Method and device for generating post-pci rehabilitation exercise program based on machine learning

Abstract

This invention discloses a method and apparatus for generating post-PCI rehabilitation exercise programs based on machine learning, relating to the field of medical data processing. The method includes: collecting basic information of the rehabilitation subject; initializing safety constraints and prior anaerobic threshold test data for the rehabilitation exercise; collecting exercise state data, including the subject's subjective complaints and objective monitoring data, while the subject is executing the real-time rehabilitation exercise program; extracting multi-dimensional state features, including basic physiological characteristics and key feature reductions, from the exercise state data, where key feature reductions characterize the decline in physiological characteristics; and iteratively optimizing the real-time rehabilitation exercise program using an optimization algorithm, combining a machine learning-based exercise intensity evaluation model, prior anaerobic threshold test data, and multi-dimensional state features, to obtain a recommended exercise program. This solves the problems of traditional rehabilitation programs relying on human experience, delayed adjustments, and difficulty in precise quantification.

Description

Technical Field

[0001] This invention relates to the field of medical data processing, specifically to a method and apparatus for generating post-PCI rehabilitation exercise programs based on machine learning. Background Technology

[0002] Percutaneous coronary intervention (PCI) is one of the main treatments for coronary artery disease, effectively opening narrowed or blocked coronary arteries and restoring blood supply to the myocardium. However, PCI only addresses local vascular lesions and cannot reverse the pathological process of coronary atherosclerosis. Post-procedure, patients still face the risks of restenosis, new lesions, and recurrence of cardiovascular events.

[0003] In addition, most existing rehabilitation programs are static prescriptions. As the rehabilitation process progresses, the anaerobic threshold level will gradually increase, resulting in insufficient intensity in the later stages of the rehabilitation program. It is difficult to continuously and accurately anchor the anaerobic threshold intensity, and the evaluation dimensions are singular, making it impossible to dynamically adapt to the changes in rehabilitation after PCI. Summary of the Invention

[0004] This application provides a method and device for generating rehabilitation exercise programs after PCI based on machine learning, which addresses the problems of traditional rehabilitation programs relying on human experience, lagging adjustments, and difficulty in precise quantification.

[0005] In view of the above problems, this application provides a method and device for generating rehabilitation exercise programs after PCI based on machine learning.

[0006] In a first aspect, this application provides a method for generating post-PCI rehabilitation exercise programs based on machine learning, the method comprising: Collect basic information about the rehabilitation subjects, and initialize the safety constraints and prior anaerobic threshold test data for rehabilitation exercises based on the basic information. When rehabilitation subjects are implementing real-time rehabilitation exercise programs, exercise status data, including subjective complaints and objective monitoring data, are collected simultaneously. Multidimensional state features, including basic physiological characteristics and key feature reduction, are extracted from the motion state data, wherein the key feature reduction is used to characterize the decline in physiological characteristics; By combining a pre-built machine learning-based exercise intensity evaluation model, the prior anaerobic threshold test data, and the multidimensional state features, the real-time rehabilitation exercise program is iteratively optimized using an optimization algorithm to obtain a recommended exercise program.

[0007] Secondly, the present invention provides a machine learning-based device for generating post-PCI rehabilitation exercise programs, the device comprising: The basic information acquisition module is used to collect basic information of the rehabilitation subject and initialize the safety constraints of rehabilitation exercise and prior anaerobic threshold test data based on the basic information. The rehabilitation data acquisition module is used to simultaneously collect movement status data, including the rehabilitation subject's subjective complaint data and objective monitoring data, when the rehabilitation subject is performing a real-time rehabilitation exercise program; A multidimensional feature extraction module is used to extract multidimensional state features, including basic physiological features and key feature reductions, from the motion state data, wherein the key feature reductions are used to characterize the decrease in physiological features; The exercise program generation module is used to combine a pre-built machine learning-based exercise intensity evaluation model, the prior anaerobic threshold test data, and the multi-dimensional state features, and to iteratively optimize the real-time rehabilitation exercise program through an optimization algorithm to obtain a recommended exercise program.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application first initializes safety constraints and prior anaerobic threshold test data to establish initial safety boundaries and avoid risks that may arise from general solutions. Second, during the execution of the real-time rehabilitation exercise program, exercise state data is collected to integrate subjective feelings with objective physiological data, providing a rich multi-source data foundation. Third, multi-dimensional state features are acquired, and the ability to reduce the decline in physiological characteristics during rest periods is quantified through key features, providing dynamic physiological information for subsequent models. Finally, by combining a machine learning-based exercise intensity evaluation model, prior anaerobic threshold test data, and multi-dimensional state features, the real-time rehabilitation exercise program is iteratively optimized to obtain recommended exercise programs. This allows the exercise programs to dynamically approximate individualized anaerobic threshold intensity, improving the effectiveness of rehabilitation training and solving the problems of traditional rehabilitation programs relying on human experience, lagging adjustments, and difficulty in precise quantification. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating the machine learning-based method for generating post-PCI rehabilitation exercise programs in this application. Figure 2 This is a schematic diagram of the device for generating PCI postoperative rehabilitation exercise programs based on machine learning, as described in this application.

[0010] In the attached diagram, the components represented by each number are as follows: Basic information acquisition module 11, rehabilitation data acquisition module 12, multi-dimensional feature extraction module 13, exercise plan generation module 14. Detailed Implementation

[0011] This application provides a machine learning-based method for generating rehabilitation exercise programs after PCI, which specifically addresses the problems of traditional rehabilitation programs relying on human experience, lagging adjustments, and difficulty in precise quantification.

[0012] The present invention will now be described in detail with reference to the accompanying drawings.

[0013] Example 1, as Figure 1 As shown, this application provides a method for generating post-PCI rehabilitation exercise programs based on machine learning, the method comprising: S10: Collect basic information about the rehabilitation subject, and initialize the safety constraints and prior anaerobic threshold test data for rehabilitation exercises based on the basic information; Step S10 in the method provided in this application embodiment includes: Collect basic information about the rehabilitation subjects and calculate their estimated maximum heart rate using a preset maximum heart rate estimation formula; Based on the estimated maximum heart rate, the target anaerobic threshold heart rate range for the rehabilitation subject is calculated according to the preset anaerobic threshold heart rate percentage range, which serves as the initial value of the prior anaerobic threshold test data. Based on the basic information and the prior expert system, safety constraints for rehabilitation exercises are set, including at least the maximum safe heart rate threshold, the minimum blood oxygen saturation threshold, and the maximum blood pressure threshold.

[0014] In this embodiment, firstly, basic information of the patient undergoing PCI rehabilitation is collected. Then, using the patient's basic information and a preset maximum heart rate formula, the patient's estimated maximum heart rate is calculated. For example, if the preset maximum heart rate is 220 beats / minute, the maximum heart rate is estimated to be 165 beats / minute using the formula 220 minus age, resulting in a maximum heart rate of approximately 140 beats / minute.

[0015] Secondly, based on the estimated maximum heart rate, and according to the preset percentage range of heart rate corresponding to the anaerobic threshold, the target anaerobic threshold heart rate range is calculated for the rehabilitation subject. This range serves as the initial value for the prior anaerobic threshold test data, where the preset anaerobic threshold corresponds to a percentage range of heart rate. Based on exercise physiology theory and cardiac rehabilitation guidelines, the anaerobic threshold intensity of normal healthy individuals or cardiac rehabilitation patients typically falls within a certain percentage range of maximum heart rate; for example, it is 60%–70% for severely ill patients such as those with heart failure, or 70%–80% for patients with good recovery.

[0016] Specifically, based on the rehabilitation rules preset by the patient after PCI, a relatively conservative and suitable range for initial rehabilitation is set, corresponding to a percentage of heart rate corresponding to the anaerobic threshold. Then, the estimated heart rate is multiplied by this preset percentage range, resulting in two values. The lower value is taken as the lower limit, and the higher value as the upper limit, thus obtaining the target anaerobic threshold heart rate range. This range serves as the initial value for the prior anaerobic threshold test data, the starting point for subsequent optimization. For example, 165 beats / minute is multiplied by 65% ​​and 75% respectively, yielding two values. The lower value is 165 × 0.65 ≈ 107 beats / minute, and the higher value is 165 × 0.75 ≈ 124 beats / minute. Therefore, the target anaerobic threshold heart rate range is generated: 107–124 beats / minute.

[0017] Finally, based on basic information and the prior expert system, safety constraints for rehabilitation exercises are set. These constraints include at least the maximum safe heart rate threshold, the minimum blood oxygen saturation threshold, and the maximum blood pressure threshold. The prior expert system is a rule-based knowledge base that can translate expert knowledge and clinical guidelines in the field of cardiac rehabilitation into logical rules that computers can understand and execute. The safety constraints are a series of physiological parameter thresholds that must be strictly observed during exercise. The maximum safe heart rate threshold is the upper limit of the heart rate allowed during exercise. The minimum blood oxygen saturation threshold is the lowest allowed blood oxygen saturation during exercise, usually 90%–95%, below which indicates a risk of hypoxia. The maximum blood pressure threshold is the upper limit of the allowed systolic blood pressure during exercise, usually 180–200 mmHg, used to prevent the risk of cerebrovascular accidents.

[0018] Specifically, the internal expert system is invoked to set corresponding safety constraints for rehabilitation exercises based on the basic information of the rehabilitation subjects. For example, based on the patient's age, the maximum safe heart rate should not exceed 85% of the estimated maximum heart rate, and for all patients after PCI, the blood oxygen saturation should not be lower than 90% during exercise.

[0019] First, by combining the estimated maximum heart rate with the maximum safe heart rate, the product of the estimated maximum heart rate and the maximum safe heart rate is calculated to obtain the maximum safe heart rate threshold. Second, according to general rules, a minimum blood oxygen saturation threshold is set. The prior expert system may also set a maximum blood pressure threshold based on the rehabilitation subject's blood pressure, such as a systolic blood pressure not exceeding 180 mmHg. Finally, the maximum safe heart rate threshold, the minimum blood oxygen saturation threshold, and the maximum blood pressure threshold are used as safety constraints for rehabilitation exercises.

[0020] For example, based on the rule derived from the prior expert system: for patients after PCI, the maximum exercise heart rate must not exceed 85% of the estimated maximum heart rate. Combining this with the estimated maximum heart rate of 165 beats / min for a rehabilitation subject, the maximum safe heart rate threshold is calculated as 165 × 85% = 140 beats / min. According to the rule: for all rehabilitation patients, blood oxygen saturation during exercise must not be lower than 90%, so a minimum blood oxygen saturation threshold of 90% is set. The rehabilitation subject has the highest blood pressure. According to the rule: for patients with a history of hypertension, systolic blood pressure during exercise must not exceed 180 mmHg, so a maximum blood pressure threshold of 180 mmHg is set. This generates the initial safety constraints for the rehabilitation subject: heart rate must be <140 beats / min, blood oxygen saturation must be >90%, and systolic blood pressure must be <180 mmHg during exercise.

[0021] In this embodiment, the estimated maximum heart rate of the rehabilitation subject is first calculated, and the target anaerobic threshold heart rate range is calculated according to the preset anaerobic threshold heart rate percentage range as the initial value of the prior anaerobic threshold test data to obtain static physiological information. Subsequently, based on the basic information and the prior expert system, safety constraints are set to provide a scientific starting point and rigid constraints for the subsequent complex intelligent optimization process, ensuring that all subsequent adjustments are carried out within a safe range and effectively avoiding the risk of cardiovascular events induced by the deviation in exercise intensity estimation.

[0022] S20: When a rehabilitation subject is performing a real-time rehabilitation exercise program, exercise status data, including the subject's subjective complaint data and objective monitoring data, is collected simultaneously. Step S20 in the method provided in this application embodiment includes: Collect the chief complaint data input by the rehabilitation subjects before and after exercise, wherein the chief complaint data includes at least fatigue score, chest pain score and dyspnea score; The objective monitoring data of the rehabilitation subject during exercise is collected in real time by wearable devices. The objective monitoring data includes at least heart rate, heart rate variability, blood oxygen saturation, respiratory rate and exercise acceleration. The collected subjective complaint data and objective monitoring data are time-aligned using a unified timestamp, and the objective monitoring data is cleaned to obtain the motion state data.

[0023] In this embodiment of the application, the subject first collects the chief complaint data input by the rehabilitation subject before and after exercise. The chief complaint data includes at least fatigue score, chest pain score and dyspnea score.

[0024] Specifically, the chief complaint data is data that patients actively report and enter based on their own subjective feelings. It cannot be measured by instruments and can directly reflect the patient's internal feelings and potential risks. The fatigue score is an assessment of the subjective degree of fatigue of the rehabilitation subject based on the fatigue rating scale. The chest pain score is the rehabilitation subject's rating of chest discomfort, pain or pressure. The dyspnea score is the rehabilitation subject's rating of shortness of breath or difficulty breathing.

[0025] For patients after PCI, chest pain is an important warning sign of myocardial ischemia; dyspnea scores can reflect the immediate response of cardiopulmonary function to exercise load. Therefore, by using subjective data including fatigue scores, chest pain scores, and dyspnea scores, the physical indicators of rehabilitation subjects can be assessed to determine whether the exercise intensity is appropriate and whether there are potential risks. Risks can be detected through subjective feelings before abnormal changes occur in the patient's physiological indicators.

[0026] Secondly, wearable devices are used to collect objective monitoring data of rehabilitation subjects during exercise in real time. The objective monitoring data includes at least heart rate, heart rate variability, blood oxygen saturation, respiratory rate, and exercise acceleration.

[0027] Specifically, rehabilitation participants can wear smart devices that continuously monitor multiple physiological and motor parameters to collect and record objective monitoring data during exercise. Heart rate, the number of times the heart beats per minute, reflects the immediate response of the cardiovascular system to exercise load; heart rate variability, obtained by monitoring changes in the differences between successive heartbeat cycles, reflects the balance between the sympathetic and parasympathetic nervous systems in the autonomic nervous system, and assesses the body's recovery ability and fatigue level.

[0028] By monitoring the percentage of oxyhemoglobin in the blood relative to the total available hemoglobin, blood oxygen saturation is obtained, reflecting the oxygenation function of the respiratory and circulatory systems and ensuring exercise safety. Respiratory rate reflects the respiratory system's metabolic demands. Simultaneously, triaxial motion data collected by the device's built-in accelerometer acquires motion acceleration, identifies exercise status, calculates steps, and assesses exercise intensity. This allows for the capture of subtle changes and dynamic trends in physiological parameters during exercise, enabling timely detection of dangerous conditions such as arrhythmias, myocardial ischemia, and decreased blood oxygenation, ensuring the safe conduct of exercise rehabilitation.

[0029] Finally, using a unified timestamp, the chief complaint data is associated with the objective monitoring data at the same time of entry. Then, outliers in the objective monitoring data are checked and replaced with interpolation or nearest-valued valid data. The motion acceleration signal is then filtered to remove high-frequency noise and obtain motion state data, such as: [Time 10:35:20, Heart rate 118, Blood oxygen 97%, Respiratory rate 22, Motion acceleration 1, Fatigue score 12, Chest pain score 0, Dyspnea score 2].

[0030] In this embodiment, when the rehabilitation subject performs a real-time rehabilitation exercise program, the patient's subjective complaint data is collected simultaneously to capture the patient's internal feelings that cannot be measured by machines. Subsequently, objective monitoring data is collected in real time through wearable devices and time-aligned according to a unified timestamp, which solves the problem of asynchronous time for data from different sources. Furthermore, noise is removed through data cleaning, thereby providing a high-quality data foundation for subsequent feature extraction and model evaluation.

[0031] S30: Extract multidimensional state features from the motion state data, including basic physiological characteristics and key feature reduction, wherein the key feature reduction is used to characterize the decline of physiological characteristics; Step S30 in the method provided in this application embodiment includes: Trend analysis is performed on the objective monitoring data in the motion state data, and the objective monitoring data is divided into motion time windows and rest time windows; Traverse multiple exercise period windows and rest period windows to extract basic physiological characteristics; Using each rest period window as the analysis window, key physiological features are selected from the basic physiological features, and the key feature reduction of each key physiological feature is obtained through traversal analysis. The basic physiological characteristics and key characteristics under multiple exercise period windows and rest period windows are merged and output to reduce them to the multidimensional state characteristics.

[0032] In this embodiment, firstly, the inherent laws and stage changes of the data are identified by observing the trend, fluctuation and pattern of the data over time. For example, the process of heart rate gradually increasing, plateauing and then gradually decreasing over time is observed to perform trend analysis on the objective monitoring data in the exercise state data. Then, based on the results of the trend analysis, the objective monitoring data is divided into exercise period windows and rest period windows. During the exercise period window, physiological indicators such as heart rate and respiratory rate remain at a relatively high level and are relatively stable, and the exercise acceleration shows periodic fluctuations. During the rest period window, physiological indicators such as heart rate and respiratory rate continue to decrease from the exercise peak, and the exercise acceleration is close to zero.

[0033] Secondly, each exercise window and each rest window are processed individually. The raw data within each window are statistically analyzed to obtain basic physiological characteristics. For example, for the exercise window, the average heart rate, maximum heart rate, and mean heart rate variability can be calculated; for the rest window, the heart rate at the end of the rest period and the average blood oxygen saturation can be calculated.

[0034] For example, consider processing a specific exercise time window [10:30:00, 10:40:00]. Statistical analysis is performed on all heart rate data to extract basic physiological characteristics: average heart rate = 118 beats / minute, peak heart rate = 125 beats / minute, mean heart rate variability = 35ms. For the rest time window [10:40:00, 10:42:00], analysis is performed to extract basic physiological characteristics: heart rate at the end of rest = 98 beats / minute, average blood oxygen saturation during rest = 97%, lowest heart rate during rest = 95 beats / minute.

[0035] Next, using each rest period window as the analysis window, key physiological characteristics were selected from the basic physiological characteristics, and the key feature reductions for each key physiological characteristic were analyzed. Among them, the key feature reductions include the rate of decrease in heart rate during the rest period, the heart rate recovery value, and the rate of recovery of blood oxygen saturation.

[0036] Specifically, based on prior knowledge, the indicators that best reflect recovery ability are selected from numerous basic physiological characteristics as key physiological features, including the rate of heart rate decline during rest, the heart rate recovery value, and the rate of blood oxygen saturation recovery. Heart rate recovery usually follows an exponential decay law: HR(t) = HR' × e(-t / t) + HR0, where HR' is the final heart rate value and HR0 is the initial heart rate value. The extracted feature decay time constant can be used as an indicator of recovery ability, reflecting the patient's physical recovery level and closely related to exercise intensity.

[0037] Finally, using correlation analysis and feature importance assessment, the subset of features most relevant to the anaerobic threshold intensity was selected from all features to determine the features for which feature reduction needs to be calculated, ensuring that the calculation of feature reduction is based on highly relevant features.

[0038] Specifically, the Pearson coefficient can be used to calculate the Pearson correlation coefficient between each basic physiological characteristic and the anaerobic threshold intensity achievement rate. The Pearson correlation coefficient measures the degree of linear correlation between two continuous variables, and its value ranges from -1 to 1. The closer the absolute value is to 1, the stronger the linear correlation; a positive value indicates a positive correlation, and a negative value indicates a negative correlation.

[0039] First, for basic physiological characteristics and the rate of achieving the anaerobic threshold, the Pearson correlation coefficient is calculated by dividing the covariance of the two characteristics by their respective standard deviations, reflecting the strength of the linear association between the characteristics and the anaerobic threshold. Then, a screening threshold is set based on the absolute value of the correlation coefficient, for example, 0.3 is used as the inclusion threshold. All characteristics with |r|≥0.3 are retained, while those with a weak correlation to the anaerobic threshold, i.e., |r|<0.3, are removed.

[0040] Secondly, feature importance is assessed for the features after correlation analysis. For example, the importance of features is assessed using random forest, with the anaerobic threshold intensity attainment rate as the target variable, and a random forest regression model is constructed. The random forest consists of multiple decision trees. Each tree uses Bootstrap sampling to randomly sample from the original training set, and randomly selects some features as candidate split features at each node. The number of decision trees is set to 500 to 1000 to ensure the stability of feature importance estimation. The maximum depth of each tree is limited to 10 to 20 layers to prevent overfitting. The minimum number of samples required for node splitting is set to 5 to 10.

