Multi-modal medical information intelligent integration and decision support system for acupuncture rehabilitation

Abstract

The acupuncture rehabilitation-oriented multi-modal medical information intelligent integration and decision support system belongs to the technical field of medical information, comprises a data acquisition and synchronization unit, is used for acquiring physiological data, psychological data and behavior data of athletes, and is also used for time stamp alignment processing on the acquired physiological data, psychological data and behavior data, so as to generate a data stream with a global time stamp; a state representation and evaluation unit is used for extracting multi-modal features based on the data stream generated by the data acquisition and synchronization unit; the state representation and evaluation unit is also used for processing the multi-modal features through a preset disposable physiological resource dynamic evaluation model, so as to estimate a current state vector representing a current physiological state of the athletes; the state representation and evaluation unit of the system utilizes advanced algorithms such as convolutional neural network and wavelet transform to extract deep features from original data, filter out noise and redundant information, and generate highly concise feature vectors.

Description

Technical Field

[0001] This invention relates to the field of medical information, specifically to a multimodal medical information intelligent integration and decision support system for acupuncture rehabilitation. Background Technology

[0002] The core objective of postoperative rehabilitation for elite athletes is to safely return to peak performance in the shortest possible time. Existing rehabilitation decision support systems largely follow standardized operating procedures based on evidence-based medicine, i.e., procedural compliance. These systems emphasize the phased and standardized nature of the rehabilitation plan, pursuing predictability of progress through strict adherence to the protocol. However, the recovery process of an athlete's body is a highly nonlinear dynamic system, whose immediate bioefficiency is influenced in real time by multiple factors such as training, treatment, and psychological stress. Especially under the influence of high-frequency competitive pressure, the athlete's physiological adaptation threshold is significantly lowered. At this point, a destructive coupling can occur between adherence to standardized procedures and the pursuit of maximizing immediate bioefficiency. Forcing athletes to perform high-intensity standard tasks may lead to a state of implicit overtraining, a sharp decline in recovery rate, and even secondary injuries, creating a negative feedback loop. This phenomenon reveals a blind spot in the existing technology: the problem of temporal mismatch of physiological resources. Existing systems plan tasks along the time axis but ignore the energy axis of the athlete's body, forcibly scheduling high-consumption tasks during periods of low physiological resources, resulting in ineffective resource allocation. Therefore, the field needs a technical solution that can proactively manage the athlete's physiological energy cycle and dynamically optimize rehabilitation decisions to solve the defect of temporal mismatch of physiological resources.

[0003] The information disclosed in the background section above is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0004] The purpose of this invention is to provide a multimodal medical information intelligent integration and decision support system for acupuncture rehabilitation, so as to solve the problems mentioned in the background art.

[0005] The technical solution of the present invention includes:

[0006] The data acquisition and synchronization unit is used to collect athletes' physiological, psychological, and behavioral data; the data acquisition and synchronization unit is also used to perform timestamp alignment processing on the collected physiological, psychological, and behavioral data to generate a data stream with a global timestamp.

[0007] The state characterization and evaluation unit is used to extract multimodal features based on the data stream generated by the data acquisition and synchronization unit. The state characterization and evaluation unit is also used to process multimodal features through a preset dynamic evaluation model of available physiological resources in order to estimate the current state vector that represents the athlete's current physiological state.

[0008] The state evolution prediction unit is used to process the historical state vector sequence estimated by the state representation and evaluation unit using a preset physiological resource evolution prediction engine to calculate the predicted state vector sequence for future time steps.

[0009] The decision optimization and instruction generation unit is used to generate the optimal intervention sequence based on the predicted state vector sequence calculated by the state evolution prediction unit; the decision optimization and instruction generation unit is also used to parse the optimal intervention sequence into a dynamic rehabilitation plan;

[0010] The closed-loop correction unit is used to calculate the prediction error after the execution of the dynamic rehabilitation plan; the closed-loop correction unit is also used to adaptively correct the system parameters of the dynamic assessment model of available physiological resources and the physiological resource evolution prediction engine in response to the prediction error exceeding the preset error threshold.

[0011] Preferably, the physiological data collected by the data acquisition and synchronization unit includes heart rate variability data, skin conductance response data, tongue image data, and pulse waveform data; the behavioral data includes biochemical indicators and subjective recovery perception data.

[0012] Preferably, the process of extracting multimodal features by the state representation and evaluation unit is as follows:

[0013] Determine the statistical characteristics and frequency domain power distribution characteristics of heart rate variability data, and generate HRV feature vectors;

[0014] A pre-defined convolutional neural network is used to process tongue image data in order to extract quantized feature vectors;

[0015] Wavelet transform was used to analyze pulse waveform data in order to extract pulse feature vectors.

[0016] Preferably, the dynamic assessment model for available physiological resources is a linear Gaussian state-space model; the current state vector includes available physiological resources, physiological fatigue index and inflammation level; the state representation and assessment unit uses the Kalman filter algorithm to estimate the current state vector by combining the state at the previous moment with the current observation vector composed of multimodal features.

