Deep learning based motion rehabilitation data prediction system

Abstract

The application provides a deep learning-based exercise rehabilitation data prediction system, which comprises the following modules: a data processing module, which collects and pre-processes exercise rehabilitation data of a patient, performs consistency scoring on the pre-processed exercise rehabilitation data, further generates a quality index, assigns a weight to the pre-processed exercise rehabilitation data according to the quality index, and generates a corresponding quality weight matrix; a feature prediction module, which performs weighted processing on the pre-processed exercise rehabilitation data, generates a weighted training set, arranges the weighted training set according to a time sequence to form three-dimensional tensor input data; a mixed deep learning architecture is constructed and three parallel prediction branches are set to obtain three-dimensional prediction values; a personalized scheme module, which evaluates the three-dimensional confidence of the three-dimensional prediction values, calculates a comprehensive evaluation risk, presets a risk threshold, divides a risk level based on the comprehensive evaluation risk, and sets a personalized rehabilitation plan for patients in each risk level.

Description

Technical Field

[0001] This invention relates to the field of sports rehabilitation therapy technology, specifically a sports rehabilitation data prediction system based on deep learning. Background Technology

[0002] Traditional sports rehabilitation therapy typically relies on manual monitoring and assessment by experienced physicians and rehabilitation therapists. However, with the increasing number of patients and the growing demand for individualized rehabilitation, traditional methods have certain limitations, such as incomplete data collection, inconsistent assessment standards, and a lack of personalized treatment plans. Furthermore, the progress of treatment is difficult to predict accurately, leading to significant differences in rehabilitation outcomes. Current technologies for processing sports rehabilitation data often rely on simple statistical methods and empirical judgment, lacking the ability to comprehensively analyze multi-source data such as patient movement data, physiological data, treatment intervention data, and functional assessment data during the rehabilitation process. In addition, traditional rehabilitation treatment plans often lack personalization and cannot flexibly adjust treatment plans according to the patient's real-time status and rehabilitation progress, resulting in a less refined and scientific rehabilitation process. With the rapid development of deep learning technology, intelligent analysis of sports rehabilitation data has become possible. In recent years, deep learning-based medical data prediction systems have made significant progress in disease diagnosis and patient monitoring, but applying this technology to sports rehabilitation data prediction still faces many challenges, especially in data quality, feature extraction, and prediction accuracy. Current technologies have not yet effectively solved these problems.

[0003] Therefore, there is an urgent need for a system that can comprehensively consider various information such as exercise, treatment, functional assessment, and physiological data, and through optimized analysis using deep learning models, accurately predict the rehabilitation effect, progress, and risk of complications for patients, and based on these predictions, tailor a personalized rehabilitation plan for each patient. Summary of the Invention

[0004] In order to solve the technical problems mentioned in the background art, the purpose of this invention is to provide a deep learning-based sports rehabilitation data prediction system.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A deep learning-based sports rehabilitation data prediction system includes the following modules:

[0007] The data processing module collects and preprocesses the patient's exercise rehabilitation data, which includes sub-data such as exercise data, physiological data, functional assessment data, and treatment intervention data. It performs a consistency score on the preprocessed exercise rehabilitation data, further generates quality indicators, and assigns weights to the preprocessed exercise rehabilitation data based on the quality indicators to generate a corresponding quality weight matrix.

[0008] The feature prediction module performs weighted processing on the preprocessed exercise rehabilitation data according to the quality weight matrix to generate a weighted training set. The weighted training set is arranged according to the time series to form three-dimensional tensor input data. A hybrid deep learning architecture is constructed and three parallel prediction branches are set up. The three-dimensional tensor input data is input into the hybrid deep learning architecture, and three-dimensional prediction values ​​are output through the three parallel prediction branches.

[0009] The personalized solution module receives the three-dimensional predicted value and evaluates the three-dimensional confidence level of the three-dimensional predicted value. It calculates a comprehensive risk assessment based on the three-dimensional predicted value and the three-dimensional confidence level. It further presets a risk threshold, classifies the risk level based on the comprehensive risk assessment, and sets up a personalized rehabilitation plan for patients at each risk level.

