An information-based postoperative remote monitoring and rehabilitation system for patients of thoracic surgery

Abstract

The application discloses a postoperative remote monitoring and rehabilitation system for thoracic surgery patients based on informatization, and relates to the technical field of postoperative monitoring.The system solves the technical problems that traditional monitoring cannot perceive the stress problem of sternal incision caused by improper posture, lacks real-time early warning models based on electrocardio characteristics, is difficult to discover acute events in time, and is difficult to make rehabilitation plans fit the actual situation, easily causes blind decision-making and cannot be dynamically adjusted due to fixed grouping, non-causal models and lack of optimization algorithms.The system dynamically correlates the body position angle and the thoracic pressure, calculates the thoracic load index and the sternal incision stress concentration index in real time, can quantify the cumulative stress of the sternal incision caused by improper posture, identifies influencing factors through the SCM structure causal model, constructs a rehabilitation data topology graph for each group, identifies key factors affecting the rehabilitation effect through causal inference, instead of only discovering correlation, provides a scientific basis for precise intervention, and avoids wrong decisions.

Description

Technical Field

[0001] This invention belongs to the field of postoperative monitoring, specifically an information-based remote monitoring and rehabilitation system for thoracic surgery patients after surgery. Background Technology

[0002] The information-based remote monitoring and rehabilitation system for thoracic surgery patients utilizes the Internet of Things (IoT), communication technology, sensors, and a data analysis platform to monitor patients' postoperative physiological indicators (such as heart rate, blood pressure, blood oxygen saturation, and respiratory rate), changes in body position, and wound healing status in real time. This data is then transmitted remotely to medical staff terminals. The system combines intelligent algorithms to analyze patient conditions, providing decision support for medical personnel and offering personalized rehabilitation guidance to patients (such as breathing exercises and activity suggestions), thus achieving dynamic, precise, and intelligent postoperative management.

[0003] Traditional monitoring only records changes in body position, failing to establish a non-linear relationship between pressure and angle. This means that when patients are in improper positions for extended periods (such as an excessively large semi-recumbent position), stress concentration at the sternal incision site may go undetected, potentially leading to serious complications such as loosening of fixation wires and sternal fracture, increasing the risk of secondary surgery and the consumption of medical resources. Furthermore, the lack of real-time early warning models based on electrocardiogram characteristics, relying solely on periodic data checks by medical staff, makes it difficult to detect acute events such as arrhythmias or myocardial ischemia in a timely manner, especially at night or when patients are alone, potentially missing the optimal window for treatment and leading to serious consequences such as sudden death. The fixed grouping strategies used in traditional monitoring (such as grouping by postoperative days) cannot reflect abrupt changes in the recovery process due to complications and differences in compliance, resulting in a disconnect between the rehabilitation plan and the actual situation (e.g., continuing training according to the original plan may worsen the condition), affecting rehabilitation outcomes. Additionally, the absence of causal models and the reliance on correlation analysis to recommend interventions (e.g., adding a certain training) may lead to spurious associations due to confounding factors, resulting in ineffective or even harmful interventions (e.g., simply lowering blood pressure when pain causes elevated blood pressure while ignoring analgesia). Doctors struggle to identify the true influencing factors, leading to highly unpredictable decision-making. Rehabilitation plans are often fixed templates and cannot be dynamically adjusted according to the patient's real-time condition (e.g., high-intensity training is still performed even if TLI is high on a certain day), which can easily lead to overload or insufficient rehabilitation. Summary of the Invention

[0004] This invention aims to solve at least one of the technical problems existing in the prior art. To this end, this invention proposes an information-based remote monitoring and rehabilitation system for postoperative thoracic surgery patients to solve the problems of traditional monitoring failing to detect sternal incision stress caused by improper posture; lack of a real-time early warning model based on electrocardiogram characteristics, making it difficult to detect acute events in a timely manner; and fixed grouping, lack of causal models, and lack of optimization algorithms, making it difficult for rehabilitation plans to be realistic, leading to blind decision-making and an inability to be dynamically adjusted.

[0005] To address the aforementioned problems, a first aspect of the present invention provides an information-based remote monitoring and rehabilitation system for thoracic surgery patients after surgery, comprising: Edge end: The edge gateway establishes a nonlinear dynamic model of body position and thoracic pressure. Based on the rehabilitation data collected by the rehabilitation data acquisition module, it associates the patient's body position angle with the thoracic data, analyzes the thoracic load index and sternal incision stress concentration index in real time, and extracts the core features of the electrocardiogram based on the continuously collected electrocardiogram data to build an electrocardiogram abnormality early warning model. Combined with the thoracic load index and sternal incision stress concentration index, it identifies abnormal data and uploads it to the cloud platform. The cloud platform includes: a dynamic grouping module, an impact analysis module, and a rehabilitation monitoring module; Dynamic grouping module: The cloud platform uses a Hidden Markov Model (HMM) combined with real-time data streams, along with the analysis results of the chest wall load index and ECG abnormality early warning model, to analyze the patient's current rehabilitation stage and perform dynamic grouping. Impact Analysis Module: Using rehabilitation data collected from patients at different time points as nodes, a rehabilitation data topology map is built for each group. Combined with the construction of an SCM structural causal model, the influencing factors at different rehabilitation stages are identified. Rehabilitation monitoring module: Based on the influencing factors of different rehabilitation stages, combined with MPC model prediction and control, and based on the patient digital twin model, the rehabilitation plan for different groups is continuously optimized.

[0006] Optionally, in one example of the above aspects, it also includes: Rehabilitation data acquisition module: Collects patients' rehabilitation data in real time through the patient's wearable device and uploads it to the edge gateway via GPRS / 5G.

[0007] Optionally, in one example of the above aspects, the edge gateway establishes a nonlinear dynamic model of body position and intrathoracic pressure, and associates the patient's body position angle with intrathoracic data based on rehabilitation data collected by the rehabilitation data acquisition module, including the following steps: Based on the nonlinear response mechanism of body position changes to thoracic load, and combined with the dynamic correlation between body position angle, rate of change and intrathoracic pressure, a quadratic nonlinear dynamic model is constructed. Based on rehabilitation data collected by the rehabilitation data acquisition module, a nonlinear function of body position on thoracic stress is established. The patient's position angle is correlated with chest cavity data.

