A follow-up scheme dynamic adjustment method and system based on multi-dimensional data

Abstract

This invention relates to the field of rehabilitation follow-up management technology, and discloses a method and system for dynamically adjusting follow-up plans based on multidimensional data. The method includes acquiring historical rehabilitation indicator data and historical follow-up frequency records of subjects and performing time-series dynamic analysis to obtain indicator fluctuation patterns and frequency cumulative effect values; performing causal inference processing based on these to obtain causal association paths and weights; extracting real-time rehabilitation indicator status based on the paths and performing risk prediction to obtain assessment results when fluctuations exceed limits; combining the above characteristics to perform frequency arrangement sequence causal simulation to determine target follow-up sequences with target deviations less than preset thresholds; matching historical trajectories and performing node analysis and timing miss risk assessment to obtain immediate adjustment instructions; finally, retrieving resource records and performing continuity verification to output personalized follow-up plans. This method can achieve precise allocation of follow-up resources and closed-loop adaptive optimization of individual rehabilitation trajectories.

Description

Technical Field

[0001] This invention relates to the field of rehabilitation follow-up management technology, and in particular to a method and system for dynamically adjusting follow-up plans based on multidimensional data. Background Technology

[0002] Currently, chronic disease rehabilitation management, as a key branch of modern medicine, effectively reduces the risk of disease recurrence and improves prognosis through continuous patient tracking and intervention. It is widely used in scenarios such as stroke rehabilitation and postoperative functional recovery. With the development of smart healthcare, monitoring devices integrated with intelligent chips are being used to collect patients' physiological parameters in real time, providing a data foundation for personalized rehabilitation.

[0003] In one existing technology, rehabilitation follow-up typically employs a fixed-cycle scheduling model, or medical staff manually increase or decrease the follow-up frequency based on clinical indicators (such as blood pressure, heart rate, or simple functional scores) at a single point in time. In practice, the system first obtains the patient's current single-dimensional indicator value and compares it with a preset static threshold. If the indicator exceeds the normal range, the system manually shortens the interval between the next follow-up visit; otherwise, the original follow-up plan is maintained. While this method is simple to operate, it neglects the real fluctuations in the individual patient's rehabilitation progress and fails to consider the dynamic causal relationship between multidimensional historical data and past follow-up frequencies, leading to a severe mismatch between follow-up resources and actual needs. Especially when patients are at critical turning points in their rehabilitation, simple linear adjustments struggle to capture the interplay between indicators and the cumulative effect of previous interventions, easily resulting in missed intervention opportunities or wasted medical resources.

[0004] In summary, existing technologies suffer from the problem that adjustments to follow-up protocols are difficult to match with patients' actual dynamic rehabilitation needs. Summary of the Invention

[0005] This invention provides a method and system for dynamically adjusting follow-up plans based on multidimensional data, in order to solve the problem that follow-up plan adjustments are difficult to match with the actual dynamic rehabilitation needs of patients.

[0006] Firstly, to address the aforementioned technical problems, this invention provides a method for dynamically adjusting a follow-up scheme based on multidimensional data, comprising:

[0007] Historical rehabilitation index data and historical follow-up frequency records of the subjects were obtained. Time-series dynamic analysis was performed based on the historical rehabilitation index data and the historical follow-up frequency records to obtain the index fluctuation pattern and frequency cumulative effect value.

[0008] Based on the index fluctuation pattern and the frequency cumulative effect value, causal inference processing is performed to obtain the causal relationship path and the corresponding path weight;

[0009] Feature extraction is performed based on the causal relationship path to obtain real-time rehabilitation indicator status including fluctuation amplitude. When the fluctuation amplitude of the real-time rehabilitation indicator status exceeds a preset fluctuation threshold, rehabilitation reversal risk prediction is performed to obtain risk assessment results.

[0010] Based on the risk assessment results, the cumulative effect value of frequency and the path weight, a causal simulation of the frequency arrangement sequence is performed to obtain the predicted value of the future state, and the target deviation value between the predicted value of the future state and the preset rehabilitation target value is calculated to determine the target follow-up sequence in which the target deviation value is less than the preset convergence threshold.

[0011] Historical rehabilitation trajectories are extracted from the historical rehabilitation indicator data. The historical rehabilitation trajectories are matched and analyzed with the target follow-up sequence to obtain the trajectory matching degree. Intervention nodes are parsed based on the trajectory matching degree to obtain the analysis results. The timing miss risk is assessed based on the analysis results to obtain an immediate adjustment instruction containing time parameters.

[0012] Resource records are retrieved based on the real-time adjustment instructions to obtain the resource load status. The real-time adjustment instructions are then validated for continuity based on the resource load status, and a final personalized follow-up plan is output.

[0013] Secondly, the present invention provides a dynamic adjustment system for follow-up schemes based on multidimensional data, comprising:

[0014] The time-series analysis module is used to acquire the subject's historical rehabilitation index data and historical follow-up frequency records, and to perform time-series dynamic analysis based on the historical rehabilitation index data and historical follow-up frequency records to obtain the index fluctuation pattern and frequency cumulative effect value.

[0015] The causal inference module is used to perform causal inference processing based on the index fluctuation pattern and the frequency cumulative effect value to obtain the causal relationship path and the corresponding path weight.

[0016] The risk prediction module is used to extract features based on the causal relationship path to obtain a real-time rehabilitation indicator status including fluctuation amplitude, and to perform rehabilitation reversal risk prediction when the fluctuation amplitude of the real-time rehabilitation indicator status exceeds a preset fluctuation threshold, thereby obtaining a risk assessment result.

[0017] The causal simulation module is used to perform frequency arrangement sequence causal simulation based on the risk assessment results, the frequency cumulative effect value and the path weight to obtain the future state prediction value, calculate the target deviation value between the future state prediction value and the preset rehabilitation target value, and determine the target follow-up sequence where the target deviation value is less than the preset convergence threshold.

[0018] The instruction generation module is used to extract historical rehabilitation trajectories from the historical rehabilitation indicator data, perform matching analysis between the historical rehabilitation trajectories and the target follow-up sequence to obtain the trajectory matching degree, perform intervention node parsing based on the trajectory matching degree to obtain the parsing result, and perform timing miss risk assessment based on the parsing result to obtain an immediate adjustment instruction containing time parameters.

[0019] The resource verification module is used to retrieve resource records according to the real-time adjustment instruction, obtain the resource load status, verify the continuity of the real-time adjustment instruction according to the resource load status, and output the final personalized follow-up plan.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] (1) This invention uses a time-series analysis algorithm to align multidimensional rehabilitation indicators and follow-up frequency, dividing them into active and stable fluctuation stages, and extracting time-series feature vectors representing the correlation between follow-up density and indicator change rate to input into a regression analysis model. This scheme obtains the cumulative effect value of frequency on indicators and the dynamic fluctuation pattern of each stage through weighted cumulative quantization, thereby realizing the dynamic and precise adjustment of rehabilitation follow-up frequency. It effectively solves the problem of mismatch between follow-up resources and patients' actual needs in traditional methods, and significantly improves the allocation efficiency of medical resources while ensuring that key rehabilitation stages receive sufficient attention.