[0010] Furthermore, the consistency score is calculated by using a scoring method that measures the deviation between each sub-data point of the preprocessed exercise rehabilitation data and the median of all sub-data points of the preprocessed exercise rehabilitation data.

[0011] The quality index is represented by the average consistency score of all preprocessed exercise rehabilitation data;

[0012] The quality weight matrix is ​​in the form of:

[0013] ;

[0014] in, Indicates the first a Quality weight matrix of subdata of preprocessed sports rehabilitation data; Indicates the first a The first subdata of the preprocessed sports rehabilitation data t Quality weights.

[0015] Furthermore, the preprocessed exercise rehabilitation data is weighted and calculated using the following formula: ;

[0016] in, Indicates the weighted number of digits. a The t-th data point of a sub-data set of sports rehabilitation data; Indicates the first a The t-th data point of a sub-data set of sports rehabilitation data;

[0017] The form of the three-dimensional tensor input data is:

[0018] ;

[0019] ;

[0020] in, Indicates the first Three-dimensional tensor input data; Indicates the number of samples; This indicates the length of the time window for each sample; This represents the data feature dimension for each time step; Represents a three-dimensional tensor dataset; This represents the total number of three-dimensional tensor data.

[0021] Furthermore, the hybrid deep learning architecture is constructed using a combination of bidirectional long short-term memory network branches and convolutional neural network branches;

[0022] The bidirectional long short-term memory network branch includes an input layer, a bidirectional LSTM layer, and a fully connected layer;

[0023] The input layer feeds the three-dimensional tensor input data into the bidirectional LSTM layer. The bidirectional LSTM encodes the data from both the forward and backward directions simultaneously, and finally obtains the bidirectional temporal feature vector. The bidirectional temporal feature vector is then fed into the fully connected layer.

[0024] The convolutional neural network branch includes an input layer, a convolutional layer, and a fully connected layer;

[0025] The input layer feeds the three-dimensional tensor input data into the convolutional layer, which extracts spatial feature vectors through convolution operations and feeds the spatial feature vectors into the fully connected layer.

[0026] The bidirectional temporal feature vector and the spatial feature vector are concatenated in a fully connected layer according to the feature dimension to obtain a comprehensive feature vector; the comprehensive feature vector is input into the output layer, and three parallel prediction branches are set in the output layer.

[0027] Furthermore, the three-dimensional prediction values ​​include predicted values ​​for rehabilitation effects, predicted values ​​for rehabilitation progress, and predicted values ​​for the risk of complications;

[0028] The process of setting up three parallel prediction branches is as follows:

[0029] The first prediction branch uses a regression layer to predict the patient's recovery outcome, denoted as:

[0030] ;

[0031] ;

[0032] in, Indicates the first One predicted value for rehabilitation effect; Indicates a fully connected operation; Indicates the first One comprehensive feature vector; This represents the learnable parameters of the first predictive branch regression layer;

[0033] The second prediction branch uses a regression layer to predict the patient's recovery progress, represented as:

[0034] ;

[0035] ;

[0036] in, Indicates the first One predicted value for recovery progress; This represents the learnable parameters of the second predictive branch regression layer;

[0037] The third predictive branch uses a categorical layer to predict the risk of patient complications, including muscle atrophy, joint stiffness, and increased pain, as shown below:

[0038] ;

[0039] ;

[0040] in, Indicates the first m One complication risk prediction value; This represents the softmax activation function; This represents the learnable parameters of the classification layer.

[0041] Furthermore, the formula for calculating the confidence level of each dimension of the three-dimensional confidence level is as follows:

[0042] ;

[0043] in, Indicates the first d Dimensional confidence level; Indicates the first d Variance of the predicted values;

[0044] The steps for calculating the comprehensive risk assessment are as follows:

[0045] 1) Construct an AHP judgment matrix based on clinical priority: complication risk > rehabilitation effect > rehabilitation progress.