[0008] Optionally, in one example of the above aspects, real-time analysis of the thoracic load index includes the following steps: Let the patient's body position angle at time t be θ(t), and the pressure value measured by the thoracic pressure sensor be P(t). Establish a formula for analyzing the thoracic load index.

[0009] Optionally, in one example of the above aspects, real-time analysis of the sternal incision stress concentration index includes the following steps: A formula for analyzing the stress concentration factor of the sternal incision was established to analyze the stress concentration index of the sternal incision in real time and determine the risk level.

[0010] Optionally, in one example of the above aspects, an ECG abnormality early warning model is constructed, which combines the thoracic load index and the sternal incision stress concentration index to identify abnormal data and upload it to the cloud platform, including the following steps: Based on continuously acquired electrocardiogram (ECG) data, core ECG features are extracted, including heart rate, heart rhythm, and ST segment deviation, to construct an ECG abnormality early warning model and identify abnormal conditions such as arrhythmia and myocardial ischemia in real time.

[0011] Optionally, in one example of the above aspects, the cloud platform uses an Hidden Markov Model (HMM) combined with real-time data streams, along with the analysis results of the chest wall load index and ECG abnormality early warning model, to analyze the patient's current rehabilitation stage and perform dynamic grouping, including the following steps: The rehabilitation phase is divided into acute phase, subacute phase, recovery phase and long-term maintenance phase; the hidden state set S={s1,s2,s3,s4} is defined, where s1,s2,s3,s4 correspond to the acute phase, subacute phase, recovery phase and long-term maintenance phase of the rehabilitation phase, respectively. The observed variable Ot is set as a multidimensional feature vector at time t, including: the thoracic load index and the ECG abnormality early warning score analyzed by the ECG abnormality early warning model, as well as the patient's rehabilitation data; Set the HMM parameters to λ=(A,B,π) and establish the state transition probability matrix: A=[aij], where aij=P(s(t+1)=sj|s(t)=si), and aij is the probability of transitioning to state sj at the next time t+1 given that the current time t is in state si; Analyze the observation probability B=[bj(ot)], assuming the observations follow a Gaussian mixture model: , Let t be the probability of observing the observation value ot when the state is sj at time t, where ot is the observation value at time t, which is the data that can be actually measured in the model; M is the number of Gaussian components in the Gaussian mixture model. Let be the mixing coefficient of the m-th Gaussian component in the j-th state, satisfying , representing the weight of each Gaussian component in the mixture model; The mean is The covariance matrix is The probability density function of the multivariate Gaussian distribution; set up The initial state distribution π is given, where π is the initial observation probability of the patient being in a given hidden state sj at the start of the rehabilitation process. The HMM parameters of the Hidden Markov Model (HMM) are updated in real time using the online EM algorithm, and A, B, and π are updated using the online EM algorithm. The current state after updating the model parameters is decoded using the Viterbi algorithm, with the following formula: ;in, This represents the most likely state at time t, obtained through decoding using the Viterbi algorithm. To represent the observation sequence from time 1 to time t, λ is the set of parameters of the model; Using the Viterbi algorithm on a given observation sequence Given the model parameters λ, find the most likely hidden state sequence at time t to determine the most likely rehabilitation stage the patient is in at the current time. Based on the most likely state at time t Patients are categorized into corresponding rehabilitation stage groups, with different rehabilitation stage groups having different rehabilitation plan templates. Based on the corresponding rehabilitation plan templates and the patients' rehabilitation data, a large AI model is used to generate personalized intervention plans.

[0012] Optionally, in one example of the above aspects, using rehabilitation data collected from patients at different time points as nodes, a rehabilitation data topology graph is established for each group. Combined with the construction of an SCM structural causal model, the influencing factors at different rehabilitation stages are identified, including the following steps: The rehabilitation data collected from each patient at different time points are used as nodes, and the patients are grouped according to their rehabilitation stage to form different rehabilitation stage groups; For each rehabilitation stage group, a rehabilitation data topology graph is constructed. Temporal edges are established based on the time sequence between the corresponding patient rehabilitation data nodes. Similarity edges are established between nodes whose semantic similarity is greater than a threshold.

[0013] Optionally, in one example of the above aspects, by constructing a structural causal model of SCM to identify influencing factors at different stages of rehabilitation, the following steps are included: For the rehabilitation data topology diagram constructed for different rehabilitation stages, influencing factor nodes are added. Influencing factor nodes include: static feature nodes and dynamic feature nodes. Static feature nodes include: patient age, gender, surgical method, and underlying disease; dynamic feature nodes include: intervention plan that changes over time. Calculate the mean of each parameter in the rehabilitation data node in the rehabilitation data topology graph, and select rehabilitation data nodes whose data difference from the mean exceeds a preset threshold as key nodes. Establish causal edges between the nearest dynamic feature nodes before the timestamp corresponding to the key nodes. Sort the rehabilitation data nodes in the rehabilitation data topology graph from front to back according to the timestamp, and establish causal edges between static feature nodes and rehabilitation data nodes located in the top 10% of the sorted nodes. Construct a causal model of the SCM structure, using causal edges between nodes as the causal relationships between nodes; The SCM structural causal model is parameterized, and the causal relationship is quantified using the ATE average treatment effect formula: ATE = E[Y|do(X=x)] - E[Y|do(X=x′)], where ATE is the average treatment effect, representing the expected difference in outcome variable Y between the intervention and non-intervention states; E[ [] represents the expectation operation, used to calculate the average of a random variable; Y is the outcome variable; do(X=x) is to intervene in variable X to fix its value at x, do(X=x′) is to intervene in variable X to fix its value at x′; E[Y|do(X=x)] is the expected value of Y under the intervention X=x; In the rehabilitation data topology diagram constructed by grouping different rehabilitation stages, the average treatment effect of each patient's influencing factor node is sorted from high to low. The influencing factor nodes corresponding to the top 20% of the average treatment effects are selected as the main influencing factors for different rehabilitation stages.