[0022] (2) This invention constructs a directed acyclic graph of multidimensional data and frequency variables using the PC algorithm, and obtains a debiased balanced sample set by blocking confounding factors based on the potential causal topology. Then, it analyzes and calculates the path weights and locks the node connection sequence with the highest value. This scheme eliminates the confounding bias in traditional correlation analysis by identifying the most decisive causal path for individual recovery (such as "increased frequency → decreased anxiety → improved function"), accurately identifies the key causal chain that dominates recovery, makes follow-up adjustments more scientifically based, and significantly enhances the pertinence and effectiveness of rehabilitation interventions.

[0023] (3) This invention uses a dynamic programming algorithm to extract the historical time-series matching segment with the highest overlap with the current trajectory, and combines patient compliance data and rehabilitation rate decay values ​​to calculate the risk of missed opportunities, thereby generating adjustment instruction codes containing compensatory frequencies and mapping them to the action library. This scheme effectively offsets the negative impact of rehabilitation rate decay by densely inserting real-time adjustment instructions within the remaining effective time window, prevents the risk of rehabilitation trajectory deviation caused by missed intervention opportunities, and ultimately significantly improves the achievement rate of patients' rehabilitation goals and maintains the continuity of services. Attached Figure Description

[0024] Figure 1This is a schematic diagram of the dynamic adjustment method for follow-up schemes based on multidimensional data provided in the first embodiment of the present invention;

[0025] Figure 2 This is a schematic diagram of the dynamic adjustment system structure of the follow-up scheme based on multidimensional data provided in the second embodiment of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0027] Reference Figure 1 The first embodiment of the present invention provides a method for dynamically adjusting a follow-up scheme based on multidimensional data, comprising the following steps:

[0028] S11, Obtain the subject's historical rehabilitation index data and historical follow-up frequency records, and perform time-series dynamic analysis based on the historical rehabilitation index data and historical follow-up frequency records to obtain the index fluctuation pattern and frequency cumulative effect value.

[0029] S12, perform causal inference processing based on the index fluctuation pattern and the frequency cumulative effect value to obtain the causal relationship path and the corresponding path weight;

[0030] S13, feature extraction is performed based on the causal relationship path to obtain the real-time rehabilitation indicator status including fluctuation amplitude, and when the fluctuation amplitude of the real-time rehabilitation indicator status exceeds the preset fluctuation threshold, rehabilitation reversal risk prediction is performed to obtain risk assessment results.

[0031] S14. Based on the risk assessment results, the cumulative effect value of frequency and the path weight, a frequency arrangement sequence causal simulation is performed to obtain the predicted value of the future state, and the target deviation value between the predicted value of the future state and the preset rehabilitation target value is calculated. A target follow-up sequence in which the target deviation value is less than the preset convergence threshold is determined.

[0032] S15, extract historical rehabilitation trajectory from the historical rehabilitation indicator data, perform matching analysis between the historical rehabilitation trajectory and the target follow-up sequence to obtain trajectory matching degree, perform intervention node analysis based on the trajectory matching degree to obtain analysis results, and perform timing miss risk assessment based on the analysis results to obtain immediate adjustment instructions containing time parameters;

[0033] S16, retrieve resource records according to the real-time adjustment instruction, obtain resource load status, verify the continuity of the real-time adjustment instruction according to the resource load status, and output the final personalized follow-up plan.

[0034] In step S11, historical rehabilitation index data and historical follow-up frequency records of the subjects are acquired. Time-series dynamic analysis is performed based on the historical rehabilitation index data and the historical follow-up frequency records to obtain the index fluctuation pattern and frequency cumulative effect value, including:

[0035] The historical follow-up frequency records are time-aligned to obtain a follow-up time interval sequence. The follow-up time interval sequence and the historical rehabilitation index data are then divided into intervals to obtain an active fluctuation phase and a stable maintenance phase.

[0036] Based on the active fluctuation phase and the stable maintenance phase, extract the time-series feature vectors that correlate follow-up density with the rate of change of indicators;

[0037] The time-series feature vector is processed using a pre-trained regression analysis model to obtain the frequency influence factor, and the frequency influence factor is weighted and accumulated to obtain the index fluctuation pattern and frequency cumulative effect value.

[0038] First, the acquired historical follow-up frequency records of the subjects are time-aligned. These records contain precise date stamps of each follow-up visit during the subject's past rehabilitation cycles. Linear interpolation is used to resample these date stamps, aligning irregular follow-up time points to a uniform weekly or monthly time grid, resulting in aligned follow-up records. The follow-up time interval sequence is obtained by calculating the difference in days between two adjacent aligned follow-up date stamps. In one implementation, if the subject's original follow-up dates in the past six months were week 1, week 3, week 8, and week 12, the resulting follow-up time interval sequence after alignment would be 14 days, 35 days, and 28 days.

[0039] Next, the data is divided into intervals based on the follow-up time interval sequence and the historical rehabilitation indicator data. The historical rehabilitation indicator data includes the subject's physiological parameters, psychological scale scores, and functional recovery indicators at the corresponding follow-up nodes. Using the follow-up time axis as a benchmark, periods where the fluctuation value of the rehabilitation indicator exceeds a preset fluctuation deviation threshold are defined as active fluctuation phases, and periods where the fluctuation value is lower than or equal to the preset fluctuation deviation threshold are defined as stable maintenance phases, thus obtaining active fluctuation phases and stable maintenance phases. For example, if the subject's joint range of motion changes by more than five degrees between the third and eighth weeks, this interval is determined to be an active fluctuation phase; if the change remains within two degrees after the twelfth week, it is determined to be a stable maintenance phase. It should be noted that the preset fluctuation deviation threshold needs to be set according to the clinical nursing standards of the specific rehabilitation disease and the basic physiological characteristics of the subject. The setting basis is usually based on the industry-standard rehabilitation assessment criteria (such as the Fugl-Meyer assessment method for stroke rehabilitation). In this embodiment, for joint function rehabilitation, the preset fluctuation deviation threshold is usually set between three and five degrees. This value is obtained by statistically analyzing the subject's routine measurement errors and physiological fluctuation range over the past three months and selecting the 95th percentile as the benchmark. Those skilled in the art will understand that the preset fluctuation deviation threshold is not a fixed constant and can be dynamically adjusted according to the accuracy specifications of different monitoring devices.

