A patient real-time remote monitoring and rehabilitation monitoring system

Abstract

This invention discloses a real-time remote patient monitoring and rehabilitation monitoring system, relating to the field of remote medical monitoring technology. The system includes cleaning and organizing collected multi-source sensor data. The organized data is imported into a virtual task scenario model for simulation, generating simulation records and performing compliance checks based on a medical knowledge graph, marking actions and physiological deviations. By fusing marked deviations with the patient's subjective feelings, a multi-dimensional state snapshot sequence is generated, and a personal rehabilitation state transition network is constructed. Based on the network's connection strength and state dwell time, the stability and risk probability of the rehabilitation path are quantified, rehabilitation stages are automatically divided, and suggestions are output. The system also integrates location monitoring functionality; when an abnormal risk exceeds a threshold, it can generate a risk alarm based on the patient's location information and send it to the monitoring provider.

Description

Technical Field

[0001] This invention belongs to the field of remote medical monitoring technology, specifically a system for real-time remote monitoring and rehabilitation of patients. Background Technology

[0002] Existing remote monitoring and rehabilitation monitoring technologies primarily collect patients' physiological and motor data directly through sensors and utilize algorithms for threshold comparisons or trend analysis. These technologies lack scenario-based simulations of the rehabilitation task execution process, and assessments rely on comparing raw data with simple models, failing to reflect a patient's ability to complete specific tasks in a controlled environment. Raw data is susceptible to environmental noise interference, and the diversity of rehabilitation tasks makes it difficult to standardize assessment criteria, resulting in discrepancies between assessment results and actual rehabilitation needs.

[0003] Existing technical solutions typically analyze data at discrete time points or single-dimensional trends, failing to integrate multimodal information such as movement deviations, physiological fluctuations, and patient subjective feelings. The dynamic evolutionary characteristics of the rehabilitation process are ignored, and assessments are limited to static snapshots, making it difficult to identify personalized patterns of state transitions. This results in the inability to quantify the stability of rehabilitation pathways, reliance on experience-based judgments for potential risk warnings, and a lack of objective evidence based on historical evolutionary data.

[0004] Furthermore, existing remote monitoring solutions often lack real-time location tracking with patients after identifying potential risks, making it difficult to provide monitoring personnel with the critical location information needed for timely intervention and affecting the efficiency of emergency response. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art;

[0006] Therefore, this invention proposes a real-time remote patient monitoring and rehabilitation monitoring system, comprising:

[0007] The data preprocessing module collects multi-source heterogeneous sensor data streams from patients, performs hierarchical cleaning operations on the multi-source heterogeneous sensor data streams, removes noise segments and fills in missing data segments, and outputs a normalized sensor data stream.

[0008] The task simulation module constructs a virtual task scene model associated with the patient's rehabilitation task, imports the normalized sensor data stream into the virtual task scene model for simulation and deduction, and generates a simulation record of the patient's movement trajectory and physiological response during the execution of the simulated task.

[0009] The compliance analysis module performs compliance verification on the simulated action trajectories and physiological responses based on a medical knowledge graph, marking action units and physiological fluctuation ranges that deviate from the standard execution paradigm.

[0010] The evolutionary analysis module integrates the labeled action units, physiological fluctuation ranges, and the subjective feelings reported by the patient in real time to generate a multidimensional rehabilitation state snapshot. Based on a time sliding window, it analyzes the evolution sequence of the multidimensional rehabilitation state snapshot and constructs a patient's personal rehabilitation state transfer network.

[0011] The comprehensive assessment module calculates the rehabilitation path stability index and the probability of abnormal state risk based on the node connection strength and state dwell time in the patient's personal rehabilitation state transfer network. It also automatically divides the current rehabilitation stage using the rehabilitation path stability index and the probability of abnormal state risk, and outputs a set of monitoring and training suggestions, in conjunction with preset staged rehabilitation milestones.

[0012] The positioning monitoring module receives the abnormal state risk probability and monitoring and training suggestion set output by the comprehensive evaluation module. When the abnormal state risk probability exceeds a preset threshold, it generates a risk alarm containing the patient's location information based on the high-risk action units or physiological parameters marked in the monitoring and training suggestion set, combined with the real-time location data of the positioning beacon worn by the patient, and sends it to the monitoring terminal.

[0013] Furthermore, the hierarchical cleaning operation performed on the multi-source heterogeneous sensor data stream includes:

[0014] The multi-source heterogeneous sensor data stream includes electromyographic signal stream, joint angle sequence and ambient temperature and humidity readings;

[0015] By setting a dynamic filtering threshold based on signal quality, the electromyography signal stream is adaptively filtered to separate the effective electromyography activity signal from the power frequency interference signal.

[0016] The continuity of the joint angle sequence is checked, and the angle jump points caused by signal loss are repaired by the interpolation algorithm to generate a smooth joint angle sequence.

[0017] By comparing the temporal logic of environmental temperature and humidity readings, abnormal readings that clearly exceed the reasonable range of physiological environment are eliminated, and the average of the data before and after the time are used for replacement.

[0018] The processed effective electromyographic activity signals, smoothed joint angle sequences, and corrected ambient temperature and humidity readings are re-aligned and packaged according to the original time axis to form the normalized sensor data stream.

[0019] Furthermore, the construction of the virtual task scenario model associated with the patient's rehabilitation task includes:

[0020] Define the core action elements of rehabilitation tasks, the environmental objects for task execution, and the expected task completion standards;

[0021] A virtual scene containing the environmental objects is constructed using 3D modeling technology, and a biomechanical skeletal and muscular model is bound to the patient's virtual avatar in the virtual scene;

[0022] A series of triggering events are preset in the virtual task scenario model, and the triggering events correspond to the key nodes or potential risk actions of task execution.

[0023] Furthermore, the step of importing the normalized sensor data stream into the virtual task scenario model for simulation includes:

[0024] The smooth joint angle sequence in the normalized sensor data stream is used to drive the skeletal model of the patient avatar to calculate and generate a continuous sequence of motion postures.

[0025] The effective electromyographic activity signals in the normalized sensor data stream are input into the muscle model of the patient's virtual avatar to simulate and calculate the muscle activation level and fatigue level.

[0026] Based on the aforementioned action posture sequence and muscle activation level, combined with environmental objects in the virtual scene, physical collision detection and interaction simulation are performed to determine the success or failure status of task execution.

[0027] The sequence of movements and postures, muscle activation levels, occurrence of triggering events, and task execution status throughout the entire simulation process are recorded and summarized into the simulation record of movement trajectory and physiological response.