[0041] Subsequently, the feature importance of the random forest is calculated based on the reduction in Gini importance. For each node of each decision tree, the reduction in the Gini coefficient before and after the split is calculated, and the reduction is attributed to the feature used for the split. The reduction values ​​for all trees and all nodes are summed by feature and averaged by the number of trees to obtain the Gini importance score for each feature. Simultaneously, precision importance is used for cross-validation to calculate the model prediction error on the original validation set; then, the feature values ​​are randomly permuted, and the prediction error is calculated again; the increase in error is used as the precision importance of the feature, and this process is repeated multiple times, with the average value taken to obtain a robust importance estimate.

[0042] Finally, the Gini importance and precision importance are averaged to obtain the final importance score for each feature. By setting an importance score threshold, for example, retaining features with scores higher than 50% of the average score, key feature reduction is finally obtained. The basic physiological features extracted from all motion windows and the basic physiological features extracted from all rest windows, along with the key feature reduction, are merged and concatenated into a fixed-length vector as the final output multidimensional state features.

[0043] In this embodiment, trend analysis is performed on the objective monitoring data in the exercise state data to divide the exercise period window and the rest period window. By extracting key features of the rest period between exercise intervals, the characteristics of the load period and the recovery period are extracted and fused separately, so that the final multidimensional state features include multiple states. This provides more discriminative and physiologically significant feature inputs for the subsequent exercise intensity evaluation model, and enhances the accuracy of identifying whether the patient has reached the anaerobic threshold.

[0044] S40: Combining the pre-built machine learning-based exercise intensity evaluation model, the prior anaerobic threshold test data, and the multidimensional state features, the real-time rehabilitation exercise program is iteratively optimized using an optimization algorithm to obtain a recommended exercise program.

[0045] In step S40 of the method provided in this application embodiment, the construction of the exercise intensity evaluation model includes: Obtain publicly available anaerobic threshold assessment sets and patient historical training data, and map them to a unified feature dimension to obtain publicly available sample sets and patient historical sample sets; Based on the publicly available sample set, a benchmark motion intensity evaluation model based on a self-attention neural network model is constructed and trained, wherein the benchmark motion intensity evaluation model takes the publicly available multidimensional state features corresponding to the publicly available sample set as input and the intensity label supervision is fixed at 1. Based on the prior anaerobic threshold test data, the intensity label is fixed at 1, and the benchmark exercise intensity evaluation model is fine-tuned once in an iterative manner. Based on the patient historical sample set, unsupervised clustering is performed on the relationships between exercise intensity labeled by experts to obtain pseudo-labels. Then, the patient historical sample set and the pseudo-labels are combined to perform a second fine-tuning training on the results of the first fine-tuning training to obtain the exercise intensity evaluation model.

[0046] In this embodiment, firstly, the anaerobic threshold evaluations of other individuals are recorded as training samples to obtain a publicly available anaerobic threshold evaluation set, ensuring that the desired output is uniformly 1. Simultaneously, data accumulated from other patients after PCI in the past are collected as historical patient training data. Subsequently, the publicly available anaerobic threshold evaluation set and the historical patient training data are uniformly transformed into feature vectors with the same meaning and scale through methods such as feature selection, feature transformation, and normalization, and mapped to a unified dimension to obtain the publicly available sample set and the historical patient sample set.

[0047] Secondly, a self-attention neural network model was selected as the basic architecture. Based on the public sample set, a benchmark motion intensity evaluation model based on the self-attention neural network model was constructed and trained. The benchmark motion intensity evaluation model takes the public multi-dimensional state features corresponding to the public sample set as input and the intensity label supervision is fixed at 1.

[0048] Specifically, for each sample in the publicly available sample set, since the collection time corresponds to the point where the subject just reached the anaerobic threshold, all samples are uniformly assigned a label value of 1. The current exercise intensity is exactly equal to the subject's anaerobic threshold intensity, i.e., the achievement rate is 1. By setting labels, when the physiological pattern of reaching the anaerobic threshold is input as a multidimensional state feature, the model output value should approach 1. Subsequently, through repeated iterations, hundreds of millions of parameters within the network are adjusted. Through learning, the benchmark exercise intensity evaluation model outputs a value close to 1 when the input feature combination [average heart rate, heart rate recovery value] is applied.

[0049] For example, a self-attention neural network is first constructed. The structure of a self-attention neural network includes: an input embedding layer, a multi-head attention layer, a fully connected layer, and an output layer. For the input feature vector, the model first maps the input to the embedding space through a linear transformation, obtaining an embedding matrix. Treating the input features as a sequence, each feature dimension corresponds to a position in the sequence. To preserve the order relationship between features, the model uses periodic variations of sine and cosine functions to add positional encoding information to correspond to the relative positional relationships between different features. The input layer has two nodes, corresponding to the two feature dimensions: [average heart rate] and [heart rate recovery value].

[0050] In the multi-head attention layer, a multi-head attention mechanism is used to partition the query, key, and value matrix according to the number of heads. Each attention head independently performs scaled dot product attention calculations. Subsequently, the attention scores are scaled and normalized so that the sum of the attention weights of each feature to all other features is 1. Each multi-head attention layer is followed by residual connections and layer normalization. Residual connections directly add the layer's input and output, allowing gradients to bypass the current layer and propagate directly to shallower layers, helping to solve the gradient vanishing problem in deep networks.

[0051] The fully connected layer includes two hidden layers with 128 and 64 neurons respectively, using ReLU activation. A Dropout layer is inserted in between to prevent overfitting. The output layer is a single neuron that maps high-dimensional features to a one-dimensional output space, then compresses them to (0,1) using a Sigmoid activation function. The output value is the percentage of the current exercise intensity relative to the anaerobic threshold. An output value of 1 indicates that the anaerobic threshold has been reached, meaning the current intensity is close to the anaerobic threshold; an output value <1 indicates that the current intensity is below the anaerobic threshold; and an output value >1 indicates that the current intensity is above the anaerobic threshold.

[0052] Subsequently, model training was performed. The input public sample set was divided into a training set and a validation set in an 8:2 ratio, and the training set was used for model training. The mean squared error loss function was used to calculate the square of the difference between the predicted value and the target value of 1 for each sample. Then, the average of all samples was taken to obtain the difference value. The training aimed to minimize the difference between the predicted value and the target value of 1, and label smoothing technology was used to avoid overfitting.

[0053] The Adam optimizer is employed with an initial learning rate of 0.0001. The Adam optimizer calculates an adaptive learning rate for each parameter individually, balancing the update step sizes for different parameters. Forward propagation is used to calculate the predicted values ​​for all samples. After calculating the loss, the gradient of the loss with respect to each parameter is calculated using the backpropagation algorithm. Backpropagation starts from the output layer and propagates backward along the computation graph, applying the chain rule to calculate the partial derivative of the loss with respect to each parameter layer by layer. Simultaneously, the gradients for the query, key, and value are calculated and passed to the embedding layer and positional encoding layer using the chain rule. After gradient calculation, the optimizer updates the model parameters based on the gradient values ​​and the adaptive learning rate.