[0046] 2) After performing a consistency check on the judgment matrix, calculate the weights of the judgment matrix;

[0047] 3) Then, use the AHP weighted summation to calculate the comprehensive assessment risk, as shown in the formula:

[0048] ;

[0049] in, This indicates a comprehensive risk assessment; , and This indicates the weights of the predicted values ​​for rehabilitation outcomes, rehabilitation progress, and complication risk. , and This indicates the confidence level of the predicted values ​​for rehabilitation effect, rehabilitation progress, and complication risk.

[0050] Furthermore, the risk threshold includes a low-risk threshold. and high risk threshold ;

[0051] Risk levels are classified according to risk thresholds as follows:

[0052] like If so, it is judged as low risk;

[0053] like If so, it is determined to be of medium risk;

[0054] like If so, it is judged as high risk;

[0055] The rehabilitation plan includes exercise intensity, training frequency, and medical intervention.

[0056] Compared with the prior art, the advantages of the present invention are as follows:

[0057] 1. This invention employs a hybrid deep learning architecture, combining two branches: a bidirectional long short-term memory network and a convolutional neural network. This architecture can simultaneously process time-series features and spatial features, thereby improving the accuracy of predicting sports rehabilitation data.

[0058] 2. This invention provides a personalized rehabilitation plan based on risk level by comprehensively assessing the patient's three-dimensional prediction values, and can customize specific rehabilitation strategies for patients with different risk levels;

[0059] 3. This invention ensures the reliability and validity of input data and improves the credibility of prediction results by performing consistency scoring, quality index generation and weighting on sports rehabilitation data. Attached Figure Description

[0060] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 This is a schematic diagram of the overall system workflow of the present invention;

[0062] Figure 2 This is a flowchart of the feature prediction module of the present invention;

[0063] Figure 3 This is a flowchart of the personalized solution module of the present invention. Detailed Implementation

[0064] To achieve the above objectives, the present invention provides a deep learning-based sports rehabilitation data prediction system. Please refer to the provided technical solution. Figures 1 to 3 The system includes the following modules:

[0065] The data processing module collects and preprocesses the patient's exercise rehabilitation data, which includes exercise data, physiological data, functional assessment data, and treatment intervention data; it performs a consistency score on the preprocessed exercise rehabilitation data, further generates quality indicators, and assigns weights to the preprocessed exercise rehabilitation data according to the quality indicators to generate a corresponding quality weight matrix.

[0066] Wearable sensors, scales and questionnaires, and electronic medical records are used to collect exercise rehabilitation data. Exercise rehabilitation data includes exercise data, physiological data, functional assessment data, and treatment intervention data, which are sub-data of exercise rehabilitation data.

[0067] Exercise data is data directly related to the patient's exercise behavior and is acquired through wearable devices; physiological data helps assess the patient's physical response during exercise rehabilitation and is acquired through wearable devices; functional assessment data helps assess the patient's rehabilitation progress and functional recovery and is obtained through scales and questionnaires; treatment intervention data records the various rehabilitation treatments and interventions received by the patient, helps assess the effectiveness of treatment and adjust rehabilitation plans, and is acquired through electronic medical records.

[0068] In this embodiment, the motion data collected from a 45-year-old male patient 6 weeks after right knee anterior cruciate ligament reconstruction were as follows: knee flexion angles of 30°, 45°, and 60°; angular velocity of knee flexion and extension of 15° / second; and vertical reaction force on the patient's side in a standing position of [missing data]. Physiological data includes the root mean square electromyography (EMG) value of the quadriceps femoris muscle. During training, the real-time heart rate was 95 beats / minute; blood oxygen saturation was 98%; functional assessment data showed a Lysholm score of 65 for the knee joint; active range of motion included flexion of 110 degrees and extension difference of 5 degrees, passive flexion of 120 degrees; and quadriceps long contraction strength was... Treatment intervention data: This training employed a combination of manual muscle strengthening exercises and machine-assisted joint range of motion training; the resistance level for machine resistance training was... Each group was repeated 12 times, for a total of 3 groups; the total duration of rehabilitation treatment was 45 minutes; treatment was administered for 3 consecutive weeks starting from the first week after surgery. , ;

[0069] In the data preprocessing stage, the sports rehabilitation data needs to be cleaned, including removing noise, filling in missing values, and removing outliers. In this embodiment, a z-score-based method is used for outlier detection, and interpolation is used to fill in missing values.