[0040] Subsequently, a time-series feature vector relating follow-up density and indicator change rate is extracted based on the active fluctuation phase and the stable maintenance phase. The follow-up density is the reciprocal of the number of follow-ups per unit time, and the indicator change rate is the slope of the rehabilitation indicator value over time. These two are combined in a time series to generate a multi-dimensional vector containing the average follow-up interval and the peak value of the change rate, thus obtaining the time-series feature vector. Next, a pre-trained regression analysis model is used to process the time-series feature vector. This regression analysis model is trained based on historical rehabilitation data from a large number of patients with similar diseases. By inputting the time-series feature vector into this model, a quantified frequency influence factor is output.

[0041] It should be noted that the regression analysis model is used to predict the corresponding frequency influence factor based on the input time-series feature vector. The model structure can be flexibly selected according to the data characteristics. For example, linear regression can be used to handle linear relationships, or nonlinear models such as random forest regression can be used to handle complex associations. For example, a random forest regression model can be used, and the model structure consists of multiple decision trees, such as constructing 100 regression trees. The maximum depth of each tree is set to 10, the minimum number of samples required for node splitting is set to 5, and the minimum number of samples for leaf nodes is set to 2. The training data comes from the complete follow-up records of 500 similar rehabilitation patients stored in the hospital information system. The time-series feature vectors of each follow-up stage are extracted from them, and each feature vector is labeled with a frequency influence factor between 0 and 1 as a label value according to the patient's subsequent rehabilitation effect. During the training process, the model uses out-of-bag error as the convergence condition. The training terminates when the out-of-bag error no longer decreases after 10 consecutive iterations.

[0042] Preferably, the frequency influence factors are weighted and accumulated to obtain the index fluctuation pattern and the frequency cumulative effect value. Specifically, weights are assigned according to the distance of the follow-up node from the current time. The weight allocation follows an exponential decay logic, that is, the weight decays by the power of the decay coefficient, where the exponent is the interval number from the current node. The closer the follow-up node is to the current time, the higher its weight value. In one implementation, the decay coefficient is set to 0.85. If the weight of the current follow-up node is set to 1, then the weight of the previous follow-up node is 1 multiplied by 0.85, the weight of the node before that is 1 multiplied by the square of 0.85, and so on, to obtain the weight value of each follow-up node. The product of the frequency influence factor of each node and the corresponding weight is summed to obtain the frequency cumulative effect value, and the index fluctuation pattern representing the patient's characteristics at different recovery stages is output simultaneously. It is worth noting that the cumulative effect value of frequency provides a quantitative reference indicator that reflects the contribution of past follow-up interventions to the current recovery status. Its essence is an intervention utility assessment based on the time decay logic. In practical applications, this value can be used to measure whether the current follow-up frequency is sufficient to support the patient's recovery rate.

[0043] It should be noted that the indicator fluctuation pattern described in this embodiment refers to the trend characteristics of indicator changes exhibited by the subjects at different stages of rehabilitation. Its technical essence is a health status trajectory representation obtained based on time-series feature vector clustering analysis.

[0044] In step S12, causal inference is performed based on the index fluctuation pattern and the frequency cumulative effect value to obtain the causal relationship path and the corresponding path weight, including:

[0045] A multidimensional causal analysis dataset is constructed based on the index fluctuation pattern and the frequency cumulative effect value.

[0046] The PC algorithm was used to construct the topology of the multidimensional causal analysis dataset to obtain a directed acyclic graph;

[0047] The obfuscation factor blocking process is performed on the directed acyclic graph to obtain a debiased balanced sample set, and the path parsing is performed on the debiased balanced sample set to obtain the causal association path and the corresponding path weight.

[0048] First, a multidimensional causal analysis dataset is constructed based on the fluctuation patterns of the indicators and the cumulative effect values ​​of the frequencies. Specifically, the fluctuation patterns of the indicators corresponding to each rehabilitation stage of the subjects are used as classification labels, and these labels are matrix-integrated with the corresponding cumulative effect values ​​and multidimensional rehabilitation indicators such as heart rate, blood pressure, anxiety scores, and joint range of motion acquired simultaneously. In one implementation, for a patient who has undergone knee surgery, the above indicators at the fourth, tenth, and eighteenth weeks post-surgery are arranged in rows, and the follow-up frequency variables for the corresponding time periods are embedded as independent columns, thereby forming a multidimensional causal analysis dataset representing the interaction between rehabilitation characteristics and follow-up intensity. It should be noted that the initialization of the multidimensional causal analysis dataset relies on historical data generated by the subjects during diagnosis and treatment, which can be extracted according to the data specifications of medical institutions by those skilled in the art.

[0049] Next, the PC algorithm is used to construct a topology for the multidimensional causal analysis dataset, resulting in a directed acyclic graph (DAG). The PC algorithm, also known as the Peter-Clark algorithm, is a causal structure learning algorithm based on conditional independence testing. It is a statistical method for learning causal structures, primarily used to construct DAGs from observational data to reveal causal relationships between variables. This algorithm progressively simplifies the graph structure through conditional independence testing, ultimately obtaining a graph model that reflects the causal relationships between variables. Specifically, first, a complete undirected graph containing all variable nodes is constructed based on the multidimensional causal analysis dataset. Then, by performing conditional independence testing on each pair of connected nodes in the graph, edges that do not satisfy statistical significance are removed. Finally, the remaining edges are oriented according to collision structure identification and cycle avoidance rules, thereby generating a DAG that reveals hierarchical dependencies between variables. For example, this graph structure might show a directional path where follow-up frequency points to a decrease in anxiety scores, and the anxiety scores further point to an increase in joint mobility.

[0050] Subsequently, a debiased balanced sample set is obtained by blocking confounding factors based on the directed acyclic graph (DAG). Specifically, by locating the backdoor path connecting follow-up frequency and rehabilitation indicators in the DAG, confounding variables such as subject age, severity of underlying diseases, and preoperative functional level are identified. These variables are then controlled using a stratified matching method to eliminate spurious causal associations caused by differences in sample baseline characteristics, thereby obtaining a debiased balanced sample set with balanced distribution of all variables.

[0051] Next, path analysis is performed on the debiased balanced sample set to obtain the causal relationship paths and their corresponding path weights. Preferably, path analysis is used to perform regression calculations on the debiased balanced sample set, calculating the standardized partial regression coefficients between adjacent nodes in the causal chain as the direct effect values. Specifically, the deviation values ​​are obtained by subtracting the average values ​​from the dependent and independent variables in the causal chain, and then standardized by dividing the deviation values ​​by their respective standard deviations. Subsequently, the covariance between the standardized dependent and independent variables is calculated and divided by the standardized variance of the independent variables. The resulting value is the standardized partial regression coefficient representing the intensity of direct intervention between two adjacent nodes. Then, the partial regression coefficients between all adjacent nodes in a specific causal chain are multiplied together, and the product is the path weight of that path. In one possible implementation, in the path from follow-up frequency via anxiety score to joint mobility, if the partial regression coefficient of frequency on anxiety is -0.65 and the partial regression coefficient of anxiety on joint mobility is -0.48, then the total weight of the causal relationship path is 0.312.