[0054] During training, the training loss and validation loss are continuously monitored, and a validation set is used to evaluate the model's generalization ability. When the validation loss stops decreasing or begins to increase, an early stopping mechanism can be triggered to stop training and save the model parameters with the lowest validation loss. For example, when both the training loss and validation loss tend to stabilize, the loss decrease is less than a preset threshold over several consecutive iterations, and the prediction mean on the validation set is close to 1 and the prediction variance is small, the benchmark motion intensity evaluation model can be considered to have converged, and the benchmark motion intensity evaluation model is obtained.

[0055] Secondly, based on prior anaerobic threshold test data, the intensity label was fixed at 1. On the baseline exercise intensity evaluation model, a small number of network parameters were slightly adjusted using simulated post-PCI patient data generated from the prior anaerobic threshold test data. By slightly adjusting the model's parameters, it was made more adaptable to the new scenario, guiding the model to focus on the possible performance range of post-PCI patients, making its evaluation criteria more aligned with this specific group.

[0056] For example, based on prior anaerobic threshold heart rate test data, several simulated samples of patients reaching the anaerobic threshold after PCI are generated. All simulated samples are labeled as 1, and a previously trained baseline exercise intensity evaluation model is loaded as input for training. The model parameters are then slightly adjusted. After training, the model's output of 0.9 for a certain prior anaerobic threshold heart rate test data is adjusted to 0.98, which is closer to 1.

[0057] Finally, based on the patient historical sample set, some data was annotated by experts. This annotated data was then used as seeds for unsupervised clustering analysis of the remaining historical data. Several data points with similar characteristics to the seed data were also grouped into the "reaching anaerobic threshold" cluster and given a pseudo-label of 1. The remaining data were given pseudo-labels 0 (below the anaerobic threshold) or 2 (above the anaerobic threshold), forming three clusters: below the anaerobic threshold, reaching the anaerobic threshold, and above the anaerobic threshold. Finally, the entire patient historical sample set with pseudo-labels was input into the model after one fine-tuning training for a second fine-tuning training. Through this second fine-tuning, the final exercise intensity evaluation model was constructed.

[0058] Step S40 in the method provided in this application embodiment includes: Establish an empirical model between exercise program parameters and exercise intensity evaluation values ​​based on anaerobic threshold; Based on prior anaerobic threshold test data and typical anaerobic domain test data, calculate the individual correction coefficient corresponding to the rehabilitation subject; By combining the empirical model with the individual correction coefficient, an individual strength mapping model for the rehabilitation subject is obtained; Based on the individual intensity mapping model, a cost function is defined, and the real-time rehabilitation exercise program is iteratively optimized according to the exercise intensity evaluation value output by the exercise intensity evaluation model to obtain a recommended exercise program.

[0059] In this embodiment, firstly, based on the output value of the exercise intensity evaluation model, the percentage of the current exercise intensity equivalent to the individual's anaerobic threshold intensity is determined. For example, an exercise intensity evaluation value of 1 indicates that the anaerobic threshold has been reached, 1.2 indicates that the anaerobic threshold has been exceeded by 20%, and 0.8 indicates that the anaerobic threshold has been exceeded by 20%. Subsequently, based on knowledge of exercise physiology, a simple empirical model is established, for example, assuming a linear relationship between the exercise intensity evaluation value and the exercise speed v: Exercise intensity evaluation value = k × v + b. Here, k and b are constants set based on average population data.

[0060] Secondly, prior anaerobic threshold test data are initial estimates of a patient's anaerobic threshold level based on basic patient information; while typical anaerobic threshold test data are reference values ​​representing the anaerobic threshold of typical or average recovered patients. For example, for a typical 55-year-old post-PCI patient, there is a statistically average anaerobic threshold heart rate of approximately 70% of the maximum heart rate.

[0061] Specifically, the prior anaerobic threshold test data is compared with the typical anaerobic threshold test data of patients after PCI. The individual correction coefficient is obtained by calculating the ratio or difference between the two: Individual Correction Coefficient = Prior Anaerobic Threshold Heart Rate / Typical Anaerobic Threshold Heart Rate. For example, for a rehabilitation subject, assuming a typical anaerobic threshold heart rate of 120 beats / minute and a prior anaerobic threshold test data of 115.5 beats / minute, the individual correction coefficient would be 115.5 / 120 = 0.9625.

[0062] Next, the empirical model is combined with the individual correction coefficient to generate an individual intensity mapping model. The product of the empirical model and the individual correction coefficient can be used as the new individual intensity mapping model. For example, the new model is: Exercise intensity evaluation value = (k × v + b) × individual correction coefficient.

[0063] Finally, based on the individual intensity mapping model, a cost function is defined, and multidimensional state features are input into the exercise intensity evaluation model to obtain the true exercise intensity evaluation value. Based on the individual intensity mapping model, an optimization algorithm is used to iteratively optimize the exercise intensity evaluation value according to the difference between the true exercise intensity evaluation value and 1, generate a new scheme, and finally obtain the recommended exercise scheme.

[0064] In step S40 of the method provided in this application embodiment, a cost function is defined based on the individual intensity mapping model, and the real-time rehabilitation exercise program is iteratively optimized according to the exercise intensity evaluation value output by the exercise intensity evaluation model to obtain a recommended exercise program, including: The multidimensional state features are input into the exercise intensity evaluation model to obtain the exercise intensity evaluation value, wherein the exercise intensity evaluation value is a percentage intensity based on the anaerobic threshold; Determine whether the exercise intensity evaluation value meets the preset intensity threshold. If it does, output the real-time rehabilitation exercise plan as the recommended exercise plan. If the conditions are not met, the optimization algorithm is initialized with the real-time rehabilitation exercise program, and the mean square error between the output of the individual intensity mapping model and 1 is defined as the cost function. The real-time rehabilitation exercise program is iteratively optimized to obtain a recommended exercise program.

[0065] In this embodiment, firstly, after completing a round of rehabilitation exercises, a set of multidimensional state features is extracted from the exercise state data, and then input into an exercise intensity evaluation model specifically designed for post-PCI patients. The model performs complex internal mathematical calculations and outputs an exercise intensity evaluation value, which is a percentage intensity based on the anaerobic threshold.

[0066] Secondly, it is determined whether the exercise intensity evaluation value meets the preset intensity threshold. If it does, the real-time rehabilitation exercise program is output as a recommended exercise program. The preset intensity threshold is a pre-set error range. Since both model evaluation and physiological response have certain fluctuations, as long as the evaluation value falls within a reasonable range near the target value of 1, the program can be considered to meet the standard. For example, the threshold is set to [95%, 105%].

[0067] Finally, the exercise intensity evaluation value obtained in the previous step is compared with the preset intensity threshold. If the exercise intensity evaluation value is not within the threshold range, the compliance condition is not met.

