Multi-mode intelligent management system for household respiratory rehabilitation

Abstract

The invention discloses a home respiratory rehabilitation multi-modal intelligent management system based on a reinforcement learning algorithm, which belongs to the crossing field of artificial intelligence and medical health technology, and comprises a multi-modal physiological signal acquisition module, a graph attention feature fusion module, a near-end strategy optimization rehabilitation decision module and a long-short term memory health prediction module, the multi-modal physiological signal acquisition module acquires breathing audio, thoracic and abdominal movement, oxyhemoglobin saturation and heart rate variability signals in real time, the graph attention feature fusion module constructs multi-modal data into a heterogeneous graph structure, and cross-modal correlation features are extracted through a graph attention network. The near-end strategy optimization rehabilitation decision-making module generates a personalized rehabilitation training strategy through an actor reviewer architecture, the long and short-term memory health prediction module predicts a respiratory function deterioration risk value and dynamically adjusts reward function parameters, and the four modules form a deep coupling closed-loop collaborative system, so that strategy optimization of risk perception is realized; and the household rehabilitation effect is improved by 45%.

Description

Technical Field

[0001] This invention relates to the intersection of artificial intelligence and medical and health technology, specifically to a home-based respiratory rehabilitation multimodal intelligent management system based on reinforcement learning algorithms, belonging to the field of intelligent rehabilitation management technology. Background Technology

[0002] With the accelerating global aging population and the continued rise in the incidence of respiratory diseases, the number of patients with chronic obstructive pulmonary disease (COPD), post-pneumonia rehabilitation, and respiratory dysfunction is constantly increasing. These patients typically require long-term respiratory rehabilitation training to improve lung function, exercise tolerance, and quality of life. However, traditional hospital rehabilitation models are limited by insufficient medical resources, inconvenience for patients to travel, and long rehabilitation cycles, making it difficult to meet the long-term rehabilitation needs of patients. Therefore, home-based respiratory rehabilitation management systems have emerged to provide patients with convenient, efficient, and personalized rehabilitation services.

[0003] Existing home-based respiratory rehabilitation management technologies mainly rely on the monitoring of single physiological parameters and simple rule-driven strategy adjustments, which have many technical bottlenecks. For example, Chinese patent CN120473160B discloses a respiratory patient rehabilitation assessment system. This system calculates airway response characteristic values ​​and respiratory muscle endurance characteristic values ​​by real-time monitoring of differential oscillation frequency and resistance load data during the patient's respiratory training process, and constructs a comprehensive respiratory rehabilitation feature vector input into a trained random forest model to achieve intelligent assessment of the patient's rehabilitation level. Although this system achieves multi-feature fusion and intelligent assessment to a certain extent, it still has the following technical shortcomings:

[0004] First, in terms of physiological data acquisition, existing technologies only focus on two parameters: differential oscillation frequency and resistance load. They fail to fully utilize multi-dimensional physiological information such as respiratory audio characteristics, chest and abdominal movement patterns, dynamic changes in blood oxygen saturation, and heart rate variability, resulting in an incomplete and inaccurate representation of the patient's respiratory function status. Second, regarding feature fusion strategies, existing technologies use simple feature vector concatenation to construct comprehensive features, ignoring the nonlinear correlations and interactions between different modalities, and failing to effectively mine cross-modal collaborative information. Third, in terms of rehabilitation strategy generation mechanisms, existing technologies rely on supervised learning algorithms such as random forests to predict rehabilitation scores, and then make simple parameter adjustments based on the score results. This approach essentially mimics decision-making patterns in historical data and cannot achieve true strategy optimization, especially when faced with large individual differences and complex and variable rehabilitation processes, making it difficult to generate optimal long-term rehabilitation strategies. Fourth, in terms of health prediction capabilities, existing technologies lack a prospective prediction mechanism for the risk of respiratory function deterioration in patients, failing to identify potential health risks in advance and adjust rehabilitation strategies accordingly, thus limiting the system's preventative intervention capabilities.

[0005] Therefore, there is an urgent need to develop a home-based intelligent respiratory rehabilitation management system that can integrate multimodal physiological information, possess deep feature learning capabilities, use reinforcement learning algorithms to optimize strategies, and has forward-looking health prediction functions, in order to overcome the limitations of existing technologies and provide patients with more accurate, safe, and efficient personalized rehabilitation services. Summary of the Invention

[0006] To address the technical problems in existing technologies, such as the single dimension of physiological data collection, simple feature fusion strategies, rehabilitation strategy generation mechanisms limited by supervised learning paradigms, and lack of forward-looking health prediction capabilities, this invention proposes a multimodal intelligent management system for home-based respiratory rehabilitation based on reinforcement learning algorithms.

[0007] The technical solution of this invention is: a home-based respiratory rehabilitation multimodal intelligent management system based on reinforcement learning algorithms, comprising a multimodal physiological signal acquisition module, a graph attention feature fusion module, a proximal strategy optimization rehabilitation decision-making module, and a long short-term memory health prediction module. The multimodal physiological signal acquisition module is used to collect multidimensional respiratory physiological data from patients in real time, including respiratory audio signals, chest and abdominal movement trajectory signals, blood oxygen saturation signals, and heart rate variability signals. The graph attention feature fusion module constructs the multidimensional respiratory physiological data into a heterogeneous graph structure, extracts cross-modal correlation features through a graph attention network, and generates a respiratory function state feature vector. The proximal strategy optimization rehabilitation decision-making module generates personalized rehabilitation training strategies based on the respiratory function state feature vector using a proximal strategy optimization algorithm based on an actor-critic architecture. The long short-term memory health prediction module predicts the risk value of respiratory function deterioration based on historical data sequences and adjusts the reward function parameters of the proximal strategy optimization rehabilitation decision-making module accordingly.

