Rehabilitation training device for old people after heart disease

Abstract

The invention relates to the technical field of rehabilitation training, in particular to a rehabilitation training device for old people after heart diseases. The method comprises the following steps: a health data acquisition unit acquires physiological data, motion data and subjective feeling data of an elderly patient; the health data transmission and storage unit encrypts the data acquired by the health data acquisition unit in real time by using an encryption communication module, and stores the encrypted data; the intelligent analysis and decision-making unit dynamically evaluates rehabilitation risks based on a state space sequence model and generates a personalized training scheme; and the rehabilitation intervention unit performs rehabilitation training intervention based on the rehabilitation risk assessment result. According to the method, the state space model is used for describing the potential dynamic process of the heart load, real-time physiological observation data and the historical trend are continuously fused through online Bayesian reasoning, model parameters are dynamically updated, and therefore accurate tracking of individual heart state evolution is achieved.

Description

Technical Field

[0001] This invention relates to the field of rehabilitation training technology, and more specifically, to a rehabilitation training device for elderly people after heart disease. Background Technology

[0002] With the increasing aging population and the continuous rise in the number of heart disease patients, scientific and safe cardiac rehabilitation training is of great significance for improving cardiac function and preventing recurrent events in elderly patients. However, current cardiac rehabilitation training systems still face many challenges in practical application. Traditional rehabilitation programs often adopt fixed intensity and uniform standards, lacking dynamic adaptability to individual differences among elderly patients (such as age, gender, cardiac function level, and baseline physiological state). This can easily lead to excessively high training intensity, causing cardiovascular risks such as arrhythmias and abnormal blood pressure, or insufficient intensity, affecting rehabilitation outcomes. Meanwhile, existing monitoring equipment generally relies on preset static physiological thresholds (such as upper limit of heart rate) for risk warnings, ignoring individual baseline differences and recent physiological trends (such as a gradual increase in resting heart rate), making early and accurate risk identification difficult, resulting in frequent false alarms and missed alarms. Furthermore, the rehabilitation process involves multi-source data, including physiological indicators (heart rate, blood pressure, blood oxygen), exercise parameters (type, intensity), and patients' subjective feelings (fatigue level, anxiety level, psychological state). However, existing systems often only focus on objective physiological data, lacking effective integration and quantitative analysis of subjective feedback, leading to biased assessment results. More significantly, in unsupervised environments such as homes or communities, data collection is susceptible to motion artifacts and equipment interference, resulting in high signal noise. Traditional prediction models, on the other hand, lack robustness and struggle to operate stably in complex scenarios. Furthermore, most systems remain at the "monitoring-alarm" level, lacking a closed-loop mechanism that links risk assessment, personalized plan generation, and real-time intervention. This prevents automatic adjustment of training intensity or triggering of emergency responses when high risks are detected, leading to intervention lag. Therefore, this paper proposes a rehabilitation training device for elderly individuals recovering from heart disease. Summary of the Invention

[0003] The purpose of this invention is to provide a rehabilitation training device for elderly people after heart disease, in order to solve the problems mentioned in the background art, such as low individualization, inaccurate risk warning, insufficient fusion of multi-source data, weak anti-interference ability and lack of closed-loop intervention mechanism in the existing rehabilitation training process for elderly patients with heart disease.

[0004] To achieve the above objectives, the present invention provides a rehabilitation training device for elderly people after heart disease, comprising: A health data acquisition unit, which collects physiological data, exercise data, and subjective feeling data of elderly patients; A health data transmission and storage unit, wherein the health data transmission and storage unit uses an encrypted communication module to encrypt the data collected by the health data acquisition unit in real time, and stores the encrypted data; The intelligent analysis and decision-making unit dynamically assesses rehabilitation risks and generates personalized training plans based on a state-space sequence model. Among them, in the process of assessing rehabilitation risks, a micro neural network is used to generate anti-interference weights to perform weighted modulation on the feature sequence output by the state space sequence model. A rehabilitation intervention unit, which conducts rehabilitation training interventions based on the results of rehabilitation risk assessment.

[0005] As a further improvement to this technical solution, the health data acquisition unit includes a patient physiological module, a patient movement module, and a subjective feeling module; The patient physiological module is used to collect physiological data and historical trend data of elderly patients. The physiological data includes the elderly patients' heart rate, heart rate variability, blood pressure, and blood oxygen. The historical trend data includes the elderly patients' historical resting heart rate. The patient movement module is used to collect movement data from elderly patients, including the type and intensity of their current movement. The subjective feeling module collects data on the subjective fatigue level and subjective psychological feelings of elderly patients as subjective feeling data.

[0006] As a further improvement to this technical solution, the health data transmission and storage unit includes a health data transmission module and a health data storage module; The health data transmission module uses an encrypted communication module to encrypt the physiological data, historical trend data, exercise data, and subjective feeling data of elderly patients in real time, and then transmits them to the health data storage module. The health data storage module utilizes a cloud-based distributed database to structure and manage physiological data, historical trend data, exercise data, and subjective feeling data.

[0007] As a further improvement to this technical solution, the intelligent analysis and decision-making unit includes a dynamic safety boundary module, a risk prediction and closed-loop intervention module, a personalized rehabilitation plan generation module, and a compliance and psychological analysis module. The dynamic safety boundary module preprocesses and extracts features from physiological data, historical trend data, exercise data, and subjective feeling data to form a multidimensional input feature vector. Based on the state space model and online Bayesian inference method, it outputs the probability distribution of cardiac load and compares the predicted distribution of cardiac load with the individual baseline and historical trend of elderly patients to generate real-time cardiac load status data and probabilistic safety boundary. The risk prediction and closed-loop intervention module predicts the risk index of future adverse cardiac events based on probabilistic safety boundaries. The personalized rehabilitation program generation module assesses the cardiac load tolerance under different training parameters based on the risk index and generates a rehabilitation training plan in combination with the individual characteristics of elderly patients. The compliance and psychological analysis module assesses elderly patients' adherence and psychological changes based on subjective feeling data.