[0068] If the conditions are not met, the real-time rehabilitation exercise program is used as the initial starting point for the search. A cost function is defined, and the predicted output is made based on the individual intensity mapping model. The mean square error between the model's predicted evaluation value and the target value 1 is calculated and optimized through the cost function. Through multiple trials and calculations, the minimum speed value is finally found, and the minimum speed value and its corresponding complete program are used as the final recommended exercise program.

[0069] In step S40 of the method provided in this application embodiment, which defines a cost function based on the individual intensity mapping model, the method further includes: Define safety penalty items based on safety constraints; The mean square error between the output of the individual intensity mapping model and 1 is defined as the intensity compliance term; The cost function is obtained by combining the strength compliance item and the safety penalty item.

[0070] In this embodiment, firstly, a safety penalty term is defined based on the set safety constraints. This penalty term is a penalty function; when the exercise program parameters or the predicted physiological response caused by them approach or exceed the safety constraints, the function value increases sharply, penalizing programs that may lead the patient into a dangerous state. The safety penalty term can be defined as: safety penalty term = w1 × max(0, predicted heart rate - 140). 2 +w2×max(0,90-predicted blood oxygen) 2 Where w1 and w2 are weighting coefficients.

[0071] Secondly, the square of the difference between the output of the individual strength mapping model and 1 is defined as the strength achievement term, i.e., strength achievement term = (f(v) - 1). 2 The intensity attainment metric quantifies the performance of candidate exercises in achieving the anaerobic threshold. A smaller value indicates that the predicted intensity is closer to the target.

[0072] Finally, the intensity target achievement item and the safety penalty item are combined to obtain the final cost function. The cost function is obtained by adding the intensity target achievement item and the safety penalty item, and the cost function is w1×max(0, predicted heart rate-140). 2 +w2×max(0,90-predicted blood oxygen) 2 +(f(v)-1) 2 The final recommended exercise plan is selected by using a cost function with the goal of minimizing both the intensity target and safety penalty terms.

[0073] In this embodiment, a publicly available sample set and a patient historical sample set are obtained to train a baseline exercise intensity evaluation model. Then, the baseline model is fine-tuned based on prior anaerobic threshold test data. After obtaining pseudo-labels, a second fine-tuning training is performed to obtain the final exercise intensity evaluation model, providing an accurate prediction tool for the optimization algorithm. Subsequently, an empirical model between exercise program parameters and exercise intensity evaluation values ​​is established, individual correction coefficients are calculated, and an individual intensity mapping model is obtained. Then, the exercise intensity evaluation model is used to determine whether it meets the preset intensity threshold, and iterative optimization is performed to achieve closed-loop adaptive adjustment of the program generation. This continuously tracks the time-varying nature of the patient's functional state and achieves the optimal balance between the effectiveness and safety of the recommended exercise program, improving the scientific nature of post-PCI rehabilitation exercise prescriptions.

[0074] The embodiments of this application, through the above specific implementation methods, achieve the following technical effects: In this embodiment, the estimated maximum heart rate of the rehabilitation subject is first calculated, and the target anaerobic threshold heart rate range is calculated according to the preset anaerobic threshold heart rate percentage range as the initial value of the prior anaerobic threshold test data to obtain static physiological information. Subsequently, based on the basic information and the prior expert system, safety constraints are set to provide a scientific starting point and rigid constraints for the subsequent complex intelligent optimization process, ensuring that all subsequent adjustments are carried out within a safe range and effectively avoiding the risk of cardiovascular events induced by the deviation in exercise intensity estimation.

[0075] Secondly, when rehabilitation subjects are implementing real-time rehabilitation exercise programs, subjective complaint data is collected simultaneously to capture patients' internal feelings that cannot be measured by machines. Subsequently, objective monitoring data is collected in real time through wearable devices and time-aligned according to a unified timestamp, which solves the problem of asynchronous time for data from different sources. Furthermore, noise is removed through data cleaning, thus providing a high-quality data foundation for subsequent feature extraction and model evaluation.

[0076] Furthermore, trend analysis was performed on the objective monitoring data in the exercise state data to divide the exercise period window and the rest period window. By extracting key features of the rest period between exercise intervals, the characteristics of the load period and the recovery period were extracted and fused to represent them separately. This resulted in the final multidimensional state features containing multiple states, providing more discriminative and physiologically significant feature inputs for the subsequent exercise intensity evaluation model, and enhancing the accuracy of identifying whether the patient has reached the anaerobic threshold.

[0077] Finally, a public sample set and a patient historical sample set were obtained to train the baseline exercise intensity evaluation model. Then, the baseline model was fine-tuned based on prior anaerobic threshold test data. After obtaining pseudo-labels, a second fine-tuning training was performed to obtain the final exercise intensity evaluation model, providing an accurate prediction tool for the optimization algorithm. Subsequently, an empirical model between exercise protocol parameters and exercise intensity evaluation values ​​was established, and individual correction coefficients were calculated to obtain an individual intensity mapping model. Then, the exercise intensity evaluation model was used to determine whether it met the preset intensity threshold, and iterative optimization was performed to achieve closed-loop adaptive adjustment of the protocol generation. This allows for continuous tracking of the time-varying nature of the patient's functional status, while simultaneously achieving the optimal balance between the effectiveness and safety of the recommended exercise protocol, thus improving the scientific rigor of post-PCI rehabilitation exercise prescriptions.

[0078] Example 2, as Figure 2 As shown, based on the same inventive concept as the machine learning-based PCI post-operative rehabilitation exercise program generation method provided in Embodiment 1, this embodiment of the invention also provides a machine learning-based PCI post-operative rehabilitation exercise program generation device, including: The basic information acquisition module 11 is used to collect basic information of the rehabilitation subject and initialize the safety constraints and prior anaerobic threshold test data of the rehabilitation exercise based on the basic information. The rehabilitation data acquisition module 12 is used to simultaneously collect exercise status data, including the rehabilitation subject's subjective complaint data and objective monitoring data, when the rehabilitation subject is performing a real-time rehabilitation exercise program; The multidimensional feature extraction module 13 is used to extract multidimensional state features, including basic physiological features and key feature reduction, from the motion state data, wherein the key feature reduction is used to characterize the decrease in physiological features; The exercise program generation module 14 is used to combine a pre-built machine learning-based exercise intensity evaluation model, the prior anaerobic threshold test data, and the multidimensional state features, and to iteratively optimize the real-time rehabilitation exercise program through an optimization algorithm to obtain a recommended exercise program.

[0079] In one embodiment, the basic information acquisition module 11 is used for: Collect basic information about the rehabilitation subjects and calculate their estimated maximum heart rate using a preset maximum heart rate estimation formula; Based on the estimated maximum heart rate, the target anaerobic threshold heart rate range for the rehabilitation subject is calculated according to the preset anaerobic threshold heart rate percentage range, which serves as the initial value of the prior anaerobic threshold test data. Based on the basic information and the prior expert system, safety constraints for rehabilitation exercises are set, including at least the maximum safe heart rate threshold, the minimum blood oxygen saturation threshold, and the maximum blood pressure threshold.

