A heart failure patient chronic disease management system based on multi-modal data dynamic fusion and intelligent feedback

Abstract

The present application relates to the technical field of heart failure management, in particular to a heart failure patient chronic disease management system based on multi-modal data dynamic fusion and intelligent feedback, comprising a patient terminal, a physician platform and an intelligent server, the data reliability is improved through a multi-modal data credibility fusion module; the key internal physiological parameters such as effective circulating blood volume are non-invasively inferred by using a dynamic physiological state estimation engine based on particle filtering; the prospective early warning of heart failure decompensation is realized by combining a deep time series risk prediction module; and a layered intervention strategy is generated through a reinforcement learning intelligent decision module. At the same time, the system has adaptive knowledge tracking and intelligent exercise rehabilitation guidance capabilities, realizing the whole-process closed-loop intelligent management from risk assessment, intelligent decision to personalized education and rehabilitation, and significantly improving the accuracy and prospectiveness of heart failure chronic disease management.

Description

Technical Field

[0001] This invention relates to the field of heart failure management technology, specifically to a chronic disease management system for heart failure patients based on multimodal data dynamic fusion and intelligent feedback. Background Technology

[0002] Chronic heart failure is the most complex and severe end-stage of cardiovascular disease, with its core pathophysiology stemming from circulatory congestion and insufficient tissue perfusion caused by the decline in cardiac pumping function. This disease, with its high incidence, high readmission rate, and heavy socioeconomic burden, has become a global public health challenge. Traditional chronic disease management heavily relies on intermittent outpatient follow-ups. In this model, clinicians can only make decisions based on fragmented, subjective self-reports and static examination indicators from the past few weeks or months. This management approach inherently has blind spots and lags: key pathophysiological processes such as volume overload and neuroendocrine activation that occur silently at home cannot be effectively captured. Often, by the time typical symptoms such as significant weight gain and dyspnea appear, the patient is already on the verge of acute decompensation, missing the optimal intervention period and ultimately leading to the closure of the warning-intervention window, necessitating emergency hospitalization. This not only severely impairs the patient's quality of life and prognosis but also results in enormous consumption of medical resources.

[0003] With the rise of the Internet of Things (IoT) and mobile healthcare technologies, first-generation remote patient management solutions attempted to overcome this predicament. These systems typically use devices such as smart scales and wearable wristbands to continuously or daily monitor physiological parameters such as weight, heart rate, and blood pressure. However, these technological advancements still face significant limitations. First, most remain in a state of data silos, where each parameter is presented independently and triggers simple threshold alarms. This isolated interpretation is highly prone to misjudgment: short-term fluctuations in weight may be due to food intake rather than fluid retention, and heart rate variability needs to be assessed in conjunction with activity levels and symptoms. The systems lack the ability to deeply fuse and analyze multi-source, heterogeneous data, resulting in low alarm accuracy, alarm fatigue, and ultimately weakening the trust and willingness of healthcare professionals to respond.

[0004] A deeper technological bottleneck lies in the fact that existing systems generally lack the ability to infer the patient's internal dynamic physiological state from surface-observed data. The core of heart failure management is volume management, and key physiological variables such as effective circulating blood volume, cardiac preload, and myocardial contractility cannot be directly and non-invasively measured. Current systems can only report weight gain but cannot answer whether this signifies true volume overload, its severity, or its progression. This lack of physiological insight keeps management decisions at a reactive level, rather than based on a deep understanding of the underlying pathophysiological processes for proactive intervention.

[0005] Therefore, there is an urgent need in this field for a heart failure chronic disease management system that can fundamentally overcome the above-mentioned bottlenecks and achieve full-process, closed-loop intelligent management from risk assessment and intelligent decision-making to personalized education and rehabilitation guidance. Summary of the Invention

[0006] The purpose of this invention is to provide a chronic disease management system for heart failure patients based on multimodal data dynamic fusion and intelligent feedback. Through multimodal data fusion, dynamic physiological state estimation and deep temporal risk prediction, it can achieve prospective early warning and personalized intervention for heart failure decompensation, effectively improving management results.

[0007] To achieve the above-mentioned technical objectives and effects, the present invention is implemented through the following technical solution: A chronic disease management system for heart failure patients based on multimodal data dynamic fusion and intelligent feedback includes: The patient terminal is equipped with a data acquisition and interaction interface, which is used to collect daily morning weight measured according to clinical standards, patient self-reported symptom data collected through structured electronic forms, continuous vital sign data collected through wearable devices connected via Bluetooth, and video streams of the patient performing standard rehabilitation training procedures collected through a camera, wherein the standard rehabilitation training procedures include the Chinese traditional fitness qigong Baduanjin. The physician and management platform provides medical institutions with a web-based interactive interface for real-time monitoring of patients' physiological data and early warning information, and for receiving intervention suggestions pushed by the system for clinical decision confirmation. The system server connects to the patient terminal and the physician and management platform via an encrypted communication protocol. The server includes a processor and a memory. The memory stores a computer program. When the computer program is executed by the processor, it implements the following core functional modules: The multimodal data acquisition and credibility fusion module is used to receive raw input data from the patient terminal and preprocess and fuse these multi-source heterogeneous data. The preprocessing includes using a time-series alignment algorithm to process vital sign data with non-uniform sampling frequency to generate smooth and continuous signal estimates. At the same time, the module dynamically calculates credibility weights for each data source. The weights are based on the signal-to-noise ratio index, data missing rate index, and user historical compliance score of the data source, and are obtained by normalization after nonlinear mapping through a neural network. The dynamic physiological state estimation engine is used to estimate the internal physiological state of a patient that cannot be directly measured from surface observation data. The engine establishes a state-space model based on the differential equations of the cardiovascular system. The state variables include effective circulating blood volume, cardiac filling pressure, myocardial contractility, and peripheral vascular resistance. The state-space model consists of state equations and observation equations. The state equations are constructed based on Starling's heart law and hemodynamic principles, while the observation equations map the internal state to observable physiological parameters. The engine uses a particle filtering algorithm for state estimation. By initializing a set of random particles to represent possible physiological states, an iterative process of prediction-update-resample is performed at each time step. Finally, the optimal state estimate is obtained by weighted averaging of all particles, and the patient's individualized dry weight is dynamically calculated based on this estimate. The deep temporal risk prediction module is used to predict the future risk of heart failure decompensation based on historical physiological state sequences. This module adopts a deep survival analysis network architecture, which constructs risk prediction as a survival probability prediction problem that changes over time. The network combines a temporal feature extraction layer and a fully connected layer. The temporal feature extraction layer captures the temporal dependencies of physiological states, and the fully connected layer maps the extracted temporal features to a risk scalar. This module outputs the probability that a patient will experience a heart failure decompensation event within a specific time window in the future, which is defined as a comprehensive risk index. The reinforcement learning-based intelligent decision-making and hierarchical early warning module is used to transform risk assessment into optimal clinical intervention strategies. This module models patient management as a Markov decision process, where the state space includes a comprehensive risk index, its rate of change, estimated physiological state, knowledge assessment score, and patient context information, and the action space includes multi-level clinical intervention measures from no intervention to emergency contact. The reward function is set based on clinical outcomes. This module adopts the Actor-Critic reinforcement learning framework, where the Actor network generates the probability distribution of each action to be taken in a given state, the Critic network evaluates the long-term value of the state, and the decision strategy is optimized through a policy gradient algorithm. The adaptive knowledge state tracking and education module is used to dynamically assess and improve patients' disease management knowledge level. This module is based on a dynamic Bayesian knowledge tracking model, which models the patient's mastery status of each knowledge point as a latent variable. Through Bayesian inference, it recursively updates the mastery probability based on the patient's answer history and educational intervention history, and selects and accurately pushes educational content that can minimize knowledge uncertainty based on the information gain principle. The intelligent exercise rehabilitation analysis and guidance engine is used to guide rehabilitation training for heart failure patients, especially for the Baduanjin training program. The engine maintains a standard movement database built on a probability model, extracts multi-dimensional biomechanical features from patient movement videos, and obtains a comprehensive movement quality score by calculating the difference from the standard movement. When the score is lower than the threshold, it performs error root cause inference based on a predefined biomechanical causal graph model and generates personalized feedback that includes specific text guidance and visual comparison.

[0008] Furthermore, in the multimodal data acquisition and credibility fusion module, the time-series alignment algorithm adopts the Gaussian process regression algorithm, which models the time-series correlation of physiological signals by defining the mean function and the Matern kernel function; the neural network is a three-layer fully connected neural network, and is normalized by the Sigmoid function.

[0009] Furthermore, the state-space model in the dynamic physiological state estimation engine is specifically defined as follows: The state equation is in differential form: Where X(t) = [V blood (t), P fill (t), Contractility (t), R systemic (t)] T , where is the state vector, representing effective circulating blood volume, cardiac filling pressure, myocardial contractility and peripheral vascular resistance respectively, f is a nonlinear function based on cardiovascular physiology, integrating Starling cardiac mechanism and hemodynamic equation, u(t) is the control input vector including patient-reported daily fluid intake and medication records, and w(t) is Gaussian process noise with zero mean. The observation equation is: Y(t) = h(X(t)) + v(t), where Y(t) = [W(t), HR(t), BP]. sys (t), BP dia (t)] T is the observation vector, which includes body weight, heart rate, systolic blood pressure and diastolic blood pressure; h is the observation function that maps the internal state to the observed value through physiological relationships; v(t) is Gaussian observation noise with zero mean. The specific execution steps of the particle filter algorithm include: Initialization: Generate N particles X by sampling from the prior distribution. p (0), and assign an initial weight of 1 / N to each particle; Prediction step: For each particle, state propagation is performed according to the state equation f: X p (t) = X p (t-1) + f(X)p (t-1),u(t))·Δt+w p (t) Where w p (t) represents the random disturbance sampled from the process noise distribution; Update step: When new observation data Y(t) arrives, calculate the weight w for each particle. p (t) = P(Y(t) | X p (t), which is the observation likelihood probability; Resampling step: The system is resampled according to the particle weights, low-weight particles are eliminated, high-weight particles are copied, and a new set of equal-weight particles is generated; State estimation: The final state estimate is obtained by weighted averaging of all particles. Furthermore, the deep survival analysis network in the deep temporal risk prediction module is specifically configured as follows: The risk function is defined as: λ(t|X) = λ0(t)·exp(G(X(t),H(t))) Where λ0(t) is the nonparametric baseline risk function obtained by the Breslow estimator, and G(·) is a deep neural network whose input includes the current physiological state estimate X(t) and a historical state sequence of length L H(t) = [X(t-1), X(t-2), ..., X(tL)]; The architecture of the network G(·) includes: Input layer: Receives the concatenated current state and historical sequence; Long Short-Term Memory (LSTM) network layer: Composed of multiple hidden units, its update formula is: (W) f ·[h t-1 x t ] + b f (The Gate of Oblivion) (W) i ·[h t-1 x t ] + b i (Input Gate) (W) C ·[h t-1 x t ] + b C (Candidate Memory) (Memory update) (W)o ·[h t-1 x t ] + b o (Output Gate) h t = o t ·tanh(C t (Hidden state); Fully connected layer: This layer stores the final hidden state h of the Long Short-Term Memory (LSTM) network layer. t Map the output to a scalar output using at least one fully connected layer, and then use an activation function. Output layer: Linear output of risk score; The comprehensive risk index CRI(t) is calculated using a survival function: The integral is approximated using numerical methods.

[0010] Furthermore, the Markov decision process in the reinforcement learning intelligent decision-making and hierarchical early warning module is specifically defined as follows: State vector s t for: s t =[CRI(t),d(CRI) / dt,V blood (t), P fill (t), Contractility (t), R systemic (t), K score (t), Compliance Score (t)] Where d(CRI) / dt is the numerical derivative of the risk index, Compliance Score (t) represents the compliance score calculated based on historical behavior; Action space A includes multiple discrete intervention levels, including no action, pushing educational content, nurse telephone follow-up, adjusting diuretic dosage, escalation to physician treatment, and triggering emergency contact; Reward function R(s) t a t ) is defined as: If a patient experiences a decompensated event within the next 7 days and the system does not trigger an appropriate warning: R = -10; If the patient's condition is stable and the system does not perform unnecessary interventions: R = +1; If the system triggers intervention and the patient's condition improves: R = +3; If the system triggers unnecessary emergency intervention: R = -2; Other cases: R = 0; The specific implementation of the Actor-Critic framework is as follows: Critic Network (s) t ): Employs a fully connected network, with input being state s_t and output being a scalar estimate of the state value; Actor Network π θ (a) t | s t ): It adopts a fully connected network, with the input being the state s_t and the output being the action probability distribution normalized by the Softmax function; Policy updates use the policy gradient method with baselines: (a) t | s t )·( (s) t )) Where α is the learning rate, and R0 is the learning rate. t Accumulate rewards for discounts.

[0011] Furthermore, the specific implementation of Baduanjin training in the intelligent sports rehabilitation analysis and guidance engine includes: The process of constructing the standard Baduanjin (Eight Pieces of Brocade) movement database: Calibration videos of healthy individuals and patients with stable heart failure performing various Baduanjin movements were collected and grouped according to cardiac function classification; Three-dimensional coordinate sequences of multiple joint points were extracted from the videos, and joint angle sequences, center of gravity trajectories, and joint velocity profiles were calculated; For each movement M and each cardiac function classification G in Baduanjin, a Gaussian mixture model was used to model the distribution of feature vectors. K is determined by the Bayesian information criterion; Specific methods for assessing patient movement quality: Extracting the same biomechanical feature vector X from patient videos. patient ; Calculate the negative log-likelihood of patient characteristics and the standard Gaussian mixture model as the basis for measuring the difference; Comprehensive action quality score Q score It is a weighted fusion of multiple sub-indicators, including morphological similarity: DTW stands for Dynamic Time Warping Distance, used to assess morphological differences in joint angle sequences. Motion smoothness: Among them, the power spectral entropy is calculated for the joint velocity sequence, which is used to quantify the smoothness of motion; Equilibrium stability: The area of ​​the 95% confidence ellipse was calculated using principal component analysis to determine the dispersion range of the centroid trajectory. Final Q score = w1·S1+ w2·S2+ w3·S3, the weights are determined through expert evaluation; The causal reasoning mechanism for error root cause analysis: A biomechanical causal graph for Baduanjin (Eight Pieces of Brocade) is constructed, with nodes including joint range of motion, muscle strength, coordination, and movement performance. When low-quality movements are detected, probabilistic reasoning is performed on the causal graph to calculate the posterior probability of each potential root cause. P(Root) Cause i |Observation) ∝ P(Observation | Root Cause i )·P Prior (Root) Cause i ) An electronic device includes: one or more processors; a memory; and one or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the one or more processors, and the one or more applications are configured to perform all the method steps implemented by the above-described chronic disease management system for heart failure patients.

[0012] A computer-readable storage medium having stored thereon processor-executable instructions for causing the processor to perform all the method steps implemented by the above-described chronic disease management system for heart failure patients.

[0013] The beneficial effects of this invention are: This invention's system achieves non-invasive and accurate inference of key physiological parameters in heart failure patients through dynamic fusion of multimodal data and state estimation based on physiological models, fundamentally improving the scientific rigor and foresight of volume management. Traditional systems can only monitor surface physiological indicators and cannot reveal underlying pathophysiological processes. This invention, through a multimodal data acquisition and reliability fusion module, uses a Gaussian process regression algorithm to temporally align non-uniformly sampled vital sign data, generating smooth and continuous signal estimates. It also utilizes a neural network to dynamically calculate the reliability weights of each data source, effectively solving the problems of data silos and false alarms. More importantly, the dynamic physiological state estimation engine establishes a state-space model based on the differential equations of the cardiovascular system, incorporating Starling's law of the heart and hemodynamic principles into the state equations. Real-time state estimation is performed using a particle filtering algorithm: a set of particles is initialized to represent possible physiological states; a prediction-update-resample iteration is executed at each time step; the prediction step propagates particle states according to the state equations; the update step calculates particle weights based on observed data; and the resampling step eliminates low-weight particles. Finally, the optimal estimates of key states such as effective circulating blood volume and cardiac filling pressure are obtained through weighted averaging. The system of this invention can distinguish between fluid retention components and temporary changes in weight fluctuations, and dynamically calculate individualized dry weight, providing a quantitative basis for diuretic adjustment and overcoming the limitations of traditional methods that rely on static indicators and subjective symptoms.

[0014] This invention's system, through the synergistic effect of deep temporal risk prediction and reinforcement learning intelligent decision-making, achieves proactive early warning and tiered intervention optimization for heart failure decompensation events, significantly extending the clinical intervention window and improving the efficiency of medical resource utilization. The deep temporal risk prediction module employs a deep survival analysis network architecture, constructing risk prediction as a survival probability problem that changes over time. The LSTM layer captures long-term dependencies in physiological states through a forget-gate-input-memory update mechanism, while the fully connected layer maps temporal features to a risk scalar, outputting a comprehensive risk index as a probability estimate of decompensation events within a specific future time window. It can identify dangerous temporal patterns such as continuously rising cardiac filling pressure, providing early warnings even if surface indicators do not reach thresholds. Building upon this, the reinforcement learning intelligent decision-making and tiered early warning module models patient management as a Markov decision process. The state space encompasses multi-dimensional information such as the risk index and its rate of change, and physiological state estimates. The action space includes multi-level clinical intervention measures. An Actor-Critic framework is used to optimize decision-making strategies through a policy gradient algorithm: the Actor network generates action probability distributions, the Critic network evaluates state value, and the reward function is set based on clinical outcomes to balance safety and intervention efficiency. This allows the system to dynamically select the optimal intervention level based on the individual patient's condition. For example, it can prioritize nurse follow-up for patients with high compliance but increased risk, rather than emergency intervention, thus avoiding alarm fatigue and ensuring timely response.

[0015] This invention's system, through a closed-loop design of adaptive knowledge state tracking and intelligent movement rehabilitation analysis, achieves personalized and precise management of patient education and rehabilitation training, effectively improving patient compliance and training safety. The adaptive knowledge state tracking and education module, based on a dynamic Bayesian knowledge tracking model, models the mastery status of each knowledge point as a latent variable and updates the mastery probability in real time through recursive Bayesian filtering: the prediction step infers the state evolution based on prior knowledge and educational history, the update step corrects the probability distribution based on answer data, and selects educational content based on the principle of maximizing information gain, ensuring that each intervention minimizes knowledge uncertainty. Simultaneously, the intelligent movement rehabilitation analysis and guidance engine, for training programs such as Baduanjin, constructs a standard movement database based on a Gaussian mixture model, extracts multi-dimensional biomechanical features from patient videos, and obtains a movement quality score by calculating a weighted fusion of morphological similarity, movement fluency, and balance stability. When the score is below a threshold, probabilistic inference is performed based on a biomechanical causal graph model to identify the root cause, and personalized feedback combining textual guidance and visual comparison is generated. This not only ensures that the educational content is precisely matched to the patient's knowledge gaps, but also guarantees the safety and effectiveness of rehabilitation training through real-time motion analysis and correction, forming a comprehensive management closed loop from cognition to behavior.

[0016] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Detailed Implementation

[0017] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example 1

[0018] The chronic disease management system for heart failure patients based on multimodal data dynamic fusion and intelligent feedback described in this embodiment includes: The patient terminal is equipped with a data acquisition and interaction interface. The data acquisition includes daily morning weight measured according to clinical standards, patient self-reported symptom data collected through structured electronic forms, continuous vital sign data collected through wearable devices connected via Bluetooth, and video streams of the patient performing standard rehabilitation training procedures collected through a camera. The standard rehabilitation training procedures specifically include the Chinese traditional fitness qigong Baduanjin. The physician and management platform provides medical institutions with a web-based interactive interface for real-time monitoring and management of patient population physiological data and early warning information, receiving intervention suggestions pushed by the system and confirming clinical decisions. The system server connects to the patient terminal and the physician and management platform via an encrypted communication protocol. The server includes a high-performance processor and a large-capacity memory. The memory stores computer programs, which, when executed by the processor, enable the coordinated operation of the following core functional modules: The multimodal data acquisition and credibility fusion module receives raw input data from patient terminals and preprocesses and fuses this multi-source heterogeneous data. The preprocessing includes using a Gaussian process regression algorithm to perform temporal alignment on vital sign data with non-uniform sampling frequencies. The Gaussian process regression model the temporal correlation of physiological signals by defining a mean function and a Matern kernel function, thereby generating smooth and continuous signal estimates at a unified time point. At the same time, this module dynamically calculates credibility weights for each data source. The weights are based on the signal-to-noise ratio, data missing rate, and user historical compliance score of the data source. They are obtained by nonlinear mapping through a three-layer fully connected neural network and then normalized by the Sigmoid function, so that the contribution of low-quality data in subsequent analysis is automatically reduced. The dynamic physiological state estimation engine, as the core reasoning mechanism of the system, is used to estimate the internal physiological state of the patient that cannot be directly measured from surface observation data. The engine establishes a state-space model based on the differential equations of the cardiovascular system. The state variables include four key physiological parameters: effective circulating blood volume, cardiac filling pressure, myocardial contractility, and peripheral vascular resistance. The state-space model consists of state equations and observation equations. The state equations are constructed based on Starling's heart law and hemodynamic principles, describing the evolution of state variables over time. The observation equations map the internal state to observable physiological parameters such as weight, heart rate, and blood pressure. The engine uses a particle filtering algorithm for state estimation. It initializes a set of random particles to represent possible physiological states, and then performs an iterative process of prediction-update-resampling at each time step. Finally, it obtains the optimal state estimate by weighted averaging of all particles, and dynamically calculates the patient's individualized dry weight based on this estimate. The deep temporal risk prediction module is used to predict the future risk of heart failure decompensation based on historical physiological state sequences. This module adopts a deep survival analysis network architecture, which constructs risk prediction as a survival probability prediction problem that changes over time. The network combines a long short-term memory network layer and a fully connected layer. The long short-term memory network layer is responsible for capturing the temporal dependencies of physiological states, and the fully connected layer is responsible for mapping the extracted temporal features to a risk scalar. The output of this module is the probability that a patient will experience a heart failure decompensation event within a specific time window in the future. This probability value is defined as a comprehensive risk index. The reinforcement learning-based intelligent decision-making and hierarchical early warning module is used to transform risk assessment into optimal clinical intervention strategies. This module models patient management as a Markov decision process, where the state space includes a comprehensive risk index, its rate of change, estimated physiological state, knowledge assessment score, and patient context information, and the action space includes multi-level clinical intervention measures from no intervention to emergency contact. The reward function is set based on clinical outcomes. This module adopts the Actor-Critic reinforcement learning framework, where the Actor network is responsible for generating the probability distribution of each action to be taken in a given state, and the Critic network is responsible for evaluating the long-term value of the state. The decision-making strategy is continuously optimized through the policy gradient algorithm to achieve personalized intervention. The adaptive knowledge state tracking and education module is used to dynamically assess and improve patients' disease management knowledge level. This module is based on a dynamic Bayesian knowledge tracking model, which models the patient's mastery status of each knowledge point as a latent variable. Through Bayesian inference, it recursively updates the mastery probability based on the patient's answer history and educational intervention history, and selects and accurately pushes educational content that can minimize knowledge uncertainty based on the information gain principle. The intelligent exercise rehabilitation analysis and guidance engine is specifically designed for rehabilitation training guidance for heart failure patients, with a particular focus on the Baduanjin (Eight Pieces of Brocade) training program. This engine maintains a standard movement database built on a Gaussian mixture model, extracts multi-dimensional biomechanical features such as joint angle sequences, center of gravity trajectory, and movement velocity profiles from patient movement videos, calculates the difference between the movement and the standard movement to obtain a comprehensive movement quality score, and when the score is below a threshold, it performs error root cause inference based on a predefined biomechanical causal graph model to generate personalized feedback that includes specific text guidance and visual comparison.

[0019] In this embodiment, the state-space model in the dynamic physiological state estimation engine is specifically defined as: The state equation is in differential form: Where X(t) = [V blood (t), P fill (t), Contractility (t), R systemic (t)] T , where is the state vector, representing the four core physiological parameters: effective circulating blood volume, cardiac filling pressure, myocardial contractility, and peripheral vascular resistance, respectively; f is a nonlinear function based on cardiovascular physiology, specifically integrating the Starling cardiac mechanism to describe the relationship between preload and cardiac output and the hemodynamic equation to describe the principle of blood pressure formation; u(t) is the control input vector, including the patient's self-reported daily fluid intake and medication records; w(t) is Gaussian process noise with zero mean, used to characterize model uncertainty. The observation equation is: Y(t) = h(X(t)) + v(t), where Y(t) = [W(t), HR(t), BP]. sys (t), BP dia (t)] T is the observation vector, which includes body weight, heart rate, systolic blood pressure and diastolic blood pressure; h is the observation function, which maps the internal state to the observed value through physiological relationships; v(t) is Gaussian observation noise with zero mean. The specific execution steps of the particle filter algorithm include: Initialization: Generate N particles X by sampling from the prior distribution. p (0), and assign an initial weight of 1 / N to each particle; Prediction step: For each particle, state propagation is performed according to the state equation f: X p (t) = X p (t-1) + f(X) p (t-1),u(t))·Δt+w p (t) Where w p (t) represents the random disturbance sampled from the process noise distribution; Update step: When new observation data Y(t) arrives, calculate the weight w for each particle. p (t) = P(Y(t) | X p (t), which is the observation likelihood probability, calculated using the observation equation and the observation noise distribution; Resampling step: The system is resampled according to the particle weights, low-weight particles are eliminated, high-weight particles are copied, and a new set of equal-weight particles is generated; State estimation: The final state estimate is obtained by weighted averaging of all particles. It should be understood that the nonlinear function f in the state equation is constructed based on the classical physiological principles of the cardiovascular system. f integrates Starling's law of the heart (describing the proportional relationship between cardiac preload and cardiac output) and hemodynamic equations (such as Poiseuille's law describing the relationship between vascular resistance and blood pressure). For example, the effective circulating blood volume V blood The change in (t) is affected by fluid intake and diuretic use, and its differential equation can be expressed as: Where u fluid (t) represents the patient's self-reported daily fluid intake, and k1 and k2 are individualized parameters calibrated using clinical data. This modeling approach ensures consistency between state variables and actual physiological processes, improving the clinical interpretability of state estimates.

[0020] Particle filtering was chosen as the core algorithm because it can handle non-Gaussian noise and nonlinear systems, making it suitable for the variability and uncertainty of the physiological state of heart failure patients. In the initialization step, the prior distribution is set based on historical patient data or population statistics (e.g., the prior distribution of effective circulating blood volume can be set as a normal distribution with a mean of 5L and a standard deviation of 0.5L). The process noise w(t) in the prediction step adopts a Gaussian distribution, and its covariance matrix is ​​learned from the training data through maximum likelihood estimation to capture model errors. The observation likelihood probability P(Y(t)|X) in the update step... p (t) is calculated based on a multidimensional Gaussian distribution; for example, the observed noise variance of body weight can be set to 0.1 kg. 2 This reflects measurement errors. The resampling step employs a systematic resampling method, avoiding particle degradation and ensuring the stability of the state estimate. The final state estimate X est (t) can be used not only for dry weight calculation, but also as input for other modules (such as risk prediction) to form a system-level linkage.

[0021] Based on the state estimate, the individualized dry weight is calculated using the following formula: Where V blood dry The effective circulating blood volume under dry weight (obtained through regression analysis of historical stable period data) is denoted by k, which is the conversion factor (typically 0.1-0.3 L / kg). This dynamic calculation overcomes the limitations of traditional static dry weight methods, reflecting the patient's fluid retention status in real time and providing a precise basis for diuretic adjustment.

[0022] In this embodiment, the deep survival analysis network in the deep temporal risk prediction module is specifically configured as follows: The risk function of the network is defined as: λ(t|X) = λ0(t)·exp(G(X(t),H(t))) Where λ0(t) is the nonparametric baseline risk function obtained by the Breslow estimator, and G(·) is a deep neural network whose input includes the current physiological state estimate X(t) and a historical state sequence of length L H(t) = [X(t-1), X(t-2), ..., X(tL)]; The specific architecture of the network G(·) is as follows: Input layer: Receives the concatenated current state and historical sequence; LSTM layer: Consists of 64 hidden units, used to capture temporal patterns of the state, and its update formula is: (W) f·[h t-1 x t ] + b f (The Gate of Oblivion) (W) i ·[h t-1 x t ] + b i (Input Gate) (W) C ·[h t-1 x t ] + b C (Candidate Memory) (Memory update) (W) o ·[h t-1 x t ] + b o (Output Gate) h t = o t ·tanh(C t (Hidden state); Fully connected layer: This layer stores the final hidden state h of the LSTM. t The output is mapped to a scalar output through two fully connected layers, using the ReLU activation function. Output layer: Linear output of risk score; The comprehensive risk index CRI(t) is calculated using a survival function: The integral is approximated using numerical methods.

[0023] It should be understood that this module constructs risk prediction as a survival analysis problem because it can handle right-truncated data (such as patient loss to follow-up) and time-varying covariates, which aligns with the practical scenarios of heart failure management. The baseline risk function λ0(t) in the risk function λ(t|X) is non-parametrically estimated from historical heart failure patient cohort data using a Breslow estimator, reflecting the population's baseline risk. The input to the neural network G(·) includes the current state estimate X(t) and the historical sequence H(t), where the historical sequence length L is set to 30 days based on clinical expert recommendations to balance short-term fluctuations and long-term trends.

[0024] The LSTM layer employs 64 hidden units, a number determined through grid search optimization to achieve a balance between computational efficiency and model capacity. The LSTM's forget gate mechanism allows the model to automatically ignore irrelevant historical information (such as temporary measurement errors), while the input and output gates ensure that long-term dependencies (such as the cumulative effects of fluid retention) are effectively captured. Network training uses a weighted cross-entropy loss function with L2 regularization (weight decay coefficient of 0.001) to prevent overfitting. Training data comes from a multicenter heart failure patient dataset, including at least 12 months of follow-up records, ensuring the model's generalization ability.

[0025] The integration window for CRI calculation is set to 7 days, corresponding to the typical warning period for heart failure decompensation. Numerical integration uses the trapezoidal rule, discretizing in hourly increments. The CRI output is a probability value between 0 and 1; in clinical applications, a threshold (e.g., 0.3) can be set to trigger an alert. For example, if CRI > 0.3, the system determines that the patient has a high risk of developing a decompensated event (such as acute pulmonary edema) within the next 7 days, requiring the initiation of intervention procedures. This index is significantly correlated with NYHA functional class, achieving an AUC of over 0.85 on the validation set.

[0026] In this embodiment, the Markov decision process in the reinforcement learning intelligent decision-making and hierarchical early warning module is specifically defined as: State vector s t for: s t =[CRI(t),d(CRI) / dt,V blood (t), P fill (t), Contractility (t), R systemic (t), K score (t), Compliance Score (t)] Where d(CRI) / dt is the numerical derivative of the risk index, Compliance Score (t) represents the compliance score calculated based on historical behavior; Action space A contains six discrete intervention levels: A = (no action, push educational content, nurse follow-up call, adjust diuretic dosage, escalate to physician treatment, trigger emergency contact). Reward function R(s) t a t ) is defined as: If a patient experiences a decompensated event within the next 7 days and the system does not trigger an appropriate warning: R = -10; If the patient's condition is stable and the system does not perform unnecessary interventions: R = +1; If the system triggers intervention and the patient's condition improves: R = +3; If the system triggers unnecessary emergency intervention: R = -2; Other cases: R = 0; The specific implementation of the Actor-Critic framework is as follows: Critic Network (s) t ): A three-layer fully connected network is used, with the input being state s. t The output is a scalar estimate of the state value; Actor Network π θ (a) t | s t ): A three-layer fully connected network is used, with the input being state s. t The output is an action probability distribution normalized by the Softmax function; Policy updates use the policy gradient method with baselines: (a) t | s t )·( (s) t )) Where α is the learning rate, and R0 is the learning rate. t Accumulate rewards for discounts.

[0027] It should be understood that the state vector s t Each component in the dataset has undergone feature importance analysis. For example, the risk index change rate d(CRI) / dt is calculated using the central difference method, reflecting the steepness or slowness of the risk trend; the compliance score... Score (t) is calculated based on a weighted average of the patient's data reporting rate, medication adherence, and exercise completion rate over the past 7 days (with weights of 0.4, 0.4, and 0.2, respectively). This design ensures that status information comprehensively covers physiological, behavioral, and risk dimensions.

[0028] The six intervention levels correspond one-to-one with the recommended measures in heart failure management guidelines. For example, adjusting diuretic dosage will generate specific recommendations (such as increasing furosemide by 20 mg), which require physician confirmation before implementation; triggering emergency contact will automatically send an alert to the emergency platform along with the patient's location information. The reward function's values ​​are set based on the Delphi method used by clinical experts, balancing safety (avoiding missed diagnoses) and efficiency (reducing unnecessary interventions). For example, a reward of -10 corresponds to a serious outcome caused by a missed diagnosis, while a reward of +3 encourages timely and effective intervention.

[0029] Training utilizes an offline reinforcement learning strategy, employing batch learning based on historical decision-making data (including physician intervention records and patient outcomes). Critic Network (s) tThe mean squared error loss function is used, and the optimizer is Adam (learning rate 0.001); the Actor network π θ (a) t | s t The policy gradient update uses a baseline-based variance reduction technique to accelerate convergence. During deployment, the system adopts... Greedy strategy ( A balance between exploration and utilization is maintained, ensuring that strategies are adjusted individually within safe boundaries. The model is updated quarterly to incorporate newly generated clinical data.

[0030] In this embodiment, the specific mathematical description of the dynamic Bayesian knowledge tracing model in the adaptive knowledge state tracking and education module is as follows: For each knowledge point i, define a binary latent variable K. t i ∈ (0, 1) represents the patient's mastery status at time t; Model parameters include: Prior knowledge probability: P(K0) i = 1) Learning probability: P(K) t i = 1 | K i t-1 = 0, a t-1 = e) represents the probability that a student who has not yet mastered the knowledge will acquire it after receiving education e; Probability of forgetting: P(K) t i = 0 | K i t-1 = 1, a t-1 = e) represents the probability that students forget what they have learned after receiving education; Guess the probability: P(O) t i = 1 | K t i = 0) indicates that the probability of the student guessing the question correctly is not known; Error probability: P(O) t i = 0 | K t i = 1) indicates the probability that a student answers a question incorrectly; The recursive Bayesian filtering process for knowledge state updates is as follows: Prediction step: Update steps: The strategy for selecting educational content is as follows: That is, choose the action that minimizes the expected entropy.

[0031] It should be understood that this model is based on the knowledge tracing theory in cognitive science, modeling the mastery status of each knowledge point (such as "the importance of a low-salt diet") as a latent variable, which aligns with the uncertainty in the patient's learning process. The prior mastery probability P(K0) i = 1) Initialization is based on patient demographic characteristics (such as education level) and baseline knowledge assessment scores. Learning and forgetting probabilities are estimated using historical educational intervention data; for example, for heart failure dietary knowledge points, the learning probability is typically 0.3–0.6, and the forgetting probability is 0.05–0.1.

[0032] The recursive process of the prediction and update steps allows the system to update its knowledge state in real time. For example, if the patient correctly answers a question about diuretic use (O... t i = 1), but the current mastery probability is low, so the system will reduce the influence of the guess probability and prioritize assuming that the patient's actual mastery level has improved. Maximizing information gain in the educational content selection strategy is equivalent to selecting content that can most effectively reduce knowledge state entropy, ensuring educational efficiency. In practice, the system maintains a knowledge graph containing 50-100 heart failure-related knowledge points, and each knowledge point has multiple versions of educational materials (such as text and video).

[0033] Based on the calculated optimal action a t * The system retrieves matching educational materials from its content library. For example, if the system detects that a patient lacks sufficient understanding of the importance of weight monitoring, it will push illustrated explanatory documents and videos of successful patient cases. The timing of these pushes is tailored to the patient's behavioral patterns (such as a preference for studying in the evening) to avoid disruption. Effectiveness evaluations show that this module improved patients' knowledge test scores by an average of 25% and also improved adherence.

[0034] In this embodiment, the specific implementation of Baduanjin training in the intelligent sports rehabilitation analysis and guidance engine includes: The process of building the standard Baduanjin movement database: Collect calibration videos of various forms of Baduanjin exercises from healthy individuals and patients with stable heart failure, and group them according to cardiac function classification. Extract the three-dimensional coordinate sequence of 25 joint points from the video, and calculate the joint angle sequence, center of gravity trajectory and joint velocity profile; For each movement M and each cardiac function grade G in the Eight Pieces of Brocade, a Gaussian mixture model is used to model the distribution of the feature vectors: K is determined by the Bayesian information criterion; Specific methods for assessing patient movement quality: Extract the same biomechanical feature vector X from patient videos. patient ; The negative log-likelihood of patient characteristics with the standard GMM model is used as the basis for measuring the difference. Overall motion quality score Q score It is a weighted fusion of three sub-indicators: Morphological similarity: DTW stands for Dynamic Time Warping Distance, used to assess morphological differences in joint angle sequences. Motion smoothness: Among them, the power spectral entropy is calculated for the joint velocity sequence, which is used to quantify the smoothness of motion; Equilibrium stability: The area of ​​the 95% confidence ellipse was calculated using principal component analysis to determine the dispersion range of the centroid trajectory. Final Q score = w1·S1+ w2·S2+ w3·S3, the weights are determined through expert evaluation; The causal reasoning mechanism of error root cause analysis: Construct a biomechanical cause-and-effect diagram of Baduanjin (Eight Pieces of Brocade), with nodes including physical functions and movement performance such as joint range of motion, muscle strength, and coordination; When a low-quality action is detected, probabilistic inference is performed on the causal graph to calculate the posterior probability of each potential root cause: P(Root) Cause i |Observation) ∝ P(Observation | Root Cause i )·P Prior (Root) Cause i ) The root cause with the highest posterior probability is selected as the basis for feedback generation.

[0035] It should be understood that the calibration video capture uses a standard RGB camera (such as Kinect V2) and is recorded in a controlled environment. The 3D coordinates of the joints are extracted using the OpenPose algorithm with an accuracy error of <2cm. The number of components K in the Gaussian Mixture Model (GMM) is automatically selected using the Bayesian Information Criterion (BIC), typically K=3-5, to capture normal variations in the movement. For example, for the "Hands Supporting the Sky and Regulating the Three Jiaos" pose, the GMM can cluster three patterns: standard movement, arms not fully extended, and center of gravity shift, providing a benchmark for quality assessment.

[0036] Morphological similarity S1 is calculated using Dynamic Time Warping (DTW) to determine the differences in joint angle sequences. The DTW window size is set to 10% of the sequence length to allow for slight temporal shifts. Motion smoothness S2 is calculated using the spectral entropy based on the power spectrum of the joint velocity signal, with a normalization factor H. max The values ​​are taken from reference values ​​for healthy individuals. The 95% confidence ellipse area for balance stability S3 is calculated from the centroid trajectory point cloud using principal component analysis; a larger area indicates poorer balance control. Weights w1, w2, and w3 are set based on expert surveys (typical values ​​are 0.5, 0.3, and 0.2), emphasizing morphological accuracy.

[0037] The biomechanical causal graph is constructed based on expert knowledge, with nodes including body function (such as hip joint range of motion and core muscle strength) and movement performance (such as arm elevation angle and trunk stability). The conditional probability table (CPT) is learned from historical rehabilitation data. For example, if insufficient arm elevation is observed, causal inference may calculate posterior probabilities: insufficient shoulder joint flexibility (60%), insufficient muscle strength (30%), and poor coordination (10%). The system then generates a root cause report to guide targeted interventions.

[0038] In this embodiment, the personalized feedback generation of the intelligent sports rehabilitation analysis and guidance engine specifically includes: The process of generating text instructions: Based on the identified root causes, retrieve the corresponding explanation module and correction suggestion module from the structured knowledge base; Template filling technology is used to combine specific joint names, deviation directions, and correction methods into natural language sentences; The process of generating visual feedback: Augmented reality technology is used to overlay a semi-transparent skeleton outline of a standard movement onto a video of a patient performing an action. The degree of joint deviation is displayed by color coding: green indicates compliance with standards, yellow indicates slight deviation, and red indicates severe deviation. An arrow is displayed at the joint with the largest deviation to indicate the correct direction of movement; Adaptive mechanism of exercise prescription: The system maintains difficulty-level versions of Baduanjin, including the standard standing posture version, the simplified standing posture version, and the seated posture version; If the patient's average Q from 5 consecutive training sessions score If the score is below 0.6, the system will automatically downgrade to the next difficulty level. If the patient undergoes Q training three consecutive times score All were above 0.8, and the system asked the patient if they would like to try a higher level.

[0039] It should be understood that the knowledge base is stored in XML format, with each root cause corresponding to multiple corrective suggestion modules. For example, for insufficient shoulder joint flexibility, the knowledge base includes modules such as warm-up exercises, assistive movements, and mental guidance. Template completion uses rule-based natural language generation (NLG) technology to ensure that the sentences are fluent and appropriate for the patient's reading level (e.g., using short sentences and active voice). Feedback language is localized; for example, the Chinese version uses traditional health-preserving terms such as mental guidance to enhance cultural affinity.

[0040] AR overlay utilizes the Unity engine and ARCore framework for real-time rendering on the patient's mobile device. The standard motion skeleton outline is extracted from pre-recorded expert videos, with color-coded deviation levels: green (deviation <10%), yellow (10-30%), and red (>30%). Arrow directions are generated using inverse kinematics calculations, indicating the direction and amplitude of joint adjustment. For example, in the "shaking head and wagging tail to clear heart fire" motion, if insufficient hip joint rotation is detected, a green arrow will be displayed at the hip to indicate the direction of rotation.

[0041] The difficulty rating is based on the NYHA functional classification (NYHA class I-IV), with the seated version suitable for class III-IV patients. The downgrade threshold (Q) score The sensitivity was determined to be <0.6% through ROC analysis, with a sensitivity of 85%. Before upgrades, the system assesses the patient's physiological status (e.g., heart rate variability) to ensure safety. Historical data indicates that this mechanism reduces training-related adverse events by 40% while maintaining rehabilitation outcomes. Example 2

[0042] The chronic disease management system for heart failure patients based on multimodal data dynamic fusion and intelligent feedback, as described in this embodiment, adopts a cloud-terminal collaborative architecture. The system server is deployed on a cloud server with GPU computing capabilities, using Kubernetes containerization. The patient terminal is a smartphone with a customized app installed, used in conjunction with a Bluetooth-connected smart scale, smart bracelet (for measuring heart rate and blood pressure), and the phone's built-in camera. Physicians and the management platform can access them via computer workstations or tablets within the medical institution.

[0043] Data Acquisition and Preprocessing Taking Mr. Wang, a stable heart failure patient with NYHA functional class II, as an example, he uses a smart scale to weigh himself every morning after using the toilet and before breakfast. The data is automatically synced to the app via Bluetooth. Simultaneously, the app sends a simplified symptom questionnaire, where Mr. Wang must check whether he has symptoms such as shortness of breath or lower extremity edema, and their severity (none, mild, moderate, severe). His smart bracelet continuously monitors his resting heart rate and blood pressure, sampling every 30 minutes.

[0044] Data fusion example: One day, the system discovered that Mr. Wang's weight data was missing, but his heart rate data from his wristband was normal. The multimodal data acquisition and fusion module initiated Gaussian process regression. Based on the weight trend of the previous two days and the current heart rate pattern, it estimated the most likely weight for that day to be 67.3 kg and assigned a low confidence weight to this estimate. At the same time, the system records this omission, and if it occurs repeatedly, it will lower the patient's compliance score.

[0045] Dynamic physiological state estimation The system server's dynamic physiological state estimation engine runs continuously. Taking Mr. Wang as an example, his initial clinical dry weight W... baseline The weight is 65.0 kg. The system initializes the particle filter, generating 1000 particles, representing the current physiological state of Mr. Wang (V). blood P fill Contractility, R systemic Possible guesses.

[0046] State estimation process: For a week, Mr. Wang's weight remained stable between 65.2-65.5 kg, and his symptoms were mild. In the update step of the particle filter, those guessed blood volume (V...) blood Moderate filling pressure (P) fill Stable particles have weights that continuously increase due to their high match with observed data (body weight, heart rate, asymptomatic status). After resampling, the particle swarm converges, and the system outputs an optimal state estimate: dynamic dry weight W. dry (t) = 65.3 kg, and the estimated value of myocardial contractility remained stable within the normal range.

[0047] Volume overload monitoring: Over the next two days, Mr. Wang's weight increased to 66.0 kg and 66.8 kg respectively, and he reported mild ankle edema. In the observation equation, the significant increase in weight and the appearance of edema symptoms led to the estimation of blood volume (V... blood Increased filling pressure (P) fill The weight of particles exceeding the limit increased sharply. The re-estimated dynamic dry weight remained at 65.3 kg, but the effective circulating blood volume (V) was identified. blood The reading has significantly exceeded the ideal range, indicating that the capacity load is too heavy.

[0048] Risk Prediction and Intelligent Decision Making The deep risk prediction module received the aforementioned dynamic physiological state estimation sequence. The LSTM network captured the dangerous time-series pattern of a continuous and rapid increase in filling pressure. The survival analysis model calculated that Mr. Wang's comprehensive risk index (CRI(t)) for acute decompensation within the next 7 days jumped from 5% a week ago to 42%.

[0049] The reinforcement learning intelligent decision-making module is triggered. The system's current state vector s t [CRI=0.42, d(CRI) / dt=+0.35 / day, P] fill =High, Knowledge Score =75, Compliance =Good]. The Actor network, based on the current strategy π θ The system calculates the probability distribution of each action: educational push notification: 30%, nurse follow-up: 65%, medication adjustment: 5%. Based on this distribution, the system selects the action with the highest probability—triggers a level-two warning (orange) and executes a nurse telephone follow-up.

[0050] Adaptive Education Closed Loop Almost simultaneously, the adaptive patient education module was activated. The system's knowledge graph showed that while Mr. Wang's overall knowledge score was acceptable, his grasp of the sub-item regarding coping with rapid weight gain under the volume management knowledge point was low (P=0.3). The system determined that this warning event was strongly correlated with this knowledge gap. Therefore, a 2-minute animated video was immediately and precisely pushed to Mr. Wang's app, explaining in detail the core points that one should immediately be alert and contact medical staff if their daily weight gain exceeds 1kg. 24 hours later, the system again pushed a related multiple-choice question to evaluate the effect, completing the closed loop of assessment-intervention-reassessment.

[0051] Exercise rehabilitation guidance (taking Baduanjin with hands supporting the sky to regulate the three jiaos as an example) Mr. Wang practiced the Eight-Section Brocade exercise as planned. He opened the rehabilitation module of the app and followed the video demonstration to complete the "Hands Supporting the Sky to Regulate the Three Jiaos" posture, recording the entire process with his phone's camera.

[0052] Motion assessment: The intelligent motion rehabilitation analysis and guidance engine extracts the joint sequence from the video and calculates: The joint angle sequence is significantly different from the DTW of the standard GMM model, suggesting insufficient arm elevation range.

[0053] A high shoulder joint velocity spectrum entropy indicates stiff and unsmooth movements.

[0054] The area of ​​the ellipse with the center of gravity swinging is normal.

[0055] After comprehensive calculation, the motion quality score Q is calculated as follows: score = 0.58 (threshold 0.7).

[0056] Root Cause Analysis and Feedback Generation: The system initiates causal reasoning. Insufficient shoulder joint range of motion and lack of fluidity are observed. Based on a biomechanical causal diagram, the posterior probability of the root cause being limited shoulder joint flexibility is calculated to be the highest (P=0.8). Therefore, personalized feedback is generated: Text guidance: We detected that in the "Hands Supporting the Sky and Regulating the Three Burners" posture, your arms were not fully raised to your ears, possibly due to insufficient shoulder joint flexibility. Suggestions: 1. Warm up with 5 shoulder circles before practicing; 2. During practice, imagine your hands slowly lifting a heavy object, pausing briefly at the highest point.

[0057] Visual feedback: A semi-transparent green standard skeletal outline is overlaid on Mr. Wang's practice video. His own skeletal structure is shown in red at the shoulder joints, clearly contrasting the angular differences.

[0058] After receiving a level-two alert from the system, the community nurse immediately contacted Mr. Wang by phone. Based on the weight trend, risk index, and symptom information provided by the system, the nurse confirmed after questioning Mr. Wang that his water intake control had been poor recently. According to the pre-set protocol, the nurse instructed him to temporarily increase the dosage of a small dose of diuretic and reiterated the importance of water restriction. After Mr. Wang performed targeted exercises based on the rehabilitation feedback from the app, he recorded another video the next day. The system assessment showed improvements in his shoulder joint range of motion and movement fluency. score Increased to 0.72.

[0059] Three days later, Mr. Wang's weight steadily dropped to 66.0 kg, his edema subsided, his CRI(t) decreased to 15%, and the system warning was lifted. This complete closed loop, from early risk identification and accurate assessment to effective intervention, successfully avoided a potential acute exacerbation and hospitalization.

[0060] In summary, this invention presents a chronic disease management system for heart failure patients based on multimodal data dynamic fusion and intelligent feedback. It includes a patient terminal, a physician platform, and an intelligent server. The system enhances data reliability through a multimodal data credibility fusion module; utilizes a particle-filter-based dynamic physiological state estimation engine to non-invasively infer key internal physiological parameters such as effective circulating blood volume; combines a deep temporal risk prediction module to provide prospective early warning of heart failure decompensation; and generates tiered intervention strategies through a reinforcement learning intelligent decision-making module. Furthermore, the system possesses adaptive knowledge tracking and intelligent exercise rehabilitation guidance capabilities, achieving closed-loop intelligent management throughout the entire process from risk assessment and intelligent decision-making to personalized education and rehabilitation, significantly improving the accuracy and foresight of heart failure chronic disease management.

[0061] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A heart failure patient chronic disease management system based on multi-modal data dynamic fusion and intelligent feedback, characterized in that, include: The patient terminal is equipped with a data acquisition and interaction interface, which is used to collect daily morning weight measured according to clinical standards, patient self-reported symptom data collected through structured electronic forms, continuous vital sign data collected through wearable devices connected via Bluetooth, and video streams of the patient's actions in performing standard rehabilitation training procedures collected through a camera. The physician and management platform provides medical institutions with a web-based interactive interface for real-time monitoring of patients' physiological data and early warning information, and for receiving intervention suggestions pushed by the system for clinical decision confirmation. The system server connects to the patient terminal and the physician and management platform via an encrypted communication protocol. The server includes a processor and a memory. The memory stores a computer program. When the computer program is executed by the processor, it implements the following functional modules: The multimodal data acquisition and credibility fusion module is used to receive raw input data from the patient terminal and preprocess and fuse these multi-source heterogeneous data. The preprocessing includes using a time-series alignment algorithm to process vital sign data with non-uniform sampling frequency to generate smooth and continuous signal estimates. At the same time, the module dynamically calculates credibility weights for each data source. The weights are based on the signal-to-noise ratio index, data missing rate index, and user historical compliance score of the data source, and are obtained by normalization after nonlinear mapping through a neural network. A dynamic physiological state estimation engine is used to estimate the internal physiological state of a patient that cannot be directly measured from surface observation data. It establishes a state-space model based on the differential equations of the cardiovascular system. The state variables include effective circulating blood volume, cardiac filling pressure, myocardial contractility, and peripheral vascular resistance. The state-space model consists of state equations and observation equations. The state equations are constructed based on Starling's heart law and hemodynamic principles. The observation equations map the internal state to observable physiological parameters. A particle filtering algorithm is used for state estimation. By initializing a set of random particles to represent possible physiological states, an iterative process of prediction-update-resample is performed at each time step. Finally, the optimal state estimate is obtained by weighted averaging of all particles, and the patient's individualized dry weight is dynamically calculated based on this estimate. The deep temporal risk prediction module is used to predict the future risk of heart failure decompensation based on historical physiological state sequences. It adopts a deep survival analysis network architecture, which constructs risk prediction as a survival probability prediction problem that changes over time. The network combines a temporal feature extraction layer and a fully connected layer. The temporal feature extraction layer captures the temporal dependencies of physiological states, and the fully connected layer maps the extracted temporal features to a risk scalar, outputting the probability that a patient will experience a heart failure decompensation event within a specific future time window, which is defined as a comprehensive risk index. The reinforcement learning-based intelligent decision-making and hierarchical early warning module is used to transform risk assessment into optimal clinical intervention strategies and model patient management as a Markov decision process. The state space includes a comprehensive risk index, its rate of change, estimated physiological state, knowledge assessment score, and patient context information. The action space includes multi-level clinical intervention measures from no intervention to emergency contact. The reward function is set based on clinical outcomes. The Actor-Critic reinforcement learning framework is adopted, in which the Actor network generates the probability distribution of each action to be taken in a given state, the Critic network evaluates the long-term value of the state, and the decision strategy is optimized through a policy gradient algorithm. The adaptive knowledge status tracking and education module is used to dynamically assess and improve patients' disease management knowledge level. Based on the dynamic Bayesian knowledge tracking model, the patient's mastery status of each knowledge point is modeled as a latent variable. The mastery probability is recursively updated based on the patient's answer history and educational intervention history through Bayesian inference. Based on the information gain principle, the educational content that can minimize knowledge uncertainty is selected for precise push. The intelligent exercise rehabilitation analysis and guidance engine is used to guide rehabilitation training for heart failure patients. It maintains a standard movement database based on a probability model, extracts multi-dimensional biomechanical features from patient movement videos, and obtains a comprehensive movement quality score by calculating the difference from the standard movement. When the score is lower than the threshold, it performs error root cause reasoning based on a predefined biomechanical causal graph model to generate personalized feedback that includes specific text guidance and visual comparison.

2. The heart failure patient chronic disease management system based on dynamic fusion and intelligent feedback of multi-modal data according to claim 1, characterized in that: In the multimodal data acquisition and reliability fusion module, the time alignment algorithm adopts the Gaussian process regression algorithm, which models the time correlation of physiological signals by defining the mean function and the Matern kernel function; the neural network is a three-layer fully connected neural network, and is normalized by the Sigmoid function.

3. The chronic disease management system for heart failure patients based on multimodal data dynamic fusion and intelligent feedback as described in claim 1, characterized in that: The state-space model in the dynamic physiological state estimation engine is specifically defined as follows: The state equation is in differential form: Where X(t) = [V blood (t), P fill (t), Contractility (t), R systemic (t)] T , where is the state vector, representing effective circulating blood volume, cardiac filling pressure, myocardial contractility and peripheral vascular resistance respectively, f is a nonlinear function based on cardiovascular physiology, integrating Starling cardiac mechanism and hemodynamic equation, u(t) is the control input vector including patient-reported daily fluid intake and medication records, and w(t) is Gaussian process noise with zero mean. The observation equation is: Y(t) = h(X(t)) + v(t), where Y(t) = [W(t), HR(t), BP]. sys (t), BP dia (t)] T is the observation vector, which includes body weight, heart rate, systolic blood pressure and diastolic blood pressure; h is the observation function that maps the internal state to the observed value through physiological relationships; v(t) is Gaussian observation noise with zero mean. The specific execution steps of the particle filter algorithm include: Initialization: Generate N particles X by sampling from the prior distribution. p (0), and assign an initial weight of 1 / N to each particle; Prediction step: For each particle, state propagation is performed according to the state equation f: X p (t) = X p (t-1) + f(X p (t-1), u(t)) - Δt + w p (t) where w p (t) is a random perturbation sampled from a process noise distribution; Update step: When new observation data Y(t) arrives, the weight w of each particle is calculated p (t) = P(Y(t) | X p (t)), i.e. the observation likelihood probability; Resampling step: The system is resampled according to the particle weights, low-weight particles are eliminated, high-weight particles are copied, and a new set of equal-weight particles is generated; State estimation: The final state estimate is obtained by weighted averaging of all particles. 。 4. The heart failure patient chronic disease management system based on dynamic fusion and intelligent feedback of multi-modal data according to claim 1, characterized in that: The deep survival analysis network in the deep temporal risk prediction module is specifically configured as follows: The risk function is defined as: λ(t|X) = λ0(t)·exp(G(X(t),H(t))) Where λ0(t) is the nonparametric baseline risk function obtained by the Breslow estimator, and G(·) is a deep neural network whose input includes the current physiological state estimate X(t) and a historical state sequence of length L H(t) = [X(t-1), X(t-2), ..., X(tL)]; The architecture of the network G(·) includes: Input layer: Receives the concatenated current state and historical sequence; Long Short-Term Memory (LSTM) network layer: Composed of multiple hidden units, its update formula is: (W f ·[h t-1 ,x t ] + b f )(forget gate) (W i ·[h t-1 ,x t ] + b i )(input gate) (W C ·[h t-1 ,x t ] + b C )(candidate memory) (memory update) (W o ·[h t-1 ,x t ] + b o )(output gate) h t = o t ·tanh(C t (Hidden state); Fully connected layer: maps the final hidden state h of the long short-term memory network layer t to a scalar output using an activation function through at least one fully connected layer Output layer: Linear output of risk score; The comprehensive risk index CRI(t) is calculated using a survival function: The integral is approximated using numerical methods.

5. The chronic disease management system for heart failure patients based on multimodal data dynamic fusion and intelligent feedback as described in claim 1, characterized in that: The Markov decision process in the reinforcement learning intelligent decision-making and hierarchical early warning module is specifically defined as follows: State vector s t is: s t = [CRI(t),d(CRI) / dt,V blood (t),P fill (t),Contractility(t),R systemic (t),K score (t),Compliance Score (t)] Where d(CRI) / dt is the numerical derivative of the risk index, Compliance Score (t) represents the compliance score calculated based on historical behavior; Action space A includes multiple discrete intervention levels, including no action, pushing educational content, nurse telephone follow-up, adjusting diuretic dosage, escalation to physician treatment, and triggering emergency contact; The reward function R(s t , a t ) is defined as: If a patient experiences a decompensated event within the next 7 days and the system does not trigger an appropriate warning: R = -10; If the patient's condition is stable and the system does not perform unnecessary interventions: R = +1; If the system triggers intervention and the patient's condition improves: R = +3; If the system triggers unnecessary emergency intervention: R = -2; Other cases: R = 0; The specific implementation of the Actor-Critic framework is as follows: Critic network (s t ): uses a fully connected network with input s_t and output a scalar estimate of the state value Actor network π θ (a t | s t ): uses a fully connected network with input s_t and output a probability distribution of actions normalized by a Softmax function; Policy updates use the policy gradient method with baselines: (a t | s t )·( (s t )) where a is the learning rate, R t is the discounted cumulative return.

6. The heart failure patient chronic disease management system based on dynamic fusion and intelligent feedback of multi-modal data according to claim 1, characterized in that: The standard rehabilitation training program includes the Chinese traditional fitness qigong Baduanjin. The specific implementation of Baduanjin training in the intelligent sports rehabilitation analysis and guidance engine includes: The process of constructing the standard Baduanjin (Eight Pieces of Brocade) movement database: Calibration videos of healthy individuals and patients with stable heart failure performing various Baduanjin movements were collected and grouped according to cardiac function classification; Three-dimensional coordinate sequences of multiple joint points were extracted from the videos, and joint angle sequences, center of gravity trajectories, and joint velocity profiles were calculated; For each movement M and each cardiac function classification G in Baduanjin, a Gaussian mixture model was used to model the distribution of feature vectors. K is determined by the Bayesian information criterion; Specific methods for assessing patient movement quality: Extracting the same biomechanical feature vector X from patient videos. patient ; Calculate the negative log-likelihood of patient characteristics and the standard Gaussian mixture model as the basis for measuring the difference; Comprehensive action quality score Q score It is a weighted fusion of multiple sub-indicators, including morphological similarity: DTW stands for Dynamic Time Warping Distance, used to assess morphological differences in joint angle sequences. Motion smoothness: Among them, the power spectral entropy is calculated to determine the power spectral entropy value of the joint velocity sequence, which is used to quantify the smoothness of motion; Equilibrium stability: The area of ​​the 95% confidence ellipse was calculated using principal component analysis to determine the dispersion range of the centroid trajectory. Final Q score = w1·S1 + w2·S2 + w3·S3, the weights are determined by expert evaluation; The causal reasoning mechanism for error root cause analysis: A biomechanical causal graph for Baduanjin (Eight Pieces of Brocade) is constructed, with nodes including joint range of motion, muscle strength, coordination, and movement performance. When low-quality movements are detected, probabilistic reasoning is performed on the causal graph to calculate the posterior probability of each potential root cause. P (Root Cause i |Observation) ∝ P (Observation | Root Cause i ) · P Prior ( Root Cause i ).

7. An electronic device, comprising: One or more processors; Memory; One or more applications, characterized in that: the one or more applications are stored in the memory and configured to be executed by the one or more processors, the one or more applications being configured to perform all the method steps implemented by the heart failure patient chronic disease management system as described in any one of claims 1-6.

8. A computer-readable storage medium having stored thereon processor- executable instructions, the computer-readable storage medium comprising: The instructions are used to cause the processor to execute all the method steps implemented by the heart failure patient chronic disease management system as described in any one of claims 1-6.

Images