[0070] After preprocessing, consistency scores are calculated based on the distribution relationship between data points and the overall dataset. In this embodiment, a scoring method is used to measure the degree of statistical abnormality of each point of preprocessed exercise rehabilitation data by calculating the deviation between each sub-data of preprocessed exercise rehabilitation data and the median of all sub-data of preprocessed exercise rehabilitation data.

[0071] This consistency score reflects the reliability of each preprocessed exercise rehabilitation data point; the result of the consistency score is a [missing information]. Within the range of values, if the consistency score is closer to 1, it indicates that the preprocessed exercise rehabilitation data conforms to the overall distribution and has high reliability; if the consistency score is closer to 0, it indicates that the preprocessed exercise rehabilitation data differs significantly from the overall distribution and may be abnormal data.

[0072] The quality indicators of the preprocessed sports rehabilitation data are calculated based on the consistency score. The quality indicators include data completeness, accuracy, and consistency. The specific quality indicator is expressed as the average consistency score of all preprocessed sports rehabilitation data. The quality indicator is the overall quality score of the preprocessed sports rehabilitation data, and the higher the value, the better the data quality.

[0073] Weights are assigned to the sub-data points of the preprocessed exercise rehabilitation data based on quality indicators, generating a corresponding quality weight matrix; the quality weight matrix is ​​in the following form:

[0074] ;

[0075] in, Indicates the first a Quality weight matrix of subdata of preprocessed sports rehabilitation data; Indicates the first a The first subdata of the preprocessed sports rehabilitation data t Quality weights.

[0076] The feature prediction module performs weighted processing on the preprocessed exercise rehabilitation data according to the quality weight matrix to generate a weighted training set. The weighted training set is arranged according to the time series to form three-dimensional tensor input data. A hybrid deep learning architecture is constructed and three parallel prediction branches are set up. The three-dimensional tensor input data is input into the hybrid deep learning architecture, and three-dimensional prediction values ​​are output through the three parallel prediction branches.

[0077] The preprocessed exercise rehabilitation data is weighted and calculated using the following formula:

[0078] ;

[0079] in, Indicates the weighted number of digits. a The first sub-data of the sports rehabilitation data t One data point; Indicates the first a The first sub-data of the sports rehabilitation data t One data point;

[0080] The weighted training set is sorted chronologically to ensure that the temporal sequence of the data is not disrupted; the weighted training set is then divided into sliding window segments according to time windows to generate... M Each 3D tensor is used as input data, and the window size for each 3D tensor input data is set to... T , represented as:

[0081] ;

[0082] ;

[0083] in, Indicates the first Three-dimensional tensor input data; Indicates the number of samples; This indicates the length of the time window for each sample; This represents the data feature dimension for each time step; Represents a three-dimensional tensor dataset; This represents the total number of three-dimensional tensor data.

[0084] Five-fold cross-validation was used to test the impact of different T-values ​​on the predictive performance of the deep learning model, and the T-value that optimizes the model performance was selected. In the scenario of upper limb motor function rehabilitation after stroke, the feature dimension k was set to 15 dimensions, covering kinematics including wrist flexion / extension timing angle, elbow flexion angle, upper limb movement speed, and deviation of hand trajectory from standard trajectory; electromyography including root mean square electromyography timing value of biceps brachii and wavelet entropy of surface electromyography signal; functional assessment including current Fugl-Meyer upper limb score and modified Ashworth upper limb spasticity grade; and treatment intervention including the level of assistance from training assistive devices and the intensity of electrical stimulation current.

[0085] The 3D tensor input data will be used as input for training and prediction to build a hybrid deep learning architecture;

[0086] A hybrid deep learning architecture is constructed by combining a bidirectional long short-term memory network branch and a convolutional neural network branch. The bidirectional long short-term memory network branch encodes the time dimension to extract temporal feature vectors, while the convolutional neural network branch performs convolution operations on the feature dimension to extract spatial feature vectors. The two types of features are concatenated and fused to obtain a comprehensive feature representation.

