CN118571415A - Optimization method of personalized postoperative rehabilitation training program for breast cancer based on machine learning - Google Patents

Abstract

The present invention relates to personalized breast cancer postoperative rehabilitation training, specifically, a personalized breast cancer postoperative rehabilitation training program optimization method based on machine learning. First, data collection and preprocessing are performed, including clinical data, physiological data, psychological state data and rehabilitation feedback data, which are processed by data cleaning, normalization and data enhancement technology; then feature engineering is performed, including feature selection and feature construction, and features that significantly affect the rehabilitation effect, including activity ability and emotional state, are selected, and composite features are constructed in combination with medical knowledge; then a hybrid model is developed, combining deep learning and traditional machine learning methods, using a multi-task learning algorithm, and adjusting the algorithm weight according to individual differences of patients to predict rehabilitation indicators; finally, dynamic adjustment and optimization are implemented, a real-time feedback system is designed, and a reinforcement learning algorithm is applied to dynamically optimize the training program according to the patient's rehabilitation effect to improve the personalization and effect of rehabilitation training.

Description

Optimization method of personalized postoperative rehabilitation training program for breast cancer based on machine learning

Technical Field

The present invention relates to personalized breast cancer postoperative rehabilitation training, and specifically to a personalized breast cancer postoperative rehabilitation training program optimization method based on machine learning.

Background Art

Although the current optimization methods for personalized breast cancer postoperative rehabilitation training programs have made some progress, there are still a series of shortcomings and drawbacks, which may limit the effectiveness of rehabilitation programs and patient satisfaction. First, many existing rehabilitation training programs lack sufficient personalization and flexibility. Although some programs attempt to adjust rehabilitation plans according to the general conditions of patients, they often fail to fully take into account changes in patients' specific physiological and psychological conditions. For example, rehabilitation plans may not take into account small changes in patients' daily activities or fluctuations in emotional states, resulting in the inability of rehabilitation programs to adjust in real time to adapt to patients' current specific needs. This lack of personalized approach may lead to inefficient rehabilitation and patients feel that the program does not pay enough attention to their personal characteristics and needs. Second, existing training programs often rely on traditional rehabilitation models that may not fully utilize the latest technology or data analysis tools. Many rehabilitation programs are still based on outdated medical theories and practices and fail to integrate modern machine learning techniques or big data analysis, which limits the ability of rehabilitation programs to process complex data and generate accurate predictions. For example, failure to use real-time data to monitor patients' rehabilitation progress or failure to fine-tune rehabilitation activities through data-driven methods may result in the inability of programs to effectively respond to rapid changes in patients' conditions. In addition, existing programs are also weak in multi-task management and information sharing. The rehabilitation process involves the management of multiple aspects, such as pain control, emotional adjustment, physical recovery, etc., and many existing rehabilitation programs have limited ability to handle these multidimensional needs. These programs may fail to consider the correlation between different rehabilitation tasks when they are designed, resulting in the inability to effectively coordinate various tasks to achieve the best rehabilitation effect. For example, a program may perform well in controlling pain but not in improving the patient's emotional state. This single optimization focus ignores the holistic and interconnected nature of rehabilitation.

More importantly, existing rehabilitation training programs often lack effective feedback mechanisms. In many cases, adjustments to rehabilitation programs are not based on system feedback or direct patient input, but rather on occasional medical assessments or regular health checks. The lack of a continuous, dynamic feedback loop means that rehabilitation programs may not respond to changes in the patient's condition in a timely manner, or fail to fully utilize the patient's personal experience and feedback to optimize the rehabilitation path. This can cause rehabilitation programs to appear too rigid or unrealistic in some cases, reducing patients' motivation to follow the rehabilitation plan.

Summary of the invention

The purpose of the present invention is to provide a personalized breast cancer postoperative rehabilitation training program optimization method based on machine learning, so as to solve some of the drawbacks and shortcomings pointed out in the background technology.

The present invention solves the above-mentioned technical problems by adopting the following technical solutions: comprising: firstly collecting and preprocessing data, including clinical data, physiological data, psychological state data and rehabilitation feedback data, and processing them through data cleaning, normalization and data enhancement technology;

Then feature engineering is performed, including feature selection and feature construction. Features that significantly affect the rehabilitation effect, including activity ability and emotional state, are selected, and composite features are constructed in combination with medical knowledge.

Then, a hybrid model was developed, combining deep learning and traditional machine learning methods, using a multi-task learning algorithm and adjusting the algorithm weights according to individual differences of patients to predict rehabilitation indicators;

Finally, dynamic adjustment and optimization are implemented, a real-time feedback system is designed, and a reinforcement learning algorithm is applied to dynamically optimize the training plan according to the patient's rehabilitation effect to improve the personalization and effectiveness of rehabilitation training.

Furthermore, the data collection and preprocessing steps include:

S1. Use dynamic adjustment filtering technology based on changes in the patient's rehabilitation status, targeting the characteristics of breast cancer rehabilitation data, including rehabilitation progress and physiological feedback; apply adjustment filtering functions to optimize the processing process:

Where x represents the original data including activity and heart rate, μ _i and σ _i are the mean and standard deviation of each data dimension, respectively, and _wi is the weight adaptively adjusted based on the rehabilitation status;

S2. Adopt a dynamic normalization strategy based on the rehabilitation stage, combine physical and physiological parameters, use formulas to maintain the validity of the data and add nonlinear features to adapt to different rehabilitation stages:

Where L, k, c are adjustment parameters, and ω(t) is a weight function involving rehabilitation time, which is used to enhance the sensitivity to early and late rehabilitation data;

S3. Generative adversarial networks (GANs) are used to simulate different rehabilitation scenarios that patients may experience. The following generative model functions are used to enhance the data:

Where x represents the measured data during the rehabilitation process, z is the random input of the generative network, and f and _vj are model parameters used to generate synthetic data that conform to the actual changes in rehabilitation progress.

Furthermore, the feature selection and feature construction include:

First, we use feature association analysis based on the graph model to select key features through the connection strength and path analysis of the nodes in the graph, using the formula:

Where x _i ,x _j represent different rehabilitation characteristics, λ _k , _ak ,b _k are the parameters for adaptive adjustment, Represents a specific operator to achieve significance evaluation of deep correlations between features;

Secondly, use the deep learning network to fuse features and construct composite features:

Where x and y represent the original features, including activity and emotional state, and the nonlinearity and depth of the processing are increased by the integral form of the time variable t;

Finally, we combined the medical expert knowledge to create new features and applied the dynamic system model:

Among them, α and β are parameters adjusted based on the rehabilitation process and are used to generate features that reflect the dynamic process of rehabilitation.