[0028] Furthermore, the compliance verification of the motion trajectory and physiological response simulation records based on the medical knowledge graph includes:

[0029] Access to a medical knowledge graph that includes a standard rehabilitation exercise library and safe physiological parameter ranges;

[0030] The motion trajectory and the motion posture sequence in the physiological response simulation record are compared frame by frame with the standard rehabilitation motion library to identify the start and end frames of the motion with deviations in kinematic parameters and define them as a deviated motion unit.

[0031] The physiological response data in the motion trajectory and physiological response simulation record are compared with the safe physiological parameter range, and the time interval that continuously exceeds the range is marked as the physiological fluctuation interval.

[0032] Associate each deviating action unit with its deviation type, and associate each physiological fluctuation range with the type of physiological parameter that exceeds the limit.

[0033] Furthermore, the fusion of the tagged action units, physiological fluctuation ranges, and the patient's real-time reported subjective feelings includes:

[0034] Natural language processing was performed on the subjective feelings text input by the patient through the terminal to extract keywords and quantify the discomfort intensity level and emotional tendency value represented by them.

[0035] Establish a time alignment index to associate the keywords and their quantified values ​​with the deviation action units and physiological fluctuation ranges that occurred within the same time period;

[0036] Create a structured record table indexed by time, synchronously recording whether there are deviations in action units, physiological fluctuation ranges, and corresponding subjective feeling keywords and quantitative values ​​at each time point. This structured record table is the snapshot of the multidimensional rehabilitation state.

[0037] Furthermore, the step of analyzing the evolution sequence of the multidimensional rehabilitation state snapshots based on a time-sliding window to construct a patient's individual rehabilitation state transition network includes:

[0038] Set an analysis window with a fixed time length, and slide the analysis window along the time axis to capture continuous segments of the multidimensional rehabilitation state snapshot;

[0039] Feature extraction was performed on the structured record table within each window segment. The features included the frequency of occurrence of various deviation action units, the proportion of the total duration of physiological fluctuation intervals, and the mean of the intensity of subjective discomfort.

[0040] The feature vector extracted from each window segment is defined as a recovery state node;

[0041] Analyze the characteristic changes between rehabilitation status nodes in adjacent time windows. If the change exceeds the preset transition threshold, it is considered that the status has transitioned, and a directed connection edge is established between the two nodes.

[0042] The network consisting of all rehabilitation status nodes and the directed connections between them is denoted as the patient's personal rehabilitation status transition network.

[0043] Furthermore, the calculation of the rehabilitation path stability index and the probability of abnormal state risk based on the node connection strength and state dwell time in the patient's personal rehabilitation state transfer network includes:

[0044] The in-degree and out-degree of each rehabilitation state node in the patient's personal rehabilitation state transfer network are statistically analyzed, the overall clustering coefficient and average path length of the network are calculated, and the weighted combination of in-degree, out-degree, overall clustering coefficient and average path length is used as the rehabilitation path stability index.

[0045] Identify all edges in the patient's personal rehabilitation status transition network that are directly connected to rehabilitation status nodes containing keywords of high-risk subjective feelings or severe physiological fluctuations;

[0046] The frequency of the edge is calculated as a proportion of the total number of state transitions, and the proportion is corrected by a time decay model and defined as the probability of abnormal state risk at the current moment.

[0047] Furthermore, the automatic division of the current rehabilitation stage by combining preset phased rehabilitation milestones and utilizing the stability index of the rehabilitation path and the probability of abnormal state risk includes:

[0048] Multiple consecutive phased rehabilitation milestones are pre-set for the patient's entire rehabilitation cycle. Each milestone includes the expected range of rehabilitation pathway stability indicators and the upper limit of the probability of abnormal status.

[0049] The stability index of the rehabilitation pathway and the probability of abnormal state risk obtained in real time are matched with the target range and upper limit in the phased rehabilitation milestones.

[0050] If the current indicator and probability value simultaneously meet the target requirements of a certain milestone, but do not meet the requirements of the next higher milestone, then the patient is determined to be in the rehabilitation stage corresponding to the aforementioned milestone.

[0051] If the current indicators and probability values ​​exceed the preset range of all milestones, a special assessment process will be initiated, and the patient will be placed in the assessment stage.

[0052] Furthermore, the set of output monitoring and training suggestions includes:

[0053] Based on the determined rehabilitation stage, the basic monitoring strategy template and training plan template that match the rehabilitation stage are retrieved from the preset strategy library;

[0054] Based on the current specific abnormal state risk probability value and the types of deviation action units that frequently appear in the latest multidimensional rehabilitation state snapshot, the parameters of the called basic template are fine-tuned.

[0055] Parameter fine-tuning includes adjusting the frequency of review of monitoring data, strengthening the monitoring thresholds for specific physiological parameters, and increasing or decreasing the number of exercise sets for specific movements in the training plan;

[0056] The fine-tuned monitoring strategy is integrated with the training plan to generate a set of monitoring and training recommendations that include specific implementation items, schedules, and precautions.

[0057] Compared with the prior art, the beneficial effects of the present invention are:

[0058] The construction of virtual task scenario models imports standardized multi-source sensor data into a pre-defined digital environment strictly associated with rehabilitation tasks for simulation. The simulation records of movement trajectories and physiological responses generated in this process provide the sole basis for standardized compliance verification. The verification process based on simulation records eliminates random interference from the real environment, enabling the assessment of patients' task performance capabilities to be conducted in a controllable and reproducible benchmark scenario. The assessment results directly reflect the quality of patients' performance under the standardized task paradigm, rather than simply reflecting physiological fluctuations in their daily activities, thus enhancing the targeted nature of rehabilitation training and the objectivity of assessment conclusions.

[0059] By fusing tagged movement deviations, physiological fluctuation ranges, and patient subjective text, the system generates a multidimensional snapshot of the rehabilitation state, integrating subjective and objective information. Based on the analysis of continuous snapshot sequences using a time-sliding window, a rehabilitation state transition network characterizing the evolutionary patterns of individual states is constructed. This network, in a data-driven manner, represents the rehabilitation process as a dynamic picture of state nodes and transition relationships. The connection strength between nodes and the state dwell time quantify the inertia and stability of the rehabilitation path, and the analysis of network topology and transition probabilities can identify abnormal paths that deviate from the expected evolutionary pattern. This allows the monitoring of the rehabilitation process to evolve from focusing on isolated indicators to the continuous learning and modeling of individual-specific evolutionary patterns. The division of rehabilitation stages and risk warnings are thus based on the dynamic analysis of the individual's historical behavioral patterns, and the generation of intervention suggestions is more closely aligned with the individual's actual recovery pace and potential risks. Furthermore, existing remote monitoring solutions, after identifying potential risks, often lack linkage with the patient's real-time location, making it difficult to provide monitoring personnel with the key location information needed for timely intervention, thus affecting the efficiency of emergency response. Attached Figure Description

[0060] Figure 1 This is a timing diagram of the real-time remote patient monitoring and rehabilitation monitoring system described in this invention.