[0017] Preferably, the physiological resource evolution prediction engine is a long short-term memory network based on an attention mechanism; the input of the state evolution prediction unit includes a sequence of historical state vectors and a known sequence of future inputs; the attention mechanism is used to assign weights to events in the historical sequence when predicting future states.

[0018] Preferably, the decision optimization and instruction generation unit generates the optimal intervention sequence by constructing and solving a quadratic objective function; the quadratic objective function aims to:

[0019] Minimize the weighted error between the predicted state vector sequence and the preset target state trajectory;

[0020] Minimize the penalty for controlling input;

[0021] Minimize the penalty for deviations from standard procedures.

[0022] Preferably, the decision optimization and instruction generation unit also handles multiple constraints when solving for the optimal intervention sequence; the multiple constraints include the daily total physiological load limit, which does not exceed a preset load threshold.

[0023] Preferably, the process by which the closed-loop correction unit calculates the prediction error is as follows:

[0024] Determine the actual change in the state vector after the intervention ends;

[0025] The prediction error is quantified by comparing the actual changes with the estimated recovery benefits stored in the cost-benefit knowledge base of the intervention.

[0026] Preferably, the process by which the closed-loop correction unit adaptively corrects the system parameters is as follows:

[0027] By using prediction error as a feedback signal, the state transition matrix, input matrix, and observation matrix of the dynamic assessment model of available physiological resources are fine-tuned.

[0028] The prediction error is used as a feedback signal to fine-tune the network weights of the physiological resource evolution prediction engine.

[0029] This invention provides an improved multimodal medical information intelligent integration and decision support system for acupuncture rehabilitation, which has the following improvements and advantages compared with the prior art:

[0030] 1. This invention achieves a comprehensive, profound, and quantitatively accurate assessment of athletes' physiological state. Unlike existing technologies that rely on single or discrete indicators for one-sided evaluation, this invention integrates modern physiological data reflecting the autonomic nervous system state, such as heart rate variability and skin conductance, with tongue images and pulse waveform data based on traditional Chinese medicine theory through a data acquisition and synchronization unit. This integration of multimodal information provides an unprecedentedly comprehensive and robust data foundation for subsequent assessments. The system's state representation and assessment unit uses advanced algorithms such as convolutional neural networks and wavelet transforms to extract deep features from the raw data, filter out noise and redundant information, and generate highly condensed feature vectors. This elevates the understanding of athletes' state from traditional qualitative descriptions to precise quantitative characterization.

[0031] 2. This invention represents a fundamental shift from passive response to proactive prediction. Utilizing a linear Gaussian state-space model, it abstracts and quantifies an athlete's internal state for the first time into core vectors such as available physiological resources, physiological fatigue index, and inflammation level, and performs optimal estimation using a Kalman filter algorithm. Furthermore, a state evolution prediction unit employs a long short-term memory network based on an attention mechanism as its prediction engine. This engine not only learns the long-term patterns of state evolution, but its unique attention mechanism also intelligently identifies and focuses on historical events that significantly impact the future. This enables the system to predict the peaks and troughs of future physiological resources with high quality, thereby transforming the basis for rehabilitation strategy formulation from a delayed response to past states to a proactive layout of future trends.

[0032] 3. This invention achieves a leap from experience-driven to scientifically optimal rehabilitation decision-making. The decision optimization and instruction generation unit in this invention constructs the generation of rehabilitation plans as a rigorous multi-objective, multi-constraint optimization problem. By solving a quadratic objective function, it systematically balances the cost, safety, and compliance with evidence-based medicine standard procedures of intervention measures, while ensuring that the rehabilitation effect moves towards the preset target trajectory. At the same time, the introduction of key constraints such as the daily total physiological load limit provides a solid safety guarantee for the entire rehabilitation process and effectively avoids the risk of overtraining. The dynamic rehabilitation plan generated by this mechanism is the optimal solution under all safety boundary conditions, thereby maximizing the efficiency of the rehabilitation process while ensuring safety.

[0033] 4. The system has been upgraded from a static model to an adaptive evolutionary capability. This invention uniquely incorporates a closed-loop correction unit. This unit generates a feedback signal by quantifying the error between the actual and predicted effects after the rehabilitation plan is implemented. When this error exceeds the threshold of statistical significance, the system will use this signal to adaptively fine-tune the system parameters of the internal core model, including the state transition matrix, input matrix, and observation matrix of the dynamic assessment model of available physiological resources, as well as the network weights of the physiological resource evolution prediction engine. This self-correction and evolutionary capability enables the system to continuously learn and adapt to the individual uniqueness of athletes. Over time, the accuracy of its assessment, prediction, and decision-making will continuously improve, achieving truly personalized and precise rehabilitation. Attached Figure Description

[0034] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0035] Figure 1 This is a flowchart of the system of the present invention. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0037] Example 1

[0038] Please see Figure 1 This invention provides a multimodal medical information intelligent integration and decision support system for acupuncture rehabilitation, comprising:

[0039] The data acquisition and synchronization unit is used to collect athletes' physiological, psychological, and behavioral data; the data acquisition and synchronization unit is also used to perform timestamp alignment processing on the collected physiological, psychological, and behavioral data to generate a data stream with a global timestamp.