Furthermore, the hybrid model construction includes:

First, a hybrid feature processing technique is used to analyze high-dimensional unstructured data such as rehabilitation training videos through a convolutional neural network (CNN). At the same time, a support vector machine (SVM) is used to process the patient's basic physiological indicators. The fully connected layer or attention mechanism is used to achieve the fusion of the two types of data features.

Secondly, a multi-task learning framework is used to predict multiple rehabilitation indicators including pain level, motor ability and emotional state. The useful information between different tasks is shared through task relevance mapping to improve the overall prediction ability of the model.

Finally, a personalized weight learning algorithm is introduced to dynamically adjust the weights of different tasks according to the patient's rehabilitation progress and individual response differences. A feedback-based loop mechanism is adopted to learn and update the task weights according to the patient's rehabilitation feedback and historical data.

Furthermore, the implementation of the hybrid feature processing technology includes:

S1. First, a two-stream hybrid network architecture is adopted. One stream processes the rehabilitation training video data through a convolutional neural network CNN, and uses a multi-scale convolution kernel W _s ⁽ⁱ⁾ to extract spatial features f _s ⁽ⁱ⁾ , according to the formula:

It is calculated that, where σ represents the nonlinear activation function, n is the number of convolutional layers, is the bias term;

S2, while the other stream processes the patient's basic physiological index x through the support vector machine SVM, outputs the weight vector v and bias c, and calculates the classification decision boundary through the multi-kernel function k. The formula is:

where α _j is the Lagrange multiplier and m is the number of support vectors;

S3. Finally, a fully connected layer with a multi-layer network structure is used to fuse the features of the two data sources. The calculation formula of the fusion feature h is:

Where _Wf and _Wp are weight matrices, and _as and _at are weights calculated by the dynamic attention mechanism.

Furthermore, the prediction of multiple rehabilitation indicators comprises:

S1. First, each rehabilitation index is processed by a neural network branch, using high-order derivatives and transformation functions _fi , and adjusting weights _Wi and bias _bi according to the patient's personalized data. The specific formula is:

Where γ _k is the adjustment parameter and n is the number of variables;

S2. Then, we introduce a graph-based attention mechanism to optimize information sharing between tasks and use adaptive weights α _ij to map task relevance. The formula is:

Where W _rel is the relationship weight matrix, β _k is the adjustment coefficient associated with each task, and m is the number of tasks;

S3. Finally, combine the prediction results and shared information of all tasks to optimize the comprehensive loss function:

Where λ _i is the weight of task i, ρ _ij is the correlation adjustment factor between tasks, and the optimization goal is to reduce the gradient difference between different tasks.

Furthermore, the personalized weight learning algorithm is implemented by the function:

Update the task weights, where ω(t) represents the task weight at time t, η is the learning rate parameter dynamically adjusted by patient feedback, and δL(t) is the rate of change of the loss function calculated based on patient feedback;

Then, a feedback-based cyclic learning mechanism is adopted, using the function:

to calculate the effective weight adjustment for each cycle T, where α(t) is the factor for adjusting the weight, λ is the attenuation coefficient to reflect the rate at which the impact of historical data weakens, and δF(t) represents the functional improvement indicator obtained from patient feedback.

Furthermore, the dynamic adjustment and optimization are achieved by the following steps:

S1. First, implement a real-time feedback system using a nonlinear dynamic system model:

Where s is the comprehensive rehabilitation state collected from the sensor, t is time, α, β, γ are parameters for adjusting the response speed and sensitivity of the model to sense the acceleration and speed of the rehabilitation state and provide data processing support for dynamic adjustment;

S2. Then define the reward function:

Where f(s,a) is the rehabilitation effect function that depends on the patient's state and treatment action, _f0 is the rehabilitation objective function, λ(s) is the dynamic coefficient adjusted based on the patient's current state, and T is the observation period; by establishing an exponential relationship between the expected rehabilitation effect and the actual rehabilitation effect, it allows the training program to be optimized within a wider dynamic range.

Beneficial effects of the present invention:

1. Personalized rehabilitation plan: By utilizing machine learning algorithms, the present invention is able to adjust and optimize the rehabilitation training plan based on the specific conditions of each patient (such as physiological data, rehabilitation progress, and personal feedback). This highly personalized approach ensures that the training program matches the patient's specific needs and abilities, thereby improving the effectiveness of rehabilitation.

2. Dynamic adjustment capability: During the rehabilitation process, the patient's condition will continue to change. The present invention enables flexible adjustment of the rehabilitation program through real-time data monitoring and feedback loop. This dynamic adjustment capability ensures that the rehabilitation program always adapts to the patient's current condition and optimizes rehabilitation efficiency.

3. Optimize rehabilitation results: Through multi-task learning and feature fusion technology, the present invention can simultaneously optimize multiple rehabilitation goals, such as pain management, emotional adjustment, and physical rehabilitation. This comprehensive optimization method improves the comprehensiveness of rehabilitation and helps patients achieve better rehabilitation results in all aspects.

4. Scientific decision support: The present invention provides scientific decision support to medical providers through advanced data analysis and predictive models, enabling them to formulate or adjust rehabilitation plans based on data-driven insights. This not only enhances the scientific nature of rehabilitation plans, but also reduces the uncertainty of medical decision-making.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a method for optimizing a personalized breast cancer postoperative rehabilitation training program based on machine learning according to the present invention.

FIG. 2 is a flow chart of building a hybrid model of the present invention.

DETAILED DESCRIPTION

The specific implementation modes of the present invention will be described in detail below with reference to the accompanying drawings.

The optimization method of personalized postoperative rehabilitation training program for breast cancer based on machine learning first involves data collection and preprocessing. This process includes the patient's clinical data, physiological data, psychological state data, and feedback data during the rehabilitation process. These data are to ensure that the training program can be adjusted according to the patient's specific situation; after data collection, preprocessing is performed to improve data quality and lay the foundation for subsequent analysis and learning algorithms. The preprocessing steps include data cleaning (removing errors, missing or outliers), data normalization (adjusting the data scale to make it on the same standard for easy comparison and processing), and data enhancement (increasing the diversity and quantity of data through technical means, such as generating more image data through rotation and scaling, or generating more physiological signal data through simulation). These steps work together to make the collected data more accurate, consistent, and useful, providing reliable input for subsequent machine learning models, thereby optimizing personalized rehabilitation training programs after breast cancer surgery.