[0061] Figure 2 A flowchart for the simulation and derivation of a regularized sensor data stream;

[0062] Figure 3 A graph showing the dynamic monitoring of the weighted norm of joint angle deviation during rehabilitation training and a comparison of compliance thresholds.

[0063] Figure 4 A dynamic evolution correlation diagram of path stability and anomaly risk probability;

[0064] Figure 5 This is a time-series graph showing the dual-axis evolution of the stability of the rehabilitation pathway and the probability of abnormal state risks. Detailed Implementation

[0065] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] See Figure 1 A method for real-time remote monitoring and rehabilitation of patients is proposed, the overall implementation of which is as follows: Multi-source heterogeneous sensor data streams are collected from various sensors attached to the patient's body and environmental monitoring devices. A hierarchical cleaning operation is performed on the multi-source heterogeneous sensor data streams, removing noise segments and filling in missing data segments through a multi-level processing flow, outputting a normalized sensor data stream. A virtual task scenario model associated with the patient's rehabilitation task is constructed, established using computer simulation technology. The normalized sensor data stream is imported into the virtual task scenario model for simulation, simulating the patient performing the task and generating simulation records of the patient's movement trajectory and physiological responses during the simulated task execution. The simulation records of movement trajectory and physiological responses are validated for compliance based on a medical knowledge graph, which stores standard rehabilitation movements and safe physiological parameters. The validation process marks movement units and physiological fluctuation ranges that deviate from the standard execution paradigm. The marked movement units, physiological fluctuation ranges, and subjective feeling text reported by the patient in real time via a mobile terminal are integrated to generate a multi-dimensional rehabilitation status snapshot.

[0067] The evolution sequence of multidimensional rehabilitation status snapshots is analyzed using a time-sliding window approach. The time-sliding window slides along the time axis to extract data segments; a patient-specific rehabilitation status transition network is constructed by analyzing segment characteristics. Based on the node connection strength and status dwell time in the patient-specific rehabilitation status transition network, rehabilitation path stability indicators and abnormal status risk probabilities are calculated. Combined with preset staged rehabilitation milestones, the current rehabilitation stage is automatically defined using the rehabilitation path stability indicators and abnormal status risk probabilities, and a set of monitoring and training suggestions matching the current rehabilitation stage is output. The location monitoring module connects to the abnormal status risk probabilities and monitoring and training suggestion set output by the comprehensive evaluation module. When the abnormal status risk probability exceeds a preset threshold, based on high-risk action units or physiological parameters marked in the monitoring and training suggestion set, combined with real-time location data from the patient's worn location beacon, a risk alarm containing the patient's location information is generated and sent to the monitoring terminal.

[0068] In one embodiment of the present invention, the hierarchical cleaning operation is performed sequentially on a multi-source heterogeneous sensor data stream containing electromyographic signal streams, joint angle sequences, and ambient temperature and humidity readings. Specifically, for the electromyographic signal stream, a dynamic filtering threshold based on signal quality is set. The calculation of this dynamic filtering threshold depends on the evaluation of the real-time signal-to-noise ratio of the electromyographic signal stream. Signal-to-noise ratio of electromyographic signal stream The relationship between them is defined by the following formula:

[0069]

[0070] in: , , These are constants pre-calibrated based on sensor characteristics and typical electromyographic signal spectra. The base of the natural logarithm, signal-to-noise ratio. The ratio of background noise power to signal power is estimated by performing a short-time Fourier transform on the electromyographic signal stream, and the dynamic filtering threshold is obtained based on the calculated value. Adaptive filtering is applied to the electromyographic signal stream using a Butterworth bandpass filter with an adjustable cutoff frequency. The lower passband limit is fixed at 20 Hz to filter out baseline drift, while the upper passband limit is determined based on a dynamic filtering threshold. Dynamic adjustment, the principle of upper limit frequency adjustment is based on the signal-to-noise ratio. At lower frequencies, a lower passband upper limit frequency is used to more aggressively filter out high-frequency interference, thus improving the signal-to-noise ratio. When the frequency is high, a higher passband upper limit frequency is used to retain more details of the electromyographic signal. After filtering, the effective electromyographic activity signal and the power frequency interference signal are separated.

[0071] In some embodiments, for a joint angle sequence, the hierarchical cleaning operation performs a continuity check. The continuity check is performed by calculating the absolute value of the joint angle difference between adjacent sampling points and comparing it with a preset angle jump threshold. This is achieved through comparison; when the absolute value of the joint angle difference continuously exceeds the angle jump threshold... When the region is identified as an angle jump point caused by signal loss, these angle jump points are repaired by an interpolation algorithm. The interpolation algorithm uses cubic spline interpolation, and uses N valid data points before and after the angle jump point as interpolation nodes to generate a smooth interpolation curve, thereby replacing the angle jump point sequence in the original data and generating a smooth joint angle sequence.

[0072] In some embodiments, for ambient temperature and humidity readings, a hierarchical cleaning operation compares the timing logic of the ambient temperature and humidity readings. The comparison of the timing logic is based on the physical continuity and rationality of the changes in environmental parameters. Specifically, a reasonable range for the physiological environment is set, and the rate of change between adjacent ambient temperature and humidity readings is calculated. When an ambient temperature and humidity reading simultaneously exceeds the reasonable range and its rate of change with the previous point exceeds the maximum allowable rate of change, the operation proceeds. When the environmental temperature and humidity reading is determined to be an abnormal reading that is significantly outside the reasonable range of the physiological environment, the abnormal reading is removed and replaced with the arithmetic mean of the environmental temperature and humidity readings at each of the M valid times before and after the abnormal reading.

