A multi-dimensional heart attack health monitoring system based on multi-source heterogeneous physiological data

Abstract

The application relates to the technical field of intelligent medical health monitoring, and discloses a multi-dimensional heart attack health monitoring system based on multi-source heterogeneous physiological data, which comprises an edge computing node, a cloud analysis engine, an early warning service module and a visualization platform; the edge computing node collects data through a sliding window mechanism to generate an effective RR interval sequence; the cloud analysis engine extracts multi-dimensional heart rate variability features such as time domain and nonlinear dynamics, and calculates a risk probability by using a parallel integrated learning model containing a long short-term memory network and XGBoost; the early warning service module triggers tactile feedback or emergency alarm and other graded interventions according to a probability threshold; and the visualization platform generates a psychophysiological stress index and a trend report. The application realizes high-precision processing and deep feature mining of physiological signals, provides time-sensitive dynamic risk prediction and closed-loop intervention mechanism, and improves the accuracy and effectiveness of heart attack monitoring.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical health monitoring technology, specifically a multi-dimensional myocardial infarction health monitoring system based on multi-source heterogeneous physiological data. Background Technology

[0002] Cardiovascular and cerebrovascular diseases are the leading cause of death worldwide, characterized by their sudden onset and rapid progression. Therefore, early risk identification and continuous physiological monitoring are crucial for reducing the risk of sudden onset. Currently, with advancements in sensor technology, smart wearable devices such as ECG wristbands and smart ECG patches are widely used for real-time collection of users' daily physiological data.

[0003] However, while existing smart wearable devices have solved the problem of portable data collection, they still have limitations in practical monitoring and early warning applications. First, data quality is highly dependent on the acquisition environment. Raw electrophysiological signals are easily affected by power frequency interference, baseline drift, and motion artifacts caused by human movement. Traditional filtering methods are usually designed for specific frequencies and are not adaptable enough to complex and variable uncontrolled environmental noise, resulting in low signal-to-noise ratios in the acquired data, making it difficult to meet the requirements of high-precision analysis.

[0004] Secondly, in terms of feature mining, existing solutions mostly remain at the stage of superficial analysis. Most monitoring systems rely only on basic statistical indicators such as average heart rate and simple time-domain variability, failing to fully mine the autonomic nervous regulation information and nonlinear dynamic information reflecting system complexity contained in the heart rate variability signal, resulting in an incomplete assessment of the balance state of the cardiac autonomic nervous system.

[0005] Meanwhile, existing risk prediction methods are often static and simplistic. Current technologies lack the ability to model the time-dependent nature of physiological parameters and the nonlinear interactions of multiple factors, making it difficult to capture the dynamic patterns of physiological state evolution over time, thus hindering the achievement of early, dynamic, and personalized risk warnings.

[0006] Furthermore, current monitoring systems have not yet formed a complete closed-loop management mechanism. Most existing products only record data or provide simple threshold alarms, lacking integration of the entire process of monitoring, assessment, early warning, and intervention. In particular, they lack quantitative assessment of psychophysiological burden and risk-level-based tiered intervention methods, which limits the translational value of monitoring data in clinical decision support and emergency rescue. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a multi-dimensional myocardial infarction health monitoring system based on multi-source heterogeneous physiological data. It aims to solve the problems of large noise interference in raw data, limited feature mining, lack of time-series dynamic risk prediction, and lack of closed-loop intervention in existing cardiovascular and cerebrovascular health monitoring technologies.

[0008] To achieve the above objectives, the present invention provides a multi-dimensional myocardial infarction health monitoring system based on multi-source heterogeneous physiological data. The system includes an edge computing node, a cloud analysis engine, an early warning service module, and a visualization platform.

[0009] The edge computing node is configured to collect surface electrophysiological data at a preset sampling frequency through an intelligent acquisition terminal, and to divide the collected data into streaming data using a sliding time window mechanism; the edge computing node is also configured to perform adaptive signal denoising and multi-stage hybrid quality control on the streaming data in sequence, generate effective RR interval sequences and upload them.

[0010] The cloud analytics engine communicates with the edge computing node and is configured to receive the valid RR interval sequence and extract multi-dimensional heart rate variability features, which at least cover time-domain features and nonlinear dynamic features. The cloud analytics engine is also configured to construct a high-dimensional feature space based on the multi-dimensional heart rate variability features, basic heart rate indicators and user attribute data, and input it into a preset time-sensitive ensemble learning model to calculate the risk probability value of future time windows.

[0011] The early warning service module is connected to the cloud analysis engine and is configured to obtain the risk probability value, compare the risk probability value with the preset early warning threshold and emergency threshold respectively, and trigger graded intervention instructions based on the comparison results.

[0012] The visualization platform is connected to the cloud-based analysis engine and configured to calculate the psychophysiological stress index based on the time-domain features, and generate a longitudinal health trend report based on historical data.

[0013] Preferably, the adaptive signal denoising process performed by the edge computing node specifically includes: The streaming data is bandpass filtered using a finite impulse response filter to remove DC components and power frequency interference; the bandpass-filtered data is decomposed into multiple layers using Daubechies wavelet systems to obtain detail coefficients at each scale; the detail coefficients at each scale are processed using a soft threshold calculated based on the noise standard deviation estimate and the signal length; and the denoised signal is reconstructed using inverse wavelet transform.

[0014] Preferably, the multi-stage hybrid quality control performed by the edge computing node uses a one-dimensional convolutional neural network model to discriminate the truncated waveform segments. This one-dimensional convolutional neural network model employs a depthwise separable convolutional structure, with its input vector being time window data truncated centered on the peak time of the R-wave. Its output layer uses a sigmoid activation function to output waveform quality probabilities. The edge computing node only retains the RR intervals corresponding to waveforms whose waveform quality probabilities are greater than or equal to a preset quality threshold.

