A Real-Time Cardiac Risk Assessment and Early Warning System Based on Edge Computing and Deep Learning

Abstract

This invention discloses a real-time cardiac risk assessment and early warning system based on edge computing and deep learning. Deployed on a computing-constrained edge terminal, the system includes a signal acquisition and spatial decomposition module, an edge-side feature recognition and segmentation module, an automated biomechanical index calculation module, an edge-side adaptive computing power allocation module, and a multi-dimensional risk scoring and graded early warning module. The system establishes a lightweight data foundation by decomposing ultrasound radiofrequency signals and extracts cardiac chamber edge contour data using a lightweight neural network. Subsequently, strain characteristic parameters are calculated and three-dimensional spatial volume reconstruction is performed to obtain atrioventricular coupling characteristic parameters. The system monitors the time consumption in real time and adaptively allocates computing power. Finally, it fuses various characteristic parameters and clinical correction factors to generate a quantitative risk score and execute graded alerts. This invention can monitor and quantitatively assess diastolic dysfunction and remodeling status in real time before a decrease in ejection fraction using limited computing power.

Description

Technical Field

[0001] This invention relates to the field of medical image processing and artificial intelligence interdisciplinary application technology, and in particular to a real-time cardiac risk assessment and early warning system based on edge computing and deep learning. Background Technology

[0002] Cardiac risk assessment, especially the quantitative analysis of diastolic function and remodeling status, plays a fundamental role in early screening for cardiovascular diseases. Clinically, echocardiography is typically used to obtain cardiac structural and kinematic data. With the widespread availability of portable ultrasound devices and the development of point-of-care ultrasound (POC) testing, rapid acquisition and analysis of cardiac ultrasound signals in medical settings or pre-hospital emergency environments has become a growing trend. Currently, cardiac structural segmentation and deformation analysis largely rely on manual delineation or semi-automatic measurement by physicians. This process is time-consuming, and the stability of the results is somewhat limited by the operator's experience.

[0003] In recent years, deep learning and biomechanical computing technologies have been gradually applied to the automated analysis of medical images, providing a technological foundation for solving the aforementioned time-consuming problems. Conventional deep learning network models and 3D volume reconstruction algorithms typically have a large number of parameters and high computational complexity, often requiring deployment on cloud servers or large workstations with high-performance processors to ensure operational efficiency. However, in practical portable detection scenarios, the computing power, memory bandwidth, and power consumption of edge terminal devices are all limited.

[0004] When performing complex tasks such as cardiac cavity feature extraction, optical flow tracing calculation, and 3D volume reconstruction directly on edge terminals with limited computing power, the relatively high computational load can easily lead to increased system processing latency, resulting in dropped images or sluggish interface operation. Simultaneously, ultrasound radio frequency signals involve large amounts of data. Without lightweight data processing and dynamic computing power scheduling mechanisms tailored to the hardware characteristics of edge devices, the system struggles to synchronously complete data flow and index calculation within the set time window. Therefore, balancing algorithm execution capability and end-to-end real-time performance under limited edge computing power conditions, and achieving objective risk warnings based on multi-dimensional biomechanical parameters, are technical challenges that require further research in this field. Summary of the Invention

[0005] To address the technical challenges of limited computing power at the edge terminal and the difficulty in real-time calculation and comprehensive early warning of complex cardiac biomechanical indicators in portable ultrasound testing scenarios, this invention provides a real-time cardiac risk assessment and early warning system based on edge computing and deep learning.

[0006] This invention provides a real-time cardiac risk assessment and early warning system based on edge computing and deep learning, deployed on an edge terminal with limited computing power. The system includes: The signal acquisition and spatial decomposition module is used to acquire the raw ultrasound radio frequency signal and decompose it into sub-data blocks according to spatial frequency in order to establish a lightweight data foundation. The edge feature recognition and segmentation module is used to receive the sub-data block through a lightweight neural network model and extract the edge contour data of the inner cavity of the same cardiac cycle on multiple orthogonal planes in real time and synchronously. The automated biomechanical index calculation module is used to calculate the instantaneous perimeter deformation of the edge contour data based on optical flow tracing technology to obtain strain characteristic parameters, and to perform three-dimensional spatial volume reconstruction based on the cross-sectional diameter and dynamic slice layer number on the multiple orthogonal planes to obtain compartment coupling characteristic parameters. The edge-side adaptive computing power allocation module is configured to monitor the segmentation time, computation time, and rendering time of a single frame of data in the edge-side terminal in real time, and construct a computing power constraint model based on the maximum allowable total time threshold of the target frame rate. When the real-time total time approaches the maximum allowable total time threshold, the computing power allocation module prioritizes reducing the network depth or input resolution of the lightweight neural network model to compress the segmentation time, and secondarily reduces the number of dynamic slice layers in the three-dimensional spatial volume reconstruction to compress the computation time, thereby ensuring that the end-to-end latency is controlled within a preset range. The multi-dimensional risk scoring and graded early warning module is used to perform nonlinear multi-source fusion of the extracted feature parameters and clinical correction factors, map them to generate quantitative risk scores, and execute graded alerts.

[0007] Optionally, the strain characteristic parameters include the three-phase strain index of the left atrium; The automated biomechanical index calculation module extracts feature pixels on the edge contour data based on the assumption of constant brightness, calculates the two-dimensional displacement vector field and performs filtering and reconstruction to obtain the instantaneous perimeter trajectory between consecutive frames.

[0008] Optionally, the automated biomechanical index calculation module, when acquiring the three-phase strain index of the left atrium: The end-diastolic endothelial circumference of the ventricle was calibrated as the baseline length. Extract the proportionality coefficient of the difference between the instantaneous circumference and the reference length relative to the reference length, and use it as the real-time strain curve data; Based on the slope changes of the electrocardiogram signal and the real-time strain curve data, the maximum strain peak value during the reserve period, the strain difference during the pipeline period, and the strain value before active contraction are automatically extracted in stages.

[0009] Optionally, the atrioventricular coupling characteristic parameters include the left atrioventricular coupling index; The automated biomechanical index calculation module calculates the instantaneous volume of the heart chamber by accumulating the cross-sectional diameter and the number of dynamic slice layers using a dual-plane disk accumulation algorithm. By capturing extreme values ​​within a specific cardiac cycle window, the maximum volume of the left atrium and the end-diastolic volume of the left ventricle are obtained respectively, and the ratio of the two is defined as the left atrioventricular coupling index.

[0010] Optionally, when monitoring and adjusting, the edge-side adaptive computing power allocation module aims to maximize the feature point sampling frequency per unit area to ensure that the highest precision biomechanical feature curve is output within the currently available terminal computing power boundary.

[0011] Optionally, the clinical correction factors include obesity factors characterized by body mass index, age-related decline factors, and atrial fibrillation history factors.

[0012] Optionally, the multi-dimensional risk scoring and graded early warning module is configured with a scoring function, which calculates the first deviation between the currently measured strain characteristic parameters and the corresponding patient's health reference values, and calculates the second deviation between the measured atrioventricular coupling characteristic parameters and the health reference values. The strain characteristic parameters are used as negative index weights, the atrioventricular coupling characteristic parameters are used as positive index weights, and the clinical correction factors are combined and weighted to output a single comprehensive risk score ranging from 0 to 100.

[0013] Optionally, the tiered alarm is executed based on a set three-level threshold logic: A green indicator suggesting good coupling is displayed when the value is below the first threshold. A yellow indicator is displayed between the first and second thresholds, prompting a detailed inspection. A red flashing warning is displayed if the value exceeds the second threshold, indicating a high likelihood of heart failure.

[0014] Optionally, the system has a built-in self-verification mechanism. The maximum strain peak value extracted by the biomechanical index automated calculation module during the reserve period must be equal to the sum of the strain difference during the pipeline period and the strain value during the contraction period, so as to be used for background physical verification of the algorithm results.

[0015] Optionally, the edge-side terminal is wirelessly connected to a smart ultrasound probe to monitor and quantify cardiac diastolic dysfunction and remodeling status in real time before the traditional ejection fraction decline.

[0016] The present invention has achieved the following beneficial effects: This invention reduces the dimensionality load of subsequent processing by deploying a signal acquisition and spatial decomposition module to spatially decompose the raw ultrasound radio frequency signal and construct sub-data blocks at the front end, thus adapting to the lightweight requirements of portable devices. The system uses an edge-side feature recognition and segmentation module to simultaneously extract the edge contour data of the heart chambers on multiple orthogonal planes, and combines it with a biomechanical index automated calculation module to obtain strain characteristic parameters and atrioventricular coupling characteristic parameters, realizing the automated calculation of cardiac deformation and volume three-dimensional reconstruction indicators. To address the limited computing power of terminal devices, the system introduces an edge-side adaptive computing power allocation module, which monitors the time consumption of each processing step in real time and dynamically adjusts the structural complexity of the lightweight neural network model and the number of three-dimensional reconstruction slice layers based on the target frame rate constraint, ensuring end-to-end latency control and maintaining the smooth operation of the system and the real-time output of results. The multi-dimensional risk scoring and graded early warning module fuses and maps biomechanical parameters with clinical correction factors to generate quantitative risk scores and trigger graded alarm mechanisms, providing objective data references and warnings for assessing cardiac diastolic dysfunction and remodeling status before the ejection fraction decreases.

[0017] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a system module structure diagram of a real-time cardiac risk assessment and early warning system based on edge computing and deep learning, as described in an embodiment of the present invention. Figure 2 This is a flowchart of the signal acquisition and spatial decomposition module in an embodiment of the present invention; Figure 3 This is a flowchart of the edge-side feature recognition and segmentation module in an embodiment of the present invention; Figure 4 This is a flowchart of the automated calculation module for biomechanical indicators in an embodiment of the present invention. Detailed Implementation

[0020] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] This invention provides a real-time cardiac risk assessment and early warning system based on edge computing and deep learning. The system is deployed in an edge terminal in the form of software, hardware, or a combination of both. The edge terminal includes, but is not limited to, portable computing devices with independent processing capabilities, storage media, and graphics display output interfaces, such as smartphones, tablets, or dedicated handheld medical testing devices. The edge terminal establishes a data transmission link with an external intelligent ultrasound probe via a wireless communication interface (e.g., Wi-Fi, Bluetooth, or other radio frequency communication protocols).

[0022] The system is configured to receive continuous ultrasound signal data streams and process them in a multi-level pipeline, ultimately outputting quantified cardiac physiological evaluation parameters and corresponding status indication signals. To achieve this data flow and parameter processing, such as... Figure 1 As shown, the system is logically divided into five main execution modules: signal acquisition and spatial decomposition module, edge-side feature recognition and segmentation module, biomechanical index automated calculation module, edge-side adaptive computing power allocation module, and multi-dimensional risk scoring and graded early warning module.

[0023] Specifically, such as Figure 2 As shown, the execution logic of the signal acquisition and spatial decomposition module is configured to acquire the raw ultrasound radio frequency signal and perform pre-processing data format conversion and spatial feature stripping. During operation, the intelligent ultrasound probe emits an ultrasound beam towards the target area and receives the echo, forming the raw ultrasound radio frequency signal. This raw ultrasound radio frequency signal is transmitted to the edge terminal in real time via a wireless communication interface in the form of a one-dimensional time series or a two-dimensional array electrical signal. After receiving the raw ultrasound radio frequency signal, the signal acquisition and spatial decomposition module first writes the analog signal or unformatted digital signal sequence into a preset circular memory buffer. Subsequently, the module executes a spatial frequency decomposition algorithm. The execution process of this algorithm includes: performing a fast Fourier transform on the raw ultrasound radio frequency signal written to the buffer to convert the time-domain signal to the spatial frequency domain; applying a set of preset bandpass filters in the spatial frequency domain to extract specific frequency band data corresponding to the reflection characteristics of myocardial tissue and blood; and performing an inverse Fourier transform on the extracted frequency band data to reconstruct a dimension-reduced signal matrix containing only the basic physical features of the target anatomical structure.

[0024] In this embodiment, the specific cutoff range of the specific frequency band is set to the echocardiographic harmonic frequency band with a center frequency of 2.5MHz to 3.5MHz. The preset time window size is specifically set to 1.5 seconds to ensure that at least one complete cardiac cycle is covered; the spatial resolution parameter is set to downsample to a 128×128 grid matrix.

[0025] Furthermore, this module divides the dimensionality-reduced signal matrix into multiple independent sub-data blocks according to preset time window size and spatial resolution parameters. Each sub-data block corresponds to a specific physical spatial region and a very short time slice during the ultrasound scanning process. Simultaneously, this module is configured to synchronously receive externally input electrocardiogram (ECG) electrophysiological signal data streams. The system performs first-order and second-order derivative feature analysis on the ECG signal to calibrate the R-wave peak timestamp (corresponding to end-diastole of the ventricle) and P-wave initiation timestamp (corresponding to the initiation point before active atrial contraction) within the current cardiac cycle in real time. In the above derivative feature analysis, to eliminate misjudgments caused by baseline drift and electromyographic noise, the system sets a fixed voltage judgment threshold: only when the first derivative crosses zero and the synchronously extracted raw ECG voltage amplitude is greater than 0.5mV is it calibrated as a valid R-wave peak timestamp; similarly, the P-wave initiation timestamp calibration must meet the baseline threshold condition of a voltage amplitude exceeding 0.1mV. The signal acquisition and spatial decomposition module uses the calibrated timestamp information as an additional dimension data label and embeds it into the header descriptor of the corresponding sub-data block, thereby establishing a lightweight data foundation and outputting the sub-data block carrying the timestamp label to the next level processing module.

[0026] Furthermore, such as Figure 3As shown, the edge-side feature recognition and segmentation module receives sub-data blocks from the signal acquisition and spatial decomposition module and extracts edge contour data of the intracardiac cavity in multiple orthogonal planes in real time synchronously in the processor of the edge-side terminal. Specifically, this module instantiates a lightweight neural network model in memory. This model adopts an encoder-decoder architecture, and the depth and number of channels of its convolutional and pooling layers are configured to adapt to the memory read / write bandwidth of the edge-side terminal. Specifically, this lightweight neural network model adopts a U-Net variant based on the MobileNetV2 backbone architecture, including a 4-layer downsampling encoder composed of depthwise separable convolutions and a corresponding upsampling decoder, with an initial feature channel count of 16 and a maximum channel count of 128. The lightweight neural network model is trained offline before deployment. The training dataset consists of a reduced-dimensional matrix sample set of echocardiograms annotated by medical image experts; the training loss function is a weighted combination of cross-entropy loss and Dice loss in a 0.5:0.5 ratio; the model optimizer is the Adam optimizer, with an initial learning rate set to [value missing]. The batch size is set to 32. Once the validation set loss function stops decreasing and tends to converge, the model parameters are fixed and edge loading is performed.

[0027] After the sub-data blocks are input into the lightweight neural network model, the model sequentially performs multi-scale two-dimensional convolution calculations, nonlinear activation processing, and batch normalization. The encoder extracts global morphological features from the sub-data blocks through layer-by-layer downsampling; the decoder maps the global morphological features back to a spatial resolution of the same size as the input sub-data blocks through upsampling combined with cross-layer skip connections. At the network output, a pixel-level probability distribution matrix is ​​generated through a Softmax or Sigmoid function layer. Each element in this probability distribution matrix represents the posterior probability that the corresponding coordinate point belongs to a specific endocardial tissue. This module sets a fixed binarization threshold and performs element-by-element logical judgment operations on the probability distribution matrix: matrix element coordinates with probability values ​​greater than or equal to the threshold are marked as foreground mask pixels, and matrix element coordinates with probability values ​​less than the threshold are marked as background pixels. The fixed binarization threshold is specifically set to 0.65. Subsequently, this module applies a morphological edge tracking algorithm (such as Moore's neighborhood tracking algorithm) to search for and close the boundaries of connected components in the foreground mask, generating a series of continuous two-dimensional spatial coordinate sets. These sets of two-dimensional spatial coordinates constitute the edge contour data. In this embodiment, the edge feature recognition and segmentation module processes data streams from different scanning angles to simultaneously generate edge contour data corresponding to the apical four-chamber and apical two-chamber cardiac sections, transforming unstructured sub-data blocks into structured geometric data columns composed of discrete coordinate points, and transmitting these geometric data columns to the biomechanical index automated calculation module.

[0028] like Figure 4 As shown, the automated biomechanical index calculation module is configured to receive structured geometric edge data and, through continuous mathematical mapping and substitution of physical formulas, execute the calculation process of strain characteristic parameters and compartment coupling characteristic parameters. Logically, this module is divided into a two-dimensional deformation analysis subunit based on optical flow tracing technology and a three-dimensional volume reconstruction subunit based on a dual-plane disk accumulation algorithm.

[0029] In the two-dimensional deformation analysis subunit based on optical flow tracing technology, the system first executes a feature point selection command. This subunit reads the edge contour data of a specific frame (such as the initial frame corresponding to the R-wave timestamp) and extracts multiple discrete coordinate points as feature pixels on the left atrial endocardial boundary coordinate set, according to a preset arc length interval or pixel interval. Specifically, the preset pixel interval is set to extract one feature pixel every 5 pixels along the endocardial contour at equal pixel distances, ensuring that the total number of feature pixels is initially controlled between 30 and 50. This initial interval serves as the feature point sampling frequency. The initial default value is then used. Subsequently, the subunit constructs a fixed-size two-dimensional pixel tracking window in the image coordinate system for each feature pixel. For adjacent frames k and k+1 in the time series, the subunit establishes an optical flow constraint equation matrix for the tracking window of each feature pixel based on the brightness constancy assumption. By solving the least-squares solution of this constraint equation matrix, the subunit calculates the horizontal and vertical displacement components of each feature pixel from frame k to frame k+1, thus synthesizing a two-dimensional pixel displacement vector field. To eliminate tracking drift caused by local image noise, the subunit applies a median filtering algorithm to the two-dimensional pixel displacement vector field, replacing abnormal vectors whose direction or magnitude difference exceeds a set variance threshold within the neighborhood with local median vectors. Here, the neighborhood is set to a 3×3 pixel window, and the set variance threshold is specifically set to 1.5 times the standard deviation of the magnitude of all displacement vectors within this local window. After filtering, the subunit performs a frame-by-frame accumulation operation of the displacement vectors along the time axis, generating the complete spatial motion trajectory of each feature pixel throughout the entire cardiac cycle. Based on the current coordinates of each feature pixel, the sub-unit calculates the Euclidean distance between adjacent feature pixels and sums up the distances between all adjacent pixels to obtain the instantaneous circumference of the left atrial endocardium corresponding to the current frame. This instantaneous circumference is a scalar function that varies with time.

[0030] After acquiring instantaneous perimeter data, the system incorporates Lagrange strain calculation logic to map geometric distance data into biomechanical strain rate data. The subunit extracts the ECG R-wave timestamp provided by the signal acquisition and spatial decomposition module, marking the frame corresponding to this timestamp as the end-diastolic reference frame. The system stores the calculated left atrial endocardial perimeter value obtained under this reference frame in a specific register and calibrates it as a baseline length. For each subsequent time point, the subunit extracts the proportionality coefficient of the difference between the instantaneous perimeter and the baseline length relative to the baseline length, and outputs this coefficient sequence as real-time strain curve data. The data structure of this real-time strain curve data is a two-dimensional floating-point array containing time series indices and corresponding strain values.

[0031] In this calculation process, the mathematical model for calculating Lagrange strain corresponds to the following relationship: ; in, Defined as The instantaneous Lagrangian strain value of the left atrial endothelium at time t is a dimensionless floating-point number, representing the unit deformation of the myocardial fiber relative to its initial state. Defined as the reference length, which is the reference circumference of the left atrial endocardium extracted by the system based on the R-wave timestamp in the synchronized electrocardiogram signal at the end-diastolic reference frame, in millimeters or pixels as the physical unit; Defined as The instantaneous circumference extracted at each moment, that is, the real-time stretching circumference of the myocardium that changes over time, is compared with... The same unit of measurement; the above parameters, through arithmetic logic operations of subtraction and division, eliminate displacement errors caused by the overall spatial translation of the heart chambers.

[0032] Based on the generated real-time strain curve data, the automated biomechanical index calculation module further executes automatic staged extraction logic to obtain the three-stage strain indexes of the left atrium. These three-stage strain indexes specifically include the maximum strain peak value during the reserve period, the strain difference during the conduit period, and the strain value before active contraction. The subunit divides the cardiac cycle into three different time windows based on the input timestamp labels.

[0033] In response to the characteristics of the reserve period, the system performs a global optimization search algorithm on the real-time strain curve data sequence within the time window between the mitral valve closure and the mitral valve opening, traversing all array elements to extract the maximum value, which is defined as the maximum strain peak value during the reserve period.

[0034] The relationship corresponding to this feature is calculated as follows: ; in, Defined as the maximum strain peak during the left atrial reserve period, the data type is a floating point number, representing the maximum blood storage capacity of the left atrium during ventricular systole; Defined as a sequence of instantaneous Lagrange strain values ​​that are continuously updated within a reserve period time window; Defined as an extremum search function in a system call, it operates within a given closed time interval and returns the global maximum strain value within that interval.

[0035] To address the characteristics of the conduit phase, the system analyzes the slope changes of real-time strain curve data and performs calculations within the time window between the mitral valve opening and the onset of active atrial contraction. The subunit reads the maximum strain peak value obtained in the previous steps and the instantaneous strain value corresponding to the onset of active atrial contraction (based on the P-wave initiation timestamp). The system performs a floating-point subtraction operation on the two values ​​to obtain the strain difference during the conduit phase.

[0036] The relationship corresponding to this feature is calculated as follows: ; in, Defined as the strain difference during the left atrial canal phase, it characterizes the passive emptying capacity of the left atrium after the mitral valve opens; Defined as peak strain, which is the maximum Lagrange strain reached during the entire cardiac cycle, its numerical value is similar to... equal; Defined as the strain value before active atrial contraction, which is the instantaneous strain sequence value that the system matches and extracts in real time based on the timestamp of the P wave start point of the synchronously input electrocardiogram signal.

[0037] For the active contraction period characteristics, the system directly reads the instantaneous strain values ​​captured at the time corresponding to the P-wave start timestamp from the memory, without performing additional arithmetic transformations, and directly defines them as contraction period strain values.

[0038] The relationship corresponding to this feature is calculated as follows: ; in, Defined as the systolic strain value of the left atrium, it is a parameter characterizing the contribution of active left atrial contraction; As defined above, it is the instantaneous Lagrange strain value at the beginning of the P wave on the electrocardiogram before the active contraction of the atrium.

[0039] Furthermore, the automated biomechanical index calculation module is internally configured with a fixed self-verification algorithm process. After the above three strain indices are calculated, the system processor automatically adds the strain difference during the pipeline period to the strain value during the contraction period using a floating-point addition instruction, and compares the sum with the difference between the sum and the maximum strain peak value during the reserve period. When the absolute value of the difference is less than the minimum machine precision value set by the system, the verification passes, and the system latches the result data and sends it to the output bus; the minimum machine precision value is specifically set here as the absolute error tolerance. If the absolute value of the difference exceeds this minimum value, it is determined that there is frame loss or data breakpoint in the optical flow tracing process of the current cycle. The system will discard the strain calculation results of this cycle and return an error code to the call stack.

[0040] The physical relationships corresponding to the built-in self-verification mechanism of this system are as follows: ; in, Defined as the calculated maximum strain peak value during the reserve period; Defined as the calculated strain difference value during pipeline operation; Defined as the directly extracted shrinkage strain value; plus sign " "and equal sign" "This represents the arithmetic equality logic check instruction operator that is forcibly executed in the background by the module processor. It is used to perform background physical verification of the results of the algorithm output data matrix of the two-dimensional deformation analysis subunit.

[0041] In the 3D volume reconstruction subunit based on the dual-plane disk accumulation algorithm, the system generates atrioventricular coupling feature parameters, specifically the left atrioventricular coupling index. The subunit synchronously receives the edge contour coordinate sets of the heart chambers on two orthogonal planes (the apical four-chamber view and the apical two-chamber view) from the edge-side feature recognition and segmentation module at the same time. This subunit first executes the major axis calibration and slice division logic: calculating the Euclidean distance from the center coordinates of the bottom of the heart chamber to the coordinates of the highest point of the apex, and defining this straight-line distance as the length of the major axis of the heart chamber. The automatic extraction step for the coordinates of the highest point of the apex involves traversing the current edge contour coordinate set and calibrating the discrete point with the largest Y-axis coordinate value as the highest point of the apex. The automatic extraction step for the center coordinates of the bottom of the heart chamber involves extracting the coordinates of two anatomical feature points belonging to the attachment points on both sides of the mitral valve annulus from the edge contour coordinate set, calculating the geometric midpoint coordinates of the line segment connecting these two points, and defining this as the center coordinates of the bottom of the heart chamber. Subsequently, based on the dynamic slice layer number parameters currently issued by the system, the subunit divides the major axis into equally spaced line segments, forming multiple mutually parallel normal plane coordinate systems. For each normal plane, the subunit calculates the distance between its intersection point and the edge contour of the apical four-chamber sectional plane, obtaining the cross-sectional diameter on the first orthogonal plane; simultaneously, it calculates the distance between its intersection point and the edge contour of the apical two-chamber sectional plane, obtaining the cross-sectional diameter on the second orthogonal plane. The subunit assumes that the three-dimensional cross-section of the heart cavity at each slice height is an ellipse, and calculates the infinitesimal volume of this truncated elliptic slice using the constant pi and the two cross-sectional diameters. Finally, the subunit discretly accumulates and sums the infinitesimal volumes of all slices along the major axis to obtain the instantaneous volume value of the heart cavity at the current time point.

[0042] The discrete integral relationship corresponding to this volume calculation is as follows: ; in, Defined as The three-dimensional instantaneous volume data of the cardiac cavity at any given time is represented by a floating-point number in milliliters. Defined as the number of dynamic slice layers, this parameter is a positive integer variable, and its specific value is controlled by the real-time scheduling output of the edge-side adaptive computing power allocation module; Defined as an incrementing loop variable in the discrete accumulation loop logic, its value range is... Increase to ; Defined as a fixed constant coefficient for converting the disk area pre-set in system memory; and Defined respectively Time of the first The physical diameter of the cross section of the layer slice on the two orthogonal cutting planes mentioned above is derived from the geometric measurement of the edge contour data; Defined as The linear length of the major axis of the heart chamber is calculated at each moment; symbol Defined as a multi-round iterative floating-point adder operation process executed in the processor.

[0043] After acquiring the instantaneous volume data series across continuous time frames, this subunit executes extreme value capture and characteristic ratio calculation instructions. The system reads the volume data series within a specified cardiac cycle window before mitral valve opening, compares the data, and obtains the maximum left atrial volume value during this phase, storing it as the dividend. Simultaneously, the system reads the left ventricular volume value at the corresponding timestamp at end-diastole and stores it as the divisor. The subunit performs single-precision floating-point division, dividing the dividend by the divisor to obtain atrioventricular coupling characteristic parameters representing atrial compensation and transchamberal spatial coupling.

[0044] The relationship corresponding to this feature is calculated as follows: ; in, Defined as a characteristic parameter of the left atrioventricular coupling index obtained through calculation, it is expressed as a dimensionless floating-point ratio; Defined as the maximum left atrial volume obtained by the system through a global maximum optimization instruction in the real-time volume curve sequence; Defined as the left ventricular end-diastolic volume value extracted by the system after identifying the time corresponding to the R wave in the electrocardiogram input stream; the result of this division expression characterizes the volume ratio coupling coefficient of two different physiological chambers in three-dimensional spatial geometry.

[0045] To ensure the real-time and continuous operation of segmentation, tracking, and reconstruction on edge terminals with limited hardware resources, the system incorporates an edge-side adaptive computing power allocation module. This module is an independent daemon process running in a high-priority system thread of the operating system. It is configured to continuously monitor the time consumption of each stage in the current frame processing lifecycle and dynamically adjust the calculation precision parameters of subsequent data streams based on closed-loop control logic. Specifically, the module has three system-level timestamp recorders that capture the clock ticks of the CPU and GPU at the entry and exit points of specific subroutines. The first recorder captures the clock difference between the lightweight neural network model reading sub-data blocks and outputting the edge contour coordinate set, calculating the segmentation time. The second recorder captures the time difference between the automatic biomechanical index calculation module receiving contour data and outputting the mathematical calculation of various feature parameters, calculating the computation time. The third recorder captures the time difference between numerical generation and writing the values ​​to system video memory for screen rendering, calculating the rendering time.

[0046] During system initialization, the adaptive computing power allocation module reads a fixed target frame rate configuration parameter and calculates the reciprocal of this target frame rate, setting it as the maximum allowable total time threshold. Simultaneously, the module monitors the feature point sampling frequency in the two-dimensional deformation analysis.

[0047] The relationships corresponding to the computing power constraint model are as follows: ; in, Defined as the density of feature pixels selected within a unit perimeter or area of ​​the edge contour data, i.e., the sampling frequency, the goal of the system's constraint optimization is to maximize the possible value of this parameter while satisfying the time consumption inequality constraint. Identify the maximization operation on the target variable; "Subject to" is an abbreviation for the English word "subject to". It defines the subsequent logical expression as a rigid boundary constraint condition that must be satisfied. Defined as the real-time measured floating-point variable of single-frame segmentation time obtained from real-time monitoring; Defined as the actual floating-point variable used for real-time monitoring and acquisition of feature tracking and volume integral calculation; Defined as a floating-point variable representing the actual rendering time measured by real-time monitoring and data acquisition; Defined as a system-preset target frame rate integer constant; the reciprocal on the right side of the equation is defined as the maximum total time threshold allowed for a single frame of data.

[0048] Within a control cycle, this module sums the three measured time-consuming variables to obtain the real-time total time, and compares the real-time total time with the maximum allowable total time threshold. When the real-time total time exceeds a preset convergence limit (e.g., 90% of the maximum threshold), the computing power allocation module activates a hierarchical parameter degradation protocol. The protocol first triggers a compression operation for segmentation time: the module sends a control word to the lightweight neural network model, instructing the network to skip some preset deep residual blocks (i.e., reduce network depth) during the next forward propagation, or to call a bilinear interpolation function to reduce the input resolution size of the tensor before reading the sub-data block. Specifically, the degradation operation is as follows: the third and fourth layer residual blocks of the encoder are turned off sequentially from deep to shallow, and the input resolution size is proportionally reduced to 75% of the original size. After performing the initial degradation operation, the system enters the next control cycle. If the total real-time latency still does not drop below the threshold safety line, the system triggers a suboptimal compression operation targeting computational latency: the module modifies the control variable pointer output to the 3D volume reconstruction subunit, decreasing the number of dynamic slice layers in the above discrete integral formula by an integer step size (reducing...). The value of is used to reduce the number of iterations and loops of the floating-point multiplier. The initial value of the dynamic slice layer is set to . The decreasing integer step size is set to 2, and the system has a built-in hard lower limit value. Once this lower limit is reached, the parameter reduction stops to ensure the minimum clinical accuracy of volumetric calculations. Through the implementation of this degradation adjustment strategy, the module can significantly reduce processing latency, thereby ensuring that the end-to-end latency from data acquisition to the display of warning indicators is controlled within a preset safe millisecond range. When the module detects an increase in system idle resources and a significant drop in total real-time latency below the threshold, it executes a reverse parameter recovery logic, prioritizing an increase in sampling frequency, followed by restoring network depth and the number of slice layers.

[0049] After data flow transformation and computational extraction by the aforementioned modules, biomechanical strain characteristics and volumetric coupling characteristics are converted into standardized floating-point feature parameter arrays and transmitted to the multi-dimensional risk scoring and graded early warning module via the internal data bus. The logical function of this module is to establish a mathematical fusion model, convert multi-source heterogeneous parameters into a single indicator variable, and execute the drive to alarm hardware.

[0050] This module includes a statically configured baseline reference database containing multiple lookup tables generated from a large amount of healthy sample data. The system reads externally entered patient baseline information, performs interpolation retrieval in the lookup tables, and obtains health reference values ​​that precisely match the age, gender, and body surface area of ​​the current user. This data includes reserve strain reference values ​​and left atrioventricular coupling index reference values. As a partial structural example of the aforementioned baseline reference database: for healthy males with a body surface area between 1.8 and 2.0 square meters, the baseline data recorded in the lookup table for the 20-29 age group are as follows: reserve strain reference value set at 39.0%, left atrioventricular coupling index reference value set at 15.2%; this lookup table performs segmented mapping and decreasing for the age dimension in 5-year increments. Specifically, the full baseline reference value dictionary covers five age groups: 20-29 years, 30-39 years, 40-49 years, 50-59 years, and 60 years and above; two genders: male and female; and three intervals for body surface area: less than 1.6 m², 1.6-1.8 m², and greater than 1.8 m². The specific mathematical rule for the decrease is as follows: for each increase in age group, the reserve period strain reference value decreases by 5% from the previous level baseline, and the left atrioventricular coupling index reference value increases by 3% from the previous level baseline. The complete correspondence between other genders and body surface area combinations is stored in this dictionary in the form of a static lookup table for direct retrieval.

[0051] Simultaneously, the module receives clinical correction factors pushed from external electronic medical record data sources. These factors include multiple discrete quantification terms: the system calculates the body mass index (BMI) from the input height and weight (floating-point values) to represent the obesity factor; it calculates the age integer value from the input date of birth as a representation of the age-related decline factor; and it reads the Boolean-type atrial fibrillation history field as a representation of the atrial fibrillation history factor. The module inputs these correction factors into a non-linear mapping function and outputs normalized scalar values ​​for the clinical correction terms.

[0052] This module is configured with a scoring function parser based on weighted polynomial logic to generate a single comprehensive risk score. The operational mapping relationship of this scoring function is as follows: ; in, Defined as the final calculated risk score, its data type is limited to a single-precision floating-point number within the range of [0, 100]. , , Defined as constant floating-point weighting coefficients fixed in system memory, the three coefficients are set by strict normalization constraints, and their sum is equal to a constant 1; Defined as the left atrial reserve strain characteristic parameter generated in real time by the biomechanics module within the current calculation cycle; Defined as the corresponding vital sign health reference value retrieved from the lookup table, this fraction is calculated by subtraction and normalized by proportion to form the weight of the negative deviation evaluation index in the system calculation; Defined as a characteristic parameter of the left atrioventricular coupling index generated in real time by the biomechanical module; Defined as the corresponding health reference value in the coupling index lookup table, this fraction is directly calculated as a ratio, constituting the weight of the positive deviation evaluation index in the system calculation; Defined as a nonlinear composite correction term that encapsulates the aforementioned clinically modified variables, including obesity, age-related decline, and atrial fibrillation history; the system uses a floating-point multiplier and adder to weightedly sum the three product terms to obtain the final value. variable; It is a mathematical model that extracts and transforms classic features (body mass index, age, history of atrial fibrillation) used in clinical practice to assess the risk of heart failure into nonlinear functions, thereby dynamically adjusting for individual physiological differences among patients.

[0053] The above weighting coefficients are determined based on the priority of clinical manifestations of cardiac remodeling in different dimensions and the needs of technical scenarios: the left atrioventricular coupling index, as a positive indicator, represents the macroscopic cross-chamber volume compensation and anatomical spatial coupling state, and is the earliest morphologically sensitive feature of diastolic dysfunction, so it is given the highest weight. Left atrial reserve strain characteristics, as a negative indicator, directly reflect the degree of damage to myocardial fibers at the microscopic biomechanical level, and serve as a secondary core assessment dimension. The nonlinear comprehensive correction term represents the basic physiological differences of individual patients and serves as an auxiliary bias term for the baseline. This weighted distribution framework ensures that the system's risk quantification is primarily based on real-time collected objective biomechanical dynamic changes, supplemented by static medical history and physical examination corrections, thus minimizing the false positive rate of the early warning system.

[0054] Based on typical clinical data calibration in this field, the specific values ​​of the constant floating-point weighting coefficients in the above items are: , , .

[0055] Furthermore, the nonlinear synthesis correction term The specific mathematical mapping rules are calculated by the following variant of the Logistic Function: ; in, This is a floating-point value for the body mass index. The age is an integer value. The Boolean integer variable represents the history of atrial fibrillation (1 for those with a history, 0 for those without); the mapping weight constants for each clinical correction factor are specifically set as follows: , , constant bias .

[0056] In the aforementioned nonlinear mapping scenario, the weight distribution of each clinical correction factor is strictly set based on its independent clinical risk contribution to the deterioration of ventricular diastolic function: history of atrial fibrillation ( As an acute triggering factor, obesity can lead to loss of atrial active contraction function and accelerate fibrotic remodeling, exhibiting a step-amplifying effect on cardiac risk and thus being assigned a very high nonlinear activation weight; in contrast, obesity ( ) and aging ( This is a chronic, progressive, cumulative factor, and its weight has been deliberately suppressed. The consideration behind this parameter combination is to ensure that the final comprehensive risk score generated by the system is not artificially high simply because the subject is older or simply obese, thereby ensuring the clinical accuracy of the tiered alarm signal triggering.

[0057] After generating the quantified risk score, the module's execution process proceeds to the tiered alarm comparison logic. The system memory has two preset progressive floating-point threshold constants, defined as a first threshold and a second threshold, respectively. For quantified risk scores within the range of 0 to 100, the first threshold is specifically set to 40.0, and the second threshold is specifically set to 75.0. The module calls the comparator interface, continuously inputting the currently generated risk score variable and comparing it arithmetically with the two threshold constants.

[0058] When the comparator outputs a Boolean true signal with a risk score strictly less than the first threshold, the module's logic branch points to a safe state. The system calls the graphics application programming interface of the display device driver to write a specific byte instruction to the preset instruction address segment of the video memory, causing the display interface to display a green glowing pixel indicator array in a specific status bar area, and refreshing the corresponding identifier code in the status console.

[0059] When the comparator outputs a Boolean true signal with a risk score greater than or equal to the first threshold and strictly less than the second threshold, the module's logic branch points to the intermediate state. The system calls the aforementioned application programming interface and changes the write parameters, writing different byte instructions to the video memory address to drive the rendering of the corresponding interface area, presenting the yellow pixel indicator array, and issuing the corresponding display signal variable.

[0060] When the comparator outputs a Boolean true signal indicating a risk score greater than or equal to the second threshold, the module's logic branch points to the threshold over-limit alarm state. The system will immediately suspend the current low-priority display thread and start a high-priority asynchronous alarm drive thread, alternately writing red color matrix code and full-transparency matrix code to the video memory at a fixed frequency. This creates a strong red flashing alarm pixel effect on the physical screen of the display terminal, indicating the current data's over-limit status in the form of a visual square wave pulse.

[0061] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A real-time cardiac risk assessment and early warning system based on edge computing and deep learning, deployed on edge terminals with limited computing power, characterized in that, The system includes: The signal acquisition and spatial decomposition module is used to acquire the raw ultrasound radio frequency signal and decompose it into sub-data blocks according to spatial frequency in order to establish a lightweight data foundation. The edge feature recognition and segmentation module is used to receive the sub-data block through a lightweight neural network model and extract the edge contour data of the inner cavity of the same cardiac cycle on multiple orthogonal planes in real time and synchronously. The biomechanical index automated calculation module is used to calculate the instantaneous perimeter deformation of the edge contour data based on optical flow tracing technology to obtain strain characteristic parameters, and to perform three-dimensional spatial volume reconstruction based on the cross-sectional diameter and dynamic slice layer number on the multiple orthogonal planes to obtain compartment coupling characteristic parameters. The edge-side adaptive computing power allocation module is configured to monitor the segmentation time, computation time, and rendering time of a single frame of data in the edge-side terminal in real time, and construct a computing power constraint model based on the maximum allowable total time threshold of the target frame rate. When the real-time total time approaches the maximum allowable total time threshold, the computing power allocation module prioritizes reducing the network depth or input resolution of the lightweight neural network model to compress the segmentation time, and secondarily reduces the number of dynamic slice layers in the three-dimensional spatial volume reconstruction to compress the computation time, thereby ensuring that the end-to-end latency is controlled within a preset range. The multi-dimensional risk scoring and graded early warning module is used to perform nonlinear multi-source fusion of the extracted feature parameters and clinical correction factors, map them to generate quantitative risk scores, and execute graded alerts.

2. The real-time cardiac risk assessment and early warning system according to claim 1, characterized in that, The strain characteristic parameters include the three-phase strain index of the left atrium; The automated biomechanical index calculation module extracts feature pixels on the edge contour data based on the assumption of constant brightness, calculates the two-dimensional displacement vector field and performs filtering and reconstruction to obtain the instantaneous perimeter trajectory between consecutive frames.

3. The real-time cardiac risk assessment and early warning system according to claim 2, characterized in that, When the automated biomechanical index calculation module obtains the three-phase strain index of the left atrium: The end-diastolic endothelial circumference of the ventricle was calibrated as the baseline length. Extract the proportionality coefficient of the difference between the instantaneous circumference and the reference length relative to the reference length, and use it as the real-time strain curve data; Based on the slope changes of the electrocardiogram signal and the real-time strain curve data, the maximum strain peak value during the reserve period, the strain difference during the pipeline period, and the strain value before active contraction are automatically extracted in stages.

4. The real-time cardiac risk assessment and early warning system according to claim 1, characterized in that, The atrioventricular coupling characteristic parameters include the left atrioventricular coupling index; The automated biomechanical index calculation module calculates the instantaneous volume of the heart chamber by accumulating the cross-sectional diameter and the number of dynamic slice layers using a dual-plane disk accumulation algorithm. By capturing extreme values ​​within a specific cardiac cycle window, the maximum volume of the left atrium and the end-diastolic volume of the left ventricle are obtained respectively, and the ratio of the two is defined as the left atrioventricular coupling index.

5. The real-time cardiac risk assessment and early warning system according to claim 1, characterized in that, When monitoring and adjusting, the edge-side adaptive computing power allocation module aims to maximize the feature point sampling frequency per unit area to ensure that the highest precision biomechanical feature curve is output within the currently available terminal computing power boundary.

6. The real-time cardiac risk assessment and early warning system according to claim 1, characterized in that, The clinical correction factors include obesity factors, age-related decline factors, and atrial fibrillation history factors, as characterized by body mass index.

7. The real-time cardiac risk assessment and early warning system according to claim 1, characterized in that, The multi-dimensional risk scoring and graded early warning module is equipped with a scoring function. The scoring function calculates the first deviation between the currently measured strain characteristic parameters and the corresponding patient's health reference value, and calculates the second deviation between the measured atrioventricular coupling characteristic parameters and the health reference value. The strain characