabilitation failure cases. It visually demonstrates the correlation between different stages of postoperative gynecological rehabilitation and five categories of influencing factors leading to pathway replanning failure. The darker the color, the more cases there are. The first 48 hours after surgery are the peak period for failure cases, especially "low patient cooperation" (9 cases) and "task delay" (8 cases), which are the most prominent factors and the focus of pathway modification. "Low patient cooperation" is the most significant cause of failure at all stages, especially at 12 and 48 hours after surgery, indicating the need to strengthen patient education and psychological support. Patient education and task execution monitoring within the first 48 hours after surgery are the core intervention links. To address the high incidence of "task delay," nurse scheduling and rehabilitation equipment allocation are optimized.

[0039] In one embodiment of the present invention, the construction and updating process of the rehabilitation knowledge graph is independent of the invocation process of the rehabilitation path planning model, and its execution is not coupled with real-time path planning requests. Complete medical records and nursing records of all gynecological postoperative patients within a specified time window are periodically extracted from the hospital's clinical data warehouse. This specified time window can be a quarter or half a year to ensure the timeliness and coverage of the knowledge. Natural language segmentation and entity recognition are performed on the extracted medical records and nursing records. Segmentation breaks down long records into semantically independent short sentences or paragraphs based on punctuation marks such as line breaks and periods in the document. Entity recognition uses a pre-trained biomedical named entity recognition model to extract postoperative days, rehabilitation action vocabulary, medical device vocabulary, and personnel role vocabulary from the segmented paragraphs.

[0040] In some embodiments, constructing a knowledge graph requires establishing relationships between entities. The identified temporal, causal, and executive membership relationships between entities are modeled as directed edges in the knowledge graph. Temporal relationships describe the sequential order of two entities in time, such as the relationship between "day 1 post-surgery" and "ankle pump exercises." Causal relationships describe how one entity's state or event leads to another entity's state or event, such as the relationship between "high pain score" and "suspending ambulation." Executive membership relationships describe the association between an action and its executor or tool, such as the relationship between "dressing change" and "responsible nurse" or "sterile dressing pack." Each directed edge in the knowledge graph is assigned a weight coefficient, determined by the frequency of co-occurrence of related entities in the medical records and nursing records. A higher frequency of co-occurrence indicates a more common relationship in historical practice, resulting in a larger weight coefficient.

[0041] It is understandable that the calculation of the weighting coefficient is key to quantifying the strength of the relationship, and its specific calculation formula can be defined as: in: Representing entities in a knowledge graph Pointing to entity The weight coefficients of the directed edges, Indicates entities in all medical and nursing records With entity The frequency of occurrence within the same segmented paragraph Representing entities Total frequency of occurrence across all records Representing entities The total frequency of occurrence in all records, function It is a normalization calculation function used to convert co-occurrence frequency and individual frequency into a standardized weight value.

[0042] In practical implementation, the knowledge graph containing entity nodes, directed edges, and weight coefficients is stored in a graph database instance, and a vectorized indexing interface is configured for the graph database instance. The vectorized indexing interface is implemented by mapping entity nodes and relationships in the knowledge graph to a low-dimensional vector space. This allows the recall step of the similar past case library to query using a hybrid mode of graph traversal and vector retrieval. Graph traversal is used to find entity nodes with the same attributes as the current patient based on relationship paths, while vector retrieval is used to quickly find the set of nodes in the vector space that are closest to the patient's current state vector. Optionally, the periodic execution of the construction and update process is triggered by a background scheduling task, for example, set to automatically execute once every Sunday morning to incorporate the latest patient recovery data. Optionally, the vectorized indexing interface configured for the graph database instance can generate vectors based on the structural information of the knowledge graph, encoding nodes and relationships into continuous vectors through graph embedding technology, so that semantically similar entities are located close to each other in the vector space.