[0040] The state characterization and evaluation unit is used to extract multimodal features based on the data stream generated by the data acquisition and synchronization unit. The state characterization and evaluation unit is also used to process multimodal features through a preset dynamic evaluation model of available physiological resources in order to estimate the current state vector that represents the athlete's current physiological state.

[0041] The state evolution prediction unit is used to process the historical state vector sequence estimated by the state representation and evaluation unit using a preset physiological resource evolution prediction engine to calculate the predicted state vector sequence for future time steps.

[0042] The decision optimization and instruction generation unit is used to generate the optimal intervention sequence based on the predicted state vector sequence calculated by the state evolution prediction unit; the decision optimization and instruction generation unit is also used to parse the optimal intervention sequence into a dynamic rehabilitation plan;

[0043] In addition, the data acquisition and synchronization unit also includes an outlier detection module. When the sensor data is detected to exceed the preset reasonable physiological range, such as a heart rate of 0 or more than 250 bpm, or when the data stream does not change for a long time, the system will trigger an alarm and process the data by interpolation or rejection of the data source to ensure the robustness of the subsequent state assessment model.

[0044] For example, this parsing process can be based on a pre-defined knowledge base of intervention measures, which will control vectors Each dimension is mapped to specific intervention measures, such as acupuncture points, physical therapy techniques, duration, and intensity. When the optimal solution is found... When a certain dimension value is a specific value, the system queries the knowledge base and generates corresponding, specific rehabilitation operation instructions;

[0045] The closed-loop correction unit is used to calculate the prediction error after the execution of the dynamic rehabilitation plan; the closed-loop correction unit is also used to adaptively correct the system parameters of the dynamic assessment model of available physiological resources and the physiological resource evolution prediction engine in response to the prediction error exceeding the preset error threshold.

[0046] This invention provides a multimodal medical information intelligent integration and decision support system for acupuncture rehabilitation. The system aims to solve the problem of physiological resource mismatch caused by adhering to standardized operations in the existing rehabilitation process. By accurately assessing, dynamically predicting and adaptively deciding on the athlete's physiological state, the system maximizes rehabilitation efficiency.

[0047] The system includes a data acquisition and synchronization unit, which aims to acquire comprehensive data reflecting the athlete's condition from multi-source heterogeneous devices. In this embodiment, this unit continuously collects various data of the athlete through wearable sensing devices, dedicated TCM four diagnostic methods quantification devices, and software interfaces. After the acquisition is completed, network time protocol is applied to all data streams for time alignment, and a global timestamp is uniformly added to form time series data required for subsequent processing.

[0048] The system further includes a state representation and evaluation unit, which aims to extract deep physiological features from raw data and estimate core physiological states that cannot be directly measured. In this embodiment, this unit performs feature engineering on the data stream with timestamps, and the processing result is input into a preset dynamic evaluation model of available physiological resources. This model can integrate multi-dimensional features and estimate the current state vector representing the athlete's current physiological state in real time.

[0049] The system further includes a state evolution prediction unit, which aims to predict the possible trajectory of changes in the athlete's physiological state over a future period of time based on historical state information. In this embodiment, this unit receives the historical state vector sequence output by the state table EPC and the evaluation unit, and processes it using a preset physiological resource evolution prediction engine to calculate the predicted state vector sequence for multiple future time steps.

[0050] The system further includes a decision optimization and instruction generation unit, which aims to calculate the optimal combination of intervention measures based on predictions of the future. In this embodiment, this unit takes the predicted state vector sequence output by the state evolution prediction unit as the core input, constructs and solves a multi-objective optimization problem to generate the optimal intervention sequence. This unit parses the sequence into a specific and executable dynamic rehabilitation plan.

[0051] The system further includes a closed-loop correction unit, which functions to reverse-correct the internal model of the system by comparing the actual effect of the plan execution with the expected effect, so that the system has self-learning and self-adaptive capabilities. In this embodiment, after the cycle of the dynamic rehabilitation plan is completed, this unit will quantify the generated prediction error. When the error exceeds the preset error threshold, this unit will be activated and use the error as a feedback signal to adaptively correct the internal system parameters of the dynamic assessment model of available physiological resources and the physiological resource evolution prediction engine.

[0052] This embodiment constructs a complete technical closed loop from data acquisition, status assessment, future prediction, optimization decision-making to closed-loop correction through the collaborative work of the above five units. It transforms the rehabilitation process from a traditional, fixed time-axis task to dynamic resource-axis management. By proactively predicting the peaks and troughs of physiological resources, it intelligently schedules rehabilitation tasks, thereby solving the risk of ineffective allocation of physiological resources or even secondary damage caused by blindly implementing standardized procedures, and significantly improving the safety and efficiency of the rehabilitation process.

[0053] Physiological data collected by the data acquisition and synchronization unit includes heart rate variability data, skin conductance response data, tongue image data, and pulse waveform data; behavioral data includes biochemical indicators and subjective recovery perception data.

[0054] In this embodiment, the data acquisition and synchronization unit is limited in its acquisition content;

[0055] The physiological data collected by this unit, in this embodiment, includes: heart rate variability data and skin conductance response data collected by wearable devices, which are used to reflect the balance and arousal level of the autonomic nervous system; tongue image data acquired by a standard light source and a high-resolution camera; and pulse waveform data acquired by a multi-array pressure sensor, which are used to quantify the body state from the perspective of traditional Chinese medicine.

[0056] In this embodiment, the behavioral data collected by this unit includes: biochemical indicators such as blood lactate and salivary cortisol entered through a professional equipment interface, as well as subjective recovery perception data, such as recovery scale scores, entered through software applications.

[0057] This embodiment achieves a deep integration of modern physiological indicators and quantitative information from the four diagnostic methods of traditional Chinese medicine by defining specific data modalities. Data such as heart rate variability and skin conductance provide objective physiological information, while tongue and pulse data provide supplementary information from a macroscopic and holistic perspective. This multimodal and multidimensional information input provides a more comprehensive and robust data foundation for subsequent state characterization and assessment units, thereby significantly improving the accuracy and depth of assessment of athletes' complex physiological states.

[0058] The process of extracting multimodal features by the state representation and evaluation unit is as follows:

[0059] Determine the statistical characteristics and frequency domain power distribution characteristics of heart rate variability data, and generate HRV feature vectors;

[0060] A pre-defined convolutional neural network is used to process tongue image data in order to extract quantized feature vectors;

[0061] Wavelet transform was used to analyze pulse waveform data in order to extract pulse feature vectors;

[0062] In this embodiment, the process of extracting multimodal features by the state representation and evaluation unit is described;

[0063] The process of generating the HRV feature vector is as follows: process the heart rate variability data, calculate its statistical characteristics, such as the standard deviation of the interval between adjacent heartbeats, and its frequency domain power distribution characteristics, such as the ratio of high-frequency power to low-frequency power, and combine them into the vector.

[0064] The process of extracting the quantized feature vector is as follows: the tongue image data is processed using a pre-set convolutional neural network; the pre-set convolutional neural network refers to a deep learning model that has been pre-trained on a dataset containing tens of thousands of tongue images annotated by senior TCM doctors; the working principle is to automatically learn and recognize visual patterns such as tongue color, tongue coating, and shape related to different physiological states through multi-layer convolution and pooling operations, and output quantized feature vectors that can represent tongue image information.

[0065] A. Convolutional neural networks for extracting tongue features:

[0066] For example, the pre-defined convolutional neural network can employ a transfer learning strategy, which involves fine-tuning a ResNet-50 model pre-trained on a large-scale ImageNet dataset. The dataset used for fine-tuning consists of tens of thousands of tongue images, whose annotation dimensions can include tongue color (e.g., pale white, red, purplish-red), tongue coating (e.g., thin white, yellowish-greasy), and tongue shape (e.g., swollen, teeth marks). During model training, a cross-entropy loss function is used, and a learning rate of [missing information - likely a percentage] is employed. The Adam optimizer is iteratively trained for 200 cycles to obtain a model that can stably extract quantized feature vectors;

[0067] The process of extracting pulse feature vectors is as follows: Wavelet transform is used to analyze pulse waveform data. Wavelet transform can effectively characterize the local characteristics of a signal in both the time and frequency domains. Through this analysis, dynamic features reflecting pulse intensity, rate, rhythm, and morphology can be extracted and combined into pulse feature vectors.

[0068] This embodiment employs specific and advanced signal processing and artificial intelligence technologies to achieve deep feature extraction from raw multimodal data. Compared to simply using the raw data, the HRV feature vector, tongue image quantification feature vector, and pulse image feature vector generated in this embodiment can more essentially and concisely reflect the athlete's internal physiological state, effectively filtering out data noise and redundant information, and providing high-quality input assurance for the accuracy of subsequent state assessment models.