[0008] As a further improvement to this technical solution, the generation of probabilistic safety boundaries based on the state-space model and online Bayesian inference method includes the following steps: The preprocessed multidimensional input feature vector is input into a pre-trained state-space sequence model, which outputs a sequence of potential state vectors and the corresponding sequence of predicted cardiac load values. The predicted cardiac load sequence and the actual collected physiological observation sequence are used to construct a Bayesian filtering problem, and the parameters of the state-space sequence model are recursively updated using an online Bayesian inference method. Based on the updated state-space sequence model, the predicted point values ​​of cardiac load and their corresponding probability distributions are output, and the uncertainty measure of the prediction results is given at the same time. Based on the predicted probability distribution, a probabilistic safety boundary is constructed, which characterizes the safety boundary of cardiac load in elderly patients as a probabilistic envelope that changes over time.

[0009] As a further improvement to this technical solution, the construction of a probabilistic safety boundary based on the predicted probability distribution specifically includes: Based on the quantile curve of the predicted probability distribution, an initial upper and lower probabilistic safety boundary is constructed, and an uncertainty weighting factor is introduced to dynamically adjust the probabilistic safety boundary according to the prediction variance. At the same time, combined with the patient's historical trend data, the probabilistic safety boundary is trend-corrected so that the probabilistic safety boundary can be dynamically adjusted with the changing trend of cardiac load.

[0010] As a further improvement to this technical solution, the method of predicting the risk index of future adverse cardiac events based on probabilistic safety boundaries includes the following steps: S1.1 Based on a pre-trained lightweight meta-learning model, the time window features of real-time cardiac load status data are used as input, and a set of parameter offsets are output. S1.2. The parameter offset is combined with the pre-trained state-space sequence model parameters to generate personalized state-space sequence model parameters for the current time step. S1.3 Input the data from another time window into the state space model after updating the parameters of the personalized state space sequence model, extract the latent state vector sequence of its output, and use the temporal consistency attention module to modulate the latent state vector sequence to generate the modulated output. S1.4 Based on the probabilistic safety boundary, the risk index of adverse cardiac events in the future time period is calculated using the modulated output, and the risk index is quantified and classified to generate a risk level.

[0011] As a further improvement to this technical solution, in step S1.3, the temporal consistency attention module is used to modulate the potential state vector sequence to generate the modulated output, including the following steps: S1.31. Within the current short-time window, extract the latent state vector sequence of the state-space sequence model. Simultaneously, the historical feature vector sequence of the previous equal-length time window is obtained; S1.32 Calculate the potential state vector sequence The dynamic time-warped distance between the sequence and the historical feature vector sequence; S1.33 Calculate the sequence of potential state vectors within the current short window. The average similarity between feature vectors at adjacent time points within the same timeframe; S1.34. The dynamic time-warped distance and average similarity are fed into the micro neural network as input, and a window-level anti-interference weight between 0 and 1 is output. ; S1.35, Utilizing window-level anti-interference weights For the sequence of potential state vectors The sequence is modulated to obtain modulated features.

[0012] As a further improvement to this technical solution, the assessment of elderly patients' performance and psychological changes based on subjective feeling data includes the following steps: S2.1 Compare the subjective fatigue level with the rehabilitation training plan and calculate the execution deviation index; S2.2. Weighted fusion of deviation indicators to generate a comprehensive execution score, reflecting the patient's adherence to the training plan; S2.3 Quantify subjective psychological feelings into psychological state indicators; S2.4 Combine the performance score with the psychological state index to form a comprehensive compliance and mental health score, and mark any abnormal situations.

[0013] As a further improvement to this technical solution, the rehabilitation intervention unit compares the predicted value of cardiac load with a probabilistic safety boundary. If it exceeds the probabilistic safety boundary, the intervention process is triggered. At the same time, the risk index is compared with a preset risk level threshold, and a risk level is generated based on the comparison result.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The rehabilitation training device for elderly patients after heart disease, as disclosed in this invention, utilizes a state-space model to characterize the potential dynamic process of cardiac load. It continuously integrates real-time physiological observation data with historical trends through online Bayesian inference, dynamically updating model parameters to achieve precise tracking of individual cardiac state evolution. Furthermore, the system constructs a probabilistic safety boundary that changes over time based on a predicted probability distribution. It introduces an uncertainty weighting factor based on prediction variance and a historical trend correction term. The boundary is tightened under high uncertainty to enhance early warning sensitivity, and moderately relaxed under low uncertainty to reduce false alarms. Simultaneously, it incorporates recent physiological trends for offset correction, identifying potential deterioration risks in advance. This probabilistic boundary mechanism, which integrates uncertainty quantification and trend perception, can identify abnormal load states in heart disease patients during rehabilitation training earlier and more accurately than traditional fixed threshold alarm methods. This effectively avoids overtraining or sudden cardiovascular events, greatly improving the safety assurance capability of the rehabilitation process.

[0015] 2. The rehabilitation training device for elderly patients after heart disease, as disclosed in this invention, introduces a pre-trained lightweight meta-learning model in the risk prediction stage. Based on physiological characteristics within a very short time window, it rapidly generates personalized parameter offsets and dynamically adjusts key parameters of the state-space model, enabling the system to quickly adapt to the real-time physiological response patterns of different patients. More importantly, a temporal consistency attention module is designed in the feature modulation stage. By calculating the dynamic time warping (DTW) distance between the current potential state sequence and historical features, as well as the average similarity of neighboring times, and inputting this into a micro-neural network to generate anti-interference weights between 0 and 1, it performs credibility-weighted attenuation of state vectors under abnormal or noise interference. This mechanism effectively suppresses false signal fluctuations caused by sensor jitter, motion artifacts, or instantaneous emotional fluctuations, ensuring that the risk index calculation is based on a stable and consistent physiological evolution pattern. It is particularly suitable for the practical challenges of unstable device wearing and numerous environmental interferences in home rehabilitation scenarios for the elderly, significantly improving the system's robustness and clinical credibility on edge devices. Attached Figure Description