[0052] It should be further explained that the path weights are calculated by multiplying standardized partial regression coefficients. The magnitude of the path weights reflects the strength of the indirect influence of the intervention variable on the outcome variable through the mediating variable, such as the strength of the indirect influence of follow-up frequency on joint mobility through anxiety scores. In practical applications, if the weight of a certain causal path is lower than the preset significance threshold (e.g., 0.1), it can be considered that the path has a weak explanatory power for rehabilitation outcomes. In subsequent frequency sequence simulations, the system can appropriately reduce the reference weight of the path or treat it as a secondary path.

[0053] It is worth noting that the causal path described in this embodiment refers to a logical chain starting from the follow-up frequency node, passing through one or more mediating indicator nodes, and ultimately pointing to the rehabilitation outcome node. It should be noted that the path weights are set according to the standardized path coefficient evaluation criterion. This criterion typically stipulates that a path coefficient with an absolute value between 0.1 and 0.3 indicates a low-intensity influence, between 0.3 and 0.5 indicates a moderate-intensity influence, and above 0.5 indicates a high-intensity influence. In practical applications, the specific method for determining the values ​​is to perform least squares regression estimation on a biased balanced sample set, select the regression coefficient that minimizes the sum of squared residuals as the initial value, and then adjust it based on the subject's historical rehabilitation response sensitivity.

[0054] In step S13, feature extraction is performed based on the causal relationship path to obtain a real-time rehabilitation indicator status including fluctuation amplitude. When the fluctuation amplitude of the real-time rehabilitation indicator status exceeds a preset fluctuation threshold, rehabilitation reversal risk prediction is performed to obtain a risk assessment result, including:

[0055] The current physiological state vector is obtained by collecting node data according to the causal relationship path, and the Euclidean distance of the current physiological state vector is calculated to obtain the fluctuation amplitude.

[0056] If the fluctuation amplitude exceeds the preset fluctuation threshold, the time series data of the rehabilitation indicators are extracted from the real-time rehabilitation indicator status, and the time series data of the rehabilitation indicators are input into the pre-trained autoregressive moving average model for prediction to obtain the probability of rehabilitation recurrence.

[0057] The probability of recurrent rehabilitation was normalized to obtain the risk assessment result.

[0058] First, feature extraction is performed based on the causal relationship path determined in the previous steps. Specifically, the subject's current physiological data is acquired in real time at key rehabilitation nodes involved in the causal relationship path using medical monitoring equipment. The acquired core indicator data, such as heart rate, anxiety score, and joint range of motion, are then vectorized and encapsulated to obtain the current physiological state vector.

[0059] Next, the Euclidean distance of the current physiological state vector is calculated to obtain the fluctuation amplitude. Specifically, the square root of the sum of the squares of the differences in each dimension between the current physiological state vector and a reference vector, such as the state vector of the previous period or the baseline vector of the ideal state for the current period obtained based on historical data statistics, is taken to obtain the original distance. To eliminate the influence of different indicator dimensions, the values ​​of each dimension of the vector need to be Z-score standardized or max-min normalized before calculation. The fluctuation amplitude is the standardized / normalized distance value. If the fluctuation amplitude exceeds a preset fluctuation threshold, it is determined that the subject's current rehabilitation status has abnormal fluctuations, and the rehabilitation reversal risk prediction process is formally triggered. In one implementation, rehabilitation indicator time-series data is first extracted from the real-time rehabilitation indicator status. Then, a pre-trained prediction model is used to extrapolate the trend of the rehabilitation indicator time-series data. Subsequently, the deviation of the predicted trend from the ideal rehabilitation trajectory is calculated, and finally, a quantitative score representing the level of risk is output. If the fluctuation amplitude does not exceed the preset fluctuation threshold, the current rehabilitation monitoring status is maintained, and the reversal risk prediction is not triggered. It should be noted that the preset fluctuation threshold needs to be set according to the physiological stability characteristics of different rehabilitation stages. The setting is usually based on the industry-standard rehabilitation monitoring standards. In this embodiment, the typical range of the preset fluctuation threshold is between 4.0 and 6.0. Specifically, the range of physiological state vector changes of the subject in the past two weeks is collected, and the 80th percentile is selected as the benchmark.

[0060] In this process, the pre-trained prediction model employs a Long Short-Term Memory (LSTM) network model, which essentially utilizes LSTM units to capture the nonlinear trend characteristics of rehabilitation indicators over a long time span. The model is trained by collecting a large amount of historical subject time-series data with similar symptoms and complete rehabilitation trajectories as the training set. The normalized indicator sequence is used as input, and the predicted indicator value for the next time step is used as the label. Gradient descent is used to minimize the mean squared error loss function, and the parameter weights of the forget gate, input gate, and output gate within the model are adjusted using the backpropagation algorithm. Regarding parameter selection, cross-validation is used to determine the number of hidden layer nodes to be between 64 and 128, and the learning rate is set between 0.001 and 0.01. In use, the subject's current and past five sampling points of rehabilitation indicator time-series data are input into the model, and the model automatically outputs the predicted rehabilitation trend value for the next follow-up period.

[0061] Subsequently, the time-series data of the rehabilitation indicators are input into a pre-trained autoregressive moving average model for prediction, yielding the probability of relapse. The training process of the autoregressive moving average model involves collecting a large number of historical rehabilitation indicator sequences from subjects with similar symptoms as a training set. Maximum likelihood estimation is used to estimate the parameters of the autoregressive coefficients and moving average coefficients in the model. The Akaike Information Criterion is used for parameter selection, choosing the order that minimizes the criterion value as the optimal model order. In use, the subject's current time-series data of rehabilitation indicators is used as input, and the model automatically captures the interaction between the linear trend of indicator fluctuations and random errors. Further, the predicted value output by the model is compared with the subject's historical lower limit of rehabilitation to obtain a prediction deviation value. This prediction deviation value is then divided by the historical standard deviation of fluctuations to obtain a standardized deviation score. This score is used as input and calculated using the logistic function. The calculation result is mapped to the interval between zero and one, thus obtaining the probability of relapse. It is worth noting that the lower limit threshold for the safe recovery status can be the mean of the subject's previous stable period indicators minus twice the standard deviation, or the minimum target value for the recovery stage specified in clinical guidelines. The autoregressive moving average model mentioned in this embodiment refers to a mathematical prediction model based on the ARMA algorithm. Its technical essence is to use a linear combination of historical observations and random error terms to fit the dynamic characteristics of time series data.