[0073] It is understandable that after completing the above-mentioned hierarchical processing, the processed effective electromyographic activity signals, smoothed joint angle sequences, and corrected environmental temperature and humidity readings need to be integrated. In specific implementation, this integration process is re-aligned and packaged according to the original timeline. The re-alignment is based on the global synchronization timestamp assigned to each data stream during acquisition. Using the timestamp as a reference, the effective electromyographic activity signals, smoothed joint angle sequences, and corrected environmental temperature and humidity readings at different sampling rates are interpolated or extracted onto a unified time grid. The packaging operation encapsulates the multi-channel data on the unified time grid into a structured data frame sequence. Each data frame contains a timestamp field and fields corresponding to the electromyographic signal amplitude, joint angle value, temperature value, and humidity value at that moment, thus forming a regularized sensor data stream. Optionally, before adaptive filtering the electromyographic signal stream, preliminary power frequency notch filtering can be performed to reduce fixed-frequency power supply interference. It is understood that during the smoothed joint angle sequence generation process, the number of nodes N used for interpolation can be adaptively adjusted according to the length of the signal loss segment. In specific implementation, for the continuity test of the joint angle sequence, the angle jump threshold... The maximum allowable rate of change can be dynamically set based on the product of the maximum physiological angular velocity of joint movement and the sampling interval. Optionally, in the processing of environmental temperature and humidity readings, this can be set. Different values ​​can be set for temperature and humidity.

[0074] In one embodiment of the present invention, see [reference] Figure 2In practice, constructing a virtual task scenario model associated with the patient's rehabilitation task begins with defining the core action elements of the rehabilitation task, the environmental objects for task execution, and the expected task completion criteria. The core action elements include a series of sub-action types necessary to complete a specific rehabilitation action, the target range of motion of each joint, and the expected activation pattern of the muscles. The environmental objects for task execution refer to the entities that interact with the patient's virtual avatar in the virtual scene, such as a virtual cup that needs to be picked up, a virtual obstacle that needs to be crossed, or a virtual button that needs to be pressed. The expected task completion criteria include the judgment conditions for successful task execution, such as the virtual cup being successfully moved to the designated position without spilling the liquid inside.

[0075] In some embodiments, a virtual scene containing environmental objects is constructed using 3D modeling technology. This 3D modeling technology uses computer-aided design software or 3D scanning technology to create 3D mesh models of the environmental objects and scene background. Within the virtual scene, a biomechanically conforming skeletal and muscle model is bound to the patient's virtual avatar. The skeletal model is a rigid body linkage system with hierarchical joint structures, whose kinematic parameters are set based on anthropometric data. The muscle model is a simulation system based on the Hill muscle mechanics model, which maps effective electromyographic activity signals to muscle excitation inputs and calculates the corresponding contractile force. The skeletal and muscle models are coupled through joint constraints and force transmission mechanisms in the physics engine. In a specific implementation, a series of trigger events are preset in the virtual task scene model. These trigger events correspond to key nodes or potentially risky actions in task execution. Key node trigger events include "hand contacting the virtual cup" or "the virtual cup being raised to a specified height." Potentially risky action trigger events include "knee flexion angle exceeding a safety threshold" or "trunk forward tilt angle exceeding a stability limit." Each trigger event is associated with a Boolean flag. When the relevant state in the virtual task scene model meets the preset conditions, the corresponding trigger event flag is set.

[0076] The process of importing the normalized sensor data stream into the virtual task scene model for simulation can be understood as follows: the smoothed joint angle sequence in the normalized sensor data stream is used as input to the skeletal model driving the patient's virtual avatar. Each angle value in the smoothed joint angle sequence is assigned in real time to the corresponding joint actuator in the skeletal model, generating a continuous sequence of motion postures. Simultaneously, the effective electromyographic activity signals in the normalized sensor data stream are input to the muscle model of the patient's virtual avatar to simulate and calculate muscle activation level and fatigue. The muscle activation level is calculated through an activation kinetic model, the output of which is a normalized muscle activation degree. Its calculation depends on the amplitude of the effective electromyographic activity signal input. and the activation and relaxation time constants of muscles. and The relationship is described by the following differential equation:

[0077]

[0078] Among them: when hour ,otherwise Muscle fatigue is based on muscle activation levels. Estimate by integrating with the duration.

[0079] In some embodiments, physical collision detection and interaction simulation are performed based on the action posture sequence and muscle activation level, combined with environmental objects in the virtual scene. Physical collision detection is achieved by a physics engine calculating the minimum distance or penetration depth between the skeletal model of the patient's virtual avatar and the 3D mesh model of the environmental objects. When the penetration depth... A value greater than zero indicates a collision. The collision force and potential changes in motion state are calculated based on the collision location, normal direction, and physical material properties. The interactive simulation determines the success or failure of the task execution based on the collision results and preset logic. For example, if the deviation between the virtual cup's trajectory and the expected trajectory remains less than the allowable error after the "hand contacts virtual cup" event is triggered, the grasping and moving sub-task is considered successful. Optionally, the entire simulation process records the motion posture sequence, muscle activation level, occurrence of triggering events, and task execution status. The motion posture sequence is stored as an array of timestamps and all joint angle values. The muscle activation level is stored as timestamps and activation values ​​of major muscle groups. The occurrence of triggering events is recorded as an event name and a list of corresponding timestamps. The task execution status is recorded as labels such as "success," "failure," or "in progress," along with their time intervals. These records are summarized and packaged into a structured motion trajectory and physiological response simulation record file. It is understandable that the computation frequency of physical collision detection can differ from the rendering frame rate of the virtual scene. A fixed physical update step size can be used to ensure the stability of the simulation. Optionally, the muscle fatigue model can further incorporate muscle energy metabolism parameters to simulate endurance changes more precisely.

[0080] In one embodiment of the present invention, in a specific implementation, compliance verification of the simulation records of motion trajectories and physiological responses based on a medical knowledge graph requires access to a medical knowledge graph containing a standard rehabilitation action library and a range of safe physiological parameters. The medical knowledge graph accesses a remote server via an application programming interface or is loaded from a locally stored graph database. Each standard rehabilitation action in the standard rehabilitation action library is defined as a set of standard joint angle vectors under a series of timestamps. The range of safe physiological parameters is stored in the form of data type, parameter name, lower safety limit, and upper safety limit. In a specific implementation, the motion posture sequence in the simulation records of motion trajectories and physiological responses is compared frame by frame with the standard rehabilitation action library. The frame-by-frame comparison process involves calculating the difference between the joint angle vector in each frame of the motion posture sequence and the standard joint angle vector under the corresponding timestamp in the standard rehabilitation action library. The difference can be quantified by the weighted norm of the deviation of each joint angle.

[0081]

[0082] in: It is the joint angle deviation vector. It is a diagonal weight matrix, whose diagonal elements Representing the importance coefficient of different joints in rehabilitation assessment, when a certain frame The value consistently exceeds the preset compliance threshold. At that time, the system identifies the start and end frames of the action that have deviations in the kinematic parameters and defines them as a deviated action unit.