[0080] In one embodiment, the rehabilitation data acquisition module 12 is used for: Collect the chief complaint data input by the rehabilitation subjects before and after exercise, wherein the chief complaint data includes at least fatigue score, chest pain score and dyspnea score; The objective monitoring data of the rehabilitation subject during exercise is collected in real time by wearable devices. The objective monitoring data includes at least heart rate, heart rate variability, blood oxygen saturation, respiratory rate and exercise acceleration. The collected subjective complaint data and objective monitoring data are time-aligned using a unified timestamp, and the objective monitoring data is cleaned to obtain the motion state data.

[0081] In one embodiment, the multidimensional feature extraction module 13 is used for: Trend analysis is performed on the objective monitoring data in the motion state data, and the objective monitoring data is divided into motion time windows and rest time windows; Traverse multiple exercise period windows and rest period windows to extract basic physiological characteristics; Using each rest period window as the analysis window, key physiological features are selected from the basic physiological features, and the key feature reduction of each key physiological feature is obtained through traversal analysis. The basic physiological characteristics and key characteristics under multiple exercise period windows and rest period windows are merged and output to reduce them to the multidimensional state characteristics.

[0082] In one embodiment, the motion scheme generation module 14 is used for: Obtain publicly available anaerobic threshold assessment sets and patient historical training data, and map them to a unified feature dimension to obtain publicly available sample sets and patient historical sample sets; Based on the publicly available sample set, a benchmark motion intensity evaluation model based on a self-attention neural network model is constructed and trained, wherein the benchmark motion intensity evaluation model takes the publicly available multidimensional state features corresponding to the publicly available sample set as input and the intensity label supervision is fixed at 1. Based on the prior anaerobic threshold test data, the intensity label is fixed at 1, and the benchmark exercise intensity evaluation model is fine-tuned once in an iterative manner. Based on the patient historical sample set, unsupervised clustering is performed on the relationships between exercise intensity labeled by experts to obtain pseudo-labels. Then, the patient historical sample set and the pseudo-labels are combined to perform a second fine-tuning training on the results of the first fine-tuning training to obtain the exercise intensity evaluation model.

[0083] In one embodiment, the motion scheme generation module 14 is further configured to: Establish an empirical model between exercise program parameters and exercise intensity evaluation values ​​based on anaerobic threshold; Based on prior anaerobic threshold test data and typical anaerobic domain test data, calculate the individual correction coefficient corresponding to the rehabilitation subject; By combining the empirical model with the individual correction coefficient, an individual strength mapping model for the rehabilitation subject is obtained; Based on the individual intensity mapping model, a cost function is defined, and the real-time rehabilitation exercise program is iteratively optimized according to the exercise intensity evaluation value output by the exercise intensity evaluation model to obtain a recommended exercise program.

[0084] Specifically, a cost function is defined based on the individual intensity mapping model, and the real-time rehabilitation exercise program is iteratively optimized based on the exercise intensity evaluation value output by the exercise intensity evaluation model to obtain a recommended exercise program, including: The multidimensional state features are input into the exercise intensity evaluation model to obtain the exercise intensity evaluation value, wherein the exercise intensity evaluation value is a percentage intensity based on the anaerobic threshold; Determine whether the exercise intensity evaluation value meets the preset intensity threshold. If it does, output the real-time rehabilitation exercise plan as the recommended exercise plan. If the conditions are not met, the optimization algorithm is initialized with the real-time rehabilitation exercise program, and the mean square error between the output of the individual intensity mapping model and 1 is defined as the cost function. The real-time rehabilitation exercise program is iteratively optimized to obtain a recommended exercise program.

[0085] The cost function defined based on the individual intensity mapping model further includes: Define safety penalty items based on safety constraints; The mean square error between the output of the individual intensity mapping model and 1 is defined as the intensity compliance term; The cost function is obtained by combining the strength compliance item and the safety penalty item.

[0086] Compared to existing technologies, this application first calculates the estimated maximum heart rate of the rehabilitation subject, and calculates the target anaerobic threshold heart rate range according to the preset anaerobic threshold heart rate percentage range, which serves as the initial value of the prior anaerobic threshold test data to obtain static physiological information. Subsequently, based on the basic information and the prior expert system, safety constraints are set, providing a scientific starting point and rigid constraints for the subsequent complex intelligent optimization process, ensuring that all subsequent adjustments are carried out within a safe range, and effectively avoiding the risk of cardiovascular events induced by the deviation in exercise intensity estimation.

[0087] Secondly, when rehabilitation subjects are implementing real-time rehabilitation exercise programs, subjective complaint data is collected simultaneously to capture patients' internal feelings that cannot be measured by machines. Subsequently, objective monitoring data is collected in real time through wearable devices and time-aligned according to a unified timestamp, which solves the problem of asynchronous time for data from different sources. Furthermore, noise is removed through data cleaning, thus providing a high-quality data foundation for subsequent feature extraction and model evaluation.

[0088] Furthermore, trend analysis was performed on the objective monitoring data in the exercise state data to divide the exercise period window and the rest period window. By extracting key features of the rest period between exercise intervals, the characteristics of the load period and the recovery period were extracted and fused to represent them separately. This resulted in the final multidimensional state features containing multiple states, providing more discriminative and physiologically significant feature inputs for the subsequent exercise intensity evaluation model, and enhancing the accuracy of identifying whether the patient has reached the anaerobic threshold.

[0089] Finally, a public sample set and a patient historical sample set were obtained to train the baseline exercise intensity evaluation model. Then, the baseline model was fine-tuned based on prior anaerobic threshold test data. After obtaining pseudo-labels, a second fine-tuning training was performed to obtain the final exercise intensity evaluation model, providing an accurate prediction tool for the optimization algorithm. Subsequently, an empirical model between exercise protocol parameters and exercise intensity evaluation values ​​was established, and individual correction coefficients were calculated to obtain an individual intensity mapping model. Then, the exercise intensity evaluation model was used to determine whether it met the preset intensity threshold, and iterative optimization was performed to achieve closed-loop adaptive adjustment of the protocol generation. This allows for continuous tracking of the time-varying nature of the patient's functional status, while simultaneously achieving the optimal balance between the effectiveness and safety of the recommended exercise protocol, thus improving the scientific rigor of post-PCI rehabilitation exercise prescriptions.

Claims

1. A method for generating post-PCI rehabilitation exercise programs based on machine learning, characterized in that, include: Collect basic information about the rehabilitation subjects, and initialize the safety constraints and prior anaerobic threshold test data for rehabilitation exercises based on the basic information; When rehabilitation subjects are implementing real-time rehabilitation exercise programs, exercise status data, including subjective complaints and objective monitoring data, are collected simultaneously. Multidimensional state features, including basic physiological characteristics and key feature reduction, are extracted from the motion state data, wherein the key feature reduction is used to characterize the decline in physiological characteristics; By combining a pre-built machine learning-based exercise intensity evaluation model, the prior anaerobic threshold test data, and the multidimensional state features, the real-time rehabilitation exercise program is iteratively optimized using an optimization algorithm to obtain a recommended exercise program.