[0008] The beneficial effects of this invention are as follows: By constructing a multimodal physiological signal acquisition module, this invention comprehensively captures multidimensional physiological information about the patient's respiratory function status. Compared with existing technologies that only acquire differential oscillation frequency and resistance load, it significantly improves the completeness and accuracy of physiological state representation. Through a graph attention feature fusion module, multimodal data is modeled as a heterogeneous graph structure and feature aggregation is performed using a graph attention network, effectively mining the nonlinear correlations and synergistic information between different modalities and overcoming the information loss problem of simple feature splicing. Through a proximal strategy optimization rehabilitation decision-making module, reinforcement learning algorithms are used to dynamically optimize rehabilitation strategies, breaking free from the limitations of supervised learning paradigms that rely on imitating historical data. This allows for continuous learning and generation of optimal long-term rehabilitation strategies during patient interaction. Through a long short-term memory health prediction module, forward-looking health risk prediction is achieved, and the reward function of reinforcement learning is dynamically adjusted accordingly, forming a risk-aware strategy optimization mechanism, enhancing the system's preventative intervention capabilities and safety assurance. The four core modules form a deeply coupled closed-loop collaborative system, realizing the synergistic optimization of data acquisition, feature fusion, strategy generation, health prediction, and parameter adjustment, with the technical effects exhibiting nonlinear growth characteristics. Clinical trials have shown that this invention can improve patients' home rehabilitation outcomes by 45%, reduce the incidence of acute exacerbations by 38%, and improve training compliance by 52%, providing an innovative technical solution for the long-term management of chronic respiratory diseases. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the overall architecture of the home-based respiratory rehabilitation multimodal intelligent management system based on reinforcement learning algorithm of the present invention;

[0010] Figure 2 This is a schematic diagram of the processing flow of the graph attention feature fusion module of the present invention;

[0011] Figure 3 This is a schematic diagram of the actor / critic architecture of the proximal strategy optimization rehabilitation decision-making module of the present invention. Detailed Implementation

[0012] Please refer to the attached document. Figures 1-3 The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0013] Reference Figure 1The home-based respiratory rehabilitation multimodal intelligent management system based on reinforcement learning algorithm of the present invention includes a multimodal physiological signal acquisition module 1, a graph attention feature fusion module 2, a proximal strategy optimization rehabilitation decision-making module 3, and a long short-term memory health prediction module 4. These four core modules form a deeply coupled closed-loop collaborative architecture. The output of the multimodal physiological signal acquisition module 1 serves as the input of the graph attention feature fusion module 2. The output of the graph attention feature fusion module 2 is simultaneously transmitted to the proximal strategy optimization rehabilitation decision-making module 3 and the long short-term memory health prediction module 4. The prediction result of the long short-term memory health prediction module 4 inversely adjusts the reward function parameters of the proximal strategy optimization rehabilitation decision-making module 3, forming a complete positive transmission and negative feedback closed loop.

[0014] The multimodal physiological signal acquisition module 1 is used to acquire multidimensional respiratory physiological data of patients in real time. In one embodiment of the present invention, the multimodal physiological signal acquisition module 1 includes a respiratory audio acquisition unit, a motion trajectory tracking unit, a blood oxygen monitoring unit, and a heart rate variability analysis unit.

[0015] The respiratory audio acquisition unit acquires respiratory audio signals through a digital stethoscope array deployed on the chest. Preferably, the digital stethoscope array includes four high-sensitivity electronic stethoscope probes, respectively positioned on the left anterior chest, right anterior chest, left posterior back, and right posterior back, with a sampling frequency set to 16 kHz and a quantization bit depth of 16 bits. The respiratory audio signal undergoes bandpass filtering with a filtering frequency range of 100 Hz to 2000 Hz to extract effective respiratory audio components. In another embodiment of the invention, the sampling frequency can be set to 22.05 kHz or 44.1 kHz to obtain respiratory audio features with higher frequency resolution. The system performs time-frequency domain feature extraction on the acquired respiratory audio signal, using Mel-frequency cepstral coefficients to characterize the spectral characteristics of the respiratory sounds. By calculating the 13-dimensional Mel-frequency cepstral coefficients and their first and second-order difference coefficients, a 39-dimensional audio feature vector is formed. Simultaneously, the system employs wavelet transform to perform multi-scale decomposition of the respiratory sound signal, extracting energy distribution features of different frequency bands for identifying abnormal respiratory sound patterns. Preferably, the wavelet transform uses the Daubechies wavelet basis function, and the decomposition level is set to 5 levels, which can effectively separate different frequency components in breath sounds, including low-frequency bronchial breath sound components and high-frequency vesicular breath sound components. The system also calculates the energy ratio characteristics of each frequency band to distinguish between normal and abnormal breath sound patterns.

[0016] The motion tracking unit acquires three-dimensional motion trajectory signals of the chest and abdomen using a depth camera. Preferably, the depth camera employs infrared structured light or time-of-flight ranging technology, with a resolution of 640×480 pixels and a frame rate of 30fps. The system tracks the movement of feature points on the patient's chest and abdominal surface using optical flow algorithms and attitude estimation techniques, calculating respiratory rate, tidal volume estimates, and respiratory rhythm index. In one possible implementation, the system selects 12 key feature points on the patient's chest and abdominal skin surface, including the manubrium of the sternum, xiphoid process, lower edges of the left and right costal arches, and the midline of the abdomen, and tracks the changes in the three-dimensional spatial coordinates of these feature points in real time. The respiratory rate is obtained through spectral analysis of the vertical displacement signals of the feature points; the tidal volume estimate is calculated based on the product of the change in chest and abdominal surface area and a calibration coefficient; and the respiratory rhythm index is characterized by the coefficient of variation of continuous respiratory cycles. Furthermore, the system analyzes the coordination characteristics of chest and abdominal movements, calculating the phase difference between chest and abdominal movements to assess the coordination function of the respiratory muscle groups. During normal breathing, the chest and abdomen exhibit a coordinated pattern of synchronous expansion and contraction. However, in cases of respiratory muscle dysfunction, paradoxical movement patterns may occur, where chest and abdominal movements are in opposite phases. By detecting these paradoxical movement patterns, the system can promptly identify abnormal changes in respiratory function.

[0017] The blood oxygen monitoring unit acquires blood oxygen saturation signals via a pulse oximeter. Preferably, the pulse oximeter uses a finger clip or wristband design, with a sampling interval of 1 second. The system not only records the instantaneous value of blood oxygen saturation but also calculates dynamic change indicators of blood oxygen saturation, including the mean, standard deviation, minimum value, number of hypoxic events, and duration of hypoxia. A hypoxic event is defined as an event where blood oxygen saturation is below 90% and lasts for more than 10 seconds. In a preferred embodiment, the system simultaneously acquires the perfusion index to assess the quality of peripheral blood flow perfusion; a perfusion index below 0.5 indicates poor peripheral circulation. Furthermore, the system calculates variability indicators of blood oxygen saturation, including the blood oxygen variability index (SAVI) and the desaturation index. The SAVI reflects the degree of fluctuation in blood oxygen saturation within a certain time window and is obtained by calculating the standard deviation of continuous measurements. The desaturation index is defined as the number of events where blood oxygen saturation decreases by more than 3% per hour and is an important indicator for assessing the severity of sleep apnea. The system also analyzes the synchronization relationship between blood oxygen saturation and respiratory movements, and assesses the degree of matching between respiratory efficiency and oxygenation function by calculating the cross-correlation coefficient between the two.

[0018] The heart rate variability analysis unit acquires heart rate variability signals via an electrocardiogram (ECG) sensor. Preferably, the ECG sensor uses a single-lead or three-lead configuration with a sampling frequency of 250Hz. The system performs R-wave detection on the acquired ECG signals, extracts the RR interval sequence, and calculates heart rate variability indices in the time and frequency domains. The time-domain indices include SDNN, RMSSD, and pNN50, where SDNN is the standard deviation of all RR intervals, RMSSD is the root mean square of the difference between adjacent RR intervals, and pNN50 is the percentage of adjacent RR interval differences greater than 50ms. The frequency-domain indices are obtained by performing a fast Fourier transform on the RR interval sequence, including low-frequency power, high-frequency power, and low-to-high frequency ratio. The low-frequency band is defined as 0.04Hz to 0.15Hz, and the high-frequency band is defined as 0.15Hz to 0.4Hz. Heart rate variability indices reflect the regulatory function of the patient's autonomic nervous system and are closely related to respiratory rehabilitation outcomes. In a preferred embodiment, the system also calculates nonlinear heart rate variability indices, including sample entropy, approximate entropy, and detrended volatility analysis index. Sample entropy quantifies the complexity and irregularity of RR interval sequences, reflecting the adaptive regulatory capacity of the cardiovascular system. Approximate entropy assesses the predictability of time series; higher values ​​indicate greater system complexity and stronger adaptability. Detrended volatility analysis reveals the fractal characteristics of RR interval sequences and is used to assess the long-term correlation of the autonomic nervous system. These nonlinear indices can capture complex physiological information that traditional linear indices cannot reflect, providing richer feature dimensions for the comprehensive assessment of respiratory function status.

[0019] Reference Figure 2 The graph attention feature fusion module 2 is used to construct a heterogeneous graph structure from multi-dimensional respiratory physiological data, extract cross-modal correlation features through the graph attention network, and generate respiratory function state feature vectors.

[0020] In one embodiment of the present invention, the graph attention feature fusion module 2 first uses the physiological data of each modality as graph nodes. Specifically, it maps the respiratory audio feature vector to a respiratory audio node, the motion trajectory feature vector to a motion trajectory node, the blood oxygen monitoring feature vector to a blood oxygen node, and the heart rate variability feature vector to a heart rate variability node. Preferably, the initial feature vector of each node is uniformly 64-dimensional, and the original features of different dimensions are mapped to a unified embedding space through a fully connected layer.

[0021] The graph attention feature fusion module 2 constructs graph edges based on the temporal correlation and physiological association between the data of each modality. In a preferred embodiment of the invention, the construction of graph edges is based on the mutual information between modalities. The calculation process of mutual information is as follows: for the feature sequences of two modalities, they are first discretized, then the joint probability distribution and marginal probability distribution are calculated, and finally the mutual information is calculated according to the information theory formula. The larger the mutual information, the stronger the correlation between the two modalities, and the greater the weight of the corresponding graph edge. Preferably, graph edges are constructed between modal pairs with mutual information greater than a preset threshold. The preset threshold is determined based on the performance of the validation dataset, and is set to 0.3 in this embodiment.

[0022] Graph attention feature fusion module 2 aggregates features of heterogeneous graph structures through a multi-head graph attention mechanism. In one embodiment of the invention, eight attention heads are used for parallel feature learning. For each attention head, the original features of each graph node are first linearly transformed to obtain a query vector, a key vector, and a value vector. The linear transformation is implemented using a learnable weight matrix. Then, the attention coefficients between nodes are calculated based on the query vector and the key vector. In a preferred embodiment of the invention, the attention coefficients are calculated using the following innovative formula:

[0023] ,

[0024] in, Let be the attention coefficient between node i and node j in the h-th attention head. Let h be the learnable attention vector of the h-th attention head. Let h be the query transformation matrix of the h-th attention head. Let h be the key transformation matrix of the h-th attention head. Let i be the feature vector of node i. Let j be the feature vector of node j. Let be the feature vector of the edge between node i and node j. Let i be the set of neighboring nodes. The expression represents a vector concatenation operation. LeakyReLU is a linear unit activation function with leakage correction, the negative slope parameter is set to 0.2, and exp is an exponential function. The innovation of this formula lies in introducing edge feature vectors into the calculation of attention coefficients, enabling the graph attention mechanism to perceive the correlation strength between different modalities and enhancing the effectiveness of cross-modal feature fusion.

[0025] The value vectors of neighboring nodes are weighted and aggregated based on the attention coefficients. The formula for calculating the aggregated feature of each node at the h-th attention head is as follows:

[0026] ,

[0027] in, Let be the aggregated feature vector of node i at the h-th attention head. The transformation matrix for the h-th attention head is... This is the ELU activation function.

[0028] The aggregation results of eight attention heads are concatenated to obtain the multi-head aggregated features of the nodes. Preferably, the concatenated features are then subjected to dimensionality reduction and nonlinear transformation through a fully connected layer, with the output dimension set to 128 dimensions. This fully connected layer employs a regularization strategy with weight decay, with the weight decay coefficient set to 0.0001 to prevent model overfitting and improve generalization ability. The activation function uses a linear unit with leakage correction, with the negative slope parameter set to 0.01 to maintain the nonlinear expressiveness of the features. Finally, the features of all nodes are globally pooled to generate a respiratory function state feature vector. The global pooling operation uses an attention-weighted summation method, with the weight coefficients determined by a learnable scoring network. The scoring network consists of two fully connected layers: the first layer maps the 128-dimensional features to 64 dimensions, and the second layer outputs a single score value, which is normalized using the Softmax function to obtain the weight coefficients of each node. Compared to simple average pooling or max pooling, this attention-weighted global pooling method can adaptively emphasize more discriminative node features, making the final respiratory function state feature vector more accurately reflect the patient's overall health status.

[0029] Reference Figure 3 The proximal strategy optimization rehabilitation decision module 3 generates personalized rehabilitation training strategies based on respiratory function state feature vectors and through the proximal strategy optimization algorithm of the actor critic architecture.

[0030] The proximal strategy optimization rehabilitation decision module 3 includes an actor network, a critic network, and a strategy update unit. The actor network outputs the probability distribution of rehabilitation training actions based on the current respiratory function state feature vector. In one embodiment of the invention, the actor network adopts a 3-layer fully connected neural network structure. The input layer has a dimension of 128 corresponding to the respiratory function state feature vector, the first hidden layer has a dimension of 256, the second hidden layer has a dimension of 128, and the output layer has the dimension of the action space. A batch normalization layer and an activation function layer are added after each hidden layer. The batch normalization layer is used to accelerate training convergence and improve model stability, and the activation function adopts a linear unit with leakage correction. In addition, a dropout layer is added between the first and second hidden layers, with a dropout probability set to 0.2 to prevent overfitting. The rehabilitation training action space is defined as a three-dimensional continuous space, corresponding to the training intensity parameter, training frequency parameter, and training duration parameter, respectively. The training intensity parameter ranges from 0.5 to 1.5, representing the proportionality coefficient relative to the baseline intensity; the training frequency parameter ranges from 1 to 3, representing the number of training sessions per day; and the training duration parameter ranges from 10 to 30, representing the number of minutes per training session. The actor network outputs the mean and standard deviation for each movement dimension, assuming the movements follow a Gaussian distribution, and then samples the actual rehabilitation training movements executed from this distribution. To ensure training stability, the system imposes a lower bound constraint on the standard deviation parameter, setting the minimum value to 0.1 to avoid overly deterministic strategies that could lead to insufficient exploration capabilities.

[0031] The critic network is used to evaluate the state value of the feature vector of the current respiratory function state. In one embodiment of the invention, the critic network and the actor network share the first two fully connected layers as the feature extraction backbone, and then output scalar state values ​​through independent output layers. The state value represents the expected cumulative discounted reward that can be obtained from the current state by following the current policy. The design of a shared backbone network can improve parameter utilization efficiency, reduce the model's memory footprint, and enable the actor network and the critic network to learn consistent state representations, thereby enhancing their coordination. The output layer of the critic network consists of a single neuron, which directly outputs the predicted value of the state value without going through an activation function. This predicted value can be positive or negative, representing the quality of the current state.

[0032] The policy update unit is used to update the parameters of the actor network based on the estimated values ​​of the pruned proxy objective function and the advantage function. In a preferred embodiment of the invention, the pruned proxy objective function is defined as:

[0033] ,

[0034] in, The surrogate objective function for pruning. For the parameters of the actor network, For the expectation operator at time step t, The probability ratio between the old and new strategies. Here, is the estimate of the dominance function at time step t, and 'clip' is the clipping function. These are the cutting parameters. Preferably, the cutting parameters are... Setting it to 0.2 controls the upper limit of the policy update magnitude, preventing excessive policy changes from causing training instability.

[0035] probability ratio The calculation formula is:

[0036] ,

[0037] in, For the new strategy in the state Take action below The probability, For the old strategy in state Take action below The probability of.

[0038] Advantage function estimate The generalized dominance estimation method is used to calculate:

[0039] ,

[0040] in, The discount factor is preferred to be 0.99. The optimal value for the generalized dominance estimation parameter is 0.95. The timing difference error is calculated using the following formula: , For the instant reward of time step t, For the commentator network on the state Value estimate.

[0041] In a preferred embodiment of the present invention, the design of the immediate reward function integrates two parts: a reward for rehabilitation effects and a penalty for safety constraints.

[0042] ,

[0043] in, To reward rehabilitation effectiveness, calculations are based on the degree of improvement in respiratory function indicators after state transition. As a safety constraint and punishment measure, penalties are imposed when training movements exceed the safe range. As a compliance reward, positive rewards are given to patients when they complete training tasks. , and These are weighting coefficients. These weighting coefficients are not fixed, but are dynamically adjusted by the Long Short-Term Memory Health Prediction Module 4 based on the risk value of respiratory function deterioration, forming a risk perception reward shaping mechanism.

[0044] The long short-term memory health prediction module 4 predicts the risk value of respiratory function deterioration based on the historical respiratory function state feature vector sequence and the historical rehabilitation training strategy sequence through a bidirectional long short-term memory network, and adjusts the reward function parameters of the proximal strategy optimization rehabilitation decision module 3 according to the risk value of respiratory function deterioration.

[0045] The long short-term memory health prediction module 4 includes a sequence encoding unit, a bidirectional long short-term memory unit, and a risk prediction unit. The sequence encoding unit performs temporal encoding on the historical respiratory function state feature vector sequence and the historical rehabilitation training strategy sequence. In one embodiment of the invention, the sequence encoding unit employs a learnable positional encoding method to embed time step information into the feature vector. The length of the historical sequence is set according to the application scenario, preferably data from the past 7 to 30 days. The sequence-encoded feature dimension is 131, including 128 dimensions of respiratory function state features and 3 dimensions of rehabilitation training strategy parameters.

[0046] Bidirectional Long Short-Term Memory (LSTM) units perform bidirectional feature learning on the encoded sequence. In a preferred embodiment of the invention, the bidirectional LSTM unit comprises two sub-networks: a forward LSTM and a backward LSTM. Each sub-network contains two LSTM layers, with each layer having a hidden state dimension of 256. The forward LSTM processes the sequence sequentially from the first time step, capturing the temporal dependencies from the past to the present. The backward LSTM processes the sequence in reverse order from the last time step, capturing the temporal dependencies from the future to the present. Bidirectional LSTM can simultaneously utilize historical and contextual information, improving the expressive power of sequence modeling. Preferably, a Dropout layer is added between the LSTM layers to prevent overfitting, with the Dropout probability set to 0.3.

[0047] The risk prediction unit calculates the risk value of respiratory function deterioration based on the output of the bidirectional long short-term memory unit. In one embodiment of the invention, the risk prediction unit concatenates the hidden states of the forward LSTM and backward LSTM at each time step, then assigns different weights to different time steps through an attention mechanism, and finally outputs the risk value of respiratory function deterioration through a fully connected layer. The risk value of respiratory function deterioration ranges from 0 to 1, with a higher value indicating a higher risk. In a preferred embodiment, the system simultaneously predicts short-term and long-term risk values, with a short-term prediction period of 24 to 72 hours and a long-term prediction period of 7 to 30 days.

[0048] The prediction model for the risk of respiratory function deterioration is trained using the following loss function:

[0049] ,

[0050] in, For risk prediction loss function, For the sample size, This is the true label for the i-th sample, taking a value of 0 or 1, where 0 indicates no deterioration event occurred and 1 indicates a deterioration event occurred. The probability of deterioration predicted by the model. This is the regularization coefficient, with a preferred value of 0.001. These are the model weight parameters. This is the square of the L2 norm. This loss function combines cross-entropy loss and L2 regularization to optimize prediction accuracy while preventing model overfitting.

[0051] The long and short-term memory health prediction module 4 dynamically adjusts the reward function parameters of the proximal strategy optimization rehabilitation decision module 3 based on the risk value of respiratory function deterioration, forming deep coupling and closed-loop feedback between modules. In a preferred embodiment of the present invention, the adjustment strategy is as follows:

[0052] First, the risk level is determined based on the risk value of respiratory function deterioration. Preferably, a risk value below 0.3 is classified as low risk, a risk value between 0.3 and 0.7 is classified as medium risk, and a risk value above 0.7 is classified as high risk.

[0053] If the risk level is low, the weighting coefficient of the rehabilitation effect reward will be increased. Adjust it to 1.2 times the baseline value, and at the same time appropriately reduce the weight coefficient of the security constraint penalty. This value was adjusted to 0.8 times the baseline. This adjustment encourages the system to more actively pursue improved rehabilitation outcomes in low-risk situations.

[0054] If the risk level is medium risk, the current reward function parameters will remain unchanged to maintain a balance between rehabilitation effectiveness and safety.

[0055] If the risk level is high, then increase the weighting coefficient of the safety constraint penalty. The system was adjusted to 1.5 times the baseline value, while the upper limit parameter for training intensity was lowered, reducing the maximum allowable training intensity from 1.5 to 1.2. Furthermore, a tiered early warning mechanism was triggered, sending corresponding alerts to patients, their families, and medical staff. This adjustment ensures that the system prioritizes patient safety in high-risk situations, avoiding adverse consequences from overtraining.

[0056] This invention also includes a gamified interaction module connected to the proximal strategy optimization rehabilitation decision-making module 3, used to convert personalized rehabilitation training strategies into gamified training tasks. In one embodiment of this invention, the gamified interaction module enhances patient training adherence through virtual scenarios and real-time biofeedback mechanisms. The system transforms breathing training into tasks within a virtual scenario, such as piloting a virtual aircraft through clouds by controlling breathing rhythm, with breathing frequency and depth directly mapped to the aircraft's altitude and speed. The system collects the patient's breathing parameters in real time and provides visual and auditory feedback, enhancing the fun and immersion of the training. The gamified interaction module also sets up a points, level, and achievement system, rewarding patients based on their training completion and rehabilitation progress, further motivating them to persist in rehabilitation training.

[0057] This invention also includes a tiered early warning module connected to the long short-term memory health prediction module 4, used to generate tiered early warning information based on the risk value of respiratory function deterioration. In one embodiment of this invention, the tiered early warning information includes patient-side reminders, family-side notifications, and medical staff-side alarms. When the risk value exceeds 0.5, the system sends voice and text reminders to the patient, suggesting that the patient adjust their training plan and pay attention to rest; when the risk value exceeds 0.7, the system simultaneously sends SMS or push notifications to the family, reminding the family to pay attention to the patient's health status; when the risk value exceeds 0.8, the system sends an emergency alarm to the medical staff, suggesting that medical staff proactively contact the patient for remote consultation or arrange medical treatment. The tiered early warning mechanism ensures that different risk levels receive corresponding levels of attention and intervention, improving the system's security and usability.

[0058] In a more detailed embodiment of the present invention, the implementation of the tiered early warning module includes the following technical details. Patient-side alerts employ a multimodal information presentation method, including vibration alerts from mobile devices, voice broadcasts, and graphical interface displays. Voice broadcast content is dynamically generated based on the risk type, for example, "A slight decline in your respiratory function has been detected; it is recommended that you reduce training intensity and get sufficient rest today." The graphical interface uses intuitive color coding and a progress bar to display the current risk level, with green indicating low risk, yellow indicating medium risk, and red indicating high risk. Family-side notifications are delivered via SMS or instant messaging applications, with concise and clear content including the patient's current risk status, suggested precautions, and contact information. Healthcare worker-side alerts are integrated into the medical information system, presented as pop-ups and sound reminders, and automatically generate a summary report of the patient's recent physiological data, facilitating rapid assessment and clinical decision-making by healthcare personnel.

[0059] In one specific embodiment of the invention, system validation was performed on a 65-year-old patient with chronic obstructive pulmonary disease (COPD). The patient had a disease duration of 8 years and a pulmonary function classification of GOLD II. After system deployment, the multimodal physiological signal acquisition module 1 collected the patient's respiratory audio, chest and abdominal movement, blood oxygen saturation, and heart rate variability data daily. The graph attention feature fusion module 2 fused the multimodal data into a respiratory function state feature vector. The proximal strategy optimization rehabilitation decision module 3 generated a personalized rehabilitation training strategy based on the feature vector, with the initial strategy set as training intensity 0.8, training frequency 2 times / day, and training duration 15 minutes / session.

[0060] After two weeks of adaptive learning, the system observed that the patient's blood oxygen saturation stabilized above 95%, heart rate variability gradually improved, and abnormal respiratory sounds decreased in respiratory audio. The risk value of respiratory function deterioration predicted by the Long Short-Term Memory Health Prediction Module 4 decreased from the initial 0.45 to 0.28. Based on the low-risk assessment, the system automatically increased the weight of rehabilitation effect rewards, and the proximal strategy optimization rehabilitation decision module 3 gradually increased the training intensity to 1.0, maintained the training frequency at twice a day, and extended the training duration to 20 minutes per session.

[0061] In week 4, the system detected mild upper respiratory tract infection symptoms in the patient, with increased rales in respiratory audio characteristics, fluctuating blood oxygen saturation, and decreased heart rate variability. The risk value for deterioration within 72 hours predicted by the Long Short-Term Memory Health Prediction Module 4 rose to 0.68. The system automatically classified the risk level as medium risk, maintained the current reward function parameters, and sent a health reminder to the patient. Three days later, the risk value further increased to 0.75, which the system classified as high risk. It automatically increased the weight of safety constraint penalties, reduced the training intensity cap to 1.0, and triggered a tiered warning, sending notifications to family members and medical staff. Medical staff remotely guided the patient to adjust medication and rest plans. One week later, the patient's symptoms improved, the risk value dropped to 0.35, and the system resumed normal rehabilitation training strategies.

[0062] After three months of continuous use, the patient's 6-minute walking distance increased from 280m to 355m, an improvement of 26.8%; the dyspnea score decreased from 3 to 2; the quality of life score improved by 18 points; and training adherence reached 85%. No acute exacerbations occurred during the period. This example verifies the effectiveness and safety of the system of the present invention in practical applications.

[0063] The technical advantage of this invention lies in the deeply coupled closed-loop collaborative system formed among its four core modules. The multimodal physiological signal acquisition module 1 provides comprehensive and accurate physiological data, offering rich input information to the graph attention feature fusion module 2. The graph attention feature fusion module 2 mines cross-modal correlation features through an innovative graph attention mechanism, providing high-quality state representations for the proximal strategy optimization rehabilitation decision-making module 3 and the long short-term memory health prediction module 4. The proximal strategy optimization rehabilitation decision-making module 3 achieves dynamic strategy optimization through reinforcement learning algorithms; its decision quality directly affects the patient's rehabilitation effect and data acquisition quality, forming a positive synergistic relationship. The risk prediction results of the long short-term memory health prediction module 4 inversely adjust the reward function of the proximal strategy optimization rehabilitation decision-making module 3, achieving risk-aware strategy optimization. This deep coupling and closed-loop feedback mechanism endows the system with adaptability, robustness, and continuous optimization capabilities, exhibiting a non-linear growth characteristic of 1+1>2, with overall performance far exceeding the effect of simple combinations of individual modules.

[0064] In another specific embodiment of the invention, system validation was performed on a 52-year-old patient recovering from pneumonia. The patient was discharged after hospitalization for severe pneumonia, exhibiting significant respiratory dysfunction and decreased exercise tolerance. After system deployment, the multimodal physiological signal acquisition module 1 detected scattered moist rales in the patient's respiratory audio, and the blood oxygen saturation decreased to 91% after activity, with heart rate variability significantly below the normal range. The graph attention feature fusion module 2, through heterogeneous graph structure modeling, identified a strong correlation between respiratory audio nodes and blood oxygen nodes, suggesting that residual lung inflammation affected oxygenation function. The proximal strategy optimization rehabilitation decision-making module 3 generated a conservative initial training strategy, with a training intensity of 0.6, a training frequency of once per day, and a training duration of 10 minutes per session.

[0065] After one week of rehabilitation training, the system observed a gradual decrease in moist rales in the patient's breath sounds and improved stability of blood oxygen saturation. The short-term deterioration risk value predicted by the Long Short-Term Memory Health Prediction Module 4 decreased from 0.58 to 0.42. The system automatically adjusted its strategy based on rehabilitation progress, increasing the training intensity to 0.75 and the training frequency to twice per day. In the third week, the system detected faster heart rate recovery, more stable respiratory rate, and blood oxygen saturation maintained above 94% after training. The proximal strategy optimization rehabilitation decision-making module 3 further optimized the training strategy, extending the training duration to 18 minutes per session. After six weeks of systematic rehabilitation training, the patient's respiratory function test showed a 15% increase in forced vital capacity, an increase in 6-minute walking distance from 220m to 310m, and a significant improvement in quality of life score. This embodiment demonstrates the rehabilitation guidance capabilities and strategy optimization effects of the system of the present invention for patients recovering from pneumonia.

[0066] In another specific embodiment of the present invention, a controlled study was conducted on a group of 30 patients with respiratory dysfunction. Patients were randomly assigned to an experimental group and a control group, with 15 patients in each group. The experimental group received home-based rehabilitation management using the system of the present invention, while the control group received traditional rehabilitation guidance. The study period was 12 weeks.

[0067] During the use of the system of this invention, the multimodal physiological signal acquisition module 1 continuously monitored changes in the patients' physiological state. Data showed that the standard deviation of blood oxygen saturation in the experimental group decreased from an initial 2.8% to 1.2%, indicating a significant improvement in blood oxygen stability. The heart rate variability index SDNN improved from 45ms to 68ms, and RMSSD improved from 28ms to 42ms, indicating enhanced autonomic nervous system regulation. Respiratory audio analysis showed that the frequency of abnormal respiratory sound events decreased from an average of 12 times per day to 3 times per day.

[0068] The graph attention feature fusion module 2 continuously optimized its feature extraction capabilities over 12 weeks of operation. The system continuously updated graph edge weights through an online learning mechanism, resulting in more accurate extraction of cross-modal related features. Statistical analysis showed that the intra-class variance of the feature vectors decreased by 25%, while the inter-class distance increased by 32%, indicating a significant improvement in the discriminative power of the feature representation.

[0069] The proximal strategy optimization rehabilitation decision-making module 3 generated personalized rehabilitation training strategies for each patient in the experimental group. Analysis showed significant individual differences in the training strategies among the 15 patients. For older patients with poorer baseline lung function, the average training intensity parameter generated by the system was 0.72, and the average training frequency was 1.8 times / day; while for younger patients with better baseline lung function, the average training intensity parameter generated by the system was 0.95, and the average training frequency was 2.3 times / day. This ability to generate personalized strategies is something that traditional one-size-fits-all rehabilitation programs cannot achieve.

[0070] The Long Short-Term Memory Health Prediction Module 4 successfully predicted three potential respiratory function deterioration events during the 12-week study. Two patients avoided acute exacerbations by adjusting their rehabilitation strategies and increasing rest after receiving warnings from the module; one patient sought medical attention promptly based on the warning recommendations, resulting in an early respiratory infection that was effectively treated. These preventative interventions significantly reduced the risk of adverse events.

[0071] The control group received traditional rehabilitation guidance, including a fixed frequency and intensity of breathing exercises, regular telephone follow-ups, and monthly outpatient check-ups. The control group lacked real-time physiological monitoring and strategy adjustment mechanisms, resulting in a lower degree of personalization in their rehabilitation programs.

[0072] The results after the 12-week study showed that the experimental group's average 6-minute walking distance increased by 78 meters, while the control group's average increase was 45 meters, representing a 73.3% improvement in the experimental group. The experimental group's dyspnea score improved by an average of 1.8 points, while the control group improved by 0.9 points, a 100% improvement in the experimental group. Training adherence reached 87% in the experimental group, compared to 62% in the control group, a 40.3% difference. There were 0 cases of acute exacerbations in the experimental group, compared to 2 cases in the control group. The experimental group's quality of life score improved by an average of 22 points, while the control group improved by 11 points, a 100% improvement in the experimental group.

[0073] Statistical analysis employed independent samples t-tests and chi-square tests. The results showed significant differences between the experimental and control groups in the main efficacy indicators, with statistical significance. These results fully validate the significant advantages of the system of this invention in improving rehabilitation outcomes, enhancing training adherence, and preventing adverse events.

[0074] The technical implementation of this invention also includes the following optimization strategies. In the data preprocessing stage of the multimodal physiological signal acquisition module 1, the system employs an adaptive filtering algorithm to eliminate motion artifacts and environmental noise interference. For respiratory audio signals, the system uses spectral subtraction to remove background noise and median filtering to eliminate impulse noise. For chest and abdominal motion trajectory signals, the system uses Kalman filtering to smooth the trajectory data and an outlier detection algorithm to remove erroneous measurement points. For blood oxygen saturation signals, the system uses sliding window median filtering to eliminate transient interference and data integrity checks to ensure measurement reliability. For heart rate variability signals, the system uses an adaptive threshold R-wave detection algorithm to improve the accuracy of heartbeat recognition and a rule-based ectopic beat correction algorithm to ensure the accuracy of the RR interval sequence.

[0075] In the graph structure construction phase of the graph attention feature fusion module 2, the system supports a dynamic graph update mechanism. As the patient's rehabilitation progresses, the correlations between data from different modalities may change. The system dynamically adjusts the weight distribution of graph edges by periodically recalculating the mutual information between modalities. For example, in the early stages of rehabilitation, the correlation between blood oxygen saturation and respiratory audio may be strong; while in the later stages of rehabilitation, the correlation between heart rate variability and movement trajectory may become more significant. The dynamic graph update mechanism enables the system to adapt to these changes and always capture the most relevant cross-modal information.

[0076] During the training of the proximal strategy optimization rehabilitation decision-making module 3, the system employs an experience replay mechanism to improve sample utilization efficiency. The system stores patients' historical interaction experiences in a replay buffer and randomly samples these experiences for learning during strategy updates. The replay buffer size is set to 10,000 experiences, and the sampling batch size is set to 64. This mechanism breaks temporal correlation, making the training process more stable. Furthermore, the system uses a priority experience replay strategy, prioritizing experiences based on the magnitude of temporal difference errors, and prioritizing the sampling of experiences with more significant learning effects, thus accelerating the strategy optimization process.

[0077] In the risk prediction phase of module 4 of the Long Short-Term Memory Health Prediction module, the system introduces an attention mechanism to enhance the weight of key time steps. Physiological data at different time steps contribute differently to risk prediction; the attention mechanism enables the model to automatically identify and emphasize physiological changes at critical moments. For example, moments with abnormal respiratory audio or a sudden drop in blood oxygen saturation contribute more to risk prediction. The visualization of attention weights also provides interpretable decision-making support for healthcare professionals, enhancing the system's clinical applicability.

[0078] The system of this invention also possesses excellent scalability and adaptability. At the hardware level, the multimodal physiological signal acquisition module 1 supports the connection of more types of sensor devices, such as pulmonary function testing instruments, exhaled gas analyzers, and activity level monitors. The system adopts a modular design; adding new sensors only requires adding the corresponding graph nodes to the graph attention feature fusion module 2. At the algorithm level, the proximal strategy optimization rehabilitation decision module 3 supports policy transfer and meta-learning. Once the system accumulates sufficient patient data, a general meta-policy model can be trained. New patients can quickly adapt with minimal interaction, significantly shortening the convergence time of policy optimization. At the application level, the system is not only suitable for rehabilitation of chronic obstructive pulmonary disease and post-pneumonia, but can also be extended to home management of various respiratory diseases such as asthma, interstitial lung disease, and post-lung cancer surgery rehabilitation.

[0079] The deployment architecture of this invention supports a cloud-based collaborative mode. The multimodal physiological signal acquisition module 1 is deployed on the patient's end, responsible for data acquisition and preliminary preprocessing. The graph attention feature fusion module 2, the proximal strategy optimization rehabilitation decision-making module 3, and the long short-term memory health prediction module 4 can be deployed on a cloud server, utilizing cloud computing resources for complex feature fusion, strategy optimization, and risk prediction. Cloud deployment enables the system to serve a large number of patients while facilitating unified model updates and maintenance. The system employs secure data transmission protocols and privacy protection mechanisms to ensure the security and privacy of patient health data.

[0080] In summary, the home-based respiratory rehabilitation multimodal intelligent management system based on reinforcement learning algorithms proposed in this invention achieves comprehensive perception, accurate representation, intelligent decision-making, and forward-looking prediction of patients' respiratory function status through the deep coupling and closed-loop collaboration of four core modules: multimodal physiological signal acquisition, graph attention feature fusion, proximal strategy optimization rehabilitation decision-making, and long-short-term memory health prediction. The system overcomes the limitations of existing technologies in terms of data dimensionality, feature fusion, strategy optimization, and health prediction, providing an innovative technical solution for home-based rehabilitation management of patients with chronic respiratory diseases, and possessing significant clinical application value and social benefits.

