Rehabilitation cloud platform system for monitoring cardiopulmonary function

Abstract

The invention relates to the technical field of medical rehabilitation monitoring, in particular to a rehabilitation cloud platform system for cardiopulmonary function monitoring. According to the system, an intelligent monitoring terminal collects physiological data; the cloud computing platform adopts wavelet threshold denoising and an index moving average method to remove noise, combines dynamic time warping to achieve multi-physiological signal time sequence alignment, constructs a double-flow model, generates a health state index and a comprehensive health risk index, introduces a multi-scale branch and differential attention mechanism, and achieves the multi-scale health state index and the comprehensive health risk index. High-precision anomaly detection is realized through prediction deviation analysis, multi-target reinforcement learning is adopted to balance health improvement and risk control, and a personalized training scheme is generated; the remote medical support module supports a doctor to monitor patient data in real time and intervene in an adjustment scheme; the user interaction end provides a real-time monitoring data visualization and individuation rehabilitation plan. Through multi-modal data fusion, time sequence synchronous optimization and adaptive reinforcement learning, accurate monitoring, dynamic rehabilitation optimization and remote medical collaboration are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of medical rehabilitation monitoring, and particularly relates to a rehabilitation cloud platform system for cardiopulmonary function monitoring. Background Art

[0002] With the rapid development of information technology and the Internet of Things in the medical field, the high incidence of chronic cardiopulmonary diseases has put forward higher requirements for remote monitoring and rehabilitation management. However, traditional hospital monitoring methods are limited by resources and time and cannot meet the needs of continuous, real-time, and personalized health management. Therefore, the integrated application of cardiopulmonary function monitoring based on wearable sensors and cloud computing platforms has become an inevitable trend.

[0003] A Chinese invention patent with the publication number CN118105051B discloses a rehabilitation cloud platform system for cardiopulmonary function monitoring. The system includes an acquisition module for acquiring heart rate data and other physical characteristics data that affect heart rate changes, and determining the noise level of heart rate data at each moment; a heart rate prediction module for determining the prediction error coefficient at each moment except the first moment, and an analysis module for constructing a time series heart rate fluctuation model and determining the predicted heart rate at the next moment based on the neural network model and the time series heart rate fluctuation model; a monitoring module for determining the corrected predicted heart rate and using the corrected predicted heart rate as the cardiopulmonary function monitoring result; by combining the changes in other physical characteristics data, the neighborhood value fluctuation of the heart rate data itself, and the fluctuation analysis of the overall heart rate data, the accuracy of the corrected predicted heart rate is improved, the objectivity and reliability of the cardiopulmonary function monitoring result are enhanced, and the cardiopulmonary function monitoring effect is enhanced.

[0004] Although some existing technologies currently involve physiological signal acquisition and preliminary data processing, there are often problems such as discontinuous monitoring data, untimely warnings, and difficulty in personalizing rehabilitation programs, and systematic solutions have not yet been formed for key technologies such as signal denoising, data alignment, and health risk assessment. Summary of the Invention

[0005] The object of the present invention is to propose a rehabilitation cloud platform system for cardiopulmonary function monitoring in view of the problems in the background art.

[0006] The technical solution of the present invention: A rehabilitation cloud platform system for cardiopulmonary function monitoring includes:

[0007] An intelligent monitoring terminal for collecting physiological data by means of a wearable sensor and transmitting the data to a cloud computing platform;

[0008] A cloud computing platform for performing health assessment, abnormal warning, and generating a personalized rehabilitation plan;

[0009] The cloud computing platform includes:

[0010] A data storage and management unit for storing long-term monitoring data based on a big data storage architecture;

[0011] A data processing unit that establishes a real-time session channel for data transmission, uses wavelet transform and exponential moving average to remove noise, combines dynamic time warping and sliding window for signal alignment, and finally normalizes the data scale through mean standard deviation normalization and extreme value scaling;

[0012] An AI analysis engine that extracts time-domain, frequency-domain, and non-linear features of cardiopulmonary signals, constructs a two-stream health assessment model to obtain health indicators and risk indices, and uses a statistical and improved long short-term memory network prediction model for real-time anomaly detection, and finally transmits the evaluation results to the personalized rehabilitation recommendation unit;

[0013] A personalized rehabilitation recommendation unit that uses a reinforcement learning model with a dual-objective optimization and adaptive exploration mechanism to dynamically adjust the rehabilitation training plan according to the patient's current state and update the rehabilitation strategy;

[0014] A remote medical support module for supporting doctors to monitor the patient's state in real time and give early warnings before abnormal situations occur;

[0015] A user interaction terminal for supporting patients to view real-time monitoring data, rehabilitation suggestions, and health reports, and view personalized training plans and video tutorials based on AI analysis.

[0016] Preferably, the establishment process of the real-time session channel is as follows:

[0017] S21. The intelligent monitoring terminal User of any patient i i sends a session connection application, and the data processing unit receives the session connection application and sends an application confirmation parameter P to the intelligent monitoring terminal User i ;

[0018] wherein, P is a random number;

[0019] S22. The intelligent monitoring terminal User i obtains its own device identifier ID i , calculates the secondary application parameter ASp = H L (ID i ||P), and accordingly sends the application packet {ID i , Asp} to the data processing unit;

[0020] wherein, H() represents a predefined hash function that outputs a 256-bit hash value; H L is the left half of the predefined hash function H, that is, 128 bits; || represents a concatenation operation;

[0021] S23. The data processing unit receives the application packet {ID i , Asp}, extracts the device identifier ID i , and the secondary application parameter Asp, and searches for the value matching the device identifier ID i in the authorized devices in a traversing manner. If the search is successful, the session confirmation parameter SCp = H R (ID i ⊕P) is calculated using the application confirmation parameter P, and the session confirmation parameter SCp is sent to the intelligent monitoring terminal User i of patient i accordingly;

[0022] wherein, H R is the right half part of the predefined hash function H, that is, 128 bits; ⊕ represents the exclusive OR operation;

[0023] S24. The intelligent monitoring terminal User i of patient i receives the session confirmation parameter SCp, and compares whether SCP = H R (ID i ⊕P) holds; if it holds, it is confirmed that a real-time session channel between the intelligent monitoring terminal User i of patient i and the data processing unit is established. Accordingly, the intelligent monitoring terminal User i of patient i uploads the physiological signal to the data processing unit.

[0024] Preferably, the noise removal process of wavelet transform and exponential moving average is as follows:

[0025] S31. The discrete wavelet transform DWT is used to decompose the signal into different frequency bands, and threshold processing is applied to the high-frequency noise part:

[0026] In the formula, S denoise represents the denoised signal; S raw represents the original physiological signal; DWT i () represents the i-th level wavelet decomposition of the signal; w i represents the denoising weight;

[0027] S32. Based on the exponential moving average EMA smoothing, the low-frequency drift is removed:

[0028] S EMA (t) = αS denoise (t)+(1-α)S EMA (t-1);

[0029] In the formula, S EMA (t) represents the smoothed signal at time t; S denoise(t) represents the denoised signal at time t; α represents the smoothing coefficient, where α ∈ [0.1, 0.3];

[0030] S33. Align non-equal-length time series through Euclidean distance, optimize signal matching by combining the sliding window method, and calculate the optimal window size to perform time alignment on physiological signals.

[0031] Preferably, the alignment process for time-aligning physiological signals is as follows:

[0032] S41. Dynamic Time Warping DWT:

[0033] In the formula, Q and C represent two sets of time series; d(Q i , C j ) represents the Euclidean distance between the i-th and j-th data points;

[0034] S42. Optimize signal matching using the sliding window method:

[0035] In the formula, W opt represents the optimal alignment window size; X t , Y t represent two sets of physiological signals; T represents the time step.

[0036] Preferably, the construction process of the two-stream health assessment model is as follows:

[0037] S51. Extract cardiopulmonary function features:

[0038] Time domain features: Calculate the root mean square difference of adjacent R-R intervals RMSSD:

[0039]

[0040] Frequency domain features: Convert the time domain signal to the frequency domain through the Fast Fourier Transform FFT: P(f) = |F(S(t))| 2 ;

[0041] Nonlinear features: Predetermine the embedding dimension m and the similarity threshold r, and calculate the sample entropy SampEn

[0042]

[0043] In the formula, RR i represents the i-th R-R interval; N represents the total number of R-R intervals; F(S(t)) represents the Fourier transform of the signal S(t); P(f) represents the power spectral density at frequency f; m represents the embedding dimension; r represents the similarity threshold; N' represents the signal length; A represents the number of similar sequence pairs of length m + 1; B represents the number of similar sequence pairs of length m;

[0044] S52. Consider the time-domain and frequency-domain features simultaneously, and fuse the patient's historical information to construct the health status indicator Y health : Y health = α·LSTM(X time ) + β·CNN(X freq ) + γ·KG(X history );

[0045] In the formula, LSTM(X time ) represents using the long short-term memory network LSTM to process the time-domain data feature X time , that is, RMSSD; CNN(X freq ) represents using the convolutional neural network CNN to process the frequency-domain feature X freq , that is, P(f); KG(X history ) is the historical information X of the patient integrated based on the knowledge graph technology history ;

[0046] S53. Generate a comprehensive health risk index HRI:

[0047] HRI = w1·HRV + w2·SpO2 + w3·BR + w4·SampEn;

[0048] In the formula, HRV represents the heart rate variability index RMSSD; SpO2 represents the blood oxygen saturation; BR represents the respiratory rate; SampEn represents the sample entropy value; w1, w2, w3, w4 represent the weights of each index

[0049] Preferably, the prediction process of the improved long short-term memory network is as follows:

[0050] S61. Obtain the standardized time series data X, that is, physiological data

[0051] S62. Construct a multi-scale branch. For each scale branch, use an independent LSTM encoder for processing:

[0052] Among them, s represents the scale branch; τ s represents the time window length of branch s represents the hidden state vector output by scale s at time step t

[0053] S63. Introduce an attention mechanism for weighted fusion:

[0054] Preliminary attention weight calculation: Calculate the preliminary attention weights for the output of each scale:

[0055]

[0056] Differential attention calculation and adjustment:

[0057]

[0058] Multi-scale feature fusion:

[0059] In the formula, f() represents a feed-forward neural network; represents the preliminary attention weight at scale s at time step t; exp() represents the exponential function; M represents the total number of scales; represents the change in attention weight between scale s at time step t and the previous time step t-1; λ” represents the differential adjustment coefficient; represents the finally adjusted attention weight; H t represents the feature representation that fuses all scales at time step t;

[0060] S64. Output the predicted value of the cardiopulmonary function index through the fully connected layer, calculate the prediction deviation, and set a dynamic threshold based on the mean and standard deviation for anomaly detection.

[0061] Preferably, the process of anomaly detection is as follows:

[0062] S71. Prediction output:

[0063] In the formula, g() represents a fully connected layer for mapping the fused feature H t to the prediction value space;

[0064] S72. Prediction deviation calculation:

[0065] In the formula, y t represents the actual observed value at the t-th moment;

[0066] S73. Statistical analysis and anomaly threshold setting:

[0067]

[0068] In the formula, N' represents the total number of samples used to calculate the mean and standard deviation; μ e represents the mean of the prediction deviation e t ; σ e represents the standard deviation of the prediction deviation;

[0069] S74. Define the threshold τ2 = δσ e , when |e t -μ e |, it is considered that the prediction at the t-th moment is abnormal.

[0070] Preferably, the policy update strategy process of the reinforcement learning model with dual-objective optimization and adaptive exploration mechanism is as follows:

[0071] S81. Define the state space S t : S t = {Y health , HRI, data};

[0072] Define the action space A t , that is, the rehabilitation training plan: A t = {a1, a2,..., a i ,..., a n};

[0073] Define the reward function R(S t , A t ):

[0074] R(S t , A t ) = w′1·ΔY health - w'2·ΔHRI + w'3·Δdata - w'4·C(A t );

[0075] Wherein, Y health represents the health status index; HRI represents the comprehensive health risk index; data represents the real-time physiological data; a i represents any one of the training methods; ΔY health represents the improvement value of the health status index; ΔHRI represents the change in health risk; C(A t ) represents the training consumption cost; w′1, w'2, w'3, w'4 represent the weight parameters; Δdata represents the sum of the change values of each index in the physiological data;

[0076] S82. Adopt the deep deterministic policy gradient for deep decision-making;

[0077] S83. Adopt the multi-objective reinforcement learning method and use the Pareto front optimization to find the optimal rehabilitation plan:

[0078] maxF(π) = λ1·G1(π) - λ2·G2(π);

[0079]

[0080] Wherein, F(π) represents the comprehensive evaluation index; λ1 represents the health improvement weight; λ2 represents the risk control weight; G1(π) represents the cumulative health benefit; represents the health status index of the patient at time t; T represents the number of training iteration steps; γ represents the discount factor; HRI tDenote the comprehensive health risk index at time t;

[0081] S84. Introduce the upper confidence bound + gradually decaying exploration mechanism to dynamically adjust the training strategy:

[0082]

[0083] wherein, Q(S t , a) denotes the expected cumulative reward obtained by selecting action a in the current state S t ; c represents the exploration factor; lnt represents the natural logarithm of the current time step t; N(a) represents the number of times action a is selected.

[0084] Preferably, the deep decision-making process is as follows:

[0085] Actor network, i.e., the policy network: generate the optimal training plan: A t = π(S t |θ π );

[0086] Critic network, i.e., the value network: evaluate the value of the current policy:

[0087] Q(S t , A t ) = R(S t , A t ) + γ'Q(S t+1 , A t+1 );

[0088] Optimization objective, maximize the long-term health benefit: ▽ θπ J(π) = E[▽ θπ Q(S, A|θ Q )▽ θπ π(S|θ π )];

[0089] wherein, A t denotes the best rehabilitation training action output by the reinforcement learning model, i.e., the training plan that the current patient should execute; π(S t |θ π ) denotes the output of the Actor network, select the optimal action A t based on the state S t ; θ π denotes the parameters of the policy network; Q(S t , A t ) denotes the state-action value function; γ' denotes the discount factor; J(π) denotes the objective function of the policy π; ▽ θπ J(π) denotes the gradient of the objective function J(π) with respect to the policy parameters θ πThe gradient; E() represents the expectation operator; Q(S,A|θ Q ) represents the Q function; ▽ θπ Q(S,A|θ Q ) represents the gradient of the Q function with respect to the Actor network parameters θ π ; ▽ θπ π(S|θ π ) represents the gradient of the policy function with respect to its parameters θ π .

[0090] Compared with the prior art, the above technical solution of the present invention has the following beneficial technical effects:

[0091] The present invention designs a rehabilitation cloud platform system for cardiopulmonary function monitoring, which realizes precise monitoring, dynamic rehabilitation optimization and remote medical collaboration through multi-modal data fusion, time series synchronization optimization and adaptive reinforcement learning, improving the cardiopulmonary rehabilitation effect and safety:

[0092] (1) High-precision data processing: Combining wavelet threshold denoising, exponential moving average (EMA) and dynamic time warping (DTW), effectively eliminates high-frequency noise and low-frequency drift of physiological signals, improves the time series alignment accuracy of multi-modal data (such as electrocardiogram, blood oxygen, respiratory rate), and ensures the reliability of joint analysis;

[0093] (2) Comprehensive health assessment and early warning: Based on the fusion of a two-stream AI model (LSTM time-domain analysis + CNN frequency-domain features) and knowledge graph historical data, realizes multi-dimensional health status assessment; The improved LSTM prediction model enhances the time series prediction accuracy through multi-scale branches and differential attention mechanisms, and combines statistical methods (Z-score) and prediction deviation analysis to achieve double anomaly detection, significantly improving the timeliness of cardiovascular event early warning;

[0094] (3) Dynamic personalized rehabilitation optimization: Adopts multi-objective reinforcement learning (MORL) to balance health improvement and risk control, dynamically adjusts the training plan through Pareto front optimization and UCB+Epsilon-Decay mechanisms, combines real-time physiological data and patient historical information, and generates an adaptive rehabilitation strategy to improve the rehabilitation effect and safety;

[0095] (4) Remote collaboration and user interaction: The remote medical module supports doctors to monitor patient data in real time and remotely adjust the plan, and the user side provides visual monitoring data, personalized training plans and video guidance, enhancing the doctor-patient collaboration efficiency and patient compliance;

[0096] (5) Security and computational efficiency: Through the session channel authentication mechanism of hash function and exclusive OR operation, systematically solves the contradiction between security authentication and real-time communication in the medical Internet of Things scenario. BRIEF DESCRIPTION OF THE DRAWINGS

[0097] Figure 1 This is the system architecture diagram of a rehabilitation cloud platform system for cardiopulmonary function monitoring proposed by the present invention. Specific implementation manners

[0098] Example 1, as Figure 1 shown, a rehabilitation cloud platform system for cardiopulmonary function monitoring proposed by the present invention includes: an intelligent monitoring terminal, a cloud computing platform, a remote medical support module, and a user interaction terminal.

[0099] The intelligent monitoring terminal collects physiological data with wearable sensors, including but not limited to heart rate, blood oxygen, respiratory rate, electrocardiogram (ECG), and transmits the data to the cloud computing platform;

[0100] The cloud computing platform conducts health assessment, anomaly warning, and generates personalized rehabilitation plans;

[0101] The cloud computing platform includes: a data storage and management unit, a data processing unit, an AI analysis engine, and a personalized rehabilitation recommendation unit;

[0102] The data storage and management unit stores long-term monitoring data based on a big data storage architecture (including but not limited to InfluxDB);

[0103] The data processing unit preprocesses the collected physiological signals;

[0104] The AI analysis engine conducts cardiopulmonary health assessment, anomaly detection, and trend prediction based on a machine learning model;

[0105] The personalized rehabilitation recommendation unit generates personalized rehabilitation plans according to the user's health condition;

[0106] The remote medical support module supports doctors to monitor the patient's status in real time, give early warnings before abnormal situations occur, and reduce the risk of sudden cardiovascular events;

[0107] The user interaction terminal supports patients to view real-time monitoring data, rehabilitation suggestions, and health reports through the App, and view personalized training plans and video tutorials based on AI analysis.

[0108] Example 2, the working mode of a rehabilitation cloud platform system for cardiopulmonary function monitoring proposed by the present invention is specifically as follows:

[0109] S1. The intelligent monitoring terminal User of any patient i collects the patient's physiological signals (including but not limited to heart rate, blood oxygen saturation, respiratory rate, electrocardiogram (ECG)) through wearable devices (including but not limited to smart watches, electrocardiogram patches, respiratory sensors);

[0110] S2. The data processing unit preprocesses the physiological signals, specifically as follows:

[0111] S21. Establish a real-time session channel with the intelligent monitoring terminal User i . The process of establishing the real-time session channel is as follows:

[0112] S2101. The intelligent monitoring terminal User of any patient i i sends a session connection application. The data processing unit receives the session connection application and sends an application confirmation parameter P (random number) to the intelligent monitoring terminal User i ;

[0113] S2102. The intelligent monitoring terminal User i obtains its own device identifier ID i , calculates the secondary application parameter ASp = H L (ID i ||P);

[0114] where H() represents a predefined hash function that outputs a 256-bit hash value; H L is the left half of the predefined hash function H, i.e., 128 bits; || represents a concatenation operation;

[0115] Accordingly, the application packet {ID i , Asp} is sent to the data processing unit;

[0116] S2103. The data processing unit receives the application packet {ID i , Asp}, extracts the device identifier ID i , the secondary application parameter Asp, and searches for a value that matches the device identifier ID i in a traversing manner;

[0117] If the search is successful, the session confirmation parameter SCp = H R (ID i ⊕P) is calculated using the application confirmation parameter P;

[0118] where H R is the right half of the predefined hash function H, i.e., 128 bits; ⊕ represents an exclusive OR operation;

[0119] Accordingly, the session confirmation parameter SCp is sent to the intelligent monitoring terminal User of patient i i ;

[0120] S2104. The intelligent monitoring terminal User of patient i i receives the session confirmation parameter SCp and compares whether SCP = H R (ID i ⊕P) holds;

[0121] If it is established, confirm the establishment of the real-time session channel between the intelligent monitoring terminal User of patient i and the data processing unit. Accordingly, the intelligent monitoring terminal User of patient i i uploads the physiological signals to the cloud computing platform; i

[0122]

[0123] S22. Combine wavelet threshold denoising and exponential moving average method to adaptively remove high-frequency interference while retaining the physiological signal characteristics. Specifically:

[0124]

[0125] In the formula, S denoise represents the denoised signal; S raw represents the original physiological signal; DWT i () represents the i-th level wavelet decomposition of the signal; w i represents the denoising weight, which depends on the signal noise level;

[0126]

[0127] S2202. Based on exponential moving average (EMA) smoothing, remove low-frequency drift and enhance signal stability:

[0128] S EMA (t) = αS denoise (t) + (1 - α)S EMA (t - 1);

[0129] In the formula, S EMA (t) represents the smoothed signal at time t; S denoise (t) represents the denoised signal at time t; α represents the smoothing coefficient, α ∈ [0.1, 0.3];

[0130] S23. Align the signals of ECG, SpO2, BR, etc. in time to ensure the time sequence synchronization of multiple physiological parameters and improve the accuracy of joint analysis. Specifically: S2301. Dynamic time warping (DWT): Used for non-equal length signal matching to improve the signal time sequence alignment accuracy:

[0131]

[0132] In the formula, Q and C represent two groups of time series; d(Q i , C j ) represents the Euclidean distance between the i-th and j-th data points;

[0132] S2302. Timing Window Optimization: The sliding window method is used to optimize signal matching, improve the calculation efficiency, and output the aligned data:

[0133]

[0134] In the formula, W opt represents the optimal alignment window size; X t , Y t represent two groups of physiological signals; T represents the time step;

[0135] S24. Normalize and feature scale the data to ensure that the numerical scales of different signals are consistent and improve the calculation stability of the subsequent AI model. Specifically:

[0136] S2401. Perform normalization processing using Z-score normalization:

[0137]

[0138] In the formula, X represents the aligned data; μ and σ represent the mean and standard deviation;

[0139] S2402. Perform feature scaling processing using Min-Max scaling:

[0140]

[0141] In the formula, X scaled represents the scaled data; X min , X max represent the minimum and maximum values of the data;

[0142] S3. The AI analysis engine realizes high-precision health monitoring through multi-modal data fusion + two-stream AI evaluation + reinforcement learning rehabilitation optimization. The specific implementation process is as follows

[0143] S31. After data cleaning, the feature extraction for cardiopulmonary function is divided into three major parts: time domain, frequency domain, and non-linearity. Specifically:

[0144] Time domain feature (HRV analysis): HRV (heart rate variability) reflects the state of the autonomic nervous system. Calculate the root mean square difference of adjacent R-R intervals (RMSSD) to sensitively capture short-term variability:

[0145]

[0146] In the formula, RR i represents the i-th R-R interval (unit: millisecond), which is obtained by detecting the R wave in the ECG signal; N represents the total number of R-R intervals;

[0147] Frequency domain features (FFT analysis): By performing a fast Fourier transform (FFT) on the time-domain signal to convert it to the frequency domain, the main frequency components in the signal can be identified (including but not limited to the dominant heart rate frequency and respiratory rate): P(f) = |F(S(t))| 2 ;

[0148] Wherein, F(S(t)) represents the Fourier transform of the signal S(t); P(f) represents the power spectral density at frequency f;

[0149] Accordingly: The FFT algorithm quickly calculates the frequency domain features, and the power spectral density helps to determine the abnormal frequency bands in the physiological signal (including but not limited to arrhythmia, respiratory disorders);

[0150] Nonlinear features (sample entropy SampEn): Sample entropy reflects the complexity of the signal, reveals the nonlinear dynamic characteristics of cardiopulmonary regulation, and pre-determines the embedding dimension m and similarity threshold r:

[0151]

[0152] Wherein, m represents the embedding dimension, and in this embodiment, the value is taken as 2 or 3; r represents the similarity threshold, and in this embodiment, the value is 0.1 to 0.2 times the standard deviation of the signal; N' represents the signal length; A represents the number of similar sequence pairs of length m + 1; B represents the number of similar sequence pairs of length m;

[0153] It should be noted that the selection of SampEn parameters is based on statistical experience and literature reports, quantifies the chaotic characteristics of cardiopulmonary signals, and provides supplementary information that cannot be obtained by traditional time-domain and frequency-domain methods;

[0154] S32. Combine the time-domain and frequency-domain features, and introduce the patient's historical data to construct a two-stream health assessment model to achieve personalized health status scoring, specifically:

[0155] S3201. Construct a two-stream health assessment model, use the two-stream model to consider the time-domain and frequency-domain features simultaneously, and fuse the patient's historical information to construct the health status index Y health :

[0156] Y health = α·LSTM(X time ) + β·CNN(X freq ) + γ·KG(X history );

[0157] Wherein, LSTM(X time ) represents using the long short-term memory network (LSTM) to process the time-domain data feature X time , that is, RMSSD, to capture the time dynamic changes; CNN(X freq ) represents using the convolutional neural network (CNN) to process the frequency-domain feature Xfreq , namely P(f), identifies the local pattern of the signal; KG(X history ) Based on knowledge graph technology, integrates the historical information X of the patient history ;

[0158] S3202. Combine multiple physiological indicators with weights to obtain a comprehensive health risk index HRI:

[0159] HRI = w1·HRV + w2·SpO2 + w3·BR + w4·SampEn;

[0160] In the formula, HRV represents the heart rate variability index (RMSSD), reflecting the autonomic nerve function; SpO2 represents the blood oxygen saturation, reflecting the blood oxygen supply state; BR represents the respiratory rate, reflecting the working condition of the respiratory system; SampEn represents the sample entropy value, reflecting the complexity of the signal; w1, w2, w3, and w4 represent the weights of each indicator, obtained by fitting historical data and model training to ensure higher prediction accuracy after the indicators are combined;

[0161] Accordingly: Quantify multiple indicators into a single health risk value through HRI, facilitating the system to automatically identify potential health abnormalities;

[0162] Accordingly: During the entire health status assessment stage, Y health provides a global and dynamic health status indicator, while HRI quantifies individual physiological parameters as a specific risk score. The two complement each other and jointly describe the patient's current health level;

[0163] S33. Based on the health assessment, implement real-time anomaly detection through statistical and deep learning methods to ensure early warning and output the anomaly detection result C. Specifically:

[0164] Method 1: Anomaly detection based on statistical methods: Use Z-score to detect the deviation Z of the current data from the mean. When |Z| exceeds the set threshold τ1 (set to 2.5 in this embodiment), the system considers this data point to be abnormal;

[0165] Method 2: Combine an improved LSTM prediction model to predict physiological indicators within a future period of time, compare with the actual collected values, and calculate the prediction deviation to judge the anomaly situation, and calculate the prediction deviation score e t : e t = |X pred - X real |;

[0166] Among them, X pred represents the future physiological indicator value predicted by the LSTM model; X real represents the actual real-time collected physiological indicator value;

[0167] Accordingly: If the model predicts an anomaly, the system deems the current state to be at risk and triggers an anomaly warning.

[0168] Accordingly: Based on the health assessment, potential problems are further identified in advance through model prediction. The deviation between the prediction and the real-time data provides a dynamic supplement for anomaly detection, forming a dual guarantee with statistical methods.

[0169] S34. Transmit the health status indicator Y health , the comprehensive health risk index HRI, and the anomaly detection result C to the personalized rehabilitation recommendation unit.

[0170] S4. Once the personalized rehabilitation recommendation unit receives the anomaly detection result C, by comparing the prediction with the actual result, continuously adjust the strategy to enable the rehabilitation training plan to self-improve in a closed loop. Specifically:

[0171] Utilize a reinforcement learning model with a dual-objective optimization (health status optimization + risk control) and an adaptive exploration mechanism to dynamically adjust the rehabilitation training plan according to the patient's current state, achieve adaptive optimization, and update the rehabilitation strategy.

[0172] S5. The remote medical support module supports doctors to remotely access patient data, conduct online consultations and rehabilitation plan adjustments, and support remote rehabilitation guidance for patients and doctors.

[0173] S6. The user interaction terminal supports patients to view real-time monitoring data, rehabilitation suggestions, and health reports, and provides rehabilitation training guidance, including but not limited to providing personalized training plans and video tutorials.

[0174] S7. The data storage and management unit stores long-term monitoring data and the formulated rehabilitation strategies based on a big data storage architecture.

[0175] Embodiment 3. A rehabilitation cloud platform system for cardiopulmonary function monitoring proposed by the present invention, whose working mode further includes a reinforcement learning model with a dual-objective optimization (health status optimization + risk control) and an adaptive exploration mechanism. Specifically:

[0176] The core goal of reinforcement learning is to find the optimal policy π * , maximize the long-term health benefit, and at the same time minimize the health risk:

[0177] Define the state space S t : S t = {Y health , HRI, data};

[0178] Define the action space A t (rehabilitation training plan): A t = {a1, a2, …, ai ,…,a n};

[0179] Define the reward function R(S t , A t ), so as to take into account both health improvement and risk control:

[0180] R(S t , A t ) = w′1·ΔY health - w'2·ΔHRI + w'3·Δdata - w'4·C(A t );

[0181] Among them, Y health represents the health status index; HRI represents the comprehensive health risk index; data represents the real-time physiological data; a i represents any training method (including but not limited to low-intensity walking, medium-intensity cycling, high-intensity running); ΔY health represents the improvement value of the health status index, and the goal is to increase this value as much as possible; ΔHRI represents the change in health risk, and the goal is to reduce this value as much as possible; C(A t ) represents the training consumption cost (including but not limited to fatigue, cardiovascular load, training time), to avoid overtraining; w′1, w'2, w'3, w'4 represent the weight parameters to balance the health benefits and training risks; Δdata represents the sum of the change values of each index in the physiological data;

[0182] S2. In order to learn the optimal rehabilitation training strategy, the Deep Deterministic Policy Gradient (DDPG) is adopted to achieve efficient decision-making:

[0183] The Actor network, that is, the policy network: generates the optimal training plan: A t = π(S t |θ π );

[0184] The Critic network, that is, the value network: evaluates the value of the current policy:

[0185] Q(S t , A t ) = R(S t , A t ) + γ'Q(S t+1 , A t+1 );

[0186] Optimization goal, maximize the long-term health benefits: ▽ θπ J(π) = E[▽ θπ Q(S, A|θ Q )▽ θπ π(S|θ π)];

[0187] Among them, A t represents the best rehabilitation training action output by the reinforcement learning model, that is, the training plan that the current patient should execute; π(S t |θ π ) represents the output of the policy network (Actor network), and based on the state S t selects the optimal action A t ; θ π represents the parameters of the policy network, which are continuously optimized during the training process; Q(S t , A t ) represents the state-action value function (Q-value), which is used to evaluate the long-term return of taking action A t in the current state; γ' represents the discount factor, with a value range of [0, 1], which is used to balance immediate rewards and long-term rewards; J(π) represents the objective function of the policy π, that is, the long-term cumulative return expected by the system, and the goal is to maximize J(π) through parameter updates; ▽ θπ J(π) represents the gradient of the objective function J(π) with respect to the policy parameter θ π , which is used to guide the direction of parameter updates; E() represents the expectation operator, that is, averaging over the distributions of states and actions; Q(S, A|θ Q ) represents the Q function, which is defined by the Critic network parameter θ Q , and represents the estimated value of the future cumulative return after taking action A in state S; represents the gradient of the Q function with respect to the Actor network parameter θ π ; represents the gradient of the policy function (Actor network) with respect to its parameter θ π , that is, the sensitivity of generating actions in state S;

[0188] S3. Rehabilitation training needs to seek a balance between health improvement (Y health ) and risk control (HRI). By using the multi-objective reinforcement learning (MORL) method and optimizing with the Pareto front, the optimal rehabilitation plan is found:

[0189] maxF(π) = λ1·G1(π) - λ2·G2(π);

[0190]

[0191] Among them, F(π) represents a comprehensive evaluation index, which is used to measure the trade-off effect between health improvement and risk control achieved by the rehabilitation training program in the long term under the strategy π; λ1 represents the health improvement weight, which is used to measure and adjust the importance of the health improvement goal in the overall goal; λ2 represents the risk control weight, which is used to measure and adjust the importance of the health risk control goal in the overall goal; G1(π) represents the cumulative health benefit; represents the health status index of the patient at time t (the value ranges from 0 to 1, and the higher the value, the better the rehabilitation effect); T represents the number of training iteration steps; γ represents the discount factor; HRI t represents the comprehensive health risk index at time t (the higher the value, the greater the risk);

[0192] S4. Introduce the Upper Confidence Bound (UCB) + Epsilon-Decay mechanism to dynamically adjust the training strategy:

[0193]

[0194] In the formula, Q(S t , a) represents the expected cumulative reward obtained by selecting action a in the current state S t ; c represents the exploration factor; lnt represents the natural logarithm of the current time step t; N(a) represents the number of times action a is selected, that is, the total number of times action a is selected from the start of the algorithm to the current time step;

[0195] Accordingly, the Epsilon-Decay mechanism gradually reduces the random exploration probability, with more exploration in the early stage and more stability in the later stage;

[0196] Accordingly, an intelligent and adaptive rehabilitation training system is formed, which can formulate the optimal plan for different patients and dynamically adjust the training strategy according to the real-time health status.

[0197] Example 4. A rehabilitation cloud platform system for cardiopulmonary function monitoring proposed by the present invention also includes an improved LSTM prediction model in its working mode. The multi-scale learning mechanism and the differential attention mechanism are introduced to better capture the complex temporal features in the cardiopulmonary function data and improve the prediction accuracy and robustness. Specifically:

[0198] S1. Obtain the standardized time series data X, that is, physiological data;

[0199] S2. Construct multi-scale branches, and each scale branch corresponds to a different time window length. For each scale branch, an independent LSTM encoder is used for processing:

[0200] Among them, s represents the scale branch; τ sDenote the time window length of branch s; Denote the hidden state vector output by scale s at time step t, that is, the feature representation at this scale;

[0201] Accordingly: After independent training of each scale of LSTM, data representations at different time resolutions are obtained, which provides multi-dimensional information for the subsequent attention mechanism;

[0202] For example: Short-scale branch: Adopt a smaller time window (including but not limited to 5 - 10 time steps) to capture short-term dynamic changes;

[0203] Long-scale branch: Adopt a larger time window (including but not limited to 20 - 30 time steps) to capture long-term trend information;

[0204] It should be noted that the feature representations of different scales have their own emphases: The short scale is more sensitive to local fluctuations, instantaneous anomalies and mutation features; The long scale is more suitable for capturing overall trends and stability changes; Through the multi-scale mechanism, the model can consider data features at both the global and local levels simultaneously, ensuring that information will not lose details or global trends due to single-scale processing;

[0205] S3. After obtaining the hidden state vectors of each scale, introduce the attention mechanism to perform weighted fusion on them;

[0206] Initial attention weight calculation: Calculate the initial attention weight for the output of each scale:

[0207]

[0208] In the formula, f() represents a feed-forward neural network, which is used to convert the hidden state of each scale into a scalar importance score, which reflects the importance of this scale at time step t; Denote the initial attention weight of scale s at time step t. After normalization, the sum of all scale weights is 1, which reflects the importance degree of each scale feature; exp() represents the exponential function, which is used to map the score to the positive domain to facilitate subsequent normalization; M represents the total number of scales;

[0209] Differential attention calculation and adjustment:

[0210]

[0211] In the formula, Denote the change in attention weight between scale s at time step t and the previous moment t - 1, which captures the mutation of the attention distribution and reflects the feature change at the critical moment; λ” represents the differential adjustment coefficient, which is used to control the intensity of attention weight adjustment; Denotes the finally adjusted attention weights, which integrate the original weights and the information of their temporal variations;

[0212] Multi-scale feature fusion:

[0213] In the formula, H t Denotes the final feature vector after fusing the feature representations of all scales at time step t, which is the weighted final feature vector, integrating the important information of each scale and providing comprehensive data for subsequent predictions; Denotes the hidden state vector output by scale s at time step t;

[0214] S4. Prediction output and anomaly detection:

[0215] S41. Prediction output:

[0216] In the formula, g() represents a fully connected layer used to map the fused feature H t To the prediction value space, determining the output accuracy and non-linear fitting ability; Denotes the prediction output at the t-th moment, representing the estimated value of the cardiopulmonary function index (including but not limited to heart rate, blood oxygen, etc.) by the model;

[0217] S42. Prediction deviation calculation:

[0218] In the formula, y t Denotes the actual observed value at the t-th moment, representing the real cardiopulmonary function data; e t Denotes the prediction deviation, reflecting the gap between the model prediction and the actual situation;

[0219] S43. Statistical analysis and anomaly threshold setting:

[0220]

[0221] In the formula, N' represents the total number of samples used to calculate the mean and standard deviation, which is the number of prediction deviations within a certain time window; μ e Denotes the mean of the prediction deviation e t Reflecting the central position of the overall prediction deviation of the model; σ e Denotes the standard deviation of the prediction deviation, measuring the dispersion degree of the deviation distribution;

[0222] S5. Define the threshold τ2 = δσ e , when |e t - μ e |, it is considered that the prediction at the t-th moment is abnormal;

[0223] Among them, δ represents the threshold coefficient, used to control the sensitivity of anomaly detection, and the value in this embodiment is 2.

[0224] The embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings. However, the present invention is not limited thereto, and various changes can be made without departing from the spirit of the present invention within the scope of knowledge possessed by those skilled in the art.

Claims

1. A rehabilitation cloud platform system for cardiopulmonary function monitoring, characterized in that, Including: An intelligent monitoring terminal, which is used to collect physiological data by means of wearable sensors and transmit the data to the cloud computing platform; The cloud computing platform, which is used for health assessment, anomaly warning, and generating personalized rehabilitation plans; The cloud computing platform includes: A data storage and management unit, which is used to store long-term monitoring data based on a big data storage architecture; A data processing unit, which establishes a real-time session channel for data transmission, uses wavelet transform and exponential moving average to remove noise, combines dynamic time warping and sliding window for signal alignment, and finally unifies the data scale through mean standard deviation normalization and extreme value scaling; An AI analysis engine, which extracts time-domain, frequency-domain, and non-linear features of cardiopulmonary signals, constructs a two-stream health assessment model to obtain health indicators and risk indexes, and uses a statistical and improved long short-term memory network prediction model for real-time anomaly detection, and finally transmits the evaluation results to the personalized rehabilitation recommendation unit; A personalized rehabilitation recommendation unit, which is used to utilize a reinforcement learning model with a dual-objective optimization and adaptive exploration mechanism to dynamically adjust the rehabilitation training plan according to the patient's current state and update the rehabilitation strategy; A remote medical support module, which is used to support doctors to monitor the patient's state in real time and give early warnings before abnormal situations occur; A user interaction terminal, which is used to support patients to view real-time monitoring data, rehabilitation suggestions, and health reports, and view personalized training plans and video tutorials based on AI analysis.

2. The rehabilitation cloud platform system for cardiopulmonary function monitoring according to claim 1, characterized in that, The establishment process of the real-time session channel is as follows: S21. The intelligent monitoring terminal User of any patient i i sends a session connection application. The data processing unit receives the session connection application and sends an application confirmation parameter P to the intelligent monitoring terminal User i ; Wherein, P is a random number; S22, Intelligent Monitoring Terminal User i Obtain its own device identification ID i , calculate the secondary application parameter ASp = H L (ID i ||P), and accordingly send the application packet {ID i , Asp} to the data processing unit; Among them, H() represents a predefined hash function that outputs a 256-bit hash value; H L is the left half of the predefined hash function H, that is, 128 bits; || represents the concatenation operation; S23. The data processing unit receives the application packet {ID i , Asp}, extracts the device identifier ID i , and the secondary application parameter Asp, and searches for the value matching the device identifier ID i in the authorized devices in a traversing manner. If the search is successful, the session confirmation parameter is calculated using the application confirmation parameter P , and the session confirmation parameter SCp is sent to the intelligent monitoring terminal User of patient i accordingly i ; Among them, H R is the right half of the predefined hash function H, that is, 128 bits; represents the exclusive OR operation; S24. The intelligent monitoring terminal User of patient i i Receives the session confirmation parameter SCp and compares Whether it holds; if it holds, it is confirmed that the real-time session channel between the intelligent monitoring terminal User of patient i i And the data processing unit is established. Accordingly, the intelligent monitoring terminal User of patient i i Uploads the physiological signals to the data processing unit.

3. The rehabilitation cloud platform system for cardiopulmonary function monitoring according to claim 1, characterized in that, The process of removing noise by wavelet transform and exponential moving average is as follows: S31. Use discrete wavelet transform (DWT) to decompose the signal into different frequency bands, and apply threshold processing to the high-frequency noise part: Where S denoise represents the denoised signal; S raw represents the original physiological signal. DWT i () represents the i-th level wavelet decomposition of the signal; w i represents the denoising weight; S32. Based on exponential moving average (EMA) smoothing, remove low-frequency drift: S EMA S(t) = αS denoise (t) + (1 - α)S EMA (t - 1); where S EMA (t) represents the smoothed signal at time t; S denoise (t) represents the denoised signal at time t; α represents the smoothing coefficient, and α ∈ [0.1, 0.3]; S33. Align non-equal-length time series through Euclidean distance, combine the sliding window method to optimize signal matching and calculate the optimal window size, and perform time alignment on physiological signals.

4. The rehabilitation cloud platform system for cardiopulmonary function monitoring according to claim 3, characterized in that, The alignment process of time alignment for physiological signals is as follows: S41. Dynamic Time Warping DWT: where Q and C represent two sets of time series; d(Q i , C j ) represents the Euclidean distance between the i-th and j-th data points; S42. Optimize signal matching using the sliding window method: Where, W opt represents the optimal alignment window size; X t , Y t represent two sets of physiological signals; T represents the time step.

5. The rehabilitation cloud platform system for cardiopulmonary function monitoring according to claim 1, wherein The construction process of the two-stream health assessment model is as follows: S51. Extract cardiopulmonary function characteristics: Time-domain characteristics: Calculate the root mean square difference of adjacent R-R intervals (RMSSD); Frequency domain feature: The time-domain signal is transformed into the frequency domain through the Fast Fourier Transform (FFT): P(f) = |F(S(t))| 2 ; Non-linear characteristics: Predetermine the embedding dimension m and similarity threshold r, and calculate sample entropy (SampEn) where RR i represents the i-th R-R interval; N represents the total number of R-R intervals; F(S(t)) represents the Fourier transform of the signal S(t); P(f) represents the power spectral density at the frequency f; m represents the embedding dimension; r represents the similarity threshold; N' represents the signal length; A represents the number of similar sequence pairs of length m + 1; B represents the number of similar sequence pairs of length m; S52. Consider time-domain and frequency-domain features simultaneously, and fuse the patient's historical information to construct the health status indicator Y health : Y health = α·LSTM(X time ) + β·CNN(X freq ) + γ·KG(X history ); In the formula, LSTM(X time ) represents processing the time-domain data feature X using the long short-term memory network LSTM time , that is, RMSSD; CNN(X freq ) represents processing the frequency-domain feature X using the convolutional neural network CNN freq , that is, P(f); KG(X history ) is the historical information X of the patient integrated based on the knowledge graph technology history ; S53. Generate a comprehensive health risk index (HRI): HRI = w1·HRV + w2·SpO2 + w3·BR + w4·SampEn; In the formula, HRV represents the heart rate variability index RMSSD; SpO2 represents blood oxygen saturation; BR represents respiratory rate; SampEn represents the sample entropy value; w1, w2, w3, and w4 represent the weights of each index.

6. The rehabilitation cloud platform system for cardiopulmonary function monitoring according to claim 1, wherein The prediction process of the improved long short-term memory network is as follows: S61. Obtain standardized time series data X, that is, physiological data; S62. Construct multi-scale branches, and for each scale branch, use an independent LSTM encoder for processing: Among them, s represents the scale branch; τ s represents the time window length of branch s; represents the hidden state vector output by scale s at time step t; S63. Introduce an attention mechanism for weighted fusion: Preliminary attention weight calculation: Calculate the preliminary attention weight for the output of each scale: Differential attention calculation and adjustment: Multi-scale feature fusion: where f() represents a feedforward neural network; represents the preliminary attention weight of scale s at time step t; exp() represents the exponential function; M represents the total number of scales; represents the change in attention weight between scale s at time step t and the previous time step t - 1; λ” represents the differential adjustment coefficient; represents the finally adjusted attention weight; H t represents the feature representation that fuses all scales at time step t; Output the predicted value of the cardiopulmonary function index through the fully connected layer, calculate the prediction deviation, and set a dynamic threshold based on the mean and standard deviation for anomaly detection.

7. The rehabilitation cloud platform system for cardiopulmonary function monitoring according to claim 6, characterized in that, The process of anomaly detection is as follows: S71. Prediction output: where g() represents a fully connected layer for mapping the fused feature H t to the prediction value space; S72. Prediction deviation calculation: where y t represents the actual observed value at the t-th moment; S73. Statistical analysis and anomaly threshold setting: Where N' represents the total number of samples used to calculate the mean and standard deviation; μ e represents the mean of the prediction deviation e t ; σ e represents the standard deviation of the prediction deviation; Define the threshold τ2 = δσ e When |e t - μ e |, it is considered that the prediction at the t-th moment is abnormal.

8. The rehabilitation cloud platform system for cardiopulmonary function monitoring according to claim 5, characterized in that, The policy update strategy process of the reinforcement learning model with double-objective optimization and adaptive exploration mechanism is as follows: S81. Define the state space S t : S t = {Y health , HRI, data}; Define the action space A t , that is, the rehabilitation training plan: A t = {a1, a2, …, a i , …, a n}; Define the reward function R(S t , A t ): R(S t ,A t ) = w1'·ΔY health -w'2·ΔHRI + w'3·Δdata - w'4·C(A t ); Among them, Y health represents a health status indicator; HRI represents a comprehensive health risk index; data represents real-time physiological data; a i represents any one of the training methods; ΔY health represents the improvement value of the health status indicator; ΔHRI represents the change in health risk; C(A t ) represents the training consumption cost; w1', w'2, w'3, w'4 represent weight parameters; Δdata represents the sum of the change values of each indicator in the physiological data; S82. Adopt deep deterministic policy gradient for deep decision-making; S83. Adopt the multi-objective reinforcement learning method and use Pareto front optimization to find the optimal rehabilitation plan: maxF(π) = λ1·G1(π) - λ2·G2(π); Among them, F(π) represents a comprehensive evaluation index; λ1 represents the weight for health improvement; λ2 represents the weight for risk control; G1(π) represents the cumulative health benefit; represents the health status index of the patient at time t; T represents the number of training iteration steps; γ represents the discount factor; HRI t represents the comprehensive health risk index at time t; S84. Introduce the upper confidence bound + gradually decaying exploration mechanism to dynamically adjust the training strategy: where Q(S t , a) represents the expected cumulative reward obtained by selecting action a in the current state S t ; c represents the exploration factor; lnt represents the natural logarithm of the current time step t; N(a) represents the number of times action a is selected.

9. The rehabilitation cloud platform system for cardiopulmonary function monitoring according to claim 8, wherein, The deep decision-making process is as follows: The Actor network, i.e., the policy network: generates the optimal training plan: A t = π(S t |θ π ); The Critic network, that is, the value network: evaluates the value of the current strategy: Q(S t ,A t ) = R(S t ,A t ) + γ'Q(S t+1 ,A t+1 ); Optimization goal, maximize long-term health benefits: Among them, A t represents the best rehabilitation training action output by the reinforcement learning model, that is, the training plan that the current patient should execute; π(S t |θ π ) represents the output of the Actor network, and selects the optimal action A t based on the state S t ; θ π represents the parameters of the policy network; Q(S t , A t ) represents the state-action value function; γ' represents the discount factor; J(π) represents the objective function of the policy π; represents the gradient of the objective function J(π) with respect to the policy parameter θ π ; E() represents the expectation operator; Q(S, A|θ Q ) represents the Q function; represents the gradient of the Q function with respect to the Actor network parameter θ π ; represents the gradient of the policy function with respect to its parameter θ π .

Images