Feature engineering includes two main processes: feature selection and feature construction. Feature selection refers to screening out the features that have the most significant impact on the rehabilitation effect from a large amount of relevant rehabilitation data, such as the patient's mobility and emotional state. Mobility can be quantified from the physical activity data during rehabilitation, such as walking distance, activity frequency, etc., and emotional state is evaluated through psychological test results, daily feedback or physiological indicators such as heart rate variability. Feature construction involves combining these selected features with medical knowledge to construct composite features that can more comprehensively reflect the patient's rehabilitation status. For example, the "rehabilitation activity" indicator can be constructed by combining data on mobility and emotional state. Such composite features can provide richer information for machine learning models, thereby more accurately predicting rehabilitation progress and adjusting rehabilitation plans to ensure the personalization of training programs and maximize their effectiveness.

A hybrid model was developed that combines deep learning and traditional machine learning methods to fully leverage the advantages of both. The deep learning part can process large amounts of complex unstructured data such as medical images or continuous physiological monitoring data, while traditional machine learning methods are suitable for processing structured clinical data and rehabilitation assessment data. In this hybrid model, a multi-task learning algorithm is used to simultaneously predict multiple rehabilitation indicators, such as pain level, functional recovery, and emotional state, so that the algorithm can handle multiple tasks under a unified framework, improving the efficiency of the model and the consistency of predictions. In addition, this model also designs a mechanism to adjust the algorithm weights according to individual differences among patients, which means that the model can dynamically adjust the weights according to the specific rehabilitation situation and feedback of each patient, making the predicted rehabilitation indicators more realistic, thereby providing patients with more personalized and accurate rehabilitation training plans.

Dynamic adjustment and optimization are implemented by designing a real-time feedback system and applying a reinforcement learning algorithm. This system integrates real-time data collected from patients, such as physiological monitoring, activity recording, and psychological state feedback. These data are collected in real time through sensors and mobile devices and transmitted to the processing center; the reinforcement learning algorithm uses these real-time data as input to evaluate the patient's rehabilitation progress and optimize the training plan according to a predefined reward system. The reward system is based on the patient's specific rehabilitation goals, such as pain relief, improved motor skills, etc.; the algorithm learns the optimal training parameters and methods through continuous trial and error, and adjusts the intensity and type of training plans to adapt to the patient's recovery speed and ability. The core of this method is the ability to dynamically adjust the plan according to the patient's immediate rehabilitation effect, which greatly improves the personalization and actual effect of rehabilitation training, and ensures that each patient can receive the rehabilitation support that best suits their individual situation.

Embodiment 1:

Patient Ms. Zhao recently completed breast cancer surgery and started rehabilitation training. Monitoring her activity and heart rate data are key physiological indicators to measure her recovery status.

First, the collected data is preprocessed using dynamic adjustment filtering technology. This technology is used to filter and optimize the raw data to ensure the quality and consistency of the data, thereby providing reliable input for subsequent machine learning models. Use the adjustment filter function:

Where x represents the original data including activity level and heart rate, μ _i and σ _i are the mean and standard deviation of each data dimension, respectively, and _wi is the weight adaptively adjusted based on the rehabilitation status.

Assume that on a certain day, the activity and heart rate data of patient Ms. Zhao are x ₁ = 5000 steps and x ₂ = 75 bpm respectively. These data are compared with her mean μ ₁ = 4500, μ ₂ = 70 and standard deviation σ ₁ = 500, σ ₂ = 5 during her normal rehabilitation period. The weights w ₁ and w ₂ are adaptively adjusted according to her rehabilitation progress, and the current values are set to w ₁ = 0.7 and w ₂ = 0.3.

Substitute into the formula to calculate:

Ultimately, the filtered results are optimized data points for the machine learning model. These processed data can better reflect the actual recovery status of the patient, Ms. Zhao, and help the model more accurately predict and adjust the rehabilitation training plan, thereby better promoting her recovery process.

A dynamic normalization strategy based on rehabilitation stages is adopted, which combines physical and physiological parameters to adapt to different rehabilitation stages, and uses advanced mathematical formulas to maintain the validity of the data and add nonlinear features. It is assumed that the activity volume and heart rate data of the patient Ms. Zhao at different rehabilitation stages need to be properly normalized to ensure that the rehabilitation model can accurately assess her progress and make corresponding adjustments.

formula:

Used to normalize the data of patient Ms. Zhao, where L, k, and c are parameters adjusted according to the rehabilitation status, and ω(t) is a weight function that changes with rehabilitation time, designed to enhance the model's sensitivity to early and late rehabilitation data. In specific applications, it is assumed that for activity data, in early rehabilitation, because the activity is low, more sensitive adjustments are required to observe any small changes, while in late rehabilitation, the activity increases, and the sensitivity of the model can be appropriately reduced.

The activity data of Ms. Zhao in the early stage of rehabilitation is set to x = 3000 steps, and the data in the late stage of rehabilitation is set to x = 6000 steps. The normalization parameters are set to L = 1, k = 0.001, c = 4500. The weight function ω(t) is designed as Where t represents the time (in days) since the start of rehabilitation, and is intended to decrease in weight as time increases, since more attention and adjustments are required in the early stages of rehabilitation.

Calculate the normalized value:

Initial x＝3000:

Calculate the first part:

Calculate the second part (integral): ln(1+3000)≈8.006

Therefore, G(3000)≈0.182+8.006＝8.188

Late x＝6000:

Calculate the first part:

Calculate the second part (integral): ln(1+6000)≈8.699

So, G(6000)≈0.818+8.699＝9.517

Such normalization not only takes into account the different needs of the rehabilitation stages, but also provides a more detailed and sensitive data processing method by integrating nonlinear features.

To improve the quality of training data and the generalization ability of the model, generative adversarial networks (GANs) are used to simulate and generate different rehabilitation situations experienced by patients. This method is based on augmented datasets based on actual collected data, which is particularly useful in cases where certain rehabilitation data is scarce or difficult to obtain.

A specific implementation involves using a generative model function:

Among them, x represents the measured data during the rehabilitation process, such as activity or heart rate, z is the random input of the generative model, representing the potential data generation factor, and f and _vj are model parameters to control the characteristics and complexity of the generated data. This function can generate a variety of rehabilitation scenario data by combining linear and nonlinear terms, which helps the model better learn and adapt to actual rehabilitation changes.

In order to further specify this scheme, the activity volume (number of steps) data x of the patient Ms. Zhao during rehabilitation is set, and its impact on the rehabilitation effect. Assume that the actual measured number of steps on a certain day is x = 5000 steps, set the model parameter f = 0.01, the values of _vj are _v1 = 0.05, _v2 = 0.01, and consider the maximum quadratic term (i.e. m = 2). The random input z takes a random number from a normal distribution, for example z = 0.2.