[0087] Bidirectional long short-term memory (LSTM) branches can capture both past and future temporal information simultaneously through forward and reverse processing methods.

[0088] The bidirectional long short-term memory network branch includes an input layer, a bidirectional LSTM layer, and a fully connected layer;

[0089] The input layer feeds the 3D tensor input data into the bidirectional LSTM layer. Since the bidirectional LSTM encodes data from both the forward and backward directions simultaneously, it finally obtains a bidirectional temporal feature vector, which is then fed into the fully connected layer.

[0090] Convolutional neural networks consist of an input layer, convolutional layers, and fully connected layers.

[0091] The input layer feeds the 3D tensor input data into the convolutional layer, which contains multiple convolutional and pooling layers. The spatial feature vector is extracted through convolution operations and then fed into the fully connected layer.

[0092] The bidirectional temporal feature vector obtained from the bidirectional long short-term memory network branch and the spatial feature vector obtained from the convolutional neural network branch are concatenated in the fully connected layer according to the feature dimension to obtain a comprehensive feature vector; the comprehensive feature vector is input into the output layer, and three parallel prediction branches are set up in the output layer;

[0093] The first prediction branch uses a regression layer to predict the patient's rehabilitation outcome, representing the degree of recovery in joint range of motion, muscle strength, and motor coordination. The regression layer outputs a real value representing the patient's rehabilitation outcome. The formula for calculating the predicted rehabilitation outcome is as follows:

[0094] ;

[0095] ;

[0096] in, Indicates the first One predicted value for rehabilitation effect; Indicates a fully connected operation; Indicates the first One comprehensive feature vector; This represents the learnable parameters of the first predictive branch regression layer; This represents the set of predicted rehabilitation outcomes;

[0097] In this embodiment, a 45-year-old male patient 6 weeks after anterior cruciate ligament reconstruction of the knee joint was used as an example. Four types of sub-data were collected, including knee flexion angle and quadriceps electromyography value. After cleaning, a consistency score was calculated, generating a quality index of 0.89 and a weight matrix. After weighting, a three-dimensional tensor input model was formed according to the time series. The regression layer output a rehabilitation effect prediction value of 68.5 points out of 100 with a confidence level of 0.85. Based on this, the resistance of the device resistance training was reduced from 12kg to 10kg, and static straight leg raise training was added.

[0098] The second prediction branch uses a regression layer to predict the patient's recovery progress, representing the percentage of the patient's current recovery stage relative to the goal of full recovery. The regression layer outputs a real value representing the percentage of recovery progress. The formula for calculating the predicted recovery progress is:

[0099] ;

[0100] ;

[0101] in, This represents the predicted value of the i-th recovery progress; This represents the learnable parameters of the second predictive branch regression layer; This represents a set of predicted recovery progress values;

[0102] The goal of complete recovery was defined as a Lysholm score ≥90 and flexion ≥135° at 12 weeks post-surgery. Data was collected from patients at 4-6 weeks post-surgery. At 6 weeks post-surgery, the Lysholm score was 69 and the flexion was 88°. A tensor input model was constructed. The regression layer output a recovery progress prediction value of 60% with a confidence level of 0.88, indicating that the progress was reasonable. Training was maintained at 3 times per week, and a progress target of 75% was set at 8 weeks post-surgery.

[0103] The third prediction branch uses a classification layer to predict the risk of patients developing complications, including muscle atrophy, joint stiffness, and increased pain. The classification layer outputs the probability of each complication, using a softmax activation function to represent the probability distribution of each complication. The formula for calculating the predicted risk value of the complication is as follows:

[0104] ;

[0105] ;

[0106] in, Indicates the first m One complication risk prediction value; This represents the softmax activation function, ensuring that the output is a probability distribution and the sum of the probabilities of all complications is 1. This represents the learnable parameters of the classification layer; This represents the set of predicted risk values ​​for complications;

[0107] The quadriceps circumference of the above patients was collected (42cm at 1 week postoperatively and 40cm at 6 weeks postoperatively), passive flexion angle of the joint, and other data. After tensor input into the model, the classification layer output risk probabilities of muscle atrophy (15%), joint stiffness (30%), and increased pain (5%), with a confidence level of 0.90. The overall risk assessment was medium risk. Clinically, dynamic joint range of motion training was added, combined with ultrasound therapy twice a week.