[0014] Optionally, in one example of the above aspects, based on the influencing factors of different rehabilitation stages, combined with MPC model predictive control, and using a patient digital twin model as a basis, the rehabilitation plan for different groups is continuously optimized, including the following steps: The temporal edges, similarity edges, and causal edges of the rehabilitation data topology graph at different rehabilitation stages are equivalent to node edges. The degree centrality of each rehabilitation data node is calculated and sorted from high to low. The top 20% of rehabilitation data nodes are selected as high-relevance nodes. The high-relevance nodes of patients are used as termination data nodes. The rehabilitation data nodes connected by a predetermined number of temporal edges before the termination data nodes are used as starting data nodes. The rehabilitation data nodes connected between the starting and termination data nodes by temporal edges are used as intermediate data nodes. The intervention plan corresponding to the dynamic feature node at the most recent moment before the high-relevance data node is selected. Intervention schemes are extracted from the dynamic feature nodes corresponding to the main influencing factors at different rehabilitation stages. Rehabilitation data nodes that establish causal edges with the corresponding influencing factor nodes are selected as starting data nodes. Rehabilitation data nodes that establish temporal edges with the rehabilitation data nodes that establish causal edges are selected as intermediate data nodes. The rehabilitation data node preceding the rehabilitation data node that establishes causal edges in the intermediate data nodes is selected as the ending data node. By using the starting, intermediate, and ending data nodes of different rehabilitation stages, as well as the parameter vectors of the intervention plan, an LSTM neural network is trained to predict rehabilitation data under the influence of the intervention plan on the influencing factor nodes. Establish an MPC model predictive control system. For the patient's personalized intervention plan, the parameter vectors of each personalized intervention plan are combined into a parameter vector set, which serves as the control input variable of the system. Determine the prediction time domain H, and at each time t, predict the system state for the next H steps based on the patient's digital twin model. An optimization function is established, and the optimization problem is solved at each time t to obtain the optimal system control input variables. Based on the optimal system control input variables, the personalized intervention plan is optimized. This process is repeated at preset time intervals, and the personalized intervention plan is continuously optimized based on new rehabilitation references.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This invention dynamically correlates body position with intrathoracic pressure, calculating the thoracic load index and sternal incision stress concentration index in real time. This quantifies the cumulative stress on the sternal incision caused by improper posture, providing early warning of the risk of incision dehiscence or wire loosening, avoiding missed detections caused by traditional static recording. It extracts core features from continuous electrocardiogram data and combines them with indicators such as TLI to identify precursors to complications such as arrhythmias or myocardial ischemia, achieving multimodal fusion early warning and improving the sensitivity and specificity of anomaly detection.

[0016] All models in this invention run locally on the edge gateway, ensuring millisecond-level anomaly identification and local alarms, and protecting patient safety even if the network is interrupted. The invention employs a Hidden Markov Model (HMM) for dynamic grouping, which automatically identifies the patient's recovery stage based on real-time data streams and dynamically adjusts the grouping, enabling the recovery strategy to adapt to the evolution of the condition and overcoming the deficiency that fixed grouping cannot reflect changes in the individual's recovery process.

[0017] This invention identifies influencing factors using a structural causal model (SCM). A rehabilitation data topology map is constructed for each group, and key factors affecting rehabilitation outcomes are identified through causal inference, rather than simply finding correlations. This provides a scientific basis for precise intervention and avoids erroneous decisions. An MPC model is used for predictive control to optimize rehabilitation plans. Based on patient digital twin models, future rehabilitation plans are continuously optimized to maximize rehabilitation efficiency while meeting safety constraints, achieving proactive intervention rather than reactive remediation. Through causal models and reinforcement learning, the system continuously learns from individual and group data to optimize rehabilitation programs. It also supports multi-center data sharing, enabling continuous model evolution and improved generalization ability. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the system framework of the present invention. Detailed Implementation

[0020] 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.

[0021] Please see Figure 1 The first aspect of this invention provides an information-based remote monitoring and rehabilitation system for thoracic surgery patients after surgery, comprising: Edge layer: The edge gateway establishes a nonlinear dynamic model of body position and thoracic pressure. Based on the rehabilitation data collected by the rehabilitation data acquisition module, it correlates the patient's body position angle (acquired through an inertial measurement unit IMU) with thoracic data (through a wearable piezoelectric sensor), analyzes the thoracic load index and sternal incision stress concentration index in real time, and extracts core ECG features based on continuously acquired ECG data, including heart rate, heart rhythm, and ST segment deviation, to construct an ECG abnormality early warning model. Combining the thoracic load index and sternal incision stress concentration index, it identifies abnormal data and uploads it to the cloud platform. The cloud platform includes: a dynamic grouping module, an impact analysis module, and a rehabilitation monitoring module; Dynamic grouping module: The cloud platform uses a Hidden Markov Model (HMM) combined with real-time data streams, and combines the analysis results of the chest wall load index and ECG abnormality early warning model to analyze the patient's current rehabilitation stage (including: acute phase, subacute phase, recovery phase and long-term maintenance phase) and perform dynamic grouping; Impact Analysis Module: Using rehabilitation data collected from patients at different time points as nodes, a rehabilitation data topology map is built for each group. Combined with the construction of an SCM structural causal model, the influencing factors at different rehabilitation stages are identified. Rehabilitation monitoring module: Based on the influencing factors of different rehabilitation stages, combined with MPC model prediction and control, and based on the patient digital twin model, the rehabilitation plan for different groups is continuously optimized.

[0022] Specifically, in this embodiment, the body position angle is dynamically correlated with the intrathoracic pressure, and the thoracic load index (TLI) and sternal incision stress concentration index are calculated in real time. This can quantify the cumulative stress on the sternal incision caused by improper posture, provide early warning of the risk of incision cracking or wire loosening, and avoid the underreporting caused by traditional static recording.

[0023] Extracting core features (such as heart rate variability and ST segment changes) from continuous electrocardiogram data and combining them with indicators such as TLI, we can identify precursors to complications such as arrhythmia or myocardial ischemia, achieve multimodal fusion early warning, and improve the sensitivity and specificity of abnormality detection.

[0024] All models run locally on the edge gateway, ensuring millisecond-level anomaly detection and local alarms. Even if the network is interrupted, patient safety can be guaranteed, solving the response delay problem caused by relying on cloud processing.

[0025] The Hidden Markov Model (HMM) is used for dynamic grouping. Based on real-time data streams (TLI, ECG early warning results), the patient's recovery stage (such as acute phase and recovery phase) is automatically identified, and the grouping is dynamically adjusted so that the rehabilitation strategy adapts to the evolution of the condition, overcoming the deficiency of fixed grouping that cannot reflect changes in the individual recovery process.