[0016] Figure 1 This is an overall flowchart of the present invention; The meanings of the labels in the diagram are as follows: 1. Health data collection unit; 2. Health data transmission and storage unit; 3. Intelligent analysis and decision-making unit; 31. Dynamic safety boundary module; 32. Risk prediction and closed-loop intervention module; 33. Personalized rehabilitation plan generation module; 34. Compliance and psychological analysis module; 4. Rehabilitation intervention unit. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0018] Example: Please refer to Figure 1 As shown, a rehabilitation training device for elderly people after heart disease is provided, comprising: Health data collection unit 1 collects physiological data, exercise data, and subjective feeling data from elderly patients; In this embodiment, the rehabilitation training device for elderly people after heart disease consists of hardware devices worn on the patient and a mobile application. Among them, the health data collection unit 1 includes a patient physiological module, a patient movement module, and a subjective feeling module; The patient physiological module is used to collect physiological data and historical trend data of elderly patients. The physiological data includes the elderly patient's heart rate, heart rate variability (HRV), blood pressure, and blood oxygen. The historical trend data includes the elderly patient's historical (past 24 hours) resting heart rate. The patient physiological module is an embedded optical heart rate sensor (for measuring heart rate and blood oxygen) and a wrist-worn sphygmomanometer (for measuring blood pressure) on the device. Specifically, in smart wearable hardware devices (taking smartwatches as an example), the health data acquisition unit 1 consists of a photoplethysmography (PPG) sensor and an infrared blood oxygen detection sensor on the back of the watch face, used to collect heart rate, blood oxygen saturation and heart rate variability; and an airbag-type blood pressure sensor serves as an extension of the patient's physiological module to acquire blood pressure data.

[0019] The patient motion module is used to collect motion data from elderly patients, including the type and intensity of their current exercise. The patient motion module directly collects the user's exercise intensity and type through a built-in power meter and speed sensor. The subjective feeling module collects subjective fatigue level and subjective psychological feeling data (including emotion score, anxiety level, and stress feeling, input through a short scale or mobile self-assessment) of elderly patients as subjective feeling data. The subjective feeling module is a touch screen on the front of the device. During or after training, the screen will pop up the Borg scale (6-20 points) and an emotion expression selection interface for users to report their subjective fatigue level and psychological feeling.

[0020] The health data transmission and storage unit 2 uses an encrypted communication module to encrypt the data collected by the health data acquisition unit 1 in real time, and then stores the encrypted data. In this embodiment, after the health data transmission and storage unit 2 uploads the encrypted data to the cloud, the cloud part of the intelligent analysis and decision-making unit 3 receives the data and processes it; the local embedded part is responsible for real-time data preprocessing and anti-interference weight calculation, and uploads the corrected feature data to the cloud to support the accurate decision-making of the cloud module. Health data transmission and storage unit 2 includes a health data transmission module and a health data storage module; The health data transmission module uses an encrypted communication module to encrypt the physiological data, historical trend data, exercise data, and subjective feeling data of elderly patients in real time and transmit them to the health data storage module. The encrypted communication module is a hardware or software unit used for data encryption and secure transmission to ensure that health data is not stolen or tampered with during collection and storage. The health data transmission module is a dedicated security encryption chip (such as TPM) on the device's motherboard. It encrypts all data collected by the sensors and touch screen in real time and then transmits it out through the device's Wi-Fi / 5G module. The health data storage module utilizes a cloud-based distributed database (a cloud-based distributed database is a database system deployed on a cloud platform to distribute data across multiple nodes, which complies with HIPAA medical data security standards) to perform structured storage and management of physiological data, historical trend data, exercise data, and subjective feeling data.