[0069] The dynamic assessment model for available physiological resources is a linear Gaussian state-space model; the current state vector includes available physiological resources, physiological fatigue index, and inflammation level; the state representation and assessment unit uses the Kalman filter algorithm, which combines the state at the previous moment with the current observation vector composed of multimodal features to estimate the current state vector;

[0070] In this embodiment, the core model and algorithm used by the state representation and evaluation unit are defined;

[0071] The dynamic assessment model for available physiological resources used in this unit, in this embodiment, is a linear Gaussian state-space model. The reason for choosing this model is that, as those skilled in the art should understand, although an athlete's physiological system is inherently nonlinear, within a specific, small working range, using a linear Gaussian state-space model is an effective and computationally feasible approximation method. This system uses a closed-loop correction unit to adjust the model parameters. The continuous fine-tuning partially compensates for the deviation between the linear model and the real scene, and can optimally estimate the intrinsic, low-dimensional physiological state that cannot be directly measured from noisy multidimensional observation data; the model consists of state equations and observation equations.

[0072] The state equation is:

[0073]

[0074] in, yes The current state vector at time t is defined as a three-dimensional vector, which contains three core potential states: available physiological resources, physiological fatigue index, and inflammation level. : A point in time in a time series;

[0075] To enable those skilled in the art to understand and implement this, the state vector is now explained. The three components are illustrated by example:

[0076] It should be noted that this three-dimensional state vector is a simplified representation of the athlete's core physiological state, aiming to capture the main issues. Other important factors, such as psychological stress and sleep duration, are considered external influences and are included in the input vector. This indirectly affects the evolution of the core state vector;

[0077] Available physiological resources: These can be quantified as a normalized composite score, for example, primarily composed of high-frequency power in heart rate variability. And subjective recovery perception scales, such as RPE score, are used to determine the more abundant the resources;

[0078] Physiological fatigue index: This can be quantified as an indicator related to exhaustion, for example, primarily related to the ratio of low-frequency to high-frequency power in heart rate variability. It is positively correlated with blood lactate concentration. The higher the index, the deeper the degree of fatigue;

[0079] Inflammation level: can be quantified using surrogate indicators, such as establishing correlations with the quantitative values ​​of tongue color redness and tongue coating thickness in tongue image features, as well as the fluctuation characteristics of skin conductance response;

[0080] It is the state vector from the previous moment; yes The input vector at any given moment is derived from behavioral data collected by the data acquisition and synchronization unit, as well as known intervention measures, including external influencing factors such as the daily training load and psychological stress scores. It is a state transition matrix that describes the natural evolution of physiological states over time; This is the input matrix, used to linearly quantify the direct impact of different interventions on the rate of change of each state. This linearity assumption is made under the closed-loop correction mechanism of this system by adjusting the matrix. The parameters are adaptively adjusted to approximate the mean dose-response relationship under different conditions; It is process noise, assumed to be Gaussian white noise;

[0081] The observation equation is:

[0082]

[0083] in, It is an observation vector composed of multimodal features extracted by the above method; It is an observation matrix that links internal physiological states with externally measurable multimodal features; This is measurement noise, also assumed to be Gaussian white noise; matrix The initial values ​​are set based on prior physiological knowledge and are continuously optimized in subsequent closed-loop calibration.

[0084] A. Regarding the state-space model matrices A, B, and C:

[0085] For example, the initialization of these matrices can follow these principles:

[0086] State transition matrix This describes the natural evolution of states, with its diagonal elements typically close to 1, indicating a certain degree of persistence in the state. The off-diagonal elements reflect the interactions between states; for example, an increase in the physiological fatigue index might lead to a slight increase in inflammation levels in the next moment. The element at the corresponding position in the matrix can be set to a small positive value;

[0087] Input matrix Quantifying external intervention The impact, if If it includes a high-intensity training load, then in the matrix In this input, the coefficient corresponding to the state of available physiological resources should be negative, and the coefficient corresponding to the state of physiological fatigue index should be positive.

[0088] Observation matrix : Associated internal state External observation For example, an increase in the inflammation level in the state vector is expected to lead to an increase in the quantified value of tongue color redness in the tongue appearance observation features, then the matrix The coefficients of these two variables at corresponding positions should be positive.

[0089] To solve this model, the state representation and evaluation unit in this embodiment employs the Kalman filter algorithm; this algorithm is a recursive estimation algorithm that combines the state from the previous time step. With the current observation vector composed of multimodal features It can estimate the current state vector in real time. The optimal value;

[0090] This embodiment constructs a clear state-space model based on modern control theory, abstracting and quantifying the athlete's rehabilitation state into core indicators such as available physiological resources. By utilizing the Kalman filter algorithm, it can accurately reveal the athlete's internal state from noisy multi-source data, providing a solid and quantitative foundation for subsequent prediction and decision-making, and realizing a leap from qualitative description to dynamic quantitative modeling of athlete state assessment.