[0015] Preferably, the nonlinear dynamic features extracted by the cloud analysis engine include short-term and long-term variation indices obtained through Poincaré scatter plot analysis, and sample entropy calculated through vector similarity.

[0016] The short-term variation index is determined based on the variance of the first projection component perpendicular to the identity line in the scatter plot, and the long-term variation index is determined based on the variance of the second projection component along the identity line in the scatter plot; the sample entropy is the negative natural logarithm of the probability ratio of the distance between vectors being less than the similarity tolerance threshold when the vector dimension increases.

[0017] Preferably, the high-dimensional feature space constructed by the cloud-based analytics engine is composed of high-dimensional feature vectors, and the construction process includes: The time-domain features, nonlinear dynamic features, baseline heart rate indicators, and static user attribute data within the current time window are normalized and then vectorized and concatenated in a preset order. The user attribute data includes age, gender encoding, body mass index, and past medical history markers. Furthermore, the cloud-based analysis engine uses a tree-based algorithm to calculate the importance score of each feature variable in the high-dimensional feature vector and removes features with importance scores below a preset retention threshold, generating a simplified feature vector.

[0018] Preferably, the time-sensitive ensemble learning model employs a parallel processing architecture including a first channel and a second channel. The first channel uses a Long Short-Term Memory (LSTM) network to recursively process the feature matrix sequence containing historical time steps to output a hidden layer state vector representing the time evolution features. The second channel uses an XGBoost model to process the feature vector at the current time step, mapping the input sample to leaf node index encodings or feature embedding vectors representing multi-factor interaction features through decision tree splitting. The hidden layer state vector and the feature embedding vector are concatenated in a fusion layer. The model receives the concatenated fused feature vector through the output layer and calculates the risk probability value, normalized to the 0-1 interval, using a linear weighted unit and a sigmoid activation function. The risk probability value is generated continuously as the time window slides and the feature matrix sequence is updated.

[0019] Preferably, the visualization platform calculates the psychophysiological stress index in the following way: Calculate the product of the natural logarithm of the standard deviation of all RR intervals and the first weighting coefficient, and the product of the natural logarithm of the root mean square of the difference between adjacent RR intervals and the second weighting coefficient. Add the two products together and divide by the natural logarithm of the reference constant to obtain the ratio. Then calculate the product of 1 minus the ratio and 100. The sum of the first weighting coefficient and the second weighting coefficient is 1.

[0020] Preferably, the logic for the early warning service module to trigger tiered intervention commands includes: When the risk probability value is between the warning threshold and the emergency threshold, it is determined to be a Level 1 warning state and the edge computing node is driven to perform tactile feedback; when the risk probability value is greater than or equal to the emergency threshold, it is determined to be a Level 2 emergency state and the communication interface is called to send alarm information containing the current geographical location information to the preset emergency contact.

[0021] A second aspect of this invention provides a multi-dimensional myocardial infarction health monitoring method based on multi-source heterogeneous physiological data, the method comprising the following steps: Electrophysiological data of the body surface are collected by an intelligent acquisition terminal at a preset sampling frequency, and the collected data are divided into streaming data using a sliding time window mechanism. The convective data is sequentially subjected to adaptive signal denoising and multi-stage hybrid quality control to generate an effective RR interval sequence. Receive valid RR interval sequences and extract multidimensional heart rate variability features, which at least cover time-domain features and nonlinear dynamic features; A high-dimensional feature space is constructed based on multi-dimensional heart rate variability features, baseline heart rate indicators, and user attribute data, and then input into a pre-set time-sensitive ensemble learning model to calculate the risk probability value for future time windows. Obtain the risk probability value and compare it with the preset warning threshold and emergency threshold respectively. Trial intervention instructions are triggered based on the comparison results. The psychophysiological stress index is calculated based on time-domain characteristics, and a longitudinal health trend report is generated based on historical data.

[0022] This invention provides a multi-dimensional myocardial infarction health monitoring system based on multi-source heterogeneous physiological data. It has the following beneficial effects: 1. This invention introduces an adaptive signal denoising and multi-stage hybrid quality control mechanism at the edge computing node, effectively solving the problem of wearable devices being easily affected by noise interference in daily activities. By employing a finite impulse response filter combined with Daubechies wavelet soft thresholding, power frequency interference and baseline drift are filtered out. A one-dimensional convolutional neural network based on depthwise separable convolution is used to perform morphological verification of the waveform, eliminating motion artifacts while preserving the true pathological waveform. This cascaded processing strategy ensures the accuracy of the RR interval sequence input to the cloud analysis engine and reduces the false alarm rate caused by poor signal quality.

[0023] 2. This invention overcomes the limitation of traditional static models in capturing time dependencies by constructing a time-sensitive parallel ensemble learning architecture. On one hand, it utilizes Long Short-Term Memory (LSTM) channels to extract the historical evolution patterns of physiological features over time; on the other hand, it uses XGBoost channels to mine multi-factor nonlinear interaction relationships in a high-dimensional feature space. Combined with multi-dimensional feature inputs encompassing both temporal and nonlinear dynamics, it can output risk probability curves reflecting future time window states in real time, thereby more sensitively identifying subtle physiological abnormal fluctuations before a myocardial infarction.

[0024] 3. This invention enhances the clinical practical value of the monitoring system through the collaborative operation of the early warning service module and the visualization platform. The system calculates the psychophysiological stress index (PSI) based on heart rate variability characteristics, enabling quantitative tracking of the user's mental burden. Simultaneously, based on the comparison between the risk probability value and a preset threshold, differentiated intervention actions are automatically triggered: background recording occurs in the low-risk stage, tactile feedback reminders are provided via the terminal in the medium-risk stage, and emergency alarm information including geographical location is automatically sent in the high-risk stage. This tiered mechanism ensures timely notification of external rescue in sudden high-risk situations, while also supporting long-term health management. Attached Figure Description

[0025] Figure 1 This is a system framework diagram of the present invention; Figure 2 This is a schematic diagram of the method flow of the present invention.

[0026] 110. Edge computing node; 120. Cloud analytics engine; 130. Early warning service module; 140. Visualization platform. Detailed Implementation

[0027] The technical solutions in 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.

[0028] See attached document Figure 1 This invention provides a multi-dimensional myocardial infarction health monitoring system based on multi-source heterogeneous physiological data. The system is designed based on a cloud-based collaborative microservice architecture to achieve closed-loop management from data perception to risk warning.

[0029] The multi-dimensional myocardial infarction health monitoring system based on multi-source heterogeneous physiological data includes: an edge computing node 110, a cloud analysis engine 120, an early warning service module 130, and a visualization platform 140. The edge computing node 110, the cloud analysis engine 120, the early warning service module 130, and the visualization platform 140 interact with each other and transmit commands through a communication network.

[0030] Edge computing node 110 is deployed on the intelligent acquisition terminal side and configured to perform physiological data acquisition and preliminary processing. Edge computing node 110 continuously acquires surface electrophysiological data at a sampling rate of no less than 250Hz and divides the continuous data into streaming data blocks based on a preset sliding time window mechanism. Edge computing node 110 integrates an adaptive filter and a hybrid quality control unit to perform wavelet denoising and bandpass filtering on the raw signal, and uses rule-based filtering and a lightweight convolutional neural network waveform classification model to remove low-quality segments containing motion artifacts, generating and uploading high-quality RR interval sequences and feature waveform data.

[0031] The cloud analytics engine 120 is deployed on a cloud server and configured to receive data uploaded by the edge computing node 110 and perform deep learning computations. The cloud analytics engine 120 constructs a multi-dimensional heart rate variability feature extraction module and a dynamic risk prediction model. Based on the RR interval sequence, the cloud analytics engine 120 calculates temporal and nonlinear dynamic features. Temporal features include the standard deviation of all RR intervals (SDNN) and the root mean square (RMSSD) of the difference between adjacent RR intervals. Nonlinear dynamic features include the Poincaré scatter plot index and multi-scale entropy. The cloud analytics engine 120 further integrates these features with user attributes and baseline heart rate, inputting them into a time-sensitive model based on ensemble learning and long short-term memory networks, outputting the probability of myocardial infarction or other risk events occurring within a specific future time window.

[0032] The early warning service module 130 is deployed independently based on a microservice architecture and is configured to monitor the probability values ​​output by the cloud analysis engine 120 in real time. The early warning service module 130 has preset multi-level thresholds for low risk, medium risk, and high risk. When the probability value exceeds the corresponding threshold, the early warning service module 130 triggers a tiered early warning process and sends early warning information to the designated terminal.

[0033] The visualization platform 140 is configured to receive and display the psychophysiological stress index and risk trends calculated by the cloud analytics engine 120. The visualization platform 140 generates trend reports including a longitudinal timeline to help identify the user's periodic physiological patterns and long-term health status evolution.

[0034] See attached document Figure 2 The present invention provides a multi-dimensional myocardial infarction health monitoring method based on multi-source heterogeneous physiological data. The method is executed by a multi-dimensional myocardial infarction health monitoring system based on multi-source heterogeneous physiological data, and specifically includes steps S1 to S4.

[0035] Step S1: Synchronous acquisition and high-throughput preprocessing of multi-source heterogeneous physiological data; Edge computing node 110 acquires surface electrophysiological data at a high sampling rate through a smart wearable device, and uses a sliding time window to achieve data streaming. Within edge computing node 110, adaptive filters suppress power frequency interference and baseline drift, and a high-precision QRS detection algorithm is used to extract RR intervals from the signal. Subsequently, a multi-stage hybrid quality control strategy, including rule-based initial screening and lightweight convolutional neural network waveform classification, is used to identify and filter abnormal waveforms and artifacts, and the quality-controlled valid data is transmitted to the cloud analysis engine 120.

[0036] Step S2: Multidimensional Heart Rate Variability Deep Feature Extraction; After receiving valid data, the cloud analysis engine 120 extracts time-domain features reflecting autonomic nerve function, calculates SDNN to assess overall autonomic nerve tension, and calculates RMSSD to assess vagal nerve activity. Simultaneously, the cloud analysis engine 120 extracts nonlinear dynamic features, constructs a Poincaré scatter plot to calculate SD1 and SD2 indices, and calculates sample entropy or multi-scale entropy to quantify the complexity, stability, and nonlinear dynamic information of the cardiac regulatory system.

[0037] Step S3: Multi-factor dynamic risk prediction; The cloud-based analytics engine 120 fuses the multi-dimensional features extracted in Step S2 with baseline heart rate and user attribute data to construct a high-dimensional feature vector. A time-sensitive ensemble learning model is then used to process this high-dimensional feature vector. This model integrates the XGBoost algorithm for handling multi-factor interactions and the Long Short-Term Memory (LSTM) network for capturing time-dependent factors. This step ultimately outputs a value between 0 and 1, representing the probability of future cardiovascular and cerebrovascular risk events.

[0038] Step S4: Quantification of psychophysiological load and closed-loop intervention; a psychophysiological stress index is constructed based on SDNN and RMSSD, and the visualization platform 140 continuously tracks the index and generates a visualization report. The early warning service module 130 executes a graded intervention strategy based on the comparison result between the probability value output in step S3 and the preset threshold; when the probability value triggers the high-risk threshold, it immediately executes low-latency early warning push and emergency contact notification operations, forming a closed loop from monitoring to intervention.

[0039] Based on the aforementioned architecture and methodology of the multi-dimensional myocardial infarction health monitoring system based on multi-source heterogeneous physiological data, the data acquisition and preprocessing process in step S1 is mainly performed by the edge computing node 110, aiming to provide high signal-to-noise ratio time-series data for subsequent feature extraction.

[0040] In step S101, high-frequency signal acquisition and streaming window segmentation are performed. Edge computing node 110 is configured with an analog-to-digital conversion interface to perform the sampling at a set frequency. Discretized acquisition of simulated body surface electrophysiological signals.

[0041] In this embodiment, the sampling frequency The frequency is set to be no less than 250Hz. The technical basis for setting this threshold is that heart rate variability (HRV) analysis is highly dependent on the accuracy of the R-wave peak location; if the sampling rate is too low, the discretization position of the R-wave peak will produce a large quantization error, that is, introduce sampling jitter, which directly leads to a significant deviation in the subsequent calculation of time-domain indicators (such as RMSSD).

[0042] To meet real-time monitoring requirements, edge computing node 110 employs a sliding time window mechanism to process continuous data streams; the time window length is set to... The sliding step size is At any moment The data segments captured by the system can be represented as vectors. ,in, The time is usually set to 30 to 60 seconds to cover enough physiological cycles for short-term HRV analysis; The interval is set from 1 to 5 seconds to achieve near real-time data updates and overlapping calculations, ensuring that sudden abnormal waveforms are not missed.

[0043] In sub-step S102, adaptive signal denoising based on a cascaded architecture is performed. Edge computing node 110 processes the original signal... Bandpass filtering and wavelet thresholding are performed sequentially. First, a passband range of [value missing] is used. Up to 45 A finite impulse response (FIR) bandpass filter is used to process the original signal to remove DC components, low-frequency respiratory baseline drift, and high-frequency power line interference. Subsequently, wavelet transform is used to process the residual non-stationary noise (mainly broadband electromyographic noise) after bandpass filtering. Specifically, Daubechies wavelet systems (such as db4 or db6) are selected as basis functions to transform the signal. Layer decomposition ( Typically, a threshold of 4 to 6 is used. For the detail coefficients obtained from the decomposition at each scale, a soft thresholding function is applied. The threshold of the soft thresholding function is... The threshold is calculated using a general threshold formula, which is as follows: ;in This is an estimate of the noise standard deviation. This is the signal length.

[0044] Finally, the denoised signal is reconstructed using inverse wavelet transform. .

[0045] In sub-step S103, high-precision QRS group detection and RR interval sequence generation are performed. This is done to obtain the clean signal. Subsequently, edge computing node 110 applies a differential threshold-based detection algorithm (such as the improved Pan-Tompkins algorithm) to locate the R-wave peak. This process involves differentiating the signal to highlight the slope characteristics of the QRS complex, followed by squaring to nonlinearly amplify high-frequency components, and smoothing the waveform by moving an integral window to obtain the signal envelope. An adaptive threshold is set to perform peak search on the envelope signal to determine the peak value. The moment when the R wave appears Based on the R-wave time, generate the RR interval sequence. , of which The formula for calculating the RR interval is as follows: ; In the formula, Indicates the first The next heartbeat and the first The time interval between heartbeats and These are the detection times of two consecutive R-wave peaks, in milliseconds (ms).

[0046] The generated RR interval sequence will be used as the basic data input to the cloud analysis engine 120 for subsequent multi-dimensional feature extraction. The specific algorithm implementation details of QRS detection belong to digital signal processing technology well known to those skilled in the art, and will not be elaborated on here. However, its localized execution at the edge significantly reduces the bandwidth requirements for uploading raw data to the cloud.

[0047] Following high-precision QRS complex detection and RR interval sequence generation, to ensure that the data uploaded to the cloud analysis engine 120 accurately reflects the physiological state of the heart and eliminates complex interferences introduced by factors such as wearer movement and poor electrode contact, the edge computing node 110 further executes a multi-stage hybrid quality control strategy. This strategy aims to solve the problem that traditional single-rule filtering is unable to identify motion artifacts with morphological similarities to QRS waves, and is specifically implemented through sub-steps S104 to S105.

[0048] In sub-step S104, a preliminary rule screening based on physiological prior knowledge is performed. Edge computing node 110 reads the previously generated RR interval sequence. For each RR interval element Numerical logic verification is performed. This verification process includes two dimensions: absolute range constraints and relative rate of change constraints. For the absolute range constraints, a physiologically valid lower limit is set. (e.g., 300 milliseconds) and upper limit (e.g., 2000 milliseconds), when the condition is met or If the data point is an outlier, it is considered an outlier.

[0049] For the relative rate of change constraint, given the continuous nature of sinus rhythm changes over a short period, the ratio of the current interval to the previous effective interval is calculated. If the ratio exceeds a preset stable range (e.g., [0.8, 1.2]), it indicates a non-physiological change in heart rate, and the system marks the data point as an object to be removed. This step utilizes low-computational-cost logical judgment to quickly filter out significant outliers.

[0050] In sub-step S105, a depth waveform verification based on a lightweight convolutional neural network is performed. For data segments that have passed the initial rule screening but may contain morphological artifacts, the edge computing node 110 calls a pre-built lightweight one-dimensional convolutional neural network (1D-CNN) model for secondary discrimination. The construction of this model fully considers the computing power limitations of edge devices, and uses depthwise separable convolution or reduces the layer depth to reduce the number of parameters. In specific implementation, the peak time of the detected R-wave is used as the reference. Centered on the denoised signal The length of the cut-off is A time window segment (e.g., waveform data extracted 100 milliseconds before and after the R-wave) is used as the model's input vector. This input vector is then processed by multiple one-dimensional convolutional layers within the model for feature extraction. The convolutional layers use sliding convolution kernels to capture local morphological features of the waveform, such as the steepness, width, and symmetry of the QRS complex wave. Subsequently, dimensionality reduction is performed through pooling layers to preserve key features and obtain translation invariance. Finally, the extracted high-dimensional features are mapped to the classification space through fully connected layers.

[0051] The output layer of this lightweight CNN model uses either the Softmax or Sigmoid function to output a quality confidence score between 0 and 1. This represents the probability that the input waveform belongs to a real heartbeat waveform. Edge computing node 110 has a set quality threshold. (e.g., 0.8), when calculated At that time, the currently detected R wave is determined to be a motion artifact or a noise false detection, and the corresponding R wave is then removed from the RR interval sequence. The sequence is then interpolated or marked with breakpoints. Only data that passes both the initial screening in sub-step S104 and the in-depth verification in sub-step S105 is confirmed as a valid data packet and uploaded to the cloud analysis engine 120 via the communication network for subsequent advanced feature analysis. This hybrid quality control mechanism effectively balances the false negative rate and false positive rate while ensuring data purity, forming a data foundation for the system's high reliability.

[0052] When the cloud analytics engine 120 receives a valid RR interval sequence uploaded via the edge computing node 110 and passed hybrid quality control, in order to quantify the basic regulatory function of the autonomic nervous system from a linear statistical dimension, the system executes the time-domain feature extraction process in step S2.

[0053] In sub-step S201, the standard deviation (SDNN) of all RR intervals is calculated. The cloud analysis engine 120 analyzes the current analysis window... Statistical analysis was performed on the effective RR interval data. The SDNN index reflects the overall dispersion of heart rate within the monitoring period and is a key indicator for assessing the total tension of the cardiac autonomic nervous system (including the combined effects of the sympathetic and parasympathetic nervous systems). Its calculation formula is expressed as follows: ; In the formula, This indicates the total number of valid RR intervals within the current time window. Indicates the first The value of each RR interval. This represents the arithmetic mean of all RR intervals within the window.

[0054] The calculated SDNN values ​​characterize the fluctuation range of the heart rate interval around the mean. When the SDNN value decreases significantly, it indicates that the autonomic nervous system's ability to regulate the heart is weakened. This decrease in regulatory ability is pathophysiologically associated with the stress state of the cardiovascular system and the risk of myocardial infarction.

[0055] In substep S202, the root mean square (RMSSD) of the difference between adjacent RR intervals is calculated. Unlike SDNN, which reflects overall variability, RMSSD focuses on capturing rapid, beat-by-beat changes in the heartbeat cycle. Because the release and clearance of neurotransmitters in the sympathetic nervous system are slow, its regulatory effects are unlikely to cause rapid, beat-by-beat fluctuations in heart rate. In contrast, the vagus nerve's regulatory response is rapid. Therefore, RMSSD is used as a specific indicator for assessing parasympathetic (i.e., vagus nerve) tone, as it is highly sensitive to high-frequency changes in heart rate and less affected by slow baseline trends. The cloud analytics engine 120 performs the calculation using the following formula: ; In the formula, and These represent the values ​​of two temporally adjacent RR intervals in the sequence. This represents the difference between adjacent intervals. By calculating the square root of the squared mean of this difference, the amplitude of rapid heart rate modulation caused by physiological processes such as respiratory sinus arrhythmia can be effectively quantified. The cloud-based analysis engine 120 uses the calculated SDNN and RMSSD values ​​as basic time-domain feature vectors, temporarily storing them in the feature database, to be fused with the subsequently extracted nonlinear dynamic features in multiple dimensions.

[0056] While the aforementioned time-domain indicators provide a basic assessment of autonomic nervous system tension, the human cardiovascular system possesses inherently nonlinear and chaotic characteristics. Relying solely on linear statistical indicators is insufficient to comprehensively capture the subtle pathological changes in the cardiac dynamics system during the latency period of a myocardial infarction. Therefore, after completing the time-domain analysis, the cloud-based analysis engine 120 further performs nonlinear dynamic feature extraction. This process is specifically implemented through sub-steps S203 and S204, aiming to quantify the geometric structure and information complexity of heart rate regulation in multiple dimensions.

[0057] In sub-step S203, a Poincaré scatter plot analysis is performed to extract geometric features SD1 and SD2. The cloud-based analytics engine 120 uses the current RR interval sequence... The x-axis is... Using the ordinate as the vertical axis, a regression plot is constructed in a two-dimensional Cartesian coordinate system. The geometric distribution of this scatter plot visually reflects the correlation structure of heart rate variability. To quantify this structure, the system calculates ellipse fitting parameters through coordinate axis rotation transformation. A line perpendicular to the identity line (i.e., ...) is defined. The direction of the line is the short axis direction, and the direction along the identity line is the long axis direction.

[0058] The system first maps the original RR interval coordinates to the new principal axis coordinate system using the following linear transformation formula: ; ; In the formula, The projection component of the data point in the direction perpendicular to the identity line. This represents the projection component of the data point along the identity line.

[0059] Based on the above projection components, calculate the vertical axis standard deviation SD1 and the axial standard deviation SD2: ; ; In the formula, This represents variance calculation. SD1 primarily reflects short-term rapid variability in heart rate, i.e., the difference between beats, and its physiological significance is positively correlated with the transient modulation effect of the vagus nerve on heart rhythm; SD2 reflects long-term variability in heart rate, encompassing the combined slow regulatory effects of the sympathetic and parasympathetic nervous systems. The ratio of SD1 to SD2 (SD1 / SD2) is further calculated to assess the relative balance between sympathetic and vagal activity in the autonomic nervous system.

[0060] In sub-step S204, entropy analysis is performed to quantify the complexity and stability of the time series. For short data segments and noise interference commonly found in physiological signals, this embodiment preferably employs the sample entropy algorithm, which offers better consistency than approximate entropy and does not calculate self-matching.

[0061] Cloud analytics engine 120 setting embedding dimensions (Usually a value of 2) and similarity tolerance threshold (Typically, the value is taken as 0.15 to 0.25 times the sequence standard deviation SDNN). The algorithm first reconstructs the RR interval sequence into a series of m-dimensional vectors. ... ,in Define two vectors. and Distance between This is the maximum absolute value of the difference between corresponding elements of the two pairs (i.e., the Chebyshev distance). For each Statistical satisfaction of distance of The number of (excluding) (The situation), denoted as And calculate its average value. .

[0062] Subsequently, the dimension was increased to Repeat the above process to statistically satisfy the condition that the distance is less than Count the number of vector pairs and calculate the average. Finally, the formula for calculating sample entropy is as follows: ; In the formula, It is the natural logarithm function. The calculated value is... Numerical values ​​are used to characterize the probability of new patterns emerging from the RR interval sequence. Lower sample entropy values ​​indicate that the heart rate sequence has higher regularity and predictability, which is usually associated with the "rigidity" of the cardiac regulatory system or pathological monochromatic patterns (such as heart failure or age-related heart rate monotonicity); higher sample entropy values ​​indicate that the heart rate sequence has higher complexity and randomness, reflecting the cardiovascular system's good adaptability to changes in the external environment and a healthy nonlinear regulatory mechanism.

[0063] Through the above calculations, the system transforms the chaotic features hidden in the RR interval sequence into measurable numerical indicators, providing an independent feature dimension in addition to linear statistics for subsequent risk prediction.

[0064] After completing the deep extraction of multi-dimensional HRV features in step S2, in order to overcome the limitations of a single data source in cardiovascular disease prediction and achieve accurate quantification of myocardial infarction risk, the cloud analysis engine 120 further executes step S3 to construct a multi-factor dynamic risk prediction engine.

[0065] In sub-step S301, vectorized concatenation and spatial construction of multi-source heterogeneous data are performed. The cloud analytics engine 120 first retrieves the current time window from the database. The various physiological indicators calculated internally include: the time-domain feature set obtained in step S2. The nonlinear dynamic feature set obtained from step S2 ; and the baseline heart rate index obtained from the preprocessing stage in step S1. ,in , and These represent the average heart rate, minimum heart rate, and maximum heart rate within the window, respectively.

[0066] In addition, to capture the impact of individual differences on the baseline risk of myocardial infarction, the system incorporates static user attribute data. The dataset includes age, gender codes, body mass index (BMI), and medical history markers (such as hypertension and diabetes, encoded using 0 / 1 binary codes). Because these features have different physical dimensions and numerical ranges, the cloud-based analytics engine 120 first performs Z-score standardization on the continuous variable features, making their mean 0 and standard deviation 1. Subsequently, the system concatenates the standardized features in a preset order to construct the time-series data. high-dimensional feature vectors : ; The vector It integrates dynamic temporal features reflecting the instantaneous state of the heart with static attribute features reflecting the individual's health background, forming a multi-source heterogeneous high-dimensional space that describes the user's current cardiovascular health status.

[0067] In sub-step S302, feature selection optimization based on importance ranking is performed. Given the initially constructed high-dimensional feature vector... The feature space may contain redundant information or noisy features (such as certain highly collinear time-domain indicators), and directly inputting it into the prediction model may lead to overfitting or wasted computational resources. The cloud-based analytics engine 120 uses a tree-based feature selection strategy to reduce the dimensionality of the feature space.

[0068] Specifically, the importance of features is evaluated using pre-trained random forest or gradient boosting tree models, and the average contribution of each feature to the reduction of Gini impurity during the decision tree split is calculated.

[0069] Based on the calculated importance score, the system sets a screening threshold and removes non-critical features with scores below the threshold, thereby obtaining an optimized and simplified feature vector. This process ensures that the retained features (such as SDNN, SampEn, and age) have a significant pathophysiological correlation with myocardial infarction risk, while removing confounding variables that contribute little to the prediction. After feature selection optimization... It will be used as standard input and fed into subsequent time-sensitive ensemble learning models for risk reasoning.

[0070] To obtain a simplified feature vector optimized by feature selection. Subsequently, the cloud-based analytics engine 120 further executes the core step in step S3, namely, constructing and running a time-sensitive ensemble learning model. This process aims to solve the technical challenge of a single model simultaneously taking into account the nonlinear interactions of multiple factors and the long-term and short-term time dependencies. Through the specific implementation of sub-steps S303 and S304, it achieves dynamic and accurate prediction of myocardial infarction risk.

[0071] In sub-step S303, a hybrid model architecture incorporating parallel processing branches is constructed. The cloud analytics engine 120 designs a topology that integrates a gradient boosting tree (XGBoost) and a long short-term memory network (LSTM) in parallel. This structure contains two independent data stream processing channels: The first channel is a time-dependent capture channel, employing an LSTM network. The input to this channel is not only the feature vector at the current time step. It is not a sequence of feature matrices containing historical time steps.

[0072] ,in This represents the number of time steps to backtrack (e.g., backtracking 30 time windows). The LSTM network uses its internal forget gate, input gate, and output gate mechanisms to process the sequence... Recursive processing can effectively memorize and transmit the evolution trend of heart rate variability over time (such as the continuous downward trend of SDNN).

[0073] This channel ultimately outputs a hidden layer state vector. This vector encodes the dynamic evolution pattern of physiological parameters over time in a high-dimensional way.