[0021] Intelligent analysis and decision-making unit 3 dynamically assesses rehabilitation risks and generates personalized training plans based on a state-space sequence model; In the process of assessing rehabilitation risks, a micro-neural network is used to generate anti-interference weights, which are then used to weight and modulate the feature sequences output by the state-space sequence model. In this embodiment, in the smart wearable hardware device (taking a smartwatch as an example), the intelligent analysis and decision-making unit 3 adopts a distributed architecture, including a local embedded part and a cloud part. The local embedded part runs on the embedded processor of the smartwatch and is used to execute the micro-neural network to calculate the anti-interference weights, realizing front-end data preprocessing and feature correction. The cloud part is responsible for complex model reasoning and decision generation, and the specific modules include a dynamic safety boundary module 31, a risk prediction and closed-loop intervention module 32, a personalized rehabilitation plan generation module 33, and a compliance and psychological analysis module 34. Furthermore, the cloud portion of the intelligent analysis and decision-making unit 3 includes a dynamic security boundary module 31 (a cloud-based State Space Model (SSM) algorithm responsible for generating probabilistic security boundaries), a risk prediction and closed-loop intervention module 32 (a cloud-based lightweight meta-learning model and temporal consistency attention calculation algorithm used to predict risk indices), a personalized rehabilitation plan generation module 33 (a cloud-based rehabilitation plan generation method that generates personalized training plans based on risk indices), and a compliance and psychological analysis module 34 (a cloud-based data analysis and report generation program that assesses patient compliance and psychological state). The device terminal uploads encrypted data to the cloud, which is then processed by the cloud algorithm service, and the decision results (including security boundaries and risk indices) are sent back to the device in real time. The dynamic safety boundary module 31 preprocesses and extracts features from physiological data, historical trend data, exercise data, and subjective feeling data. Specifically, the preprocessing and feature extraction involve: first, using a low-pass filter and sliding window method to denoise and smooth physiological signals such as heart rate, blood pressure, and blood oxygen, and correcting outliers using median filtering or threshold clipping; simultaneously, applying the exponentially weighted moving average (EWMA) method to extract short-term trends from historical trend data; second, extracting exercise types from exercise data through classification coding, and calculating exercise intensity features by combining acceleration or power signals. The process involves: quantifying and standardizing subjective fatigue levels and psychological states into numerical features; standardizing and weighting psychological data to create a psychological state index; and finally, concatenating and normalizing the processed features into a unified multidimensional input feature vector. Based on a state-space model (SSM) and online Bayesian inference, a cardiac load probability distribution is output. This predicted cardiac load distribution is then compared with the individual baseline and historical trends of elderly patients (the individual baseline and historical trends of elderly patients are stored in a manner consistent with medical safety regulations). In a fully standardized cloud database, the individual baseline for elderly patients refers to the normal range of physiological data for the patient in a resting, stable, and healthy state. This is not a fixed value, but a personalized range that takes into account factors such as personal medical history, age, gender, weight, and cardiac function class. For example, Mr. Zhang's individual baseline is 55-68 bpm, while Ms. Li's is 60-72 bpm; blood pressure control between 110 / 70 mmHg and 130 / 80 mmHg is considered a safe baseline for blood pressure; baseline blood oxygen saturation is maintained at 95%-99%; historical trend refers to the patient's recent (e.g., historical data) The changes and patterns of physiological data and exercise performance (from days to weeks) reflect the patient's recovery progress or potential problems, including whether the resting heart rate is gradually decreasing; whether the peak heart rate at the same exercise intensity is lower than a week ago (indicating improved cardiac function); whether the average nighttime heart rate over the past 24 hours is higher than in previous days; and whether HRV (heart rate variability) during sleep shows a decreasing trend (which could be an early sign of overexertion or potential problems). Real-time cardiac load status data (including predicted heart rate, blood pressure, blood oxygen, exercise intensity, and deviation from probabilistic safety thresholds) and probabilistic safety boundaries are generated. The generation of probabilistic safety boundaries based on the State-Space Model (SSM) and online Bayesian inference methods includes the following steps: The preprocessed multidimensional input feature vector is fed into a pre-trained state-space sequence model (such as the S5 or Mamba model). This model characterizes the potential dynamic process of cardiac load through state equations and maps the potential states to observable predicted cardiac load values ​​through observation equations, thereby outputting a sequence of potential state vectors and corresponding sequences of predicted cardiac load values. (The pre-trained state-space sequence model is based on a state-space modeling framework and adopts a structure combining state transition layers and observation mapping layers. Its training parameters include the state transition matrix, observation matrix, and noise covariance matrix; the model input is a multidimensional time series feature vector containing physiological data, historical trend data, etc.) The system outputs a sequence of potential state vectors representing cardiac load and corresponding predicted values, based on motion data and subjective perception data. In the intermediate layer, the temporal evolution of the state vectors is calculated recursively or through convolution. Simultaneously, attention or gating mechanisms are used to weight and modulate the input features to capture short-term fluctuations and long-term trends. Model parameters (such as the state transition matrix, observation matrix, and noise covariance) are initialized during pre-training based on multi-source data from publicly available cardiac rehabilitation databases and clinically collected data. Subsequently, they are recursively updated during individual patient use using online Bayesian inference, thus balancing prior group patterns with real-time individual differences to achieve accurate tracking and risk assessment of cardiac state evolution. The predicted cardiac load sequence and the actual collected physiological observation sequence (i.e., collected physiological data, including actual heart rate and blood pressure) are used to construct a Bayesian filtering problem. An online Bayesian inference method is used to recursively update the parameters of the state-space sequence model. Specifically: First, the predicted value sequence is used as the prior state of the state-space sequence model, and the actual observations are used as evidence for observation updates. Then, using the online Bayesian inference method, the posterior distribution is recursively calculated to update the model parameters such as the state transition matrix, observation matrix, and noise covariance in real time. At each time step, the updated posterior distribution is used as the prior distribution for the next time step to continuously fuse historical predictions and real-time observations, thereby achieving dynamic correction of the model parameters and improving the accuracy of cardiac load prediction and its adaptability to individual changes. Based on the updated state-space sequence model, the system outputs predicted point values ​​of cardiac load and their corresponding probability distributions, while also providing a measure of the uncertainty of the prediction results. Specifically: First, the multidimensional input feature vector at the current moment is input into the state-space sequence model updated by online Bayesian methods, and the evolution of the latent state vector is calculated using the state transition equation and the observation equation. Then, the latent state is mapped to predicted point values ​​of cardiac load through the observation equation, and a predicted probability distribution is generated by combining model parameters and noise covariance to reflect the uncertainty of the prediction. Finally, the prediction standard deviation or confidence interval is calculated based on the probability distribution as a measure of uncertainty, and the predicted value and its uncertainty index are output simultaneously. Based on the predicted probability distribution, a probabilistic safety boundary is constructed, which represents the safety boundary of cardiac load in elderly patients as a probabilistic envelope that changes over time. When the predicted value approaches or enters the probabilistic envelope, it can be determined that the probability of cardiac load risk has increased significantly, thereby achieving early identification and warning of potential risks. Furthermore, constructing a probabilistic safety boundary addresses the core issue that static safety thresholds in rehabilitation training for elderly cardiac patients fail to reflect the dynamic and complex physiological states of individuals. The construction process integrates predictive probability distributions, real-time uncertainties, and historical trends, transforming the safety boundary from a fixed value into a dynamic, probabilistic "safety channel." This achieves a leap from delayed, one-size-fits-all alarms to proactive, personalized risk warnings. Its function is to more sensitively identify potential risks (early warning) while effectively reducing false alarms caused by temporary or noise-induced interference (avoiding over-intervention), thus improving the continuity and effectiveness of rehabilitation training while ensuring safety. Based on the predicted probability distribution, a probabilistic safety boundary is constructed, specifically as follows: Based on the quantile curves of the predicted probability distribution (quantile curves are curves representing the changes over time of predicted values ​​calculated at each predicted time point according to a preset probability level based on the predicted probability distribution of cardiac load; the upper quantile curve represents the safe upper limit for high cardiac load, and the lower quantile curve represents the safe lower limit for low cardiac load), preliminary upper and lower probabilistic safety boundaries are constructed, and an uncertainty weighting factor is introduced. ( In the formula, The weights are adaptive and automatically adjusted based on the patient's recent state stability (the specific automatic adjustment strategy is: using the standard deviation of heart rate variability (SDNN) as a surrogate indicator of recent state stability; if the current time is...). Select the past minutes (e.g.) The standard deviation of the physiological data (heart rate) in the window is calculated. and will Normalized to the interval [0,1], the normalized standard deviation is mapped to the range using a monotonically increasing function. Value: In the formula, The preset minimum weight value (empirically taken as 0.5) is used. The preset maximum weight value (empirically, it is taken as 2.0). The slope parameter controls the rate of change. The normalized standard deviation (This is the offset, empirically taken as 0.5). For time, For a moment The variance of the predicted cardiac load probability distribution represents the magnitude of the uncertainty in the model's prediction. To obtain the standard deviation by taking the square root of the variance (used to quantify the uncertainty of the predicted distribution), the probabilistic safety boundary is dynamically adjusted based on the predicted variance (the boundary is moderately tightened when the model is under high uncertainty (large predicted variance); and moderately widened when the model is under low uncertainty (small predicted variance). Simultaneously, the probabilistic safety boundary is trend-corrected by incorporating historical patient trend data (i.e., integrating historical patient trends (such as the direction of change in heart rate / blood pressure over the past 24 hours). When a sustained upward or downward trend occurs, the dynamic boundary shifts accordingly; for example, if the resting heart rate has been consistently rising for several days, the lower boundary line shifts upward with the trend, exposing potential risks in advance). This allows the probabilistic safety boundary to be dynamically adjusted according to the changing trend of cardiac load. The probabilistic safety boundary consists of an upper boundary and a lower boundary. The upper boundary of the probabilistic safety boundary is: ; The lower boundary of the probabilistic safety boundary is: ; In the formula, For at any time The probabilistic safety upper bound represents the safe limit of cardiac load. When the predicted value exceeds this upper bound, the system determines that the risk has increased. For a moment The 95th percentile predictive value indicates that there is a 95% probability that the cardiac load at that time point will be lower than this value. It is used to characterize the probabilistic upper limit of high cardiac load. This is a trend sensitivity coefficient used to adjust the weight of the influence of historical trends on the boundary position. For a moment The historical trend correction value for cardiac load, calculated by analyzing the direction and magnitude of recent (e.g., the past 24 hours) changes in heart rate and blood pressure, is used to anticipate potential risks. Specifically: First, at time... The former Within a given time period, predicted cardiac load or physiological index values ​​for patients are acquired, and the differences between adjacent time points are calculated in chronological order. Then, a weighted average is calculated for each time difference based on predefined weighting coefficients, with the weight of more recent time points appropriately increased to enhance sensitivity to recent changes. Finally, the weighted average is used as a trend correction value. , For a moment The probabilistic safety lower boundary represents the safe lower limit of cardiac load. When the predicted value is lower than this lower boundary, the system determines that there is a risk (such as low heart rate). For a moment The 5th percentile predicted value indicates that there is a 5% probability that the cardiac load is lower than this value at that time point, and is used to characterize the probabilistic lower limit of low cardiac load.

[0022] The risk prediction and closed-loop intervention module 32 predicts the risk index of adverse cardiac events occurring in the next short period of time (e.g., the next 5 minutes) based on the probabilistic safety boundary. The method of predicting the risk index of adverse cardiac events in the near future (e.g., within the next 5 minutes) based on probabilistic safety boundaries includes the following steps: S1.1 Based on a pre-trained lightweight meta-learning model, a very short time window (e.g., a historical 20-second period) of real-time cardiac load data is used as input, and a set of parameter offsets is output. The pre-trained lightweight meta-learning model adopts a three-layer structure: input embedding layer—temporal coding layer—output conditioning layer. The temporal coding layer can be a gated recurrent unit (GRU) or a one-dimensional convolutional network, used to extract dynamic features of cardiac load within the short time window. During model training, the patient's physiological signal sequence and corresponding individualized parameter offsets are used as supervision labels. The mean squared error loss function (MSE) is used to constrain the difference between the predicted parameters and the true parameters, and L2 regularization is combined to suppress overfitting. The training method uses a task-based small sample meta-learning box. Frameworks (such as MAML) enable rapid adaptive training on data subtasks from multiple different patients, allowing the model to quickly generate personalized parameter adjustments based on inputs within extremely short time windows during the inference phase. Training parameters include weight matrices, bias vectors, and learning rate adjustment parameters. The model input is a feature vector of cardiac load status within a short time window, including information such as heart rate, blood pressure, blood oxygen saturation, exercise intensity, and subjective fatigue level. Intermediate layers extract dynamic patterns and short-term dependencies from the time series using GRU or one-dimensional convolution combined with attention mechanisms. The output layer maps the latent state to parameter offsets, used to personalize the key parameters of the pre-trained state-space sequence model, thereby enabling the model to rapidly adapt to the real-time cardiac load characteristics of different patients. S1.2. The parameter offset is combined with the pre-trained state-space sequence model parameters to generate personalized state-space sequence model parameters at the current moment, enabling the model to dynamically adapt to the real-time cardiac load characteristics of the patient. Specifically, by using the parameter offset as a dynamic correction factor, the key parameters of the state-space sequence model (including the state transition matrix, observation matrix, and noise covariance) are adjusted in real time in a personalized manner, so that the model can adaptively update its internal dynamic relationship according to the current physiological and motor characteristics of the elderly patient, generating personalized model parameters that conform to individual characteristics, thereby ensuring that the state-space model has higher accuracy and robustness in predicting cardiac load and constructing dynamic safety boundaries. S1.3 Input the data from another short time window (such as the historical 20 seconds) into the state space model after updating the parameters of the personalized state space sequence model, extract the latent state vector sequence of its output, and use the temporal consistency attention module to modulate the latent state vector sequence to generate the modulated output, that is, the modulated feature sequence. Furthermore, the use of a temporal consistency attention module to modulate the latent state vector sequence addresses the problem of instantaneous distortion and false alarms in physiological data caused by motion noise, transient abnormalities in sensor signals, or sudden interference during rehabilitation training. During modulation, the consistency between the current data and historical temporal patterns is evaluated to intelligently determine the data's reliability, thereby significantly suppressing random interference while preserving the true physiological change trend. Its role is to act as an intelligent filter, ensuring that the features upon which the subsequent risk prediction model is based are real and reliable, fundamentally reducing the risk of false alarms and misinterventions caused by poor data quality, and improving the accuracy and robustness of the entire system.

[0023] The temporal consistency attention module is used to modulate the latent state vector sequence to generate the modulated output, including the following steps: S1.31. Within the current short-time window (e.g., a historical 20 seconds), extract the latent state vector sequence of the state-space sequence model. (The state-space sequence model here and thereafter is a state-space sequence model with personalized state-space sequence model parameter configuration, and a sequence of potential state vectors.) This represents a sequence, while simultaneously acquiring the historical feature vector sequence from the previous equal-length window (e.g., 20–40 seconds ago); S1.32 Calculate the potential state vector sequence The dynamic time warping (DTW) distance between the current feature and the historical feature vector sequence is used to characterize the degree of consistency between the current feature and the historical time series pattern. Among them, Dynamic Time Warping (DTW) distance for: ; In the formula, This represents the alignment path obtained by the dynamic time warping algorithm, used to match corresponding elements in two time series. For index pairs, indicating the nth index in the current window sequence. The vector and the first vector in the historical window sequence The vectors are aligned. In the previous equal-length window (i.e., time...) The historical feature vector sequence extracted from within; S1.33 Calculate the sequence of potential state vectors within the current short window. The average similarity between feature vectors at adjacent time points within an interior space; an abnormal feature value caused by noise will be significantly inconsistent with its neighbors. Among them, average similarity for: ; In the formula, This represents the number of potential state vectors within the current short window. The first in the current window A potential state vector, The first in the current window A potential state vector, This represents the calculation of the cosine similarity between two adjacent latent state vectors, with a numerical range of [−1, 1], where the closer to 1, the more similar they are. S1.34. Combine Dynamic Time Warping (DTW) distance with average similarity. As input, it is fed into a miniature neural network (such as a two-layer perceptron), and the output is a window-level anti-interference weight between 0 and 1. ( In the formula, This is a miniature neural network employing a two-layer Multilayer Perceptron (MLP) structure (i.e., the MLP input layer has a dimension of 2 (corresponding to the two inputs, DTW distance and average similarity, where the model parameters (weight matrix and bias vector) are initialized using the Xavier normal distribution method), the hidden layer has 8 neurons, and the output layer has 1 neuron). Training parameters include the weight matrix, bias vector, and activation function parameters (including the ReLU function in the first layer and the Sigmoid activation function in the second layer). The input consists of two feature vectors: Dynamic Time Warped (DTW) distance and average similarity. The intermediate stream undergoes linear transformation through the first fully connected network and non-linear features are extracted using the ReLU function, then compressed through the second fully connected network. Finally, the output layer generates a window-level anti-interference weight value between 0 and 1 using the Sigmoid function, which is used to modulate the credibility of the latent state vector. The model training data comes from a multi-source cardiac rehabilitation dataset with noise interference and real labels. The labels are annotated by clinical experts or generated based on signal quality evaluation indicators. During training, the cross-entropy loss function is used as the optimization objective, and the Adam optimizer is used for parameter updates. The early stopping strategy and L2 regularization are combined to prevent overfitting, thereby ensuring that the model can suppress noise interference and maintain low computational cost and high robustness when running in real time on edge devices. A weight of 1 means that the current feature can be retained; a weight close to 0 means that the current feature is affected by noise interference, and its contribution to subsequent calculations will be significantly reduced. S1.35, Utilizing window-level anti-interference weights For the sequence of potential state vectors Modulation is performed to obtain the modulated features. This is used for subsequent risk calculations; S1.4 Based on the probabilistic safety boundary, the system calculates the risk index of adverse cardiac events (including arrhythmia, abnormal blood pressure, and myocardial ischemia) in the near future (e.g., 1–5 minutes) using the modulated output. Specifically, based on the probabilistic safety boundary, the system takes the modulated output as input, combines the predicted cardiac load value and its uncertainty distribution, makes a probabilistic inference on the state evolution trajectory in the near future (e.g., 1–5 minutes), and calculates the risk index of adverse cardiac events (e.g., arrhythmia, abnormal blood pressure, and myocardial ischemia). The risk index is then quantified and classified to generate risk levels (including low risk, medium risk, and high risk).

[0024] The personalized rehabilitation program generation module 33 assesses the tolerability of cardiac load under different training parameters based on a risk index, and generates a rehabilitation training plan in combination with the individual characteristics of elderly patients. The plan includes the plan goals, intensity, time, and frequency. Specifically, based on the generated probabilistic safety boundary and the real-time calculated risk index, the module simulates and predicts the cardiac load under different combinations of training parameters (such as exercise type, intensity, duration, and frequency), and calculates the probability that the cardiac load exceeds the safety boundary under each plan. Subsequently, the module combines the individual characteristics of the patient, including age, gender, weight, cardiac function class, historical resting heart rate, and exercise adaptability, to perform a weighted assessment of the risk tolerance of each training plan. Finally, the module selects training parameters within the controllable risk range and generates a personalized rehabilitation training plan, including specific training goals, exercise type, intensity level, time arrangement, and frequency. The compliance and psychological analysis module 34 assesses the elderly patients' performance and psychological changes based on subjective feeling data; The assessment of elderly patients' adherence to treatment and psychological changes based on subjective experience data includes the following steps: S2.1. Compare the subjective fatigue level with the rehabilitation training plan to calculate the execution deviation index, including intensity deviation (actual heart rate or exercise power vs. planned target), duration deviation (actual exercise time vs. planned time), and completion index (action completion rate, repetition rate, etc.). Specifically: First, obtain the subjective fatigue level score of elderly patients after rehabilitation training and compare it with the target fatigue level range in the pre-set rehabilitation training plan; then, quantify the difference between the patient's actual fatigue level and the target range. If the actual value is higher than the upper limit of the target range, it is judged that the training intensity is too high; if it is lower than the lower limit of the target range, it is judged that the training intensity is too low; finally, after normalizing the difference, an execution deviation index is generated to characterize the consistency between the patient's training execution and the expected plan. S2.2. Weighted fusion of deviation indicators to generate a comprehensive execution score that reflects the patient's adherence to the training plan. Specifically, the execution deviation indicators calculated from multiple dimensions such as subjective fatigue level deviation, heart rate load deviation, and exercise volume deviation are used as inputs. Weight parameters are set according to the importance of each dimension in rehabilitation training. The deviation indicators are weighted, summed, and normalized to obtain the comprehensive execution score. S2.3 Quantify subjective psychological feelings data (including emotion scores, anxiety levels, and stress feelings) into psychological state indicators (including psychological comfort index, anxiety index, and stress index). S2.4 Combine the performance score with psychological state indicators to form a comprehensive compliance and mental health score, and mark abnormal situations, such as high fatigue + high stress indicating possible overtraining or excessive psychological stress; low performance + low stress indicating unfamiliarity with operation or insufficient compliance.

[0025] Rehabilitation intervention unit 4 involves rehabilitation training intervention based on the results of rehabilitation risk assessment; In this embodiment, the rehabilitation intervention unit 4 is the main touch screen display interface and control system of the device. The rehabilitation intervention unit 4 compares the predicted or observed value of cardiac load with the probabilistic safety boundary. If it exceeds the probabilistic safety boundary, the intervention process is triggered. At the same time, the risk index is compared with the preset risk level threshold, and a risk level is generated based on the comparison result. The risk level includes low risk, medium risk, and high risk. Specifically, the predicted value of cardiac load or the actual observed value after filtering is compared with the probabilistic safety boundary in real time. When the detected value exceeds the upper safety boundary, it is determined that the cardiac load is too high, triggering the upper boundary flag and immediately initiating the intervention process. When the detected value is below the lower safety boundary, it is determined that the cardiac load is too low, triggering the lower boundary flag and immediately initiating the intervention process. When the detected value is between the upper and lower boundaries, it is determined that it is within the safe range, and no intervention is triggered due to boundary issues, but a comprehensive assessment is still required in conjunction with the risk index.

[0026] Furthermore, the risk index output by the risk prediction and closed-loop intervention module is compared with the preset risk level thresholds: when the risk index is less than or equal to the low-risk threshold 'a', it is determined to be a low-risk level, indicating that the probability of an adverse cardiac event occurring in the near future is extremely low; when the risk index is between the low-risk threshold 'a' and the high-risk threshold 'b', it is determined to be a medium-risk level, indicating that there is a certain possibility of an adverse event occurring in the near future, requiring continuous monitoring; when the risk index exceeds the high-risk threshold 'b', it is determined to be a high-risk level, indicating that the probability of an adverse cardiac event occurring in the near future is extremely high, requiring immediate intervention. Ultimately, the decision to intervene and the type of intervention is a comprehensive decision based on the two comparative results mentioned above. That is, any condition met can trigger the corresponding intervention: when the predicted or observed cardiac load exceeds the probabilistic safety boundary, intervention is triggered immediately regardless of the risk prediction level, and the intervention level is determined by the risk level; if the boundary is not exceeded but the risk level is high, intervention is also triggered immediately, implementing measures such as terminating training, issuing audible and visual alarms, and notifying the doctor; if the boundary is not exceeded and the risk level is medium, early warning intervention is triggered, preventing further escalation of risk by reducing training intensity, providing interface warnings, or remotely reminding the doctor; if the boundary is not exceeded and the risk level is low, the current training is maintained, the system interface displays a green safety status and provides encouraging prompts, thereby achieving coordinated judgment and intervention of immediate danger signals and short-term risk predictions.

[0027] In this embodiment, the rehabilitation training device displays the elderly patient's cardiac load status, training plan execution, and risk index in real time through a visual interface. Based on the risk level, it provides patients with real-time prompts and operational suggestions, including adjusting training intensity, rest, and movement correction. It also supports elderly patients to report their subjective state and receive system feedback. Doctors and family members can remotely view the training and psychological state and provide intervention suggestions. Furthermore, it generates personalized training and psychological support content based on historical and real-time data, and stores and shares interactive data to provide a basis for optimizing subsequent rehabilitation plans, thereby improving patient compliance and rehabilitation experience.

[0028] Furthermore, this rehabilitation training device for elderly people after heart disease is mainly applied to smart wearable devices, such as smartwatches and heart monitoring chest patches. Smart wearable devices connect to the encrypted communication module on the device side via Bluetooth, Wi-Fi or 5G modules, and upload health data to the cloud distributed database in real time via the health data transmission module, where it is uniformly stored and managed. The continuous physiological and motion data provided by the wearable device are uniformly sent to the dynamic safety boundary module, and combined with the online Bayesian inference method, the probability distribution of cardiac load changing over time is calculated. When the smartwatch receives real-time heart rate fluctuations or the chest patch receives ECG signals, anti-interference weights generated by a micro neural network are used to weaken abnormal jitters or signal drift, thereby ensuring the reliability of the risk prediction and closed-loop intervention modules. As a result, the personalized rehabilitation plan generation module can match the daily activity patterns of the elderly (such as walking and cycling) and dynamically adjust the exercise intensity and training frequency.

[0029] Wearable devices use touchscreens and speakers as intervention feedback interfaces: when the smartwatch detects that the heart rate or blood oxygen exceeds the probabilistic safety boundary generated by the system, the rehabilitation intervention unit can directly prompt the patient to slow down or rest immediately through the watch's touchscreen interface or voice broadcast; when the risk index exceeds the preset threshold, the system can link the smartwatch and mobile phone to push alarms to doctors and family members, realizing real-time intervention in home scenarios.

[0030] By integrating with smart wearable devices such as smartwatches and chest patch ECG monitors, this rehabilitation training device can maintain cardiac load tracking and risk warning capabilities in non-medical settings such as homes and communities, thereby improving the accessibility and adherence to rehabilitation training.

[0031] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A rehabilitation training device for elderly people after heart disease, characterized in that, include: Health data acquisition unit (1), wherein the health data acquisition unit (1) collects physiological data, exercise data and subjective feeling data of elderly patients; The health data transmission and storage unit (2) uses an encrypted communication module to encrypt the data collected by the health data acquisition unit (1) in real time and stores the encrypted data. Intelligent analysis and decision-making unit (3), which dynamically assesses rehabilitation risks and generates personalized training programs based on state-space sequence models; Among them, in the process of assessing rehabilitation risks, a micro neural network is used to generate anti-interference weights to perform weighted modulation on the feature sequence output by the state space sequence model. Rehabilitation intervention unit (4), which conducts rehabilitation training intervention based on the results of rehabilitation risk assessment.