[0083] In some embodiments, the physiological response data in the motion trajectory and physiological response simulation record are compared with the safe physiological parameter range. The physiological response data includes muscle activation level and simulated heart rate and blood pressure derived data. The comparison operation checks whether the time series of physiological response data continuously exceeds the lower or upper limit of the corresponding safe physiological parameter range, and marks the entire continuous time interval from the first exceedance to the last return to the range as the physiological fluctuation interval. It can be understood that each deviating motion unit is associated with its deviation type, which is selected from a predefined type set, including "insufficient joint range of motion", "excessive joint range of motion", "incorrect motion timing", and "poor motion coordination". Each physiological fluctuation interval is associated with the physiological parameter type that exceeds the limit, and the physiological parameter type that exceeds the limit is directly recorded as the data item name, such as "excessively high biceps activation level" or "simulated low diastolic blood pressure".

[0084] In some embodiments, natural language processing (NLP) is performed on the subjective feeling text input by the patient via a terminal. NLP uses a pre-trained word embedding model combined with a domain-specific sentiment dictionary to extract keywords from the subjective feeling text and quantify their representation of discomfort intensity level and sentiment tendency value. Keywords are nouns or phrases related to rehabilitation discomfort. The discomfort intensity level is mapped to a preset numerical level scale by analyzing the co-occurrence relationship between intensity adverbs and keywords in the text. The sentiment tendency value is calculated by a sentiment analysis model, outputting a continuous value between -1 and +1. In a specific implementation, a time alignment index is established to associate information from different sources. The time alignment index uses the timeline of motion trajectory and physiological response simulation records as a benchmark, associating the keywords extracted by NLP and their quantified values ​​with the deviation motion units and physiological fluctuation intervals that occurred within the same time period. The association operation is based on precise matching of timestamps or time window inclusion relationships. If the subjective feeling text does not have a specific time point, the text submission time is used as the association time point.

[0085] Optionally, a structured record table is created, indexed by time. This table synchronously records whether there are deviated action units, physiological fluctuation ranges, and corresponding subjective feeling keywords and quantitative values ​​at each time point. Specific fields in the structured record table include a timestamp field, a list of deviated action unit identifiers, a list of physiological fluctuation range identifiers, subjective feeling keywords, discomfort intensity level, and emotional tendency value. This structured record table serves as a snapshot of the multidimensional rehabilitation status. The list of deviated action unit identifiers can store identifiers for multiple units, each identifier pointing to a detailed description of the deviated action unit. Optionally, the list of physiological fluctuation range identifiers uses the start and end timestamps of the range as index keys for quick retrieval.

[0086] See Figure 3In the compliance verification of rehabilitation movements, dynamic monitoring of the weighted norm of joint angle deviation is the core basis for identifying deviating movement units. Specifically, the purple curve represents the weighted norm of joint angle deviation calculated frame-by-frame throughout the rehabilitation training. Its value is obtained by multiplying the joint angle deviation vector by the diagonal weight matrix, whose elements reflect the importance coefficients of different joints in the rehabilitation assessment. The red dashed line represents the preset compliance threshold (1.2). When the purple curve consistently exceeds this threshold, the system marks this interval as a non-compliant interval (pink-filled area) and further identifies the start and end frames of the movements within this interval to define the deviating movement units. As shown in the figure, the weighted norm of joint angle deviation significantly increases and exceeds the threshold during training periods of 10-20 minutes, 20-30 minutes, and 35-45 minutes, indicating that the patient's movements deviate from the standard paradigm during these periods. Further analysis of its impact on the stability of the rehabilitation pathway is needed, combined with multidimensional rehabilitation status snapshots.

[0087] In one embodiment of the present invention, in a specific implementation, analyzing the evolution sequence of multidimensional rehabilitation state snapshots based on a time-sliding window requires setting an analysis window of a fixed time length. The length of the analysis window... It can be set to five minutes or fifteen minutes, the specific value depending on the general cycle of the rehabilitation task and the data sampling frequency. After setting, slide the analysis window along the time axis with a fixed step size to capture continuous segments of multidimensional rehabilitation status snapshots. Each segment contains the window time length. All structured record table entries within the window segment. In some embodiments, feature extraction is performed on the structured record table within each window segment. The feature extraction operation reads all structured record table records within each window segment and calculates three types of statistical features. The first type of feature is the frequency of occurrence of action units of various deviation types. This is calculated by counting the action units of each deviation type that appear within the window and dividing by the total number of action units within the window to obtain the frequency. The second type of feature is the proportion of the total duration of physiological fluctuation intervals. This is calculated by dividing the total duration of all marked physiological fluctuation intervals within the window by the window length. The third type of feature is the mean of the subjective discomfort intensity. It is calculated by taking the arithmetic mean of the subjective discomfort intensity levels of all records within the window. The feature vector extracted from each window segment is a multidimensional array containing all the above feature values.

[0088] In practical implementation, the feature vector extracted from each window segment is defined as a rehabilitation state node. Each rehabilitation state node is represented by a unique identifier and a data structure containing the feature vector, window start time, and window end time. The feature changes between rehabilitation state nodes in adjacent time windows are analyzed. The calculation method involves comparing the absolute difference of each corresponding feature value in the feature vectors of two adjacent rehabilitation state nodes. If the absolute difference of any feature value exceeds a preset transition threshold for that feature, the state between the rehabilitation state nodes is considered to have transitioned. A directed connection edge is then established between the two rehabilitation state nodes, pointing from the earlier window to the later window. The network consisting of all rehabilitation state nodes and the directed connections between them is denoted as the patient's individual rehabilitation state transition network. See Table 1, which shows a simplified feature vector table.

[0089] Table 1: Feature Vector Table of Rehabilitation Status Nodes

[0090]

[0091] It is understandable that, based on the node connection strength and state dwell time in the patient's individual rehabilitation state transition network, the stability index of the rehabilitation path and the probability of abnormal state risk are calculated. The calculation formula is:

[0092]

[0093] in: and These represent the patient's individual recovery status transfer network. The in-degree and out-degree of each recovery state node. Represents the overall clustering coefficient of the network. Represents the average path length of the network. , , These are preset weighting coefficients used to balance the contributions of different network features to stability. In some embodiments, calculating the probability of abnormal state risk requires identifying all edges in the patient's individual rehabilitation state transition network that are directly connected to rehabilitation state nodes containing high-risk subjective feeling keywords or severe physiological fluctuations. High-risk subjective feeling keywords include "severe pain" or "dizziness," and severe physiological fluctuations are defined as fluctuations in physiological parameter values ​​exceeding the safe range by a certain proportion or duration. The frequency of these identified edges is calculated as a proportion of all state transitions; this proportion is denoted as . and the proportion After correction using the time decay model, the abnormal state risk probability at the current moment is defined as... The time decay model assigns higher weights to recent state transitions, and one implementation of this model is as follows:

[0094]

[0095] in: Iterate through all identified edges. It is the time difference between the current moment and the moment when the state transition represented by that edge occurs. This is the attenuation coefficient. Optional, it's a weighting coefficient in the formula for calculating the stability index of the rehabilitation pathway. , , This can be obtained through regression analysis of historical rehabilitation data. It can be understood that the state dwell time can be indirectly reflected by calculating the average number of neighboring windows of a rehabilitation state node in the patient's individual rehabilitation state transition network. A network with a smaller average number of neighboring windows generally indicates more frequent state changes. Optionally, the decay coefficient in the time decay model... The attenuation coefficient can be adjusted according to the rehabilitation stage. In the early rehabilitation stage, a smaller attenuation coefficient can be used to observe longer-term trends, while in the later stage, a larger attenuation coefficient can be used to focus on recent changes.