2. The method for generating a post-PCI rehabilitation exercise program based on machine learning as described in claim 1, characterized in that, Collect basic information about the rehabilitation subject, and initialize the safety constraints and prior anaerobic threshold test data for rehabilitation exercises based on the basic information, including: Collect basic information about the rehabilitation subjects and calculate their estimated maximum heart rate using a preset maximum heart rate estimation formula; Based on the estimated maximum heart rate, the target anaerobic threshold heart rate range for the rehabilitation subject is calculated according to the preset anaerobic threshold heart rate percentage range, which serves as the initial value of the prior anaerobic threshold test data. Based on the basic information and the prior expert system, safety constraints for rehabilitation exercises are set, including at least the maximum safe heart rate threshold, the minimum blood oxygen saturation threshold, and the maximum blood pressure threshold.

3. The method for generating a post-PCI rehabilitation exercise program based on machine learning as described in claim 1, characterized in that, While rehabilitation subjects are implementing real-time rehabilitation exercise programs, exercise status data, including subjective complaints and objective monitoring data, are collected simultaneously, including: Collect the chief complaint data input by the rehabilitation subjects before and after exercise, wherein the chief complaint data includes at least fatigue score, chest pain score and dyspnea score; The objective monitoring data of the rehabilitation subject during exercise is collected in real time by wearable devices. The objective monitoring data includes at least heart rate, heart rate variability, blood oxygen saturation, respiratory rate and exercise acceleration. The collected subjective complaint data and objective monitoring data are time-aligned according to a unified timestamp, and the objective monitoring data is cleaned to obtain the motion state data.

4. The method for generating a post-PCI rehabilitation exercise program based on machine learning as described in claim 1, characterized in that, From the motion state data, multidimensional state features including basic physiological characteristics and key features are extracted, including: Trend analysis is performed on the objective monitoring data in the motion state data, and the objective monitoring data is divided into motion time windows and rest time windows; Traverse multiple exercise period windows and rest period windows to extract basic physiological characteristics; Using each rest period window as the analysis window, key physiological features are selected from the basic physiological features, and the key feature reduction of each key physiological feature is obtained through traversal analysis. The basic physiological features and key features under multiple exercise period windows and rest period windows are merged and output to reduce them to the multidimensional state features.

5. The method for generating a post-PCI rehabilitation exercise program based on machine learning as described in claim 1, characterized in that, The construction of the exercise intensity evaluation model includes: Obtain publicly available anaerobic threshold assessment sets and patient historical training data, and map them to a unified feature dimension to obtain publicly available sample sets and patient historical sample sets; Based on the publicly available sample set, a benchmark motion intensity evaluation model based on a self-attention neural network model is constructed and trained, wherein the benchmark motion intensity evaluation model takes the publicly available multidimensional state features corresponding to the publicly available sample set as input and the intensity label supervision is fixed at 1; Based on the prior anaerobic threshold test data, the intensity label is fixed at 1, and the benchmark exercise intensity evaluation model is fine-tuned once in an iterative manner. Based on the patient historical sample set, unsupervised clustering is performed on the relationships between exercise intensity labeled by experts to obtain pseudo-labels. Then, the patient historical sample set and the pseudo-labels are combined to perform a second fine-tuning training on the results of the first fine-tuning training to obtain the exercise intensity evaluation model.

6. The method for generating a post-PCI rehabilitation exercise program based on machine learning as described in claim 1, characterized in that, Combining a pre-built machine learning-based exercise intensity evaluation model, the prior anaerobic threshold test data, and the multidimensional state features, the real-time rehabilitation exercise program is iteratively optimized using an optimization algorithm to obtain a recommended exercise program, including: Establish an empirical model between exercise program parameters and exercise intensity evaluation values ​​based on anaerobic threshold; Based on prior anaerobic threshold test data and typical anaerobic domain test data, calculate the individual correction coefficient corresponding to the rehabilitation subject; By combining the empirical model with the individual correction coefficient, an individual strength mapping model for the rehabilitation subject is obtained; Based on the individual intensity mapping model, a cost function is defined, and the real-time rehabilitation exercise program is iteratively optimized according to the exercise intensity evaluation value output by the exercise intensity evaluation model to obtain a recommended exercise program.

7. The method for generating a post-PCI rehabilitation exercise program based on machine learning as described in claim 6, characterized in that, Based on the individual intensity mapping model, a cost function is defined. The real-time rehabilitation exercise program is iteratively optimized according to the exercise intensity evaluation value output by the exercise intensity evaluation model to obtain a recommended exercise program, including: The multidimensional state features are input into the exercise intensity evaluation model to obtain the exercise intensity evaluation value, wherein the exercise intensity evaluation value is a percentage intensity based on the anaerobic threshold; Determine whether the exercise intensity evaluation value meets the preset intensity threshold. If it does, output the real-time rehabilitation exercise plan as the recommended exercise plan. If the conditions are not met, the optimization algorithm is initialized with the real-time rehabilitation exercise program, and the mean square error between the output of the individual intensity mapping model and 1 is defined as the cost function. The real-time rehabilitation exercise program is iteratively optimized to obtain a recommended exercise program.

8. The method for generating a post-PCI rehabilitation exercise program based on machine learning as described in claim 6, characterized in that, Based on the individual intensity mapping model, the cost function is defined, and it also includes: Define safety penalty items based on safety constraints; The mean square error between the output of the individual intensity mapping model and 1 is defined as the intensity compliance term; The cost function is obtained by combining the strength compliance item and the safety penalty item.

9. A machine learning-based device for generating post-PCI rehabilitation exercise programs, characterized in that, The apparatus for implementing the machine learning-based PCI postoperative rehabilitation exercise program generation method according to any one of claims 1-8, the apparatus comprising: The basic information acquisition module is used to collect basic information of the rehabilitation subject and initialize the safety constraints of rehabilitation exercise and prior anaerobic threshold test data based on the basic information. The rehabilitation data acquisition module is used to simultaneously collect movement status data, including the rehabilitation subject's subjective complaint data and objective monitoring data, when the rehabilitation subject is performing a real-time rehabilitation exercise program; A multidimensional feature extraction module is used to extract multidimensional state features, including basic physiological features and key feature reductions, from the motion state data, wherein the key feature reductions are used to characterize the decrease in physiological features; The exercise program generation module is used to combine a pre-built machine learning-based exercise intensity evaluation model, the prior anaerobic threshold test data, and the multi-dimensional state features, and to iteratively optimize the real-time rehabilitation exercise program through an optimization algorithm to obtain a recommended exercise program.

Images