[0108] Both the first and second prediction branches use mean squared error as the loss function in their regression layers, as shown in the formula:

[0109] ;

[0110] in, The loss function representing the mean squared error; Indicates the first i The true label of each sample; Indicates the first i The regression layer predicted values ​​for each sample;

[0111] The regression layer prediction values ​​include rehabilitation effect prediction values ​​and rehabilitation progress prediction values;

[0112] The classification layer of the third prediction branch uses cross-entropy loss as the loss function, as shown in the formula:

[0113] ;

[0114] in, Represents the cross-entropy loss function; Indicates the number of complication categories; Indicates the first i The first sample c The true labeling of similar complications; Indicates the first i The first sample c Predicted risk values ​​for complications of this type;

[0115] The final loss function is obtained by combining the loss functions of the regression layer and the classification layer. L The format is:

[0116] ;

[0117] The hybrid deep learning architecture is trained using the Adam optimizer, minimizing the loss function and updating the parameters of the hybrid deep learning architecture. , and By minimizing the loss function, the model can optimize its parameters, gradually narrowing the gap between the predicted results and the actual labels. The optimization process is carried out through backpropagation, and the gradient information of the loss function is used to adjust the model's parameters.

[0118] The three-dimensional tensor input data is fed into a hybrid deep learning architecture, and three-dimensional prediction values ​​are output through three parallel prediction branches. The three-dimensional prediction values ​​include rehabilitation effect prediction values, rehabilitation progress prediction values, and complication risk prediction values.

[0119] The personalized solution module receives the three-dimensional predicted value and evaluates the three-dimensional confidence level of the three-dimensional predicted value. It calculates a comprehensive risk assessment based on the three-dimensional predicted value and the three-dimensional confidence level. It further presets a risk threshold, classifies the risk level based on the comprehensive risk assessment, and sets up a personalized rehabilitation plan for patients at each risk level.

[0120] The predicted values ​​for rehabilitation effect, rehabilitation progress, and complication risk are synthesized, and each predicted value has a corresponding confidence level. Monte Carlo dropout technology is used to perform multiple forward propagations on the hybrid deep learning architecture, and the variance of each dimension's predicted value is calculated to assess the confidence level. A smaller variance indicates higher confidence, and vice versa. The formula for calculating the confidence level for each dimension is as follows:

[0121] ;

[0122] in, No. d Dimensional confidence level; Indicates the first d Variance of the predicted values;

[0123] Based on clinical priorities in sports rehabilitation, dynamic weights are determined using the analytic hierarchy process (AHP). Predicted values ​​are then adjusted using confidence levels, and the final comprehensive risk assessment steps are as follows:

[0124] 1) Construct an AHP judgment matrix based on clinical priority: complication risk > rehabilitation effect > rehabilitation progress.

[0125] The AHP judgment matrix is ​​constructed using the three-dimensional prediction values ​​of the system's core as evaluation dimensions, employing a 1-9 scale. 1 indicates that both dimensions are equally important, 2 indicates that the former is slightly important, 3 indicates that the former is significantly important, and so on up to 9, which indicates that the former is extremely important. The anti-diagonal is the reciprocal of the corresponding scale. Values ​​are assigned to the relative importance of each dimension. In this embodiment, the risk of complications is significantly more important than the rehabilitation effect and can be assigned a value of 3; the rehabilitation effect is slightly more important than the rehabilitation progress and can be assigned a value of 2; the risk of complications is very important than the rehabilitation progress and can be assigned a value of 4. Simultaneously, all diagonal elements of the matrix are 1, ultimately forming a 3×3 AHP judgment matrix.