[0062] Finally, the probability of recurrent rehabilitation is normalized to obtain the risk assessment result. Preferably, the max-min normalization method is used to convert the calculated probability of recurrent rehabilitation into a standardized risk assessment result. During the normalization process, adjustments are made based on the postoperative stage weights. Specifically, according to the subject's current postoperative days, the corresponding weight coefficient is retrieved from a preset stage weight table; for example, the weight for the early postoperative recovery stage is set to 1.2, and for the stable period to 0.8. The normalized score is then multiplied by the weight coefficient to obtain the final risk assessment result. It should be noted that the postoperative stage weight table is designed with reference to a large amount of clinical rehabilitation statistics, reflecting the differences in risk sensitivity among subjects at different stages.

[0063] It should be clarified that the probability of relapse and the risk of reversal of rehabilitation mentioned in this embodiment refer to the same concept in practical applications, namely, the possibility that the subject's rehabilitation status will deteriorate. In specific implementation, the probability of relapse can be normalized and used as the core quantitative indicator of the risk assessment result, and weighted and corrected according to the postoperative stage of the patient. Those skilled in the art will understand that in different application scenarios, the risk assessment result can also be integrated with other clinical indicators, such as the incidence of complications and readmission risk score, for comprehensive judgment.

[0064] In step S14, based on the risk assessment results, the cumulative effect value of frequency, and the path weight, a frequency arrangement sequence causal simulation is performed to obtain a predicted value for the future state. The target deviation value between the predicted value for the future state and a preset rehabilitation target value is calculated. A target follow-up sequence is determined where the target deviation value is less than a preset convergence threshold, including:

[0065] Based on the risk assessment results and the frequency cumulative effect value, a step size mapping is performed to obtain multiple sets of candidate frequency arrangement sequences;

[0066] Based on the path weights, the state transition probabilities of the multiple candidate frequency arrangement sequences are calculated to obtain the predicted values ​​of the future states.

[0067] The Euclidean distance between the predicted future state value and the preset rehabilitation target value is calculated to obtain the target deviation value, and the candidate frequency arrangement sequence whose target deviation value is less than the preset convergence threshold is determined as the target follow-up sequence.

[0068] First, a step-size mapping is performed between the risk assessment results and the frequency cumulative effect value to obtain multiple candidate frequency arrangement sequences. Intervention timestamp sequences and single intervention intensity values ​​from past follow-up interventions are extracted from historical records, and the cumulative effect value at the current moment is calculated. The cumulative effect value at the current moment is calculated using a time-decaying cumulative model. Specifically, each intervention intensity is exponentially decayed and weighted according to the number of days from the current moment, and then summed. In one implementation, a daily decay rate of 5% is set, and each intervention intensity is multiplied by a power of -0.05 days (the natural constant), and then all products are weighted and summed to obtain the cumulative effect value at the current moment.

[0069] Next, the cumulative effect value is fitted to the standard rehabilitation response curve to calculate the missing effect difference. The fitting process involves using the least squares method to map the actual cumulative effect value onto the standard rehabilitation response curve constructed based on a large amount of historical data from similar patients, finding the corresponding ideal time point and its corresponding ideal effect value. Then, the missing effect difference is obtained by subtracting the actual cumulative effect value from the ideal effect value. Using the missing effect difference and the risk assessment result as input parameters, the incremental coefficient of the follow-up frequency is determined, and this incremental coefficient is multiplied by the current follow-up cycle to obtain the follow-up frequency adjustment range. The obtained follow-up frequency adjustment range is mapped to a preset health status association path to generate multiple candidate frequency arrangement sequences with different time step distributions. It should be noted that the mapping logic of the incremental coefficient is based on a weighted judgment of risk level and missing effect degree.

[0070] Subsequently, the state transition probability is calculated for the multiple candidate frequency arrangement sequences based on the path weights to obtain the predicted future state value. The preset health state association path is a topological chain composed of multiple health state nodes representing the level of rehabilitation. Its setting is based on the clinical guidelines that divide the rehabilitation process into stages such as early postoperative recovery, basic functional recovery, and long-term stability, and defines the transition logic between stages. The causal inference model adopts a structural causal model. Its training process uses a historical debiased balanced sample set, determines the transition strength between each health state node through maximum likelihood estimation, and uses path coefficients as the basis for parameter selection. During extrapolation, each candidate frequency arrangement sequence is used as an intervention variable input, combined with the extracted response delay time value, and the Markov chain Monte Carlo method is used to simulate the subject's rehabilitation state score in the future preset period, thus obtaining the predicted future state value.

[0071] In one embodiment, the causal inference processing can employ a causal structure learning model based on the PC algorithm. The model input is a multidimensional causal analysis dataset constructed from index fluctuation patterns and frequency cumulative effect values. The dataset includes follow-up frequency, various rehabilitation indicators, and potential confounding variables (such as age, disease duration, etc.). The model first constructs a completely undirected graph containing all variables, and then gradually removes edges that do not have significant associations through conditional independence tests. The significance level (e.g., p-value) of the retained edges is usually set to 0.05. After the undirected graph is constructed, the model orients the edges according to collision structure identification rules and acyclic constraints, and finally outputs a directed acyclic graph reflecting the causal hierarchy between variables. Based on this graph structure, each causal path is further analyzed, and the standardized regression coefficients between adjacent variables are calculated using linear regression as direct effects. The path weight is the product of all direct effects on the path. To prevent overfitting, the model uses the BIC scoring criterion to penalize the complexity of the graph and selects the graph with the best score as the final output. The entire training process requires no manual intervention and is entirely based on data-driven autonomous discovery of causal structures.

[0072] Finally, the Euclidean distance between the predicted future state value and the preset rehabilitation target value is calculated to obtain the target deviation value. The preset rehabilitation target value is usually taken as the full score (e.g., 100 points) of the ideal rehabilitation curve at the corresponding time point or its 90th percentile value. Candidate frequency arrangement sequences with target deviation values ​​less than a preset convergence threshold are determined as target follow-up sequences. The time point data in these sequences are extracted to calculate the time difference between adjacent time points, resulting in the optimized follow-up interval sequence. It is worth noting that the preset convergence threshold is usually set based on industry-standard rehabilitation achievement assessment criteria, and in this embodiment, it is typically set between 3.0 and 5.0.

[0073] It is worth noting that the deduction of the target follow-up sequence is based on the simulation results of the structural causal model under ideal conditions. In this invention, the simulation value is mainly used to reflect the theoretical difference in the contribution of different follow-up frequencies to the achievement rate of rehabilitation goals. In the actual rehabilitation process, the physiological feedback of the subjects will be affected by multiple complex factors such as environmental interference and individual metabolic differences. Those skilled in the art will understand that, without departing from the core idea of ​​this invention, the structural causal model can be replaced with a more complex causal deduction model based on deep reinforcement learning, or a dynamic correction coefficient based on Bayesian networks can be introduced to further improve the prediction accuracy under extreme physiological fluctuations.