[0096] See Figure 4 The figure presents the dynamic correlation between the stability index of the patient's rehabilitation pathway (blue curve), the probability of abnormal state risk (red curve), and the preset rehabilitation milestones (yellow dashed lines) at the early, middle, and late stages. From the perspective of index evolution, the stability index of the rehabilitation pathway gradually increased from 0.25 in the first week of the early stage to 0.88 in the second week of the late stage, reflecting the improved clustering coefficient and optimized average path length of the patient's individual rehabilitation state transition network as the rehabilitation process progresses, thus enhancing the continuity and predictability of the rehabilitation state. The probability of abnormal state risk, on the other hand, continuously decreased from 0.75 in the first week of the early stage to 0.12 in the second week of the late stage. This trend stems from the decay of the connection frequency of high-risk state nodes in the network over time, and the use of a time decay model to assign higher weights to recent transition events, accurately reflecting the dynamic changes in risk. The yellow dashed lines in the figure represent the preset rehabilitation milestones, corresponding to the stage nodes of the first week of the middle stage and the first week of the late stage, respectively. The stability index of the rehabilitation pathway broke through the milestone threshold of 0.5 in the first week of the middle stage and reached a stable level of 0.82 in the second week. The probability of abnormal status risk dropped below 0.5 in the first week of the middle stage and further decreased to 0.19 in the second week. The synergistic changes of the two indicators verified the effectiveness of the system in automatically dividing the rehabilitation stage based on milestones.

[0097] In one embodiment of the present invention, in specific implementation, the current rehabilitation stage is automatically divided by combining preset phased rehabilitation milestones and utilizing rehabilitation pathway stability indicators and abnormal state risk probabilities. The preset phased rehabilitation milestones are set by rehabilitation physicians or system administrators at the initial stage of the patient's rehabilitation cycle. The phased rehabilitation milestones include multiple consecutive stages, such as "acute phase", "early recovery", "mid-recovery", and "late recovery". Each phased rehabilitation milestone includes the expected target range of rehabilitation pathway stability indicators and the upper limit of abnormal state risk probabilities. The target range of rehabilitation pathway stability indicators is a closed interval. For example, the target range of rehabilitation pathway stability indicators in the acute phase may be [0.2, 0.5], and the target range of rehabilitation pathway stability indicators in the late recovery phase may be [0.7, 0.9]. The upper limit of abnormal state risk probabilities is a single-value threshold. For example, the upper limit of abnormal state risk probabilities in the acute phase may be set to 0.4, and the upper limit of abnormal state risk probabilities in the late recovery phase may be set to 0.1.

[0098] In practice, the real-time calculated stability indicators of the rehabilitation pathway and the probability of abnormal states are matched with the target range and upper limit in the phased rehabilitation milestones. The matching operation is performed by calculating the current stability indicators of the rehabilitation pathway. and the probability of current abnormal state risk With each stage of rehabilitation milestone Target range of rehabilitation pathway stability indicators and the upper limit of the probability of abnormal state risk Matching degree between To achieve matching degree The calculation formula is:

[0099]

[0100] in: and These are the weighting coefficients assigned to the deviation of the stability index of the rehabilitation pathway and the portion of the probability of exceeding the limit of the abnormal state risk, respectively. The function ensures that the probability of the current abnormal state exceeds the upper limit only if the probability of the current abnormal state exceeds the upper limit. Penalty items are only generated at certain times, and the matching degree of all phased rehabilitation milestones is calculated. Then, select the matching degree. The smallest interim rehabilitation milestone is used as a candidate stage, if the current rehabilitation pathway stability index Probability of current abnormal state Simultaneously meeting the target range of the rehabilitation pathway stability index for this candidate stage of rehabilitation milestone (i.e. (within the range) and the upper limit of the probability of abnormal state risk (i.e.) If a patient does not meet the requirements of the next higher-order stage of rehabilitation milestone, then the patient is considered to be in the rehabilitation stage corresponding to that candidate stage of rehabilitation milestone.

[0101] In some embodiments, if the current rehabilitation pathway stability index and the current abnormal state risk probability value exceed the preset range of all phased rehabilitation milestones, such as the calculated matching degree... If all values ​​exceed a preset tolerance threshold, or if the current abnormal state risk probability value significantly exceeds all upper limits and the rehabilitation pathway stability index value is far below all target range lower limits, then a special assessment process will be initiated. The special assessment process includes generating a special assessment report containing detailed abnormal data and potential risk analysis, notifying remote monitoring personnel for manual review, and simultaneously classifying the patient into the pending assessment stage.

[0102] It is understandable that the output is a set of monitoring and training suggestions that match the current rehabilitation stage. Based on the determined rehabilitation stage, the basic monitoring strategy template and training plan template that match the rehabilitation stage are called from the preset strategy library. The strategy library is stored in the form of a database. Each rehabilitation stage entry is associated with one or more basic templates. The basic monitoring strategy template defines the monitoring data type, collection frequency and alarm rules, and the basic training plan template defines the training actions, number of sets, number of repetitions and rest intervals.

[0103] In practice, based on the current abnormal state risk probability value and the types of deviation movement units that frequently appear in the latest multidimensional rehabilitation status snapshot, the parameters of the basic template are fine-tuned. The parameter fine-tuning includes adjusting the review frequency of monitoring data, strengthening the monitoring threshold of specific physiological parameters, and increasing or decreasing the number of exercise sets for specific movements in the training plan. For example, if the current abnormal state risk probability value is high, the frequency of remote monitoring personnel reviewing physiological data is adjusted from once every four hours to once per hour. If deviation movement units of the type of "excessive knee joint mobility" frequently appear in the latest multidimensional rehabilitation status snapshot, the number of isometric contraction exercise sets for knee joint stability muscles is increased in the training plan.

[0104] Optionally, the fine-tuned monitoring strategy and training plan can be integrated. During the integration process, the fine-tuned parameters are filled into the corresponding fields of the basic template to generate a set of monitoring and training suggestions containing specific execution items, schedules, and precautions. The set of monitoring and training suggestions is presented to patients and rehabilitation instructors in the form of a structured document or an interactive list. It can be understood that the rules for fine-tuning the parameters of the basic template can be a rule-based expert system or a machine learning model trained based on historical adjustment records. Optionally, when generating the set of monitoring and training suggestions, the basis for each suggestion can be marked, such as "Because the probability value of the abnormal state risk exceeds 0.3, it is recommended to increase the frequency of monitoring review".

[0105] After receiving the abnormal state risk probability and monitoring and training suggestion set output by the comprehensive assessment module, the positioning monitoring module first reads the preset risk probability threshold. This threshold corresponds to the stage-specific rehabilitation milestones and is pre-set according to the risk prevention and control needs of different rehabilitation stages. The module continuously monitors the real-time output abnormal state risk probability and compares it with the preset threshold in real time. When the detected abnormal state risk probability reaches or exceeds the preset threshold, the module extracts the marked high-risk action units and corresponding associated physiological parameters from the monitoring and training suggestion set, clarifying the specific action type and physiological monitoring indicators of the current risk. Simultaneously, the module receives location data sent by the positioning beacon worn by the patient in real time via a wireless communication protocol. This positioning beacon is bound to the patient's rehabilitation training equipment or wearable device and can accurately provide feedback on the patient's specific spatial location information. The module integrates and processes the high-risk action units, related physiological parameter information, and real-time location data, and generates risk alarm information containing the patient's current location, the name of the high-risk action unit, and the type of associated physiological parameter according to the preset alarm information format. The module sends risk alarm information to pre-bound monitoring terminals via a preset communication link. These monitoring terminals include mobile terminals of rehabilitation physicians and monitoring platforms of medical institutions, ensuring that the monitoring parties can obtain the patient's risk status and location information in a timely manner so as to take rapid intervention measures.

[0106] See Figure 5The graph presents the dynamic correlation between the rehabilitation pathway stability index (line graph) and the probability of abnormal state risk (bar graph) over monitoring time points. The specific analysis is as follows: The rehabilitation pathway stability index gradually increased from 0.25 at monitoring point 1 to 0.88 at monitoring point 10, showing a continuous upward trend. This reflects the optimization of node connection strength and state retention time in the patient's rehabilitation state transition network over time, and the continuous improvement in the compliance and consistency of rehabilitation actions, conforming to the stage evolution pattern of "early recovery → mid-recovery → late recovery". The probability of abnormal state risk continuously decreased from 0.38 at monitoring point 1 to 0.08 at monitoring point 10, showing a negative correlation with the stability index. This change corresponds to the decrease in the proportion of connection frequency of high-risk nodes in the rehabilitation state transition network over time, indicating a significant reduction in the frequency and severity of patient movement deviations and physiological fluctuations. The monitoring nodes divided by the dashed lines in the graph can be matched with preset staged rehabilitation milestones (such as acute phase, early recovery, mid-recovery, and late recovery). For example, the stability index of monitoring points 1-3 is in the range of [0.2, 0.5] and the risk probability is higher than 0.2, which is consistent with the characteristics of the acute phase; the stability index of monitoring points 8-10 enters the range of [0.7, 0.9] and the risk probability is lower than 0.1, corresponding to the target requirements of the later recovery stage. The synergistic change of the dual-axis trend verifies the assessment value of the rehabilitation pathway stability index and the risk probability of abnormal states. The dynamic correlation between the two can be directly used to automatically divide the rehabilitation stages and provide data support for outputting personalized monitoring and training recommendations.

[0107] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A system for real-time remote monitoring and rehabilitation monitoring of patients, characterized in that, include: The data preprocessing module collects multi-source heterogeneous sensor data streams from patients, performs hierarchical cleaning operations on the multi-source heterogeneous sensor data streams, removes noise segments and fills in missing data segments, and outputs a normalized sensor data stream. The task simulation module constructs a virtual task scene model associated with the patient's rehabilitation task, imports the normalized sensor data stream into the virtual task scene model for simulation and deduction, and generates a simulation record of the patient's movement trajectory and physiological response during the execution of the simulated task. The compliance analysis module performs compliance verification on the simulated action trajectories and physiological responses based on a medical knowledge graph, marking action units and physiological fluctuation ranges that deviate from the standard execution paradigm. The evolutionary analysis module integrates the labeled action units, physiological fluctuation ranges, and the subjective feelings reported by the patient in real time to generate a multidimensional rehabilitation state snapshot. Based on a time sliding window, it analyzes the evolution sequence of the multidimensional rehabilitation state snapshot and constructs a patient's personal rehabilitation state transfer network. The comprehensive assessment module calculates the rehabilitation path stability index and the probability of abnormal state risk based on the node connection strength and state dwell time in the patient's personal rehabilitation state transfer network. It also automatically divides the current rehabilitation stage using the rehabilitation path stability index and the probability of abnormal state risk, and outputs a set of monitoring and training suggestions, in conjunction with preset staged rehabilitation milestones. The positioning monitoring module receives the abnormal state risk probability and monitoring and training suggestion set output by the comprehensive evaluation module. When the abnormal state risk probability exceeds a preset threshold, it generates a risk alarm containing the patient's location information based on the high-risk action units or physiological parameters marked in the monitoring and training suggestion set, combined with the real-time location data of the positioning beacon worn by the patient, and sends it to the monitoring terminal.

2. The system for real-time remote monitoring and rehabilitation monitoring of patients as described in claim 1, characterized in that, The hierarchical cleaning operation performed on the multi-source heterogeneous sensor data stream includes: The multi-source heterogeneous sensor data stream includes electromyographic signal stream, joint angle sequence and ambient temperature and humidity readings; By setting a dynamic filtering threshold based on signal quality, the electromyography signal stream is adaptively filtered to separate the effective electromyography activity signal from the power frequency interference signal. The continuity of the joint angle sequence is checked, and the angle jump points caused by signal loss are repaired by the interpolation algorithm to generate a smooth joint angle sequence. By comparing the temporal logic of environmental temperature and humidity readings, abnormal readings that clearly exceed the reasonable range of physiological environment are eliminated, and the average of the data before and after the time are used for replacement. The processed effective electromyographic activity signals, smoothed joint angle sequences, and corrected ambient temperature and humidity readings are re-aligned and packaged according to the original time axis to form the normalized sensor data stream.

3. The system for real-time remote monitoring and rehabilitation monitoring of patients as described in claim 1, characterized in that, The construction of the virtual task scenario model associated with the patient's rehabilitation task includes: Define the core action elements of rehabilitation tasks, the environmental objects for task execution, and the expected task completion standards; A virtual scene containing the environmental objects is constructed using 3D modeling technology, and a biomechanical skeletal and muscular model is bound to the patient's virtual avatar in the virtual scene; A series of triggering events are preset in the virtual task scenario model, and the triggering events correspond to the key nodes or potential risk actions of task execution.

4. The system for real-time remote monitoring and rehabilitation monitoring of patients as described in claim 3, characterized in that, The step of importing the normalized sensor data stream into the virtual task scenario model for simulation includes: The smooth joint angle sequence in the normalized sensor data stream is used to drive the skeletal model of the patient avatar to calculate and generate a continuous sequence of motion postures. The effective electromyographic activity signals in the normalized sensor data stream are input into the muscle model of the patient's virtual avatar to simulate and calculate the muscle activation level and fatigue level. Based on the aforementioned action posture sequence and muscle activation level, combined with environmental objects in the virtual scene, physical collision detection and interaction simulation are performed to determine the success or failure status of task execution. The sequence of movements and postures, muscle activation levels, occurrence of triggering events, and task execution status throughout the entire simulation process are recorded and summarized into the simulation record of movement trajectory and physiological response.

5. The system for real-time remote monitoring and rehabilitation monitoring of patients as described in claim 1, characterized in that, The compliance verification of the motion trajectory and physiological response simulation records based on the medical knowledge graph includes: Access to a medical knowledge graph that includes a standard rehabilitation exercise library and safe physiological parameter ranges; The motion trajectory and the motion posture sequence in the physiological response simulation record are compared frame by frame with the standard rehabilitation motion library to identify the start and end frames of the motion with deviations in kinematic parameters and define them as a deviated motion unit. The physiological response data in the motion trajectory and physiological response simulation record are compared with the safe physiological parameter range, and the time interval that continuously exceeds the range is marked as the physiological fluctuation interval. Associate each deviating action unit with its deviation type, and associate each physiological fluctuation range with the type of physiological parameter that exceeds the limit.

6. The system for real-time remote monitoring and rehabilitation monitoring of patients as described in claim 1, characterized in that, The fusion of labeled action units, physiological fluctuation ranges, and patient-reported subjective feelings in real time includes: Natural language processing was performed on the subjective feelings text input by the patient through the terminal to extract keywords and quantify the discomfort intensity level and emotional tendency value represented by them. Establish a time alignment index to associate the keywords and their quantified values ​​with the deviation action units and physiological fluctuation ranges that occurred within the same time period; Create a structured record table indexed by time, synchronously recording whether there are deviations in action units, physiological fluctuation ranges, and corresponding subjective feeling keywords and quantitative values ​​at each time point. This structured record table is the snapshot of the multidimensional rehabilitation state.

7. The system for real-time remote patient monitoring and rehabilitation monitoring as described in claim 1, characterized in that, The process of analyzing the evolution sequence of the multidimensional rehabilitation state snapshots based on a time-sliding window to construct a patient's individual rehabilitation state transition network includes: Set an analysis window with a fixed time length, and slide the analysis window along the time axis to capture continuous segments of the multidimensional rehabilitation state snapshot; Feature extraction was performed on the structured record table within each window segment. The features included the frequency of occurrence of various deviation action units, the proportion of the total duration of physiological fluctuation intervals, and the mean of the intensity of subjective discomfort. The feature vector extracted from each window segment is defined as a recovery state node; Analyze the characteristic changes between rehabilitation status nodes in adjacent time windows. If the change exceeds the preset transition threshold, it is considered that the status has transitioned, and a directed connection edge is established between the two nodes. The network consisting of all rehabilitation status nodes and the directed connections between them is denoted as the patient's personal rehabilitation status transition network.

8. The system for real-time remote monitoring and rehabilitation monitoring of patients as described in claim 7, characterized in that, The calculation of rehabilitation pathway stability indicators and abnormal state risk probabilities based on node connection strength and state dwell time in the patient's personal rehabilitation state transfer network includes: The in-degree and out-degree of each rehabilitation state node in the patient's personal rehabilitation state transfer network are statistically analyzed, the overall clustering coefficient and average path length of the network are calculated, and the weighted combination of in-degree, out-degree, overall clustering coefficient and average path length is used as the rehabilitation path stability index. Identify all edges in the patient's personal rehabilitation status transition network that are directly connected to rehabilitation status nodes containing high-risk subjective feeling keywords or severe physiological fluctuations; The frequency of the edge is calculated as a proportion of the total number of state transitions, and the proportion is corrected by a time decay model and defined as the probability of abnormal state risk at the current moment.

9. The system for real-time remote monitoring and rehabilitation monitoring of patients as described in claim 1, characterized in that, The method of automatically classifying the current rehabilitation stage by combining preset phased rehabilitation milestones and utilizing the stability index of the rehabilitation path and the probability of abnormal state risk includes: Multiple consecutive phased rehabilitation milestones are pre-set for the patient's entire rehabilitation cycle. Each milestone includes the expected range of rehabilitation pathway stability indicators and the upper limit of the probability of abnormal status. The stability index of the rehabilitation pathway and the probability of abnormal state risk obtained in real time are matched with the target range and upper limit in the phased rehabilitation milestones. If the current indicator and probability value simultaneously meet the target requirements of a certain milestone, but do not meet the requirements of the next higher milestone, then the patient is determined to be in the rehabilitation stage corresponding to the aforementioned milestone. If the current indicators and probability values ​​exceed the preset range of all milestones, a special assessment process will be initiated, and the patient will be placed in the assessment stage.

10. The system for real-time remote monitoring and rehabilitation monitoring of patients as described in claim 9, characterized in that, The set of output monitoring and training suggestions includes: Based on the determined rehabilitation stage, the basic monitoring strategy template and training plan template that match the rehabilitation stage are retrieved from the preset strategy library; Based on the current specific abnormal state risk probability value and the types of deviation action units that frequently appear in the latest multidimensional rehabilitation state snapshot, the parameters of the called basic template are fine-tuned. Parameter fine-tuning includes adjusting the frequency of review of monitoring data, strengthening the monitoring thresholds for specific physiological parameters, and increasing or decreasing the number of exercise sets for specific movements in the training plan; The fine-tuned monitoring strategy is integrated with the training plan to generate a set of monitoring and training recommendations that include specific implementation items, schedules, and precautions.

Images