[0126] 2) After performing a consistency check on the judgment matrix, calculate the weights of the judgment matrix;

[0127] 3) Then, use the AHP weighted summation to calculate the comprehensive risk assessment. The formula is:

[0128] ;

[0129] in, This indicates a comprehensive risk assessment; , and The weights of the predicted values ​​for rehabilitation outcomes, rehabilitation progress, and complication risks are used to balance the impact of different prediction tasks. , and The confidence level represents the predicted values ​​of rehabilitation effect, rehabilitation progress, and complication risk, and is a value in the range of [0,1], indicating the reliability of the prediction results.

[0130] Risk thresholds are set based on the mean and standard deviation of a comprehensive risk assessment, and low-risk thresholds are set accordingly. The format is:

[0131] ;

[0132] in, This represents the average of the overall risk assessment. This represents the standard deviation of the overall risk assessment.

[0133] Set a high-risk threshold The format is:

[0134] ;

[0135] Risk levels are classified according to preset risk thresholds as follows:

[0136] like If so, it is judged as low risk;

[0137] like If so, it is determined to be of medium risk;

[0138] like If so, it is judged as high risk;

[0139] Based on the patient's risk level, a personalized rehabilitation plan is customized; the specific rehabilitation plan includes exercise intensity, training frequency, and medical intervention.

[0140] For low-risk patients, rehabilitation progresses smoothly, so the rehabilitation plan is relatively easy and can adopt a conventional rehabilitation training program. The specific rehabilitation plan is as follows: moderate exercise intensity to avoid overtraining; training frequency of 3-4 times per week, adjusted according to the patient's physical condition; regular follow-up of medical interventions, and appropriate adjustments to the medical plan to ensure that no complications occur.

[0141] For patients at medium risk, more monitoring and adjustments are needed to avoid problems during the rehabilitation process. The specific rehabilitation plan is as follows: moderate exercise intensity, appropriately increasing the load on joints and muscles, but avoiding overtraining; training frequency of 4-5 times per week, monitoring rehabilitation progress, and adjusting training content as needed; medical intervention, increasing the involvement of rehabilitation doctors and therapists, regularly assessing joint range of motion and muscle strength to prevent complications.

[0142] For high-risk patients, rehabilitation is more complex and requires strict control and treatment. The specific rehabilitation plan is as follows: low to moderate exercise intensity, with a focus on monitoring and protecting joints to avoid overuse; training frequency of 2-3 times per week, with the intensity of training appropriately increased or decreased according to the patient's rehabilitation stage; frequent follow-up medical interventions, which may require drug treatment, physical therapy or other interventions, and close monitoring of complications.