[0091] Example 2

[0092] The physiological resource evolution prediction engine is a long short-term memory network based on an attention mechanism; the input of the state evolution prediction unit includes a sequence of historical state vectors and a known sequence of future inputs; the attention mechanism is used to assign weights to events in the historical sequence when predicting future states;

[0093] In this embodiment, the core prediction engine used by the state evolution prediction unit is described;

[0094] The physiological resource evolution prediction engine used in this unit, in this embodiment, is a long short-term memory network based on the attention mechanism. The reason for choosing this model is that the long short-term memory network can effectively capture long-term dependencies in time series data through its internal gating mechanism, while the attention mechanism further enhances the performance of the model. The attention mechanism refers to a mechanism that imitates human cognitive attention, and its function is to automatically assign higher computational weights to the events with the greatest impact in the historical sequence when predicting future states.

[0095] B. For the physiological resource evolution prediction engine:

[0096] For example, a long short-term memory network based on an attention mechanism can be specifically designed as a network structure containing two stacked LSTM layers, each containing 256 hidden units. During prediction at each time step, the attention mechanism will calculate weights based on the magnitude of the physiological fatigue index in the historical sequence, thereby giving higher attention to those historical state points of extreme fatigue. The network is trained using mean squared error as the loss function.

[0097] The input to the prediction engine consists of two parts: a sequence of historical state vectors output by the state representation and evaluation unit, and a known sequence of future inputs from external sources, such as training momentum or rehabilitation therapy arrangements planned for the next few days; the output of the engine is a sequence of predicted state vectors for multiple future time steps.

[0098] The prediction engine used in this embodiment can not only learn the evolutionary patterns of states like traditional time series models, but also achieve intelligent focus on key historical events through an attention mechanism. This makes the prediction results more sensitive and accurate to the individual experiences of athletes, and can foresee the potential future physiological resource troughs caused by a specific historical event. This provides high-quality prediction information for the system to make forward-looking and protective decision optimizations, and significantly improves the foresight and accuracy of decision-making.

[0099] Example 3

[0100] The decision optimization and instruction generation unit generates the optimal intervention sequence by constructing and solving a quadratic objective function; the quadratic objective function aims to:

[0101] Minimize the weighted error between the predicted state vector sequence and the preset target state trajectory;

[0102] Minimize the penalty for controlling input;

[0103] Minimize the penalty for deviation from standard procedures;

[0104] When solving for the optimal intervention sequence, the decision optimization and instruction generation unit also handles multiple constraints, including the daily total physiological load limit, which does not exceed a preset load threshold.

[0105] In this embodiment, the optimization method and constraints used by the decision optimization and instruction generation unit to generate the optimal intervention sequence are described in detail;

[0106] This unit generates the optimal intervention sequence by constructing and solving a quadratic objective function. This objective function aims to find the optimal control strategy within a finite future time domain using a model predictive control framework; its mathematical form is as follows:

[0107]

[0108] in, : This indicates that the objective of the formula is to minimize the objective function. ; It is the objective function to be minimized; : Index of time steps; It is the predicted future state vector output by the state evolution prediction unit; It is a preset target state trajectory, which is predefined by rehabilitation experts based on clinical guidelines and the athlete's rehabilitation stage goals; It is the sequence of intervention measures to be solved; It is an intervention sequence corresponding to the standard rehabilitation process obtained from an external knowledge base or standardized rehabilitation program; and These refer to the lengths of the prediction and control time domains, respectively. All are positive semi-definite symmetric weighted matrices, and their values ​​are preset tuning parameters that balance different optimization objectives; : Represents the quadratic norm of a vector, or the weighted quadratic norm, used to quantify error or cost;

[0109] Crucially, the elements of these weighted matrices are not dimensionless pure numerical values, but rather have specific units to ensure the objective function... All items have the same dimensionless cost unit, specifically:

[0110] matrix The element unit is Used to square the state error The cost of converting to dimensionless quantities;

[0111] matrix and The element unit is Used to square the control input The cost of converting to dimensionless quantities;

[0112] In this way, the objective function Only when the three components—state error cost, control input cost, and process deviation cost—are unified in terms of dimensions can meaningful addition and minimization be achieved.

[0113] B. For the decision optimization weighted matrix Q, R, S:

[0114] These weighting matrices are set according to the emphasis of the rehabilitation strategy, for example:

[0115] matrix This determines the severity of the penalty for deviating from the target trajectory. If the primary goal at this stage is to quickly eliminate inflammation, then... The weight corresponding to the error in the level of inflammation should be set relatively large;

[0116] matrix This determines the level of penalty for the cost of the intervention; if the cost of acupuncture or physical therapy is high, then... The weight corresponding to this intervention should be set relatively high to encourage the system to find lower-cost alternatives;

[0117] matrix This determines the severity of punishment for deviations from standard procedures. In the early stages of recovery, to ensure safety, a higher level of punishment can be set. The weighting is adjusted to make the plan more closely resemble the standard protocol; in the later stages of rehabilitation, the weighting can be appropriately reduced. Weighting gives the system more room for personalized optimization.

[0118] This quadratic objective function specifically aims to achieve three mutually balanced objectives:

[0119] Minimize the weighted error between the predicted state vector sequence and the preset target state trajectory, determined by the first term. Achievement; This is the core driving force, ensuring that the rehabilitation plan progresses towards the ideal state;

[0120] Minimize the control input penalty, as described in the second term. This measure aims to avoid using excessively costly or intensive interventions, reflecting considerations of both economy and safety.

[0121] Minimize the penalty for deviation from the standard process, as specified in the second item. This ensures that the dynamic plans generated by the system do not deviate indefinitely from the standard procedures validated by evidence-based medicine, thus guaranteeing the reliability of the decisions.

[0122] In solving for the optimal intervention sequence, this unit also handles multiple constraints; in this embodiment, the key constraint is the daily total physiological load limit, that is, ensuring that the total physiological cost generated by all intervention measures and training tasks within a day does not exceed a preset load threshold. This threshold is a safe upper limit set based on statistical analysis of a large amount of athletes' physiological data or authoritative sports physiology guidelines. Its purpose is to prevent overtraining or secondary injuries caused by excessive daily load.

[0123] This embodiment achieves scientific and optimal decision-making by constructing rehabilitation decision-making as a rigorous, constrained model predictive control optimization problem. It is no longer based on isolated rules or experience, but rather on a predictive framework that systematically balances multiple objectives such as rehabilitation effect, intervention cost, process compliance, and physiological safety. This method can generate a dynamic intervention sequence with optimal comprehensive benefits under all safety constraints, thereby maximizing the acceleration of the rehabilitation process while ensuring safety.

[0124] Example 4

[0125] The process of calculating the prediction error by the closed-loop correction unit is as follows:

[0126] Determine the actual change in the state vector after the intervention ends;

[0127] The prediction error is quantified by comparing the actual changes with the estimated recovery benefits stored in the cost-benefit knowledge base of the intervention measures.

[0128] The process by which the closed-loop correction unit adaptively corrects the system parameters is as follows:

[0129] By using prediction error as a feedback signal, the state transition matrix, input matrix, and observation matrix of the dynamic assessment model of available physiological resources are fine-tuned.

[0130] The prediction error is used as a feedback signal to fine-tune the network weights of the physiological resource evolution prediction engine;

[0131] In this embodiment, the working mechanism of the closed-loop correction unit is described, including the calculation method of the prediction error and the adaptive correction process of the system parameters;

[0132] The closed-loop correction unit calculates the prediction error starting after the completion of the cycle of dynamic rehabilitation plan. It acquires the actual multimodal data after the intervention through the data acquisition and synchronization unit, and estimates the actual state vector after the intervention through the state representation and evaluation unit. It then determines the actual change in the state vector after the intervention and compares this actual change with the estimated recovery benefit of the intervention stored in the system's internal intervention cost-benefit knowledge base. The difference between the two is quantified as the prediction error. The intervention cost-benefit knowledge base is a structured database that stores prior knowledge such as estimated resource costs and recovery benefits associated with various interventions.

[0133] The construction of a cost-benefit knowledge base for intervention measures may employ, but is not limited to, a combination of one or more of the following methods:

[0134] Evidence-based medicine data entry: Systematically extract quantitative data on the effects of specific interventions, such as acupuncture at specific acupoints or massage techniques, on physiological indicators, such as heart rate variability and blood lactate, from published clinical studies and medical guidelines, as initial entries for the knowledge base.

[0135] Quantification of expert experience: Design a structured questionnaire and invite senior TCM doctors or rehabilitation therapists to evaluate the resource costs of different intervention measures, such as time, consumables and estimated recovery benefits, such as the degree of improvement in the fatigue index. The evaluation can be graded and scored, and this semi-quantitative expert knowledge is transformed into data in the knowledge base.

[0136] Online learning and knowledge base: The knowledge base is designed as a dynamic database. The system continuously records each intervention during operation. Compared with the actual observed state changes The relationship between them; when enough data is accumulated, the cost-benefit values ​​of various intervention measures in the knowledge base can be automatically filled and updated through methods such as regression analysis, thereby achieving self-improvement of the knowledge base;

[0137] The process by which the closed-loop correction unit adaptively corrects the system parameters is as follows:

[0138] When the calculated prediction error exceeds a preset error threshold When the error occurs, the correction process is triggered. This threshold can be set according to the statistical distribution of historical prediction data, for example, by taking the upper bound of the 95% confidence interval, to ensure that the correction mechanism is triggered only when a statistically significant deviation occurs. The unit uses this prediction error as a feedback signal and uses algorithms such as recursive least squares or gradient descent to fine-tune the core model parameters within the system.

[0139] This unit uses prediction error as a feedback signal to evaluate the state transition matrix of the dynamic assessment model of available physiological resources. Input matrix and observation matrix Fine-tuning is possible; this allows the condition assessment model to more accurately reflect the physiological evolution of a particular athlete and their response patterns to interventions.

[0140] This unit uses prediction error as a feedback signal to fine-tune the network weights of the physiological resource evolution prediction engine; this enables the prediction model of the attention mechanism-based long short-term memory network to more accurately capture the time series characteristics of the athlete's state changes.

[0141] This embodiment establishes a complete closed-loop feedback and correction mechanism, endowing the entire system with the ability to self-evolve and adapt to individual needs. The system is no longer a static model, but can dynamically optimize its internal state assessment model and future prediction model by continuously learning the errors generated by its own decisions. This adaptive capability enables the system to become more and more in line with the uniqueness of individual athletes as the time goes by, thereby continuously improving the accuracy of its assessment, prediction and decision-making, and achieving truly individualized and precise rehabilitation.

[0142] It should be noted that the above embodiments are only used to illustrate the technical solutions 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 solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A multimodal medical information intelligent integration and decision support system for acupuncture rehabilitation, characterized in that: include: The data acquisition and synchronization unit is used to collect athletes' physiological, psychological, and behavioral data. The data acquisition and synchronization unit is also used to perform timestamp alignment processing on the acquired physiological, psychological and behavioral data to generate a data stream with a global timestamp. The state characterization and evaluation unit is used to extract multimodal features based on the data stream generated by the data acquisition and synchronization unit; The state representation and assessment unit is also used to process multimodal features through a pre-defined dynamic assessment model of available physiological resources in order to estimate the current state vector representing the athlete's current physiological state. The state evolution prediction unit is used to process the historical state vector sequence estimated by the state representation and evaluation unit using a preset physiological resource evolution prediction engine to calculate the predicted state vector sequence for future time steps. The decision optimization and instruction generation unit is used to generate the optimal intervention sequence based on the predicted state vector sequence calculated by the state evolution prediction unit; the decision optimization and instruction generation unit is also used to parse the optimal intervention sequence into a dynamic rehabilitation plan; The closed-loop correction unit is used to calculate the prediction error after the execution of the dynamic rehabilitation plan; the closed-loop correction unit is also used to adaptively correct the system parameters of the dynamic assessment model of available physiological resources and the physiological resource evolution prediction engine in response to the prediction error exceeding the preset error threshold. The dynamic assessment model for available physiological resources is a linear Gaussian state-space model; the current state vector includes available physiological resources, physiological fatigue index, and inflammation level; the state representation and assessment unit uses the Kalman filter algorithm, which combines the state at the previous moment with the current observation vector composed of multimodal features to estimate the current state vector; The physiological resource evolution prediction engine is a long short-term memory network based on an attention mechanism; the input of the state evolution prediction unit includes a sequence of historical state vectors and a known sequence of future inputs. Attention mechanisms are used to assign weights to events in a historical sequence when predicting future states; The decision optimization and instruction generation unit generates the optimal intervention sequence by constructing and solving a quadratic objective function; the quadratic objective function aims to: Minimize the weighted error between the predicted state vector sequence and the preset target state trajectory; Minimize the penalty for controlling input; Minimize the penalty for deviations from standard procedures.

2. The multimodal medical information intelligent integration and decision support system for acupuncture rehabilitation according to claim 1, characterized in that, Physiological data collected by the data acquisition and synchronization unit includes heart rate variability data, skin conductance response data, tongue image data, and pulse waveform data; behavioral data includes biochemical indicators and subjective recovery perception data.

3. The multimodal medical information intelligent integration and decision support system for acupuncture rehabilitation according to claim 1, characterized in that, The process of extracting multimodal features by the state representation and evaluation unit is as follows: Determine the statistical characteristics and frequency domain power distribution characteristics of heart rate variability data, and generate HRV feature vectors; A pre-defined convolutional neural network is used to process tongue image data in order to extract quantized feature vectors; Wavelet transform was used to analyze pulse waveform data in order to extract pulse feature vectors.

4. The multimodal medical information intelligent integration and decision support system for acupuncture rehabilitation according to claim 1, characterized in that, When solving for the optimal intervention sequence, the decision optimization and instruction generation unit also handles multiple constraints, including the daily total physiological load limit, which does not exceed a preset load threshold.

5. The multimodal medical information intelligent integration and decision support system for acupuncture rehabilitation according to claim 1, characterized in that, The process of calculating the prediction error by the closed-loop correction unit is as follows: Determine the actual change in the state vector after the intervention ends; The prediction error is quantified by comparing the actual changes with the estimated recovery benefits stored in the cost-benefit knowledge base of the intervention.

6. The multimodal medical information intelligent integration and decision support system for acupuncture rehabilitation according to claim 1, characterized in that, The process by which the closed-loop correction unit adaptively corrects the system parameters is as follows: By using prediction error as a feedback signal, the state transition matrix, input matrix, and observation matrix of the dynamic assessment model of available physiological resources are fine-tuned. The prediction error is used as a feedback signal to fine-tune the network weights of the physiological resource evolution prediction engine.

Images