Substitute into the generative model function and calculate the synthetic data:

H(5000,0.2)＝0.2·sin(0.01·5000)+0.05·5000+0.01·5000 ²

H(5000,0.2)＝0.2·sin(50)+250+250000·0.01

H(5000,0.2)＝0.2·(-0.262)+250+2500

H(5000,0.2)≈-0.0524+250+2500

H(5000,0.2)≈2749.9476

This calculation shows that the data generated by GAN is able to simulate actual rehabilitation data changes and provide a broad data foundation to help machine learning models more accurately predict patient responses in different rehabilitation situations.

Embodiment 2:

In this embodiment, feature selection and construction are key steps, aiming to accurately identify and utilize key information in the rehabilitation process through machine learning methods. This goal is achieved by using feature association analysis based on graph models to identify key features from rehabilitation data and analyze the interactions and dependencies between these features to optimize the prediction accuracy and effect of the training model.

In practical application, we set two key features of Ms. Zhao in the rehabilitation process: activity ability (x_i) and emotional state (x_j). Activity ability is quantified by the number of steps per day, while emotional state can be expressed by the score of daily psychological assessment. The following formula is used to perform correlation analysis between features:

Among them, λ _k , a _k , and b _k are model parameters, which are dynamically adjusted according to rehabilitation data to reflect the changes in the importance and interaction of features at different rehabilitation stages. Represents addition, multiplication, or other mathematical operations, depending on the type of features being analyzed and the expected model behavior.

Example of setting parameters: Consider a simple linear relationship, where λ _k = 0.5, a _k = 2, b _k = 2, and set is a multiplication operation. If the patient Ms. Zhao's activity volume on a certain day is x _i = 4000 steps, and her emotional state score is x _j = 8, then the calculation of the association analysis is as follows:

S(4000,8)＝0.5·(4000 ² ·8 ² )

S(4000,8)＝0.5·(16000000·64)

S(4000,8)＝0.5·1024000000

S(4000,8)＝512000000

The results of this calculation show that there is a strong correlation between activity and emotional state, which is quantified by the graphical model. The analysis helps understand which rehabilitation features are most important to the overall recovery process of the patient Ms. Zhao and how these features interact with each other. In this way, the training program can be optimized to make it more targeted and efficient, thereby accelerating Ms. Zhao's recovery process.

In order to evaluate and adjust the rehabilitation plan more finely, the deep learning network is used to fuse features and construct composite features. This method not only considers a single indicator, but also provides a comprehensive evaluation of rehabilitation effects by combining multiple rehabilitation indicators. Specifically, the following formula can be used to achieve the construction of composite features:

Where x and y represent two key features of Ms. Zhao’s rehabilitation process: mobility and emotional state. Mobility is quantified by the number of steps per day, while emotional state is quantified by the score of mental health assessment. The time variable t is introduced in an integral form, which increases the nonlinearity and depth of the processing, allowing the model to capture complex temporal dynamic characteristics, thereby better understanding the long-term effects and short-term fluctuations of the rehabilitation process.

In order to apply this formula in practice, we set a specific data example: on a certain day, Ms. Zhao's activity x is 5000 steps, and her emotional state y is 8 points. To simplify the calculation, the integral range is set from 0 to π (this is a commonly used range that can cover and function), and perform numerical integration:

Using numerical integration methods (such as trapezoidal rule, Simpson's rule, etc.) to solve this integral, and setting it to be calculated through a computer program or numerical integration tool, an approximate result can be obtained. This result gives a numerical value, reflecting the overall rehabilitation index of the patient Ms. Zhao under a given rehabilitation state.

The constructed composite feature C(x,y) can be directly used to train a machine learning model for predicting rehabilitation progress, or as a new indicator for evaluating rehabilitation effectiveness. In this way, the rehabilitation plan can be adjusted more accurately according to the actual rehabilitation situation of the patient Ms. Zhao, achieving more efficient personalized rehabilitation.

Combining medical expert knowledge to create new features and applying dynamic system models to increase understanding and response to the dynamic changes in rehabilitation progress are key steps. A specific mathematical model is used to express the complexity of the rehabilitation process, as follows:

In this model, x and y represent the indicators of activity and emotional state during rehabilitation, respectively. Activity is quantified by the number of steps per day, while emotional state is expressed by the score of mental health assessment. Parameters α and β are adjustment coefficients, which are adjusted according to different stages of rehabilitation to adapt to the rehabilitation needs of individuals.

Suppose there is actual data, where Ms. Zhao's daily activity x is 4000 steps and her emotional state y is 7 points. In order to apply the above model, it is necessary to estimate the rate of change of these variables over time. The integral was calculated. α = 0.1 and β = 0.05 were set based on the expert's advice and the analysis of previous similar patient data.

For calculation Assume that at the subsequent time dt, the number of steps taken by patient Ms. Zhao increased to 4100 steps and her emotional state improved to 8 points. Therefore, the rate of change can be approximated as:

Set dt to one day, then To simplify the calculation, assume that x·y increases linearly within a day, then the integral ∫x·y dt within a day can be estimated as:

∫x·y dt≈30400

Substituting these values into the model:

D(4000,7)＝0.1·4800+0.05·30400

D(4000,7)＝480+1520

D(4000,7)＝2000

The calculation result D(4000,7)=2000 provides a quantitative method to understand and predict the dynamic changes of the rehabilitation process. Through the application of this model, Ms. Zhao's rehabilitation training plan can be adjusted and optimized more effectively to ensure that it can better adapt to her rehabilitation progress, thereby improving the rehabilitation effect and personalization.

Embodiment 3:

In this embodiment, the construction of the hybrid model includes hybrid feature processing technology, which combines the methods of convolutional neural network (CNN) and support vector machine (SVM) to optimize the feature extraction and fusion process from unstructured and structured data sources.

In the specific implementation, a two-stream hybrid network architecture is adopted, in which the stream processes the rehabilitation training video data through CNN, mainly to capture the dynamic changes and important spatial features in the video. For example, the CNN stream can analyze the movement quality of the patient Ms. Zhao during physical rehabilitation exercises, such as the range of motion and speed of the arm. These video data are input into a CNN with multi-scale convolution kernels to adapt to visual patterns of different sizes and complexities. The formula used is:

Where X(s) represents the input of video data, is the i-th convolution kernel, sliding on the time dimension s, used to extract specific spatial features. Function σ is a nonlinear activation function, such as ReLU or Sigmoid, which is used to introduce nonlinear processing to enhance the expression ability of the network. is a bias term that enhances the flexibility of the model. n represents the number of convolutional layers, which determines the depth of the network and the complexity of feature extraction.