[0026] By using the SCM structural causal model to identify influencing factors, a rehabilitation data topology map is constructed for each group. Key factors affecting rehabilitation outcomes (such as postural habits and medication timing) are identified through causal inference, rather than just finding correlations. This provides a scientific basis for precise intervention and avoids erroneous decisions.

[0027] MPC model predictive control is used to optimize rehabilitation plans. Based on the patient's digital twin model, future rehabilitation plans (such as training intensity and follow-up frequency) are continuously optimized to maximize rehabilitation efficiency while meeting safety constraints, thus achieving proactive intervention rather than reactive remediation.

[0028] Through causal models and reinforcement learning, the system continuously learns from individual and group data to optimize rehabilitation plans, while supporting multi-center data sharing (such as federated learning) to enable the model to continuously evolve and improve its generalization ability.

[0029] In one embodiment of the present invention, it further includes: Rehabilitation data acquisition module: Real-time acquisition of patient rehabilitation data, including body position, pressure, and vital signs data, through the patient's wearable device (IMU, piezoelectric sensor, blood oxygen probe), and uploading to the edge gateway via GPRS / 5G.

[0030] In one embodiment of the present invention, the edge gateway establishes a nonlinear dynamic model of body position and thoracic pressure, and associates the patient's body position angle with thoracic data based on rehabilitation data collected by the rehabilitation data acquisition module, including the following steps: Based on the nonlinear response mechanism of body position changes to thoracic load, and combining the dynamic correlation between body position angle, rate of change, and intrathoracic pressure, a quadratic nonlinear dynamic model is constructed, as shown in the following formula:

[0031] in, The local pressure in the thoracic cavity at time t (in kPa); Let t be the patient's body position angle, which is the angle between the patient's torso and the horizontal plane (the angle with the horizontal direction, in degrees). The rate of change of body position at time t (in ° / s); , , and These are the model coefficients; in this embodiment, they were obtained by fitting historical data from 500 postoperative patients undergoing thoracic surgery. =-0.003, =0.12, =0.005, =0.75; This is a nonlinear interaction term, reflecting the synergistic effect of body position angle and rate of change on intrathoracic pressure; Based on rehabilitation data collected by the rehabilitation data acquisition module, a nonlinear function of body position on thoracic stress is established. The patient's position angle was correlated with chest cavity data:

[0032] in, sgn( ) represents the rate of change in body position at time t; ) is the sign function; k1 and k2 are the coefficients of the nonlinear function.

[0033] In this embodiment, the patient's body position angle is obtained through an inertial measurement unit (IMU), and the chest cavity data is obtained through a wearable piezoelectric sensor.

[0034] In one embodiment of the present invention, the real-time analysis of the thoracic load index includes the following steps: Let the patient's body position angle at time t be θ(t) (0° for supine position, 90° for upright position), and the pressure value measured by the thoracic pressure sensor be P(t). Establish the formula for the thoracic load index analysis:

[0035] in, The thorax load index is defined by β1, β2, β3, and β4, which are weighting coefficients and can be fitted using historical data. The derivative of intrathoracic pressure P(t) with respect to time t represents the rate of change of pressure; it reflects how quickly intrathoracic pressure changes with time and is important for assessing dynamic load. This is a nonlinear function of body position on thoracic stress. The integral term represents the historical pressure decay memory, T is the time window, and λ is the decay factor, simulating the tissue creep effect, that is, the influence of historical pressure on the current load gradually weakens over time.

[0036] In one embodiment of the present invention, the real-time analysis of the sternal incision stress concentration index includes the following steps: A formula for analyzing the stress concentration factor of the sternal incision is established to analyze the stress concentration index of the sternal incision in real time and determine the risk level. The formula is as follows:

[0037] in, The stress concentration index of the sternal incision. Let t be the local pressure in the thoracic cavity. E is the area of ​​the sternal incision (unit: cm², determined according to the surgical record), E is the elastic modulus of the sternal tissue (unit: kPa), and A is the area of ​​the thoracic support (unit: cm²). like If the risk is determined to be high, an early warning will be triggered immediately, reminding the patient to adjust their position and notifying the doctor.

[0038] In this embodiment, the risk assessment criteria are as follows: like It was determined to be low risk, among which, This is the preset maximum stress concentration index for the sternal incision. The maximum value of the sternal incision stress concentration index was determined by statistically analyzing historical data when patients did not seek medical treatment due to sternal incision stress concentration. like The risk level was determined to be medium, and a posture adjustment prompt was sent.

[0039] like If the risk is determined to be high, an early warning will be triggered immediately, reminding the patient to adjust their position and notifying the doctor.

[0040] In one embodiment of the present invention, an electrocardiogram abnormality early warning model is constructed, which combines the thoracic load index and the sternal incision stress concentration index to identify abnormal data and upload it to a cloud platform, including the following steps: Based on continuously acquired ECG data, core ECG features are extracted, including heart rate, heart rhythm, and ST segment deviation. An ECG abnormality early warning model is constructed to identify arrhythmias, myocardial ischemia, and other abnormalities in real time. Formula:

[0041] in, The ECG abnormality warning score at time t (normalized to [0,1]); The heart rate abnormality coefficient at time t; Let t be the coefficient of cardiac arrhythmia. The offset coefficient of segment ST at time t; , , These are the corresponding weights; in this embodiment, the weights can be adjusted according to the patient's condition. =0.3、 =0.4、 =0.3.

[0042] like It issues emergency warnings and immediately pushes data to doctors and patients; In this embodiment, the warning level is: if An emergency warning will be issued, and data will be immediately sent to doctors and patients to initiate emergency care, as the abnormal electrocardiogram is a serious condition that could endanger the patient's life; if A general warning is issued, reminding the patient to rest and upload more ECG data for further assessment of the condition; if This indicates that the electrocardiogram (ECG) is normal, and continued monitoring is sufficient.

[0043] The thoracic load index was normalized based on historical data, and the ECG abnormality early warning score was analyzed by constructing an ECG abnormality early warning model. The normalized thoracic load index, sternal incision stress concentration index, and ECG abnormality warning score are weighted and summed. If the weighted sum is greater than a threshold, the patient data is considered abnormal and uploaded to the cloud platform. In this embodiment, the normalized thoracic load index, sternal incision stress concentration index, and ECG abnormality warning score are weighted and summed. The weight of the normalized thoracic load index is set to 0.4, the weight of the sternal incision stress concentration index is set to 0.3, and the weight of the ECG abnormality warning score is set to 0.3.