2. The rehabilitation training device for elderly people after heart disease according to claim 1, characterized in that: The health data acquisition unit (1) includes a patient physiological module, a patient movement module, and a subjective feeling module; The patient physiological module is used to collect physiological data and historical trend data of elderly patients. The physiological data includes the elderly patients' heart rate, heart rate variability, blood pressure, and blood oxygen. The historical trend data includes the elderly patients' historical resting heart rate. The patient movement module is used to collect movement data from elderly patients, including the type and intensity of their current movement. The subjective feeling module collects data on the subjective fatigue level and subjective psychological feelings of elderly patients as subjective feeling data.

3. The rehabilitation training device for elderly people after heart disease according to claim 1, characterized in that: The health data transmission and storage unit (2) includes a health data transmission module and a health data storage module; The health data transmission module uses an encrypted communication module to encrypt the physiological data, historical trend data, exercise data, and subjective feeling data of elderly patients in real time, and then transmits them to the health data storage module. The health data storage module utilizes a cloud-based distributed database to structure and manage physiological data, historical trend data, exercise data, and subjective feeling data.

4. The rehabilitation training device for elderly people after heart disease according to claim 1, characterized in that: The intelligent analysis and decision-making unit (3) includes a dynamic safety boundary module (31), a risk prediction and closed-loop intervention module (32), a personalized rehabilitation plan generation module (33), and a compliance and psychological analysis module (34). The dynamic safety boundary module (31) preprocesses and extracts features from physiological data, historical trend data, exercise data and subjective feeling data to form a multidimensional input feature vector. Based on the state space model and online Bayesian inference method, it outputs the probability distribution of cardiac load and compares the predicted distribution of cardiac load with the individual baseline and historical trend of elderly patients to generate real-time cardiac load status data and probabilistic safety boundary. The risk prediction and closed-loop intervention module (32) predicts the risk index of future adverse cardiac events based on probabilistic safety boundaries; The personalized rehabilitation program generation module (33) assesses the tolerability of cardiac load under different training parameters based on the risk index and generates a rehabilitation training plan in combination with the individual characteristics of elderly patients. The compliance and psychoanalysis module (34) assesses the performance and psychological changes of elderly patients based on subjective feeling data.

5. The rehabilitation training device for elderly people after heart disease according to claim 4, characterized in that: The method for generating probabilistic safety boundaries based on the state-space model and online Bayesian inference includes the following steps: The preprocessed multidimensional input feature vector is input into a pre-trained state-space sequence model, which outputs a sequence of potential state vectors and the corresponding sequence of predicted cardiac load values. The predicted cardiac load sequence and the actual collected physiological observation sequence are used to construct a Bayesian filtering problem, and the parameters of the state-space sequence model are recursively updated using an online Bayesian inference method. Based on the updated state-space sequence model, the predicted point values ​​of cardiac load and their corresponding probability distributions are output, and the uncertainty measure of the prediction results is given at the same time. Based on the predicted probability distribution, a probabilistic safety boundary is constructed, which characterizes the safety boundary of cardiac load in elderly patients as a probabilistic envelope that changes over time.

6. The rehabilitation training device for elderly people after heart disease according to claim 5, characterized in that: The construction of a probabilistic safety boundary based on the predicted probability distribution specifically involves: Based on the quantile curve of the predicted probability distribution, an initial upper and lower probabilistic safety boundary is constructed, and an uncertainty weighting factor is introduced to dynamically adjust the probabilistic safety boundary according to the prediction variance. At the same time, combined with the patient's historical trend data, the probabilistic safety boundary is trend-corrected so that the probabilistic safety boundary can be dynamically adjusted with the changing trend of cardiac load.

7. The rehabilitation training device for elderly people after heart disease according to claim 6, characterized in that: The method for predicting the risk index of future adverse cardiac events based on probabilistic safety boundaries includes the following steps: S1.1 Based on a pre-trained lightweight meta-learning model, the time window features of real-time cardiac load status data are used as input, and a set of parameter offsets are output. S1.

2. The parameter offset is combined with the pre-trained state-space sequence model parameters to generate personalized state-space sequence model parameters for the current time step. S1.3 Input the data from another time window into the state space model after updating the parameters of the personalized state space sequence model, extract the latent state vector sequence of its output, and use the temporal consistency attention module to modulate the latent state vector sequence to generate the modulated output. S1.4 Based on the probabilistic safety boundary, the risk index of adverse cardiac events in the future time period is calculated using the modulated output, and the risk index is quantified and classified to generate a risk level.

8. The rehabilitation training device for elderly people after heart disease according to claim 7, characterized in that: In step S1.3, the temporal consistency attention module is used to modulate the potential state vector sequence to generate the modulated output, including the following steps: S1.

31. Within the current short-time window, extract the latent state vector sequence of the state-space sequence model. Simultaneously, the historical feature vector sequence of the previous equal-length time window is obtained; S1.32 Calculate the potential state vector sequence The dynamic time-warped distance between the sequence and the historical feature vector sequence; S1.33 Calculate the sequence of potential state vectors within the current short window. The average similarity between feature vectors at adjacent time points within the same timeframe; S1.

34. The dynamic time-warped distance and average similarity are fed into the micro neural network as input, and a window-level anti-interference weight between 0 and 1 is output. ; S1.35, Utilizing window-level anti-interference weights For the sequence of potential state vectors The sequence is modulated to obtain modulated features.

9. The rehabilitation training device for elderly people after heart disease according to claim 4, characterized in that: The assessment of elderly patients' performance and psychological changes based on subjective perception data includes the following steps: S2.1 Compare the subjective fatigue level with the rehabilitation training plan and calculate the execution deviation index; S2.

2. Weighted fusion of deviation indicators to generate a comprehensive execution score, reflecting the patient's adherence to the training plan; S2.3 Quantify subjective psychological feelings into psychological state indicators; S2.4 Combine the performance score with the psychological state index to form a comprehensive compliance and mental health score, and mark any abnormal situations.

10. The rehabilitation training device for elderly people after heart disease according to claim 1, characterized in that: The rehabilitation intervention unit (4) compares the predicted value of cardiac load with the probabilistic safety boundary. If it exceeds the probabilistic safety boundary, the intervention process is triggered. At the same time, the risk index is compared with the preset risk level threshold, and a risk level is generated based on the comparison result.

Images