At the same time, another stream processes basic physiological indicators such as heart rate, blood pressure, etc. through SVM. These data are usually structured and contain important physiological information that can be used to evaluate recovery status and physical response.

Use fully connected layers or attention mechanisms to achieve the fusion of these two types of data features. The fully connected layer can integrate the features extracted by CNN and SVM to form a unified feature representation, while the attention mechanism can further enhance the model's focus on key features and dynamically adjust the association weights between features according to the needs of rehabilitation training.

For example, at a certain moment, the video data processed by CNN shows that the patient Ms. Zhao's arm movement range has increased by 10% compared with the previous week, while the heart rate data processed by SVM shows that the heart rate remains stable during rehabilitation training. After this information is fused through the fully connected layer, a comprehensive feature can be generated, indicating that Ms. Zhao's rehabilitation is progressing well and her training plan needs to be adjusted to further enhance the intensity and effect of the training.

The other stream optimizes the rehabilitation model by processing its physiological indicators through support vector machine (SVM). In the hybrid model, SVM is used to analyze the patient Ms. Zhao’s physiological indicators, such as heart rate, blood pressure, etc., and helps predict the rehabilitation status and adjust the rehabilitation plan by establishing accurate classification decision boundaries.

In the SVM model, the classification decision boundary is calculated using the following formula:

Where x represents the input physiological index, α _j is the Lagrange multiplier, m is the number of support vectors, k(x, x _j ) is the kernel function used to map the input space to a high-dimensional feature space, and c is the bias term.

It is assumed that multiple physiological indicators including heart rate and blood pressure are collected during a rehabilitation training, and the radial basis function (RBF) is selected as the kernel function in the form of where γ is a parameter of the kernel function, controlling the width or “spread” of the function.

Set Ms. Zhao's heart rate data x for the day to 75bpm and blood pressure to 120/80mmHg.

The selected support vector _xj is the average heart rate 72 bpm and blood pressure 118/78 mmHg measured last week.

Set γ = 0.1, α _j = 0.5 (this is a simplified example, in practice α _j will be obtained through training data), and c = -0.1.

Assuming that only the influence of heart rate is considered, |xx _j | is simplified to |75-72|=3.

Substitute into the SVM classification decision formula:

g(x)＝0.5·e ^-0.9 -0.1

g(x)＝0.5·0.4066-0.1

g(x)＝0.2033-0.1

g(x)＝0.1033

The calculation result g(x)=0.1033 indicates that based on the currently input heart rate data and the previously trained model, the patient Ms. Zhao's rehabilitation status is predicted to be improving (because the value of g(x) is positive). This information can be used to adjust Ms. Zhao's rehabilitation training plan, such as increasing or adjusting the intensity or duration of certain activities.

The fully connected layer with a multi-layer network structure is used to achieve the fusion of the two data source features from the convolutional neural network (CNN) and the support vector machine (SVM). This fusion is achieved through complex mathematical formulas to ensure that the two types of data features can be effectively combined, thereby providing Ms. Zhao with a more accurate prediction of her rehabilitation status and adjustment of her training plan.

The specific implementation scheme involves the following calculation formula:

h=tanh(W _f ·(exp(-|a _s f _s ⁽ⁱ⁾ -a _t f _t ^(j) | ² )·(a _s f _s ⁽ⁱ⁾ +a _t f _t ^(j) ))+ W _p ·g(x))

In this formula, _fs ⁽ⁱ⁾ and _ft ^(j) are the feature vectors extracted from CNN and SVM, respectively, and _as and _at are the attention weights corresponding to these features, which are dynamically adjusted by the dynamic attention mechanism according to the current recovery state. _Wf and _Wp are the weight matrices of the fully connected layer, which are used to integrate and transform the input features.

set up:

f _s ⁽ⁱ⁾ represents the feature vector obtained by the rehabilitation training video processed by CNN, such as 0.8, 0.6;

f _t ^(j) represents the feature vector obtained by the physiological index processed by SVM, such as 0.4, 0.9;

a _s = 0.7 and a _t = 0.3 represent the current attention weights;

For simplification, the specific values of _Wf and _Wp are set to the unit matrix I.

Substitute these data and calculate the specific fusion feature h:

1. First calculate the weighted feature and the square of its difference:

Δ＝|a _s f _s ⁽ⁱ⁾ -a _t f _t ^(j) | ² =|0.7×[0.8,0.6]-0.3×[0.4,0.9]| ²

Δ＝|[0.56,0.42]-[0.12,0.27]| ² ＝|[0.44,0.15]| ² ＝0.44 ² +0.15 ² ＝0.2137

2. Use an exponential function to reduce its influence and multiply it with the weighted eigenvector:

exp(-Δ)·(a _s f _s ⁽ⁱ⁾ +a _t f _t ^(j) )=exp(-0.2137)×([0.56,0.42]+[0.12,0.27])

≈0.807×[0.68,0.69]＝[0.548,0.557]

3. Calculate the output of the fully connected layer:

h=tanh(W _f ·[0.548,0.557]+W _p ·g(x))

Set g(x) = 0.1 (the prediction value obtained from SVM), and _Wf and _Wp are the identity matrices:

h＝tanh([0.548,0.557]+[0.1,0.1])＝tanh([0.648,0.657])

h≈[0.571,0.579]

Through this calculation process, the rehabilitation training video features and physiological index features of the patient Ms. Zhao are integrated to obtain the feature vector h that can directly reflect her rehabilitation status. This feature fusion method not only improves the representation ability of the data, but also makes the rehabilitation training program more accurate and personalized.

Embodiment 4:

In this embodiment, a hybrid model structure is used, which integrates a multi-task learning framework to simultaneously predict multiple key indicators in the rehabilitation process, such as pain level, motor ability, and emotional state. This method uses machine learning technology to conduct in-depth analysis of the correlation between different rehabilitation indicators and share useful information, thereby improving the overall prediction ability of the model.

Specifically for the prediction of rehabilitation indicators, independent neural network branches are designed for each rehabilitation indicator, which process rehabilitation data through high-order derivatives and transformation functions. By introducing the nonlinear transformation function fi, more complexity and flexibility can be added to the model to more accurately simulate and predict various changes in the rehabilitation process. The core formula of the model is as follows:

In this formula:

X represents input rehabilitation data, such as daily activity data and psychological assessment results;

_Wi and _bi are the weights and biases of each task network, respectively. These parameters are optimized through training data;

γ _k is a tuning parameter used to adjust the contribution and importance of different variables;

n is the number of variables involved in the calculation, which determines the model complexity and processing power.

In order to concretize this model, a set of hypothetical parameters and data can be set for calculation examples. The focus is on the patient Ms. Zhao's motor ability and pain level:

Set the activity data X of patient Ms. Zhao to 5000 steps and pain level 3;

Set _Wi to 0.5,0.5, and consider only one layer of network and two inputs;

Set bias _bi to 0;

Set γ _k to 1,1 to treat the first-order derivative as well.

Substituting into the calculation, first find the linear combination of weights and inputs:

W _ik ·x _k +b _ik =0.5×5000+0.5×3＝2501.5

Then calculate higher-order derivatives:

Finally, the calculated output is:

This output _yi can be interpreted as the predicted value of the rehabilitation index obtained after adjustment and calculation by the neural network. In practical applications, through calculation, the patient's rehabilitation status can be dynamically evaluated and predicted, and her rehabilitation plan can be adjusted according to the model output to achieve more effective and personalized rehabilitation support.

In order to more effectively utilize the relevant information in rehabilitation data and achieve optimized synergy between tasks, a graph-based attention mechanism is introduced. This mechanism allows the model to identify and emphasize the correlation between different tasks when dealing with multi-task learning, thereby improving the overall prediction accuracy and effect.

The calculation formula set is as follows, which is used to determine the adaptive weights between tasks:

In this formula:

and are the feature representations of the i-th and j-th tasks at the k-th level respectively;

W _rel is the relationship weight matrix, which is used to adjust the relationship strength between different tasks;

β _k is the adjustment coefficient associated with each task, which determines the contribution of different task levels;

m is the total number of tasks involved in the calculation.

The setting is processing three main rehabilitation tasks: pain level assessment, motor ability monitoring, and emotional state assessment. A neural network is used to extract the features of each task, and the associations between these tasks are evaluated and strengthened through the attention mechanism. To illustrate this calculation process, some parameters are set:

The feature dimension of each task is set to k = 2, that is, each task has two feature vectors.

W _rel is taken as the unit matrix I to simplify the calculation.

β _k = 1, which means that all task levels are assumed to contribute equally.

Considering a specific example of a feature vector, let:

For pain level assessment, the eigenvector h ₁ = [0.9, 0.1]

For sports performance monitoring, the feature vector h ₂ = [0.5, 0.5]

For emotional state assessment, the feature vector h ₃ = [0.2, 0.8]

The attention weight between _h1 and _h2 is calculated using the above formula:

α ₁₂ =softmax(1·tanh(1·([0.9,0.1]+[0.5,0.5])))

α ₁₂ =softmax(1·tanh([1.4,0.6]))

α ₁₂ =softmax([0.885,0.537])

Here, the softmax function converts these values into probabilities, with higher values indicating stronger correlations.

By calculating α _ij between different tasks, the model can give more weight to features that are highly relevant to the target task when predicting. For example, if pain level is highly correlated with motor ability, the information of these two tasks will be shared more, thereby improving the accuracy of predicting pain level.

Next, a complex multi-task learning framework is used to simultaneously optimize multiple rehabilitation-related indicators, such as pain level, motor ability, and emotional state. In order to maximize the performance of the model and ensure that information between tasks can be effectively shared, a comprehensive loss function is designed, which not only considers the loss of a single task, but also includes the correlation loss between tasks, so as to reduce the gradient difference between different tasks and improve the overall synergy of the model.

The specific comprehensive loss function is expressed as:

In this formula:

λ _i is the weight of the i-th task, which determines the importance of the task in the total loss;

is the loss of the i-th task;

ρ _ij is the moderating factor of inter-task correlation, which determines the strength of the association between different tasks;

α _ij is the inter-task association weight calculated by the graph-based attention mechanism mentioned above;

and are the gradients of the i-th and j-th tasks respectively.

There are three rehabilitation tasks: pain level assessment (task 1), motor ability monitoring (task 2), and emotional state assessment (task 3).

λ ₁ = 0.3, λ ₂ = 0.4, and λ ₃ = 0.3 are set to reflect the importance of different tasks in rehabilitation training.

Setting ρ _ij = 0.1 indicates that we want to strengthen the correlation effect between tasks.

Set the loss for each task and They are 0.2, 0.1, and 0.15 respectively.

Using this data, we can calculate the value of the combined loss function. First, we calculate the loss contribution of each task:

Calculate the contribution of the gradient difference between tasks (set the gradient difference quantization value to 0.05 for any two tasks):

The final comprehensive loss function is:

Through such calculations, we can see that by adjusting the weights of tasks and the correlation adjustment factors between tasks, we can effectively control and optimize the behavior of the rehabilitation training model, making the rehabilitation process more personalized and coordinated, and ultimately improving the effect of rehabilitation training.

Embodiment 5:

In this embodiment, in order to ensure that the program can adapt to the specific situation of each patient, a dynamic weight adjustment mechanism is adopted. This mechanism continuously adjusts the task weights according to the patient's rehabilitation feedback and historical data through a feedback-based cyclic learning algorithm to maximize the rehabilitation effect.

In a specific implementation scheme, the following mathematical formula is used to update the task weight:

here:

ω(t) represents the task weight at time t;

η is a dynamically adjusted learning rate parameter, which is adjusted according to the patient's recovery speed and feedback;

δL(t) is the rate of change of the loss function, which is calculated based on the patient's most recent feedback and historical performance.

In the rehabilitation training of patient Ms. Zhao, the initial task weight ω(0) is set to 1.0, and the learning rate parameter η is set to 0.05. After the first rehabilitation training, the patient's feedback shows that her pain level has decreased, and the calculated loss function change rate δL(1) is 0.1. According to the formula, the task weight of the second training can be calculated:

ω(1)＝1.0·log(1+20·0.1564)

ω(1)＝1.0·log(4.128)

ω(1)≈1.418

Through this calculation, we can see that due to the patient's positive feedback, the task weight increases, indicating that the rehabilitation program needs to strengthen the current training plan to accelerate the rehabilitation process. This dynamic weight adjustment mechanism not only makes the rehabilitation plan more adaptive, but also promotes a more effective personalized treatment path through real-time feedback.

In order to optimize the rehabilitation progress and realize personalized rehabilitation training, a feedback-based cyclic learning mechanism is introduced. A mathematical function is used to calculate the effective weight adjustment of each cycle T, so that the rehabilitation plan not only reflects the current patient status, but also can be adjusted according to historical data.

Specifically, this mechanism is implemented through the following functions:

in:

α(t) is the adjustment weight factor at time t;

λ is the decay coefficient, which is used to control the influence of historical data in the current weight adjustment. A larger λ value means that the influence of historical data decreases faster.

δF(t) is an indicator of functional improvement obtained from patient feedback, for example, it can be an improvement in pain level, an increase in motor ability, or an improvement in emotional state.

The goal is to adjust the rehabilitation training plan of Ms. Zhao according to her daily feedback during the treatment period (T = 30 days). Set λ = 0.1, indicating that the influence of historical data decays to about 90% of the original value every day.

Suppose that within these 30 days, the rehabilitation feedback index δF(t) of the patient Ms. Zhao is the improvement of motor ability, and the improvement value gradually increases from 0.1 on the first day to 0.5 on the 30th day. The numerical integration method can be used to calculate the numerator and denominator:

1. Calculate molecules Set δF(t) to increase linearly:

Through numerical integration, we get an approximate value, such as 2.5.

2. Calculate the denominator

This is the standard exponential decay integral, with an approximate value of 9.5.

3. Calculate α(t):

This value α(30) = 0.263 indicates that at the end of the rehabilitation period, more attention should be paid to improving the patient Ms. Zhao's motor ability. This weight adjustment mechanism allows the patient Ms. Zhao's training program to be dynamically adjusted according to her real-time feedback and rehabilitation progress, thereby more accurately meeting her rehabilitation needs and optimizing the rehabilitation effect.

Embodiment 6:

In this embodiment, the personalized rehabilitation experience is further enhanced by implementing a real-time feedback system, which processes the data collected from the sensors based on a nonlinear dynamic system model. The model can monitor the rehabilitation status in real time and dynamically adjust the rehabilitation plan to ensure that the patient receives the best support throughout the rehabilitation cycle.

The specific implementation model formula is as follows:

Among them, s is the rehabilitation state, which can be a comprehensive evaluation value of muscle strength, joint range of motion, etc., collected from sensors in real time. t is time, and α, β, and γ are parameters for adjusting the response speed and sensitivity of the model.

To demonstrate the implementation and calculation of this model, a simplified scenario can be set up, where:

α=0.1, adjust the acceleration perception;

β = 0.05, adjusting speed perception;

γ = 0.01, adjusting for the baseline effect of recovery status.

Considering the patient's rehabilitation data, the change of the rehabilitation status s over time t can be approximated as s(t) = t ² , indicating that the rehabilitation status gradually improves over time. It is necessary to calculate the integral from t = 0 to t = T, and T = 10 days is set as the end time of the cycle.

Substitute s(t) and its derivative Second Derivative Into the model, the integral becomes:

Through the numerical integration method, the value of this integral can be estimated, which will give an indicator of the cumulative adjustment of the effect of the rehabilitation training program over time. If the result of this integral shows that the rehabilitation status is not as good as expected, the model can be optimized by adjusting the values of α, β, and γ to make it more sensitive or more stable, so as to better adapt to the actual rehabilitation process of the patient.

Next, an advanced reward function is used, which is designed to dynamically adjust the rehabilitation program and ensure that the patient's rehabilitation progress is consistent with the predetermined goals. This reward function establishes an exponential relationship between the expected rehabilitation effect and the actual rehabilitation effect, allowing the rehabilitation training program to be optimized within a wider dynamic range.

In a specific implementation scheme, the reward function is defined as:

in:

f(s,a) is the rehabilitation effect function that depends on the patient state s and the treatment action a;

f ₀ is the rehabilitation objective function, i.e., the target value of the ideal state or rehabilitation;

λ(s) is a dynamic coefficient adjusted based on the patient’s current state. It regulates the contribution of the rehabilitation effect to the reward function, allowing the model to be flexibly adjusted according to the patient’s actual progress.

T is the observation period, for example, it can be set to the total duration of rehabilitation training, such as one month or three months.

To demonstrate the calculation and application of this reward function, consider the following example:

The rehabilitation goal set for patient Ms. _Zhao is to achieve a certain level of activity, such as taking 8,000 steps per day.

The observation period T of the reward function is set to 30 days.

The dynamic coefficient λ(s) is set to 0.1, which means that even small changes in recovery effect can significantly affect the reward.

Suppose that during a rehabilitation cycle, the number of steps taken by the patient Ms. Zhao, f(s,a), gradually increases from 2000 to 6000. This change can be simulated and the corresponding reward function can be calculated. First, a simplified function model is set in which the patient's rehabilitation effect increases linearly:

f(s,a)＝2000+133.33t

where t varies from 0 to 30 days.

The numerator of the reward function can be calculated by numerical integration:

This integral involves the integration of an exponential function and can be solved numerically. The calculated integral value is assumed to be approximately 123.

The reward function is:

R(s,a)＝log(123)

This will give a positive value, reflecting the overall match between the patient Ms. Zhao's rehabilitation progress and the goal throughout the rehabilitation cycle. If this value is close to 0 or negative, it means that the rehabilitation progress is not good and the rehabilitation plan needs to be adjusted.

The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims (8)

1. The personalized breast cancer postoperative rehabilitation training scheme optimization method based on machine learning is characterized by comprising the following steps of: firstly, data collection and preprocessing are carried out, wherein the data collection and preprocessing comprises clinical data, physiological data, mental state data and rehabilitation feedback data, and the data collection and preprocessing are processed through data cleaning, normalization and data enhancement technologies;

Then, carrying out feature engineering, comprising feature selection and feature construction, selecting features with obvious influence on rehabilitation effect, comprising activity capacity and emotional state, and constructing composite features by combining medical knowledge;

Then developing a mixed model, combining deep learning and a traditional machine learning method, adopting a multi-task learning algorithm, and adjusting algorithm weight according to individual differences of patients so as to predict rehabilitation indexes;

And finally, implementing dynamic adjustment and optimization, designing a real-time feedback system, and applying an enhanced learning algorithm to dynamically optimize a training scheme according to the rehabilitation effect of the patient so as to improve the individuation and effect of rehabilitation training.

2. The machine learning based personalized breast cancer postoperative rehabilitation training protocol optimization method according to claim 1, wherein the data collection and preprocessing comprises the following steps:

s1, using a dynamic adjustment filtering technology based on patient rehabilitation state change, aiming at the characteristics of breast cancer rehabilitation data, including rehabilitation progress and physiological feedback; applying a tuning filter function to optimize the process:

where x represents raw data including activity and heart rate, mu _i and sigma _i are respectively the mean and standard deviation of each data dimension, and w _i is a weight adaptively adjusted based on rehabilitation status;

S2, adopting a dynamic normalization strategy based on a rehabilitation stage, combining physical and physiological parameters, using a formula to maintain the validity of data and adding nonlinear characteristics to adapt to different rehabilitation stages:

wherein L, k, c are adjustment parameters, ω (t) is a weight function related to rehabilitation time for enhancing sensitivity to rehabilitation early and late data;

S3, simulating different rehabilitation situations possibly experienced by the patient by using the generated countermeasure network GANs, and enhancing data by using the following generated model functions:

Where x represents the measured data during rehabilitation, z is the random input to the generation network, and f and v _j are model parameters to generate synthetic data that fit the actual rehabilitation progress changes.

3. The machine learning based personalized breast cancer postoperative rehabilitation training protocol optimization method according to claim 1, wherein the feature selection and feature construction comprises:

Firstly, adopting feature association analysis based on a graph model, selecting key features through connection strength and path analysis of nodes in the graph, and using a formula:

Where x _i,x_j represents the different rehabilitation characteristics, lambda _k,a_k,b_k is the adaptively adjusted parameter, Representing specific operators to achieve significance assessment of deep-level associations between features;

Secondly, feature fusion is carried out by utilizing a deep learning network, and composite features are constructed:

Where x, y represents the original features, including mobility and emotional state, increasing the nonlinearity and depth of the process by the integral form of the time variable t;

Finally, creating new features by combining medical expert knowledge, and applying a dynamic system model:

Where α, β are parameters adjusted based on the rehabilitation process for generating features reflecting the rehabilitation dynamics process.

4. The machine learning based personalized breast cancer postoperative rehabilitation training protocol optimization method according to claim 1, wherein the hybrid model construction comprises:

Firstly, adopting a mixed feature processing technology, analyzing high-dimensional unstructured data such as rehabilitation training videos through a convolutional neural network CNN, simultaneously using a support vector machine SVM to process basic physiological indexes of a patient, and realizing fusion of two types of data features by using a full-connection layer or an attention mechanism;

Secondly, predicting a plurality of rehabilitation indexes including pain level, exercise capacity and emotion state by using a multi-task learning framework, and sharing useful information among different tasks through task correlation mapping so as to improve the overall prediction capacity of the model;

finally, a personalized weight learning algorithm is introduced, weights of different tasks are dynamically adjusted according to the rehabilitation progress and individual response difference of the patient, and task weights are learned and updated according to rehabilitation feedback and historical data of the patient by adopting a feedback-based circulation mechanism.

5. The machine learning based personalized breast cancer postoperative rehabilitation training protocol optimization method according to claim 4, wherein the implementation of the hybrid feature processing technique comprises:

S1, firstly, a double-flow mixed network architecture is adopted, wherein one flow processes rehabilitation training video data through a convolutional neural network CNN, a multi-scale convolutional kernel W _s ⁽ⁱ⁾ is utilized to extract a spatial feature f _s ⁽ⁱ⁾, and the method comprises the following steps:

calculated, where σ represents the nonlinear activation function, n is the number of convolutional layers, Is a bias term;

s2, the other flow processes the basic physiological index x of the patient through a support vector machine SVM, outputs a weight vector v and a bias c, and calculates a classification decision boundary through a multi-core function k, wherein the formula is as follows:

where α _j is the Lagrangian multiplier and m is the number of support vectors;

S3, finally, a full-connection layer containing a multi-layer network structure is used for realizing fusion of two data source characteristics, and a calculation formula of the fusion characteristic h is as follows:

h＝tanh(W_f·(exp(-|a_sf_s ⁽ⁱ⁾-a_tf_t ^(j)|²)·(a_sf_s ⁽ⁱ⁾+a_tf_t ^(j)))+W_p·g(x))

Where W _f and W _p are weight matrices and a _s and a _t are weights calculated by the dynamic attention mechanism.

6. The machine learning based personalized breast cancer postoperative rehabilitation training protocol optimization method of claim 4, wherein predicting a plurality of rehabilitation metrics comprises:

S1, firstly, each rehabilitation index is branched through a neural network, and the weight W _i and the bias b _i are adjusted according to personalized data of a patient by utilizing a higher derivative and a transformation function fi, wherein the specific formula is as follows:

wherein γ _k is the adjustment parameter and n is the number of variables;

s2, optimizing information sharing among tasks by introducing a graph-based attention mechanism, and mapping task correlation by using an adaptive weight alpha _ij, wherein the formula is as follows:

Wherein W _rel is a relation weight matrix, beta _k is an adjustment coefficient associated with each task, and m is the number of tasks;

S3, finally, optimizing the comprehensive loss function by combining the prediction results and the shared information of all tasks:

Where λ _i is the weight of task i and ρ _ij is the inter-task correlation adjustment factor to optimize the goal is to reduce the gradient differences between different tasks.

7. The machine learning-based personalized breast cancer postoperative rehabilitation training scheme optimization method according to claim 4, wherein the personalized weight learning algorithm is characterized by the following functions:

Updating task weights, wherein ω (t) represents task weights at time t, η is a learning rate parameter dynamically adjusted by patient feedback, δl (t) is a loss function rate of change calculated from patient feedback;

then, using a feedback-based loop learning mechanism, the function is used:

To calculate the effective weight adjustment for each period T, where α (T) is a factor of the adjustment weight, λ is the decay factor to reflect the rate of decay of the effect of the historical data, δf (T) represents the function improvement index derived from the patient feedback.

8. The machine learning-based personalized breast cancer postoperative rehabilitation training scheme optimization method according to claim 1, wherein the dynamic adjustment and optimization are realized by the following steps:

s1, firstly, implementing a real-time feedback system, and adopting a nonlinear power system model:

Where s is the comprehensive rehabilitation state collected from the sensor, t is time, α, β, γ are parameters for adjusting the response speed and sensitivity of the model to sense the acceleration and speed of the rehabilitation state, and provide data processing support for dynamic adjustment;

S2, defining a reward function:

Where f (s, a) is a rehabilitation effect function dependent on the patient's state and the treatment action, f ₀ is a rehabilitation objective function, λ(s) is a dynamic coefficient adjusted based on the patient's current state, and T is the investigation period; by establishing an exponential relationship between the desired rehabilitation effect and the actual rehabilitation effect, it allows optimizing the training regimen over a wider dynamic range.