[0074] In step S15, historical rehabilitation trajectories are extracted from the historical rehabilitation indicator data. The historical rehabilitation trajectories are matched with the target follow-up sequence to obtain a trajectory matching degree. Intervention node analysis is performed based on the trajectory matching degree to obtain analysis results. Based on the analysis results, a timing miss risk assessment is conducted to obtain immediate adjustment instructions containing time parameters, including:

[0075] The overlap between the historical rehabilitation trajectory and the target follow-up sequence is calculated using a dynamic programming algorithm to obtain the trajectory matching degree, and time-series matching segments are extracted based on the trajectory matching degree.

[0076] Based on the time-series matching segment, node parsing is performed to obtain a preset intervention node, real-time time point is obtained, and the remaining effective duration is calculated based on the real-time time point and the preset intervention node.

[0077] If the remaining effective duration is less than the preset response limit, the subject's historical follow-up records are obtained, and compliance analysis is performed based on the historical follow-up records to obtain historical compliance data.

[0078] The timing risk value is calculated based on the historical compliance data. If the timing risk value exceeds the preset immediate intervention threshold, an immediate adjustment instruction including time parameters is output.

[0079] First, historical rehabilitation trajectories reflecting the subject's rehabilitation trend are extracted from the historical rehabilitation indicator data. The dynamic time warping algorithm in dynamic programming is used to calculate the cumulative distance between the target follow-up sequence and each trajectory in the historical rehabilitation trajectory database. A two-dimensional distance matrix is ​​constructed, where the horizontal and vertical axes correspond to the target follow-up sequence and historical trajectory points in the database, respectively. The local Euclidean distance of each grid point in the matrix is ​​calculated, and a minimum-cost path from the matrix's starting point to its ending point is found using a recursive formula. The sum of the distances of all points on this path is the cumulative distance. The trajectory segment with the smallest cumulative distance and a matching value greater than a preset matching threshold is selected as the temporal matching segment with the highest overlap, thus obtaining the trajectory matching degree. In one implementation, the preset matching threshold is set based on the distribution density of the historical similar case database. By calculating the probability density function of the cumulative distance between all pairwise trajectories in the database, the top 15% quantile is selected as the upper limit of the threshold to ensure that the matching results are statistically significant. If the calculated trajectory matching degree is lower than this threshold, the search scope is expanded.

[0080] Next, node parsing is performed on the time-series matching segments to obtain preset intervention nodes. By scanning the rate of change of indicators in the time-series matching segments, the time points with the largest slope of indicator improvement are located, and the specific intervention operations (such as intensive training or physical therapy) performed at these time points in the historical records are retrieved. These key time points are defined as preset intervention nodes. Subsequently, real-time time points are obtained, and the difference between the real-time time points and the preset intervention nodes is calculated to obtain the remaining effective duration. If the remaining effective duration is less than the preset response limit, the subject's historical follow-up records are obtained, and compliance analysis is performed based on the historical follow-up records to obtain historical compliance data.

[0081] Subsequently, the risk of missed opportunity is calculated based on the historical compliance data to obtain a missed opportunity risk value. Specifically, the recovery rate decay value, which characterizes the degree of slowdown in the subject's recovery speed, is divided by the historical compliance data to obtain the probability of the recovery window closing due to insufficient participation, which is the missed opportunity risk value. If the missed opportunity risk value exceeds a preset immediate intervention threshold, a compensatory follow-up frequency is generated based on the recovery rate decay value. The specific generation logic is to use the difference between the recovery rate decay value and the standard recovery rate as the compensation benchmark, and calculate the additional follow-up visits required within the remaining effective time using an up-rounding function to obtain the compensatory follow-up frequency. It should be noted that the preset immediate intervention threshold is usually set based on industry-standard rehabilitation risk warning criteria; in this embodiment, the threshold is usually set between 0.7 and 0.85, specifically by analyzing past cases of rehabilitation stagnation, statistically analyzing their risk score distribution before the window closed, and selecting the mean plus one standard deviation as the benchmark.

[0082] Finally, the compensatory follow-up frequency is mapped to a pre-set action library to output real-time adjustment instructions containing time parameters. The pre-set action library contains standardized intervention templates for different risk levels, such as instruction sets like "urgently increase in-hospital rehabilitation" or "increase remote video guidance"; the data in this library is obtained from clinical rehabilitation guidelines and is updated quarterly through batch import and verification of evidence-based medicine cases.

[0083] It is worth noting that the dynamic time warping algorithm described in this embodiment essentially solves the matching problem of nonlinear scaling of rehabilitation sequences on the time axis, ensuring that similar rehabilitation trajectories can be identified even if the follow-up time points are not completely consistent. It should also be noted that the generation of compensatory follow-up frequency aims to forcibly pull back the deviated rehabilitation curve before the effective duration ends through short-term, high-frequency interventions.

[0084] In step S16, resource records are retrieved according to the real-time adjustment instruction to obtain the resource load status, and the real-time adjustment instruction is continuously verified based on the resource load status to output the final personalized follow-up plan, including:

[0085] Based on the time parameter of the real-time adjustment instruction, the medical resource database is indexed to obtain resource occupancy duration data, and load rate mapping is performed based on the resource occupancy duration data to obtain the resource load status;

[0086] If the resource load status does not exceed the preset carrying limit, then the equipment distribution characteristics are analyzed to obtain the time fragmentation index;

[0087] Based on the time fragmentation index, continuity verification is performed to obtain a verification pass indicator. Based on the verification pass indicator, medical staff and equipment are matched to determine the final personalized follow-up plan.

[0088] First, resource occupancy duration data is obtained by indexing the medical resource database based on the time parameter of the real-time adjustment instruction. The medical resource allocation record database stores dynamic occupancy information for all rehabilitation-related resources within the medical institution, specifically including the physical space status of each rehabilitation clinic, appointment schedules for specific rehabilitation equipment (such as isokinetic muscle strength training devices, balance devices, and hydrotherapy equipment), and staff shift and work hour allocation records. Database initialization relies on real-time interface calls to the hospital's appointment system. In one implementation, the follow-up date and time parameters explicitly stated in the real-time adjustment instruction are extracted, and the medical resource allocation record database is searched accordingly to obtain the total occupancy duration of rehabilitation equipment and facilities within the corresponding time range and the distribution of idle time periods among the equipment.

[0089] Next, load rate mapping is performed based on the resource occupancy duration data to obtain the resource load status. Specifically, the ratio of the occupied time of resources within the target time window to the theoretical total available time is calculated, and the calculated percentage result is divided into three levels: low load, moderate load, and overload. In one implementation, if the total available time of equipment in a rehabilitation therapy room is 100 hours in the next 14 days, and the reserved time is 68 hours, then the mapped resource load status is 68%, which is in the moderate load range.