[0044] In one embodiment of the present invention, the cloud platform uses a Hidden Markov Model (HMM) combined with real-time data streams, and combines the analysis results of the chest wall load index and electrocardiogram abnormality early warning model to analyze the patient's current rehabilitation stage (including: acute phase, subacute phase, recovery phase, and long-term maintenance phase), and performs dynamic grouping, including the following steps: The rehabilitation phase is divided into acute phase, subacute phase, recovery phase and long-term maintenance phase; the hidden state set S={s1,s2,s3,s4} is defined, where s1,s2,s3,s4 correspond to the acute phase, subacute phase, recovery phase and long-term maintenance phase of the rehabilitation phase, respectively. Set the observed variable Ot as the multidimensional feature vector at time t, including: The analysis included the chest wall load index and the ECG abnormality early warning model, the ECG abnormality early warning score, and the patient's rehabilitation data, including: body position, pressure, and vital signs data; vital signs included: physiological parameters such as heart rate, blood pressure, and respiratory rate. The HMM parameters are set as λ=(A,B,π), and a state transition probability matrix is ​​established: A=[aij], where aij=P(s(t+1)=sj|s(t)=si), and aij is the probability of transitioning to state sj at the next time t+1, given that the current time t is in state si; for example, the probability of transitioning from the acute phase to the subacute phase, or from the recovery phase to the long-term maintenance phase. These probabilities can be determined through statistical analysis of a large amount of patient rehabilitation data, reflecting the natural evolutionary pattern of the rehabilitation process.

[0045] Analyze the observation probability B=[bj(ot)], assuming the observations follow a Gaussian mixture model: , Let t be the probability of observing the observation value ot when the state is sj at time t, where ot is the observation value at time t, which is the data that can be actually measured in the model; M is the number of Gaussian components in the Gaussian mixture model. Let be the mixing coefficient of the m-th Gaussian component in the j-th state, satisfying , representing the weight of each Gaussian component in the mixture model; The mean is The covariance matrix is The multivariate Gaussian distribution probability density function is used to calculate the probability density of observing ot under given mean and covariance; for example, the probability of a combination of specific vital signs, chest wall load index, and ECG abnormality warning results in the acute phase. Gaussian mixture models can flexibly fit the distribution of observed data under different states.

[0046] set up The initial state distribution π represents the initial observation probability of a patient in a given hidden state sj at the start of the rehabilitation process. For example, newly admitted patients may have a higher initial observation probability in the acute phase, while the initial observation probability of directly entering the long-term maintenance phase is lower. The parameters of the Hidden Markov Model (HMM) are updated in real time using an online EM algorithm, updating A, B, and π. Forward and backward probabilities are calculated using a forward-backward algorithm to update the model parameters, enabling the model to adapt to changes in the patient's rehabilitation status. For example, if the patient's ECG abnormality warning score remains low and the chest wall load index gradually stabilizes, the model may adjust the state transition probability, favoring the patient's transition to a later rehabilitation phase.

[0047] The current state after updating the model parameters is decoded using the Viterbi algorithm, with the following formula: ;in, This represents the most likely state at time t, obtained through decoding using the Viterbi algorithm. To represent the observation sequence from time 1 to time t, λ is the parameter set of the model; in this embodiment, the parameter set of the model includes: the state transition probability matrix A, observation probability related parameters (such as those in a Gaussian mixture model). , (e.g., Σjm, etc.) and initial state distribution π, etc.

[0048] Using the Viterbi algorithm on a given observation sequence Given the model parameters λ, find the most likely hidden state sequence at time t to determine the most likely rehabilitation stage the patient is in at the current time. Based on the most likely state at time t Patients are categorized into corresponding rehabilitation stage groups, with different rehabilitation stage groups having different rehabilitation plan templates. Based on the corresponding rehabilitation plan templates and the patients' rehabilitation data, a large AI model is used to generate personalized intervention plans.

[0049] In this embodiment, the setting of rehabilitation plan templates for different rehabilitation stages follows the following rules: Acute phase group: The rehabilitation plan may focus on stabilizing vital signs, relieving acute symptoms, and triggering interventions such as close monitoring and adjustment of drug treatment.

[0050] Subacute phase group: The plan may include gradually increasing the amount of activity, conducting simple rehabilitation training, and interventions may include arranging rehabilitation training courses and adjusting the nursing care plan.

[0051] Recovery group: The focus is on restoring physical function and improving quality of life. The rehabilitation plan may include more complex exercise training, psychological counseling, etc. Intervention measures may include developing personalized exercise plans and providing psychological support.

[0052] Long-term maintenance group: The aim is to maintain the patient's health status and prevent relapse. The rehabilitation plan may include regular follow-up, guidance on healthy lifestyles, etc. Interventions may include sending health reminders and arranging regular check-ups.

[0053] In one embodiment of the present invention, rehabilitation data collected from patients at different time points are used as nodes to establish a rehabilitation data topology graph for each group. Combined with the construction of an SCM structural causal model, the influencing factors of different rehabilitation stages are identified, including the following steps: The rehabilitation data collected for each patient at different time points are used as nodes. For example, the initial assessment data upon admission is one node, the weekly rehabilitation progress assessment data are different nodes, and the final assessment data upon discharge is also one node. Each node contains various rehabilitation-related variables for that stage of the patient. Patients are then grouped according to their rehabilitation stage, forming different rehabilitation stage groups. For each rehabilitation stage group, a rehabilitation data topology graph is constructed. Temporal edges are established based on the time sequence between corresponding patient rehabilitation data nodes. Similarity edges are established between nodes whose semantic similarity is greater than a threshold. Edges between nodes represent the association between data; for example, an edge may exist between a node representing the intensity of rehabilitation training in a previous stage and a node representing functional recovery in a subsequent stage, indicating a possible causal relationship. This relationship can be represented using a directed graph, where the direction of the edge represents the direction of causal influence. In this embodiment, the semantic similarity between rehabilitation data nodes is calculated using cosine similarity, and similarity edges are established between nodes whose semantic similarity is greater than a threshold of 0.8.