[0143] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based sports rehabilitation data prediction system, characterized in that, include: The data processing module collects and preprocesses the patient's exercise rehabilitation data, which includes sub-data such as exercise data, physiological data, functional assessment data, and treatment intervention data. A consistency score is performed on the preprocessed exercise rehabilitation data to generate quality indicators. Weights are then assigned to the preprocessed exercise rehabilitation data based on the quality indicators to generate a corresponding quality weight matrix. The feature prediction module performs weighted processing on the preprocessed exercise rehabilitation data according to the quality weight matrix to generate a weighted training set. The weighted training set is arranged according to the time series to form three-dimensional tensor input data. A hybrid deep learning architecture is constructed and three parallel prediction branches are set up. The three-dimensional tensor input data is input into the hybrid deep learning architecture, and three-dimensional prediction values ​​are output through the three parallel prediction branches. The personalized solution module receives the 3D predicted value and evaluates the 3D confidence level of the 3D predicted value, and calculates a comprehensive risk assessment based on the 3D predicted value and the 3D confidence level. Further preset risk thresholds, classify risk levels based on the comprehensive risk assessment, and set up personalized rehabilitation plans for patients at each risk level; The hybrid deep learning architecture is constructed by combining bidirectional long short-term memory network branches and convolutional neural network branches; The three-dimensional tensor input data is input into the hybrid deep learning architecture to obtain a comprehensive feature vector; the comprehensive feature vector is then input into three parallel prediction branches to obtain a three-dimensional prediction value; The three-dimensional prediction values ​​include predicted values ​​for rehabilitation effects, predicted rehabilitation progress, and predicted risk of complications. The process of setting up three parallel prediction branches is as follows: The first prediction branch uses a regression layer to predict the patient's recovery outcome, denoted as: ； ； in, Indicates the first One predicted value for rehabilitation effect; Indicates a fully connected operation; Indicates the first One comprehensive feature vector; This represents the learnable parameters of the first predictive branch regression layer; The second prediction branch uses a regression layer to predict the patient's recovery progress, represented as: ； ； in, Indicates the first One predicted value for recovery progress; This represents the learnable parameters of the second predictive branch regression layer; The third predictive branch uses a categorical layer to predict the risk of patient complications, including muscle atrophy, joint stiffness, and increased pain, as shown below: ； ； in, This represents the predicted risk value for the m-th complication; This represents the softmax activation function; This represents the learnable parameters of the classification layer; The first and second prediction branches use mean squared error as the loss function, and the formula is as follows: ； in, The loss function representing the mean squared error; This represents the true label of the i-th sample; This represents the regression layer predicted value for the i-th sample; The regression layer prediction values ​​include rehabilitation effect prediction values ​​and rehabilitation progress prediction values; The classification layer of the third prediction branch uses cross-entropy loss as the loss function, as shown in the formula: ； in, Represents the cross-entropy loss function; Indicates the number of complication categories; This represents the true label of the c-th complication in the i-th sample; This represents the predicted risk value for the c-th type of complication in the i-th sample; Combining the loss functions of the regression and classification layers, the final loss function L takes the form: ； The hybrid deep learning architecture is trained using the Adam optimizer to minimize the loss function, and the parameters of the hybrid deep learning architecture are updated through backpropagation. , and ; The formula for calculating the confidence level of each dimension of the three-dimensional confidence level is as follows: ； in, Indicates the confidence level in the d-th dimension; This represents the variance of the predicted value in the d-th dimension; The steps for calculating the comprehensive risk assessment are as follows: 1) Construct an AHP judgment matrix, based on clinical priority of complication risk > rehabilitation effect > rehabilitation progress; 2) After performing a consistency check on the judgment matrix, calculate the weights of the judgment matrix; 3) Then, use the AHP weighted summation to calculate the comprehensive assessment risk, as shown in the formula: ； in, This indicates a comprehensive risk assessment; , and This indicates the weights of the predicted values ​​for rehabilitation outcomes, rehabilitation progress, and complication risk. , and This indicates the confidence level of the predicted values ​​for rehabilitation effect, rehabilitation progress, and complication risk.

2. The system according to claim 1, characterized in that, The consistency score is calculated by measuring the deviation of each preprocessed sub-data of exercise rehabilitation data from the median of all preprocessed sub-data of exercise rehabilitation data. The quality index is represented by the average consistency score of all preprocessed exercise rehabilitation data; The quality weight matrix is ​​in the form of: ； in, This represents the quality weight matrix of the a-th preprocessed subdata of the sports rehabilitation data; This represents the t-th quality weight of the sub-data of the a-th preprocessed exercise rehabilitation data.

3. The system according to claim 2, characterized in that, The preprocessed exercise rehabilitation data is weighted and calculated using the following formula: ； in, This represents the t-th data point of the sub-data point of the weighted a-th sports rehabilitation data point; This represents the t-th data point of the sub-data point of the a-th sports rehabilitation data point; The form of the three-dimensional tensor input data is: ； ； in, Indicates the first Three-dimensional tensor input data; Indicates the number of samples; This indicates the length of the time window for each sample; This represents the data feature dimension for each time step; Represents a three-dimensional tensor dataset; This represents the total number of three-dimensional tensor data.

4. The system according to claim 3, characterized in that, The risk threshold includes a low-risk threshold. and high risk threshold Risk levels are classified according to risk thresholds as follows: like If so, it is judged as low risk; like If so, it is determined to be of medium risk; like If so, it is judged as high risk; The rehabilitation plan includes exercise intensity, training frequency, and medical intervention.

Images