[0090] Subsequently, if the resource load status does not exceed the preset carrying capacity limit, equipment distribution characteristic analysis is performed to obtain the time fragmentation index. The carrying capacity limit is usually set at 80% to 90%. Under the premise of load safety, the time fragmentation index is obtained by analyzing the dispersion of idle periods on the time-series coordinate axis, calculating the standard deviation of the time span of adjacent idle blocks, and normalizing it. The lower the index value, the more concentrated the idle time is, which is more conducive to ensuring the continuity of rehabilitation intervention. If the resource load exceeds the carrying capacity limit, the process returns to step S14, and the suboptimal target follow-up sequence is regenerated for resource verification by adjusting the convergence threshold or extending the prediction period.

[0091] Next, continuity verification is performed based on the time fragmentation index to obtain a verification pass flag. Specifically, the verification process involves determining whether the time fragmentation index is lower than a preset continuity threshold. The preset continuity threshold is set based on the average time required for a single patient to complete a full set of rehabilitation actions at a medical institution. Its purpose is to ensure that the allocated time blocks are sufficient to support the complete treatment process without interruption. In one implementation, the value is determined by calculating the reciprocal of the shortest continuous time required to complete a standard rehabilitation course within the past three months. A common method is to select the 70th percentile of the time fragmentation index in historically successful continuous intervention cases as a benchmark. If the verification passes, the system generates a verification pass flag and locks the current scheduling matrix. Finally, based on the verification pass flag, medical personnel and equipment are matched to determine the final personalized follow-up plan. Preferably, combining the subject's rehabilitation characteristic data (such as joint flexion recovery needs), therapists with orthopedic rehabilitation expertise are preferentially assigned in the locked scheduling matrix, and a plan including the specific execution time, location, and personnel list is output.

[0092] It should be noted that the preset carrying capacity limit needs to be set according to the operational scale of different medical institutions, and its setting is usually based on industry-standard hospital resource management guidelines. It is worth noting that the time fragmentation index described in this embodiment is essentially used to optimize resource scheduling efficiency to reduce the cost of switching between medical and nursing staff. It is also important to emphasize that when the resource load exceeds the preset carrying capacity limit (e.g., 90%) and cannot be optimized through the fragmentation index, the system should follow the "risk priority" principle, prioritizing the protection of subjects whose risk assessment results are in the high-risk range, and automatically triggering a resource saturation warning to the medical staff to ensure that the rehabilitation intervention time for high-risk patients is not delayed.

[0093] In summary, this invention, by combining multidimensional data time-series analysis and causal inference path locking, dynamically identifies key turning points in rehabilitation and outputs compensatory adjustment instructions, thereby achieving precise allocation of follow-up resources and closed-loop adaptive optimization of individual rehabilitation trajectories.

[0094] Reference Figure 2The second embodiment of the present invention provides a dynamic adjustment system for follow-up schemes based on multidimensional data, comprising:

[0095] The time-series analysis module is used to acquire the subject's historical rehabilitation index data and historical follow-up frequency records, and to perform time-series dynamic analysis based on the historical rehabilitation index data and historical follow-up frequency records to obtain the index fluctuation pattern and frequency cumulative effect value.

[0096] The causal inference module is used to perform causal inference processing based on the index fluctuation pattern and the frequency cumulative effect value to obtain the causal relationship path and the corresponding path weight.

[0097] The risk prediction module is used to extract features based on the causal relationship path to obtain a real-time rehabilitation indicator status including fluctuation amplitude, and to perform rehabilitation reversal risk prediction when the fluctuation amplitude of the real-time rehabilitation indicator status exceeds a preset fluctuation threshold, thereby obtaining a risk assessment result.

[0098] The causal simulation module is used to perform frequency arrangement sequence causal simulation based on the risk assessment results, the frequency cumulative effect value and the path weight to obtain the future state prediction value, calculate the target deviation value between the future state prediction value and the preset rehabilitation target value, and determine the target follow-up sequence where the target deviation value is less than the preset convergence threshold.

[0099] The instruction generation module is used to extract historical rehabilitation trajectories from the historical rehabilitation indicator data, perform matching analysis between the historical rehabilitation trajectories and the target follow-up sequence to obtain the trajectory matching degree, perform intervention node parsing based on the trajectory matching degree to obtain the parsing result, and perform timing miss risk assessment based on the parsing result to obtain an immediate adjustment instruction containing time parameters.

[0100] The resource verification module is used to retrieve resource records according to the real-time adjustment instruction, obtain the resource load status, verify the continuity of the real-time adjustment instruction according to the resource load status, and output the final personalized follow-up plan.

[0101] It should be noted that the follow-up plan dynamic adjustment system based on multidimensional data provided in this embodiment of the invention is used to execute all the process steps of the follow-up plan dynamic adjustment method based on multidimensional data in the above embodiment. The working principle and beneficial effect of the two are one-to-one, so they will not be described again.