[0054] In one embodiment of the present invention, by constructing a structural causal model of SCM, the influencing factors at different stages of rehabilitation are identified, including the following steps: For the rehabilitation data topology diagram constructed for different rehabilitation stages, influencing factor nodes are added. Influencing factor nodes include: static feature nodes and dynamic feature nodes. Static feature nodes include: patient age, gender, surgical method, and underlying disease; dynamic feature nodes include: intervention plan that changes over time. Calculate the mean of each parameter in the rehabilitation data node in the rehabilitation data topology graph, and select rehabilitation data nodes whose data difference from the mean exceeds a preset threshold as key nodes. Establish causal edges between the nearest dynamic feature nodes before the timestamp corresponding to the key nodes. Sort the rehabilitation data nodes in the rehabilitation data topology graph from front to back according to the timestamp, and establish causal edges between static feature nodes and rehabilitation data nodes located in the top 10% of the sorted nodes. Construct a causal model of the SCM structure, using causal edges between nodes as the causal relationships between nodes; The SCM structural causal model is parameterized, and the causal relationship is quantified using the ATE average treatment effect formula: ATE = E[Y|do(X=x)] - E[Y|do(X=x′)], where ATE is the average treatment effect, representing the expected difference in outcome variable Y between the intervention and non-intervention states; E[ [] represents the expectation operation, used to calculate the average of a random variable; Y is the outcome variable; do(X=x) is to intervene in variable X to fix its value at x, do(X=x′) is to intervene in variable X to fix its value at x′; E[Y|do(X=x)] is the expected value of Y under the intervention X=x; In the rehabilitation data topology diagram constructed by grouping different rehabilitation stages, the average treatment effect of each patient's influencing factor node is sorted from high to low. The influencing factor nodes corresponding to the top 20% of the average treatment effects are selected as the main influencing factors for different rehabilitation stages.

[0055] In one embodiment of the present invention, based on the influencing factors of different rehabilitation stages, combined with MPC model predictive control, and using a patient digital twin model as a basis, the rehabilitation plan for different groups is continuously optimized, including the following steps: The temporal edges, similarity edges, and causal edges of the rehabilitation data topology graph at different rehabilitation stages are equivalent to node edges. The degree centrality of each rehabilitation data node is calculated and sorted from high to low. The top 20% of rehabilitation data nodes are selected as high-relevance nodes. The high-relevance nodes of patients are used as termination data nodes. The rehabilitation data nodes connected by a predetermined number of temporal edges before the termination data nodes are used as starting data nodes. The rehabilitation data nodes connected between the starting and termination data nodes by temporal edges are used as intermediate data nodes. The intervention plan corresponding to the dynamic feature node at the most recent moment before the high-relevance data node is selected. Intervention schemes are extracted from the dynamic feature nodes corresponding to the main influencing factors at different rehabilitation stages. Rehabilitation data nodes that establish causal edges with the corresponding influencing factor nodes are selected as starting data nodes. Rehabilitation data nodes that establish temporal edges with the rehabilitation data nodes that establish causal edges are selected as intermediate data nodes. The rehabilitation data node preceding the rehabilitation data node that establishes causal edges in the intermediate data nodes is selected as the ending data node. By using the starting, intermediate, and ending data nodes of different rehabilitation stages, as well as the parameter vectors of the intervention plan, an LSTM neural network is trained to predict rehabilitation data under the influence of the intervention plan on the influencing factor nodes. An MPC model predictive control system is established. For each patient's personalized intervention plan, the parameter vectors of the individual intervention plans are combined into a parameter vector set, which serves as the system's control input variable. The prediction time domain H is determined, and at each time t, the system state for the next H steps is predicted based on the patient's digital twin model. The objective function for this finite-time domain optimization problem is set as follows:

[0056] The constraints are: , The minimum and maximum values ​​are preset for the patient's recovery data at time t+k; , The preset minimum and maximum values ​​are given for each parameter of the intervention plan for the patient at time t+k. in, This represents the control input variables ut and ut+1 for the system. The optimization problem is performed using ut+H-1 to find the set of control input values ​​that minimizes the objective function; H is the length of the finite time domain, which determines the number of time steps or stages to consider starting from the current time t; k is a counter variable used to iterate through the time steps from 0 to H-1 within the finite time domain. The data represents the patient's rehabilitation at time t+k, predicted by the LSTM neural network. It reflects certain state characteristics of the system at that time, such as position and velocity. The rehabilitation data of the patient at time t+H is predicted by the LSTM neural network. Let xref be the control input variable applied to the system at time t+k; let xref be the patient's rehabilitation reference data, i.e., the state the system is expected to reach; Q, R, P are the weight matrices. Q is a weight matrix (usually a symmetric positive semi-definite matrix) used to measure the importance of state deviations. Different state components may correspond to different weights to highlight the importance of deviations of certain state variables in optimization.

[0057] R is a weight matrix (usually a symmetric positive definite matrix) used to measure the cost of the control input, limit the size of the control input, and avoid over-control.

[0058] P is also a weighting matrix (usually a symmetric positive semi-definite matrix) that measures the significance of the deviation between the system state and the reference state at the end of the finite time domain, t+H.

[0059] The optimization problem is solved at each time t to obtain the optimal system control input variables. Based on the optimal system control input variables, the personalized intervention plan is optimized. This process is repeated at preset time intervals, and the personalized intervention plan is continuously optimized based on new rehabilitation references to adapt to changes in the patient's rehabilitation status.

[0060] 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. An information-based remote monitoring and rehabilitation system for thoracic surgery patients after surgery, characterized in that, include: Edge end: The edge gateway establishes a nonlinear dynamic model of body position and thoracic pressure. Based on the rehabilitation data collected by the rehabilitation data acquisition module, it associates the patient's body position angle with the thoracic data, analyzes the thoracic load index and sternal incision stress concentration index in real time, and extracts the core features of the electrocardiogram based on the continuously collected electrocardiogram data to build an electrocardiogram abnormality early warning model. Combined with the thoracic load index and sternal incision stress concentration index, it identifies abnormal data and uploads it to the cloud platform. The cloud platform includes: a dynamic grouping module, an impact analysis module, and a rehabilitation monitoring module; Dynamic grouping module: The cloud platform uses a Hidden Markov Model (HMM) combined with real-time data streams, along with the analysis results of the chest wall load index and ECG abnormality early warning model, to analyze the patient's current rehabilitation stage and perform dynamic grouping. Impact Analysis Module: Using rehabilitation data collected from patients at different time points as nodes, a rehabilitation data topology map is built for each group. Combined with the construction of an SCM structural causal model, the influencing factors at different rehabilitation stages are identified. Rehabilitation monitoring module: Based on the influencing factors of different rehabilitation stages, combined with MPC model prediction and control, and based on the patient digital twin model, the rehabilitation plan for different groups is continuously optimized.