[0074] The second channel is a non-linear interactive mining channel, employing the XGBoost model. This channel uses statistical feature vectors within the current timeframe or a short time window. As input, the XGBoost ensemble gradient boosting algorithm, based on decision trees, is used to automatically find non-linear combinations of different features by splitting nodes (e.g., the risk multiplication effect when "elderly" and "low RMSSD" coexist). The XGBoost model does not directly output classification results here; instead, it is configured to output the index encoding of the leaf nodes or transformed feature embedding vectors. This vector represents the complex lateral interaction characteristics of multiple factors under the current physiological state.

[0075] Subsequently, the system sets up a fully connected fusion layer to combine the vectors output from the two channels. and The features are concatenated to form a fused feature vector. This fusion vector carries both longitudinal time evolution information and lateral factor interaction information, providing comprehensive data support for the final risk judgment.

[0076] In sub-step S304, dynamic probability calculation and rolling updates are performed. The cloud analytics engine 120 will fuse the feature vectors. The input is fed to the output layer, which contains a linearly weighted unit and a non-linear activation function. The system uses the sigmoid activation function to map the model's output to the (0, 1) interval and generates standardized risk probability values ​​using a probability value formula. The probability formula is as follows: ; In the formula, Indicates at time Predict the probability of myocardial infarction or malignant cardiovascular events occurring within a specific future time window (e.g., the next hour); This is the weight matrix of the fusion layer. For bias terms, It is a natural constant.

[0077] This prediction mechanism is not executed all at once, but rather triggered in a rolling manner over time. As the edge computing node 110 continuously uploads new data streams, the time window increases in increments. (For example, every minute) slide forward. Whenever a new feature vector is generated. The model automatically updates the input sequence. Then, re-execute the above reasoning process and output the updated probability value. This mechanism means that the system outputs not a static diagnostic conclusion, but a risk probability curve that fluctuates continuously over time. It can keenly capture the accumulation process of risk from low to high, thus providing timely trigger signals for subsequent early warning modules when the probability value suddenly changes or continues to rise.

[0078] In order to transform the output of the above complex mathematical model into intuitive indicators with clinical guidance and to achieve an effective closed loop from monitoring to intervention, the multi-dimensional myocardial infarction health monitoring system based on multi-source heterogeneous physiological data further executes step S4.

[0079] In sub-step S401, the cloud analysis engine 120 constructs a psychophysiological stress index based on heart rate variability characteristics. This index aims to quantify the user's current sympathetic-vagal balance and psychophysiological load level. The cloud analysis engine 120 calls the overall autonomic nervous tension index SDNN and the vagal nerve activity index RMSSD calculated in step S2. Considering that the heart rate variability index is negatively correlated with stress level (i.e., the higher the stress, the lower the HRV), and that baseline values ​​differ among individuals, the system uses a weighted inverse normalization algorithm based on logarithmic transformation for calculation. The calculation formula is as follows: ; In the formula, This represents the natural logarithm operation, used to correct the skewed distribution of HRV data; and These are weighting coefficients, and they satisfy... , usually set Greater than This gives higher weight to vagal nerve activity, thus making it more sensitive to relaxation and recovery abilities; The reference constant is taken from the maximum logarithm of HRV in a large sample database of healthy people or the maximum historical baseline of the user.

[0080] Calculated The value is between 0 and 100. The higher the value, the heavier the user's current psychological and physiological load and the lower the autonomic nervous system regulation reserve.

[0081] In sub-step S402, the early warning service module 130 executes a hierarchical early warning and closed-loop intervention strategy based on dynamic probability. This module monitors the risk probability value output in step S3 in real time. It then compares this with preset multi-level risk thresholds. This embodiment sets two key thresholds: a warning threshold... (e.g., 0.60) and emergency threshold (e.g., 0.85).

[0082] when When the system is deemed to be in a low-risk state, it will only record data in the background without disturbing the user. when At that time, it was determined to be a medium-risk state (warning level).

[0083] The early warning service module 130 generates a first-level intervention command and sends a tactile feedback command to the edge computing node 110 through the communication network. This drives the smart wearable device to generate a vibration reminder at a specific frequency and pushes health guidance information such as "suggest rest" or "take deep breaths" through the accompanying mobile terminal application.

[0084] when At that time, the situation is determined to be high-risk (emergency response level). The early warning service module 130 immediately triggers the secondary intervention process: First, an alarm is pushed to the user's terminal with high priority and the user is required to confirm their state of consciousness. Second, if no confirmation is received from the user within a preset time (e.g., 60 seconds), or if the risk probability continues to rise, the module automatically calls the SMS or automatic voice call interface (API) to send a help request message containing the current GPS location information, real-time heart rate data, and risk probability to the preset emergency contact or medical emergency center.

[0085] This tiered mechanism ensures that intervention measures are matched with risk levels, avoiding "alarm fatigue" caused by frequent false alarms and guaranteeing timely response during acute attacks.

[0086] In sub-step S403, the visualization platform 140 generates a longitudinal tracking report and periodic pattern recognition results based on accumulated historical data. The visualization platform 140 extracts continuous PSI index sequences and risk probabilities from the database. The system generates a time-based trend curve. Its built-in periodic analysis algorithm processes long-term data to identify the user's circadian rhythm patterns.

[0087] For example, if the system detects that a user's PSI index remains consistently high during nighttime sleep and the SDNN does not show the expected nighttime rebound (Non-dipper pattern), the system will mark this abnormal pattern in the trend report, indicating that the user has a nighttime autonomic nervous system recovery dysfunction. The visualization platform 140 integrates the above quantitative data, risk event records, and trend analysis results into a visualized health dashboard, which can be accessed by clinicians to assist in diagnosis, or used by users for long-term lifestyle adjustments and health management, thus realizing a complete technological closed loop from data collection, real-time analysis, immediate early warning to long-term management.

Claims

1. A multi-dimensional myocardial infarction health monitoring system based on multi-source heterogeneous physiological data, characterized in that, include: An edge computing node is configured to perform real-time acquisition of body surface electrophysiological data and generate streaming data through a sliding time window mechanism; the edge computing node is also used to perform adaptive signal denoising and multi-stage hybrid quality control on the streaming data, generate valid RR interval sequences and upload them. The multi-stage hybrid quality control uses a one-dimensional convolutional neural network model to distinguish the extracted waveform segments. The one-dimensional convolutional neural network model adopts a depth-separable convolutional structure. The input vector is time window data extracted with the peak time of the R-wave as the center. The output layer uses the Sigmoid activation function to output the waveform quality probability. The edge computing node retains only the RR interval sequence corresponding to waveforms whose waveform quality probability is greater than or equal to a preset quality threshold; A cloud-based analytics engine is communicatively connected to the edge computing node and configured to receive the valid RR interval sequence; the cloud-based analytics engine extracts multi-dimensional heart rate variability features, which include at least time-domain features and nonlinear dynamic features; The cloud-based analytics engine is also configured to construct a high-dimensional feature space based on the multi-dimensional heart rate variability features, and to calculate the risk probability value of future time windows using a pre-set time-sensitive ensemble learning model and probability value formula. The time-sensitive ensemble learning model employs a parallel processing architecture that includes a first channel and a second channel; The first channel uses a long short-term memory network to recursively process the feature matrix sequence containing historical time steps and output the hidden layer state vector representing the time evolution features. The second channel uses the XGBoost model to process the feature vector at the current time step. It maps the input sample into a leaf node index encoding or feature embedding vector that represents the multi-factor interaction features through decision tree splitting. The hidden layer state vector and the feature embedding vector are concatenated in the fusion layer. The time-sensitive ensemble learning model receives the concatenated fused feature vector through the output layer, normalizes the output of the ensemble learning model to the interval between 0 and 1 using a linear weighted unit and a Sigmoid activation function, and calculates the risk probability value using a probability value formula. The risk probability value is generated in a rolling manner as the time window slides and the feature matrix sequence is updated, forming a continuously changing risk probability curve; The early warning service module is connected to the cloud analysis engine and is configured to obtain the risk probability value, compare the risk probability value with a preset multi-level risk threshold, and trigger corresponding intervention instructions based on the comparison result. When the risk probability value is between the warning threshold and the emergency threshold, it is determined to be a level one warning state and the edge computing node is driven to perform tactile feedback. When the risk probability value is greater than or equal to the emergency threshold, it is determined to be a level two emergency state and the communication interface is invoked to send an alarm message containing the current geographical location information to the preset emergency contact. The visualization platform is connected to the cloud-based analytics engine and configured to construct a psychophysiological stress index based on heart rate variability characteristics and generate a longitudinal health trend report.

2. The multi-dimensional myocardial infarction health monitoring system based on multi-source heterogeneous physiological data according to claim 1, characterized in that, The adaptive signal denoising process includes: First, a finite impulse response filter is used to perform bandpass filtering on the streaming data to remove DC components and power frequency interference. Subsequently, the bandpass filtered data is decomposed into multiple layers using the Daubechies wavelet system to obtain detail coefficients at each scale. The soft threshold is calculated using the threshold formula, and the detail coefficients at each scale are then processed. Finally, the denoised signal is reconstructed using inverse wavelet transform.

3. The multi-dimensional myocardial infarction health monitoring system based on multi-source heterogeneous physiological data according to claim 1, characterized in that, The nonlinear dynamic features extracted by the cloud analysis engine include short-term and long-term variation indices obtained through Poincaré scatter plot analysis, as well as sample entropy calculated using the sample entropy calculation formula. The short-term variation index is determined based on the variance of the first projection component perpendicular to the identity line in the scatter plot, and the long-term variation index is determined based on the variance of the second projection component along the identity line in the scatter plot; the sample entropy is the negative natural logarithm of the probability ratio of the distance between vectors being less than the similarity tolerance threshold when the vector dimension increases.

4. The multi-dimensional myocardial infarction health monitoring system based on multi-source heterogeneous physiological data according to claim 1, characterized in that, The high-dimensional feature space is composed of high-dimensional feature vectors, and the construction process of the high-dimensional feature vectors includes: The time-domain features, nonlinear dynamic features, baseline heart rate indicators, and static user attribute data within the current time window are normalized and then vectorized and concatenated in a preset order; wherein, the user attribute data includes age, gender code, body mass index, and past medical history markers.

5. A multi-dimensional myocardial infarction health monitoring system based on multi-source heterogeneous physiological data according to claim 4, characterized in that, The cloud-based analytics engine also includes feature selection optimization logic, which uses a tree-based algorithm to score the importance of each feature variable in the high-dimensional feature vector and removes features whose importance scores are lower than a preset retention threshold, thereby generating a simplified feature vector as input to the time-sensitive ensemble learning model.

6. The multi-dimensional myocardial infarction health monitoring system based on multi-source heterogeneous physiological data according to claim 1, characterized in that, The psychophysiological stress index is calculated as follows: Calculate the product of the natural logarithm of the standard deviation of all RR intervals and the first weighting coefficient, and the product of the natural logarithm of the root mean square of the difference between adjacent RR intervals and the second weighting coefficient. Add the two products together and divide by the natural logarithm of the reference constant to obtain the ratio. Then calculate the product of 1 minus the ratio and 100. Wherein, the sum of the first weighting coefficient and the second weighting coefficient is 1.

Images