[0102] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0103] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for dynamically adjusting a follow-up scheme based on multi-dimensional data, characterized in that, include: Historical rehabilitation index data and historical follow-up frequency records of the subjects were obtained. Time-series dynamic analysis was performed based on the historical rehabilitation index data and the historical follow-up frequency records to obtain the index fluctuation pattern and frequency cumulative effect value. Based on the index fluctuation pattern and the frequency cumulative effect value, causal inference processing is performed to obtain the causal relationship path and the corresponding path weight; Feature extraction is performed based on the causal relationship path to obtain real-time rehabilitation indicator status including fluctuation amplitude. When the fluctuation amplitude of the real-time rehabilitation indicator status exceeds a preset fluctuation threshold, rehabilitation reversal risk prediction is performed to obtain risk assessment results. Based on the risk assessment results, the cumulative effect value of frequency and the path weight, a causal simulation of the frequency arrangement sequence is performed to obtain the predicted value of the future state, and the target deviation value between the predicted value of the future state and the preset rehabilitation target value is calculated to determine the target follow-up sequence in which the target deviation value is less than the preset convergence threshold. Historical rehabilitation trajectories are extracted from the historical rehabilitation indicator data. The historical rehabilitation trajectories are matched and analyzed with the target follow-up sequence to obtain the trajectory matching degree. Intervention nodes are parsed based on the trajectory matching degree to obtain the analysis results. The timing miss risk is assessed based on the analysis results to obtain an immediate adjustment instruction containing time parameters. Resource records are retrieved according to the real-time adjustment instructions to obtain the resource load status. The real-time adjustment instructions are then continuously verified based on the resource load status, and the final personalized follow-up plan is output. The step of performing causal inference processing based on the index fluctuation pattern and the frequency cumulative effect value to obtain the causal relationship path and the corresponding path weight includes: A multidimensional causal analysis dataset is constructed based on the index fluctuation pattern and the frequency cumulative effect value. The PC algorithm was used to construct the topology of the multidimensional causal analysis dataset to obtain a directed acyclic graph; The obfuscation factor blocking process is performed on the directed acyclic graph to obtain a debiased balanced sample set, and the path parsing is performed on the debiased balanced sample set to obtain the causal association path and the corresponding path weight. The step of retrieving resource records based on the real-time adjustment instruction to obtain the resource load status, and then verifying the continuity of the real-time adjustment instruction based on the resource load status to output the final personalized follow-up plan includes: Based on the time parameter of the real-time adjustment instruction, the medical resource database is indexed to obtain resource occupancy duration data, and load rate mapping is performed based on the resource occupancy duration data to obtain the resource load status; If the resource load status does not exceed the preset carrying limit, then the equipment distribution characteristics are analyzed to obtain the time fragmentation index; Based on the time fragmentation index, continuity verification is performed to obtain a verification pass indicator. Based on the verification pass indicator, medical staff and equipment are matched to determine the final personalized follow-up plan. 2.The method of claim 1, wherein, The process of acquiring historical rehabilitation indicator data and historical follow-up frequency records of the subjects, and performing time-series dynamic analysis based on the historical rehabilitation indicator data and historical follow-up frequency records to obtain indicator fluctuation patterns and frequency cumulative effect values ​​includes: The historical follow-up frequency records are time-aligned to obtain a follow-up time interval sequence. The follow-up time interval sequence and the historical rehabilitation index data are then divided into intervals to obtain an active fluctuation phase and a stable maintenance phase. Based on the active fluctuation phase and the stable maintenance phase, extract the time-series feature vectors that correlate follow-up density with the rate of change of indicators; The time-series feature vector is processed using a pre-trained regression analysis model to obtain the frequency influence factor, and the frequency influence factor is weighted and accumulated to obtain the index fluctuation pattern and frequency cumulative effect value. 3.The method of claim 1, wherein, The process involves feature extraction based on the causal relationship path to obtain real-time rehabilitation indicator status including fluctuation amplitude. When the fluctuation amplitude of the real-time rehabilitation indicator status exceeds a preset fluctuation threshold, a rehabilitation reversal risk prediction is performed to obtain a risk assessment result, including: The current physiological state vector is obtained by collecting node data according to the causal relationship path, and the Euclidean distance of the current physiological state vector is calculated to obtain the fluctuation amplitude. If the fluctuation amplitude exceeds the preset fluctuation threshold, the time series data of the rehabilitation indicators are extracted from the real-time rehabilitation indicator status, and the time series data of the rehabilitation indicators are input into the pre-trained autoregressive moving average model for prediction to obtain the probability of rehabilitation recurrence. The probability of recurrent rehabilitation was normalized to obtain the risk assessment result. 4.The method of claim 1, wherein, The step of performing frequency-arranged sequence causal simulation based on the risk assessment results, the cumulative effect value of frequency, and the path weights to obtain future state prediction values, calculating the target deviation value between the future state prediction values ​​and the preset rehabilitation target values, and determining the target follow-up sequence where the target deviation value is less than a preset convergence threshold includes: Based on the risk assessment results and the frequency cumulative effect value, a step size mapping is performed to obtain multiple sets of candidate frequency arrangement sequences; Based on the path weights, the state transition probabilities of the multiple candidate frequency arrangement sequences are calculated to obtain the predicted values ​​of the future states. The Euclidean distance between the predicted future state value and the preset rehabilitation target value is calculated to obtain the target deviation value, and the candidate frequency arrangement sequence whose target deviation value is less than the preset convergence threshold is determined as the target follow-up sequence.

5. The method of claim 1, wherein, The process involves extracting historical rehabilitation trajectories from the historical rehabilitation indicator data, performing matching analysis between the historical rehabilitation trajectories and the target follow-up sequence to obtain a trajectory matching degree, parsing intervention nodes based on the trajectory matching degree to obtain analysis results, and assessing the risk of missed opportunities based on the analysis results to obtain immediate adjustment instructions containing time parameters, including: The overlap between the historical rehabilitation trajectory and the target follow-up sequence is calculated using a dynamic programming algorithm to obtain the trajectory matching degree, and time-series matching segments are extracted based on the trajectory matching degree. Based on the time-series matching segment, node parsing is performed to obtain a preset intervention node, real-time time point is obtained, and the remaining effective duration is calculated based on the real-time time point and the preset intervention node. If the remaining effective duration is less than the preset response limit, the subject's historical follow-up records are obtained, and compliance analysis is performed based on the historical follow-up records to obtain historical compliance data. The timing risk value is calculated based on the historical compliance data. If the timing risk value exceeds the preset immediate intervention threshold, an immediate adjustment instruction including time parameters is output. 6.The method of claim 1, wherein, Before acquiring the subject's historical rehabilitation index data and historical follow-up frequency records, the method further includes: Multidimensional raw physiological signals of the subjects are collected through monitoring devices integrated with smart chips; The multidimensional raw physiological signals are preprocessed to obtain regularized physiological signals, and the regularized physiological signals are resampled to obtain historical rehabilitation index data.

7. The method of claim 2, wherein, The historical rehabilitation indicator data includes at least one of the subject's heart rate, blood pressure, anxiety score, and joint range of motion.

8. A multi-dimensional data-based follow-up scheme dynamic adjustment system, characterized in that, For implementing the method as described in any one of claims 1-7, comprising: The time-series analysis module is used to acquire the subject's historical rehabilitation index data and historical follow-up frequency records, and to perform time-series dynamic analysis based on the historical rehabilitation index data and historical follow-up frequency records to obtain the index fluctuation pattern and frequency cumulative effect value. The causal inference module is used to perform causal inference processing based on the index fluctuation pattern and the frequency cumulative effect value to obtain the causal relationship path and the corresponding path weight. The risk prediction module is used to extract features based on the causal relationship path to obtain a real-time rehabilitation indicator status including fluctuation amplitude, and to perform rehabilitation reversal risk prediction when the fluctuation amplitude of the real-time rehabilitation indicator status exceeds a preset fluctuation threshold, thereby obtaining a risk assessment result. The causal simulation module is used to perform frequency arrangement sequence causal simulation based on the risk assessment results, the frequency cumulative effect value and the path weight to obtain the future state prediction value, calculate the target deviation value between the future state prediction value and the preset rehabilitation target value, and determine the target follow-up sequence where the target deviation value is less than the preset convergence threshold. The instruction generation module is used to extract historical rehabilitation trajectories from the historical rehabilitation indicator data, perform matching analysis between the historical rehabilitation trajectories and the target follow-up sequence to obtain the trajectory matching degree, perform intervention node parsing based on the trajectory matching degree to obtain the parsing result, and perform timing miss risk assessment based on the parsing result to obtain an immediate adjustment instruction containing time parameters. The resource verification module is used to retrieve resource records according to the real-time adjustment instruction, obtain the resource load status, verify the continuity of the real-time adjustment instruction according to the resource load status, and output the final personalized follow-up plan.

Images