2. The information-based remote monitoring and rehabilitation system for thoracic surgery patients according to claim 1, characterized in that, Also includes: Rehabilitation data acquisition module: Collects patients' rehabilitation data in real time through the patient's wearable device and uploads it to the edge gateway via GPRS / 5G.

3. The information-based remote monitoring and rehabilitation system for thoracic surgery patients according to claim 1, characterized in that, The edge gateway establishes a nonlinear dynamic model of body position and intrathoracic pressure. Based on rehabilitation data collected by the rehabilitation data acquisition module, it correlates the patient's body position angle with intrathoracic data, including the following steps: Based on the nonlinear response mechanism of body position changes to thoracic load, and combining the dynamic correlation between body position angle, rate of change, and intrathoracic pressure, a quadratic nonlinear dynamic model is constructed, as shown in the following formula: in, The local pressure in the thoracic cavity at time t (in kPa); Let t be the patient's body position angle, which is the angle between the patient's torso and the horizontal plane; Let t be the rate of change in body position at time t; , , and These are the model coefficients; It is a non-linear interaction term; Based on rehabilitation data collected by the rehabilitation data acquisition module, a nonlinear function of body position on thoracic stress is established. The patient's position angle was correlated with chest cavity data: in, sgn( ) represents the rate of change in body position at time t; ) is the sign function; k1 and k2 are the coefficients of the nonlinear function.

4. The information-based remote monitoring and rehabilitation system for thoracic surgery patients according to claim 1, characterized in that, Real-time analysis of the thoracic load index includes the following steps: Let the patient's body position angle at time t be θ(t), and the pressure value measured by the thoracic pressure sensor be P(t). Establish the formula for the thoracic load index analysis: in, The thorax load index is represented by β1, β2, β3, and β4, which are weighting coefficients. Let P(t) be the derivative of the intrathoracic pressure P(t) with respect to time t, representing the rate of change of pressure. This is a nonlinear function of body position on thoracic stress. The integral term represents the historical pressure decay memory, where T is the time window and λ is the decay factor.

5. The information-based remote monitoring and rehabilitation system for thoracic surgery patients according to claim 1, characterized in that, Real-time analysis of the sternal incision stress concentration index includes the following steps: A formula for analyzing the stress concentration factor of the sternal incision is established to analyze the stress concentration index of the sternal incision in real time and determine the risk level. The formula is as follows: in, The stress concentration index of the sternal incision. Let t be the local pressure in the thoracic cavity. E is the area of ​​the sternal incision, E is the elastic modulus of the sternal tissue, and A is the area of ​​the thoracic support. like If the risk level is determined to be high, an alert will be immediately triggered, reminding the patient to adjust their position and the doctor to be notified. This is the preset maximum value of the stress concentration index for the sternal incision.

6. The information-based remote monitoring and rehabilitation system for thoracic surgery patients according to claim 1, characterized in that, An early warning model for electrocardiogram abnormalities was constructed, combining the chest wall load index and the sternal incision stress concentration index to identify abnormal data and upload it to the cloud platform. This included the following steps: Based on continuously acquired ECG data, core ECG features are extracted, including heart rate, heart rhythm, and ST segment deviation, to construct an ECG abnormality early warning model. The formula is: in, The ECG abnormality warning score at time t (normalized to [0,1]); The heart rate abnormality coefficient at time t; Let t be the coefficient of cardiac arrhythmia. The offset coefficient of segment ST at time t; , , These are the corresponding weights; if It issues emergency warnings and immediately pushes data to doctors and patients; The thoracic load index was normalized based on historical data, and the ECG abnormality early warning score was analyzed by constructing an ECG abnormality early warning model. The normalized thoracic load index, sternal incision stress concentration index, and electrocardiogram abnormality warning score are weighted and summed. If the weighted sum is greater than the threshold, the patient data is determined to be abnormal and uploaded to the cloud platform.

7. The information-based remote monitoring and rehabilitation system for thoracic surgery patients according to claim 1, characterized in that, The cloud platform uses a Hidden Markov Model (HMM) combined with real-time data streams, along with analysis results from the chest wall load index and ECG abnormality early warning model, to analyze the patient's current rehabilitation stage and dynamically group them, including the following steps: The rehabilitation phase is divided into acute phase, subacute phase, recovery phase and long-term maintenance phase; the hidden state set S={s1,s2,s3,s4} is defined, where s1,s2,s3,s4 correspond to the acute phase, subacute phase, recovery phase and long-term maintenance phase of the rehabilitation phase, respectively. The observed variable Ot is set as a multidimensional feature vector at time t, including: the thoracic load index and the ECG abnormality early warning score analyzed by the ECG abnormality early warning model, as well as the patient's rehabilitation data; Set the HMM parameters to λ=(A,B,π) and establish the state transition probability matrix: A=[aij], where aij=P(s(t+1)=sj|s(t)=si), and aij is the probability of transitioning to state sj at the next time t+1 given that the current time t is in state si; Analyze the observation probability B=[bj(ot)], assuming the observations follow a Gaussian mixture model: , Let t be the probability of observing the observation value ot when the state is sj at time t, where ot is the observation value at time t, which is the data that can be actually measured in the model; M is the number of Gaussian components in the Gaussian mixture model. Let be the mixing coefficient of the m-th Gaussian component in the j-th state, satisfying , representing the weight of each Gaussian component in the mixture model; The mean is The covariance matrix is The probability density function of the multivariate Gaussian distribution; set up The initial state distribution π is given, where π is the initial observation probability of the patient being in a given hidden state sj at the start of the rehabilitation process. The HMM parameters of the Hidden Markov Model (HMM) are updated in real time using the online EM algorithm, and A, B, and π are updated using the online EM algorithm. The current state after updating the model parameters is decoded using the Viterbi algorithm, with the following formula: ;in, This represents the most likely state at time t, obtained through decoding using the Viterbi algorithm. To represent the observation sequence from time 1 to time t, λ is the set of parameters of the model; Using the Viterbi algorithm on a given observation sequence Given the model parameters λ, find the most likely hidden state sequence at time t to determine the most likely rehabilitation stage the patient is in at the current time. Based on the most likely state at time t Patients are categorized into corresponding rehabilitation stage groups, with different rehabilitation stage groups having different rehabilitation plan templates. Based on the corresponding rehabilitation plan templates and the patients' rehabilitation data, a large AI model is used to generate personalized intervention plans.