[0043] See Figure 5 This is a line graph showing the predictive performance of post-gynecological surgery rehabilitation, visually illustrating the changing trends of the rehabilitation pathway planning model's predictive accuracy and functional recovery score within 7 days post-surgery. The score steadily increased from 0.75 on day 1 to 0.97 on day 7, indicating that the model's grasp of the patient's condition became increasingly accurate as the rehabilitation progressed. The two curves representing the functional recovery score almost overlapped, demonstrating the model's highly reliable prediction of the patient's functional recovery. In the first 3 days post-surgery, there was a certain gap between the predicted accuracy and the actual value of the functional recovery score, due to the large fluctuations in the patient's condition and the limited observable data in the early stages. From day 4 onwards, the predictive accuracy exceeded 0.9, and the two functional recovery curves almost overlapped, indicating that the model entered a stable and reliable working state. Overall, the predictive accuracy is high, especially in the mid-to-late post-operative period, making it a reliable reference for clinical decision-making.

[0044] The above are merely preferred embodiments of the present invention and are not intended to limit the present 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 that can be applied to other fields. 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 personalized planning and management method for post-gynecological surgery rehabilitation pathways, characterized in that, Includes the following steps: Acquire all clinical records and monitoring data generated before, during and after gynecological surgery for patients, reorganize the clinical records and monitoring data according to the collection time and data type, and generate a multidimensional time-series snapshot of the patient's status with a unique identifier; The patient's multidimensional time-series snapshots were discretized and decomposed to separate a group of physiological indicators for defining the recovery stage, a group of motor characteristics for measuring activity capacity, and a group of vital sign fluctuations for assessing the risk of complications. Using the patient's multidimensional temporal state snapshot as an index, recall past case databases from the rehabilitation knowledge graph whose similarity to the patient's multidimensional temporal state snapshot exceeds a preset threshold, and extract case rehabilitation task sequences and case rehabilitation resource usage records from the past case database; The patient's multidimensional temporal state snapshot, the case rehabilitation task sequence in the previous case library, and the case rehabilitation resource occupancy record are jointly input into the rehabilitation path planning model, and the output is a rehabilitation task list and resource allocation instructions containing time anchors.

2. The method for personalized planning and management of gynecological postoperative rehabilitation pathways according to claim 1, characterized in that, The acquisition of all clinical records and monitoring data generated before, during, and after gynecological surgery is further specifically implemented as follows: The hospital information platform interface was used to extract the preoperative laboratory test results, imaging examination reports, gynecological examination records, preoperative nursing assessment forms, and anesthesia visit records of the patients who underwent gynecological surgery. The surgical anesthesia system interface is used to obtain the patient's anesthesia induction and maintenance drug dosage, fluid balance, surgical procedure name and surgeon code, intraoperative event markers, and surgical duration during the gynecological surgery. The system uses bedside monitoring equipment and nursing handheld terminal to receive in real time the dynamic vital signs, wound drainage characteristics and flow rate, pain visual analog scale self-reported values, first flatus and feeding timestamps, and first ambulation duration and number of steps of the patient undergoing gynecological surgery in the postoperative recovery room and ward. All clinical records and monitoring data are arranged with millisecond precision according to the time of data generation, and multiple indicators generated at the same time are packaged into a data package, ultimately generating a multidimensional time-series snapshot of the patient's status with the patient's hospital number as the index key and the timestamp as the sequence key.

3. The method for personalized planning and management of gynecological postoperative rehabilitation pathways according to claim 2, characterized in that, The patient's multidimensional time-series snapshots are discretized and decomposed to separate a group of physiological indicators for defining the recovery phase, a group of motor characteristics for measuring activity capacity, and a group of vital sign fluctuations for assessing the risk of complications. This is accomplished through the following sub-steps: Extract the peak and trough values ​​of the continuous body temperature curve, the dynamic change rate of white blood cell count, the absolute value of C-reactive protein, the hourly decreasing slope of postoperative drainage volume, and the rolling average value of the visual analog scale score for pain from the patient's multidimensional time-series snapshot, and package the parameter combination into a stage boundary determination feature package. Extract the number of seconds required to transition from a bedridden state to a sitting-up position, the number of seconds required to transition from a sitting-up position to a standing position at the bedside, the number of seconds required to transition from a standing position at the bedside to an indoor walking position, the duration of each walking session, and the self-rating score of fatigue after walking from the patient's multidimensional time-series snapshot, and package the parameters into a quantitative feature package of activity capacity. The extreme values ​​of the frequency domain index of heart rate variability, the amplitude of respiratory rate fluctuation, the cumulative duration of blood oxygen saturation below the threshold, and the frequency of dressing bleeding and exudation replacement are extracted from the patient's multidimensional time-series snapshot. The parameters are combined and packaged into a complication early warning feature package. The stage boundary determination feature package, the activity capacity quantification feature package, and the complication early warning feature package are stored as three types of feature vectors that are independent of each other but related through patient identifiers.

4. The method for personalized planning and management of post-gynecological postoperative rehabilitation pathways according to claim 3, characterized in that, Using the patient's multidimensional temporal state snapshot as an index, a database of past cases with a similarity exceeding a preset threshold to the patient's multidimensional temporal state snapshot is retrieved from the rehabilitation knowledge graph. Specifically, this includes: Calculate the Euclidean distance between the feature vector of the latest moment in the patient's multidimensional temporal state snapshot and the initial stage vector of each previous case in the rehabilitation knowledge graph, and select the case set with the smallest distance value as the initial screening case pool; Within the initial screening case pool, the age range, surgical incision type code, and number of previous pregnancies of the patients undergoing gynecological surgery are compared with the values ​​of the corresponding fields in previous cases. Cases with completely identical values ​​are assigned a higher-level similarity label. Retrieve complete rehabilitation path records corresponding to previous cases with higher-level similar tags. The rehabilitation path records include a list of rehabilitation action names performed daily by the case, the single execution duration and frequency of each rehabilitation action, the role category of the person in charge of the execution, and the specific model and time period of the equipment used.

5. The method for personalized planning and management of gynecological postoperative rehabilitation pathways according to claim 4, characterized in that, The patient's multidimensional temporal state snapshot, the case rehabilitation task sequence in the past case database, and the case rehabilitation resource usage record are jointly input into the rehabilitation pathway planning model. The model outputs a rehabilitation task list and resource allocation instructions containing time anchors, which are executed in the following order: The patient's multidimensional time-series snapshot and the case rehabilitation task sequence are input into the task anchor regression module. The task anchor regression module sets the time anchor points for the patient undergoing gynecological surgery, the first time anchor point for getting out of bed, the first time anchor point for pelvic floor muscle contraction training, and the first time anchor point for attempting voluntary urination, with minutes as the smallest unit, based on the type of task performed by the patient at the current time and the corresponding time in the previous cases. The case rehabilitation resource occupancy record, along with the bed location information, family member caregiver code, and the hospital's current shift available nurse schedule from the patient's multidimensional time-series snapshot, are input into the resource occupancy prediction module. The resource occupancy prediction module calculates the specific time period during which the patient undergoing gynecological surgery uses the corridor walker, lower limb pneumatic pump, and electrocardiogram monitor based on the start and end times of equipment use in the case rehabilitation resource occupancy record and the activity ability quantitative feature package of the patient. The three time anchors are combined with the specific time period to form a task list with precise execution time. At the same time, the executor code, execution location address and required equipment number of each task are written into the task instruction as auxiliary fields.

6. The method for personalized planning and management of gynecological postoperative rehabilitation pathways according to claim 5, characterized in that, The method further includes, after outputting the rehabilitation task list and resource allocation instructions containing time anchors, tracking the completion status of each task in the task list in real time: The system receives entry signals from the nurse's handheld terminal as it approaches the patient's bed area via Bluetooth beacon positioning, records the actual time when the nurse performs venous thrombosis prevention procedures, and calculates the deviation from the planned execution time in the task list to obtain the positive delay time or negative advance time. The angle change curve of the patient's lower limb lifting movement is collected by a wearable accelerometer. The number of peaks in the angle change curve is compared with the number of ankle pump repetitions required in the task list to obtain the task completion percentage. When the task completion percentage is lower than a preset completion threshold, or when the positive delay duration exceeds the allowed waiting time, a task replanning event is triggered, and the replanning event code is pushed to the rehabilitation path generation model.

7. The method for personalized planning and management of gynecological postoperative rehabilitation pathways according to claim 6, characterized in that, The method also includes performing an incremental path correction process for the patient state that triggered the replanning event: After receiving the replanning event code, the rehabilitation path generation model automatically freezes all currently unexecuted task entries and uses only the latest five-minute data fragment generated in the patient's multidimensional time-series snapshot and the completed task entries in the task list as input to re-call the previous case library. However, at this time, it only recalls rehabilitation paths of failed cases that are similar to the risk index value distribution in the current complication warning feature package of the patient who underwent gynecological surgery. Extract the key nodes that cause task delays or resource conflicts in the recovery path of the failed cases, set the key nodes as avoidance objects, and modify the execution order and executor assignment roles of the remaining tasks in the task list based on the avoidance objects; The revised rehabilitation task instructions are output and updated synchronously with the nurses' handheld terminals via the nursing station's large screen.

8. The method for personalized planning and management of gynecological postoperative rehabilitation pathways according to claim 7, characterized in that, The construction and updating process of the rehabilitation knowledge graph is independent of the invocation process of the rehabilitation path generation model: Complete medical records and nursing records of all gynecological postoperative patients within a specified time window are periodically extracted from the hospital's clinical data warehouse. Natural language segmentation and entity recognition are performed on the medical records and nursing records. The extracted entities include postoperative days, rehabilitation action vocabulary, medical device vocabulary, and personnel role vocabulary. The temporal, causal, and hierarchical relationships between the entities are modeled as directed edges in a knowledge graph, and weight coefficients are assigned to the directed edges. The weight coefficients of the directed edges are determined by the frequency of the co-occurrence of the entities in the medical records and nursing records. The knowledge graph is stored in a graph database instance, and a vectorized index interface is configured for the graph database instance so that the recall step of the similar past case library can be queried in a hybrid mode of graph traversal and vector retrieval.

9. The method for personalized planning and management of gynecological postoperative rehabilitation pathways according to claim 8, characterized in that, The construction steps of the rehabilitation pathway planning model are specifically completed in the following ways: Historical rehabilitation datasets covering gynecological surgery patients throughout the year are extracted from the hospital information platform. The preoperative basic index vector, intraoperative anesthesia record vector, postoperative daily physiological index peak sequence, daily activity ability quantitative feature package sequence, complication warning feature package sequence, daily rehabilitation task name and execution time, and functional recovery score at discharge are aligned for each patient in the historical rehabilitation dataset to generate training sample units identified by the patient's hospital number. The training sample units are truncated according to the number of days after the patient's surgery. The length of the input window is fixed as a complete state snapshot sequence of seven days after the surgery. The patient's functional recovery score at the end of the seventh day is used as the prediction target label. The truncated training sample units are input into a task temporal prediction network based on a long short-term memory network. The task temporal prediction network extracts temporal dependency features from the input state snapshot sequence to generate a hidden state vector sequence. The hidden layer state vector sequence is input into a fully connected layer for dimensionality transformation, and then a linear regression output layer is used to generate the probability distribution of the rehabilitation task categories to be executed on that day and the corresponding execution time offset. Calculate the cross-entropy loss function value between the probability distribution of the rehabilitation task category and the actual executed task category in the historical rehabilitation dataset. Simultaneously, calculate the mean squared error loss function value between the execution time offset and the actual execution time offset. The two loss function values ​​are weighted and summed to obtain the total loss value. The weight parameters of the task encoding network and the fully connected layer are updated through the backpropagation algorithm until the total loss value converges. The converged model is then saved as a rehabilitation path planning model instance.

10. A personalized planning and management system for post-gynecological surgery rehabilitation pathways, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the personalized planning and management method for gynecological postoperative rehabilitation pathways as described in any one of claims 1 to 9.

Images