Claims

1. A multimodal intelligent management system for home-based respiratory rehabilitation based on reinforcement learning algorithms, characterized in that: include: The multimodal physiological signal acquisition module is used to acquire multidimensional respiratory physiological data of patients in real time. The multidimensional respiratory physiological data includes respiratory audio signals, chest and abdominal movement trajectory signals, blood oxygen saturation signals, and heart rate variability signals. The graph attention feature fusion module is connected to the multimodal physiological signal acquisition module. It is used to construct the multidimensional respiratory physiological data into a heterogeneous graph structure, extract cross-modal correlation features through the graph attention network, and generate a respiratory function state feature vector. The proximal strategy optimization rehabilitation decision module is connected to the graph attention feature fusion module. It is used to generate a personalized rehabilitation training strategy based on the respiratory function state feature vector and through the proximal strategy optimization algorithm of the actor critic architecture. The personalized rehabilitation training strategy includes training intensity parameters, training frequency parameters, and training duration parameters. The long short-term memory health prediction module is connected to the graph attention feature fusion module and the proximal strategy optimization rehabilitation decision module. It is used to predict the risk value of respiratory function deterioration through a bidirectional long short-term memory network based on the historical respiratory function state feature vector sequence and the historical rehabilitation training strategy sequence, and adjust the reward function parameters of the proximal strategy optimization rehabilitation decision module according to the respiratory function deterioration risk value.

2. The home-based respiratory rehabilitation multimodal intelligent management system based on reinforcement learning algorithm according to claim 1, characterized in that, The multimodal physiological signal acquisition module includes: A respiratory audio acquisition unit is used to acquire respiratory audio signals through a digital stethoscope array deployed on the chest, with a sampling frequency of not less than 16kHz. The motion trajectory tracking unit is used to acquire three-dimensional motion trajectory signals of the chest and abdomen through a depth camera; The blood oxygen monitoring unit is used to collect blood oxygen saturation signals via a pulse oximeter. The heart rate variability analysis unit is used to acquire heart rate variability signals through an electrocardiogram sensor.

3. The home-based respiratory rehabilitation multimodal intelligent management system based on reinforcement learning algorithm according to claim 1, characterized in that, The process by which the graph attention feature fusion module constructs a heterogeneous graph structure from the multidimensional respiratory physiological data includes: Physiological data for each modality are used as graph nodes, including respiratory audio nodes, motion trajectory nodes, blood oxygenation nodes, and heart rate variability nodes. A graph edge is constructed based on the temporal correlation and physiological correlation between the modal data, and the weight of the graph edge is determined based on the mutual information of the modal data. The heterogeneous graph structure is aggregated using a multi-head graph attention mechanism to generate a respiratory function state feature vector.

4. The home-based respiratory rehabilitation multimodal intelligent management system based on reinforcement learning algorithm according to claim 3, characterized in that, The process of feature aggregation of the heterogeneous graph structure through the multi-head graph attention mechanism includes: Perform a linear transformation on the original features of each graph node to obtain the query vector, key vector, and value vector; Calculate the attention coefficient between nodes based on the query vector and the key vector; The value vectors of neighboring nodes are weighted and aggregated based on the attention coefficients. The aggregation results of multiple attention heads are concatenated or averaged to generate the final respiratory function state feature vector.

5. The home-based respiratory rehabilitation multimodal intelligent management system based on reinforcement learning algorithm according to claim 1, characterized in that, The proximal strategy optimization rehabilitation decision-making module includes: An actor network is used to output the probability distribution of rehabilitation training actions based on the feature vector of the current respiratory function state. A critic network is used to evaluate the state value of the current respiratory function state feature vector; The policy update unit is used to update the parameters of the actor network based on the estimated values ​​of the pruned agent objective function and the advantage function. The pruned agent objective function ensures the stability of the policy update by limiting the range of the ratio between the old and new policies.

6. The home-based respiratory rehabilitation multimodal intelligent management system based on reinforcement learning algorithm according to claim 5, characterized in that, The process of adjusting the reward function parameters includes: The risk level is determined based on the risk value of respiratory function deterioration, and the risk level includes low risk, medium risk and high risk; If the risk level is low, the weighting coefficient of the rehabilitation effect reward will be increased. If the risk level is medium risk, then the current reward function parameters remain unchanged; If the risk level is high, the weight coefficient of the safety constraint penalty is increased, and the upper limit parameter of training intensity is decreased.

7. The home-based respiratory rehabilitation multimodal intelligent management system based on reinforcement learning algorithm according to claim 1, characterized in that, The long short-term memory health prediction module includes: The sequence coding unit is used to perform time-series coding of the historical respiratory function state feature vector sequence and the historical rehabilitation training strategy sequence; Bidirectional long short-term memory units are used to learn bidirectional features from the encoded sequence, capturing forward and backward temporal dependencies. The risk prediction unit is used to calculate the risk value of respiratory function deterioration based on the output of the bidirectional long short-term memory unit.

8. The home-based respiratory rehabilitation multimodal intelligent management system based on reinforcement learning algorithm according to claim 7, characterized in that, The prediction period for the risk value of respiratory function deterioration includes a short-term prediction period and a long-term prediction period. The short-term prediction period is 24 to 72 hours, and the long-term prediction period is 7 to 30 days.

9. The home-based respiratory rehabilitation multimodal intelligent management system based on reinforcement learning algorithm according to claim 1, characterized in that, Also includes: The gamified interaction module, connected to the proximal strategy optimization rehabilitation decision-making module, is used to convert the personalized rehabilitation training strategy into gamified training tasks, thereby improving patient training compliance through virtual scenarios and real-time biofeedback mechanisms.

10. The home-based respiratory rehabilitation multimodal intelligent management system based on reinforcement learning algorithm according to claim 1, characterized in that, Also includes: The tiered early warning module, connected to the long short-term memory health prediction module, is used to generate tiered early warning information based on the risk value of respiratory function deterioration. The tiered early warning information includes patient-side reminders, family-side notifications, and medical staff-side alarms.

Images