8. The information-based remote monitoring and rehabilitation system for thoracic surgery patients according to claim 1, characterized in that, Using rehabilitation data collected from patients at different time points as nodes, a rehabilitation data topology graph is built for each group. Combined with the construction of a structural causal model of SCM, the influencing factors of different rehabilitation stages are identified, including the following steps: The rehabilitation data collected from each patient at different time points are used as nodes, and the patients are grouped according to their rehabilitation stage to form different rehabilitation stage groups; For each rehabilitation stage group, a rehabilitation data topology graph is constructed. Temporal edges are established based on the time sequence between the corresponding patient rehabilitation data nodes. Similarity edges are established between nodes whose semantic similarity is greater than a threshold.

9. The information-based remote monitoring and rehabilitation system for thoracic surgery patients according to claim 8, characterized in that, By constructing a structural causal model of SCM, influencing factors at different stages of rehabilitation are identified, including the following steps: For the rehabilitation data topology diagram constructed for different rehabilitation stages, influencing factor nodes are added. Influencing factor nodes include: static feature nodes and dynamic feature nodes. Static feature nodes include: patient age, gender, surgical method, and underlying disease; dynamic feature nodes include: intervention plan that changes over time. Calculate the mean of each parameter in the rehabilitation data node in the rehabilitation data topology graph, and select rehabilitation data nodes whose data difference from the mean exceeds a preset threshold as key nodes. Establish causal edges between the nearest dynamic feature nodes before the timestamp corresponding to the key nodes. Sort the rehabilitation data nodes in the rehabilitation data topology graph from front to back according to the timestamp, and establish causal edges between static feature nodes and rehabilitation data nodes located in the top 10% of the sorted nodes. Construct a causal model of the SCM structure, using causal edges between nodes as the causal relationships between nodes; The SCM structural causal model is parameterized, and the causal relationship is quantified using the ATE average treatment effect formula: ATE = E[Y|do(X=x)] - E[Y|do(X=x′)], where ATE is the average treatment effect, representing the expected difference in outcome variable Y between the intervention and non-intervention states; E[ [] represents the expectation operation, used to calculate the average of a random variable; Y is the outcome variable; do(X=x) is to intervene in variable X to fix its value at x, do(X=x′) is to intervene in variable X to fix its value at x′; E[Y|do(X=x)] is the expected value of Y under the intervention X=x; In the rehabilitation data topology diagram constructed by grouping different rehabilitation stages, the average treatment effect of each patient's influencing factor node is sorted from high to low. The influencing factor nodes corresponding to the top 20% of the average treatment effects are selected as the main influencing factors for different rehabilitation stages.

10. The information-based remote monitoring and rehabilitation system for thoracic surgery patients according to claim 1, characterized in that, Based on the influencing factors at different stages of rehabilitation, combined with MPC model predictive control, and using patient digital twin models as a foundation, the rehabilitation plans for different groups are continuously optimized, including the following steps: The temporal edges, similarity edges, and causal edges of the rehabilitation data topology graph at different rehabilitation stages are equivalent to node edges. The degree centrality of each rehabilitation data node is calculated and sorted from high to low. The top 20% of rehabilitation data nodes are selected as high-relevance nodes. The high-relevance nodes of patients are used as termination data nodes. The rehabilitation data nodes connected by a predetermined number of temporal edges before the termination data nodes are used as starting data nodes. The rehabilitation data nodes connected between the starting and termination data nodes by temporal edges are used as intermediate data nodes. The intervention plan corresponding to the dynamic feature node at the most recent moment before the high-relevance data node is selected. Intervention schemes are extracted from the dynamic feature nodes corresponding to the main influencing factors at different rehabilitation stages. Rehabilitation data nodes that establish causal edges with the corresponding influencing factor nodes are selected as starting data nodes. Rehabilitation data nodes that establish temporal edges with the rehabilitation data nodes that establish causal edges are selected as intermediate data nodes. The rehabilitation data node preceding the rehabilitation data node that establishes causal edges in the intermediate data nodes is selected as the ending data node. By using the starting, intermediate, and ending data nodes of different rehabilitation stages, as well as the parameter vectors of the intervention plan, an LSTM neural network is trained to predict rehabilitation data under the influence of the intervention plan on the influencing factor nodes. An MPC model predictive control system is established. For each patient's personalized intervention plan, the parameter vectors of the individual intervention plans are combined into a parameter vector set, which serves as the system's control input variable. The prediction time domain H is determined, and at each time t, the system state for the next H steps is predicted based on the patient's digital twin model. The objective function for this finite-time domain optimization problem is set as follows: The constraints are: , The minimum and maximum values ​​are preset for the patient's recovery data at time t+k; , The preset minimum and maximum values ​​are given for each parameter of the intervention plan for the patient at time t+k. in, This represents the control input variables ut and ut+1 for the system. The optimization solution is performed using ut+H-1; H is the length of the finite time domain, and k is a counter variable used to traverse the time steps from 0 to H-1 within the finite time domain. The rehabilitation data of the patient at time t+k is predicted by the LSTM neural network. The rehabilitation data of the patient at time t+H is predicted by the LSTM neural network. Let xref be the control input variable applied to the system at time t+k; let xref be the patient's rehabilitation reference data, i.e., the state the system is expected to reach; Q, R, P are the weight matrices. The optimization problem is solved at each time t to obtain the optimal system control input variables. Based on the optimal system control input variables, the personalized intervention plan is optimized. This process is repeated at preset time intervals, and the personalized intervention plan is continuously optimized based on the new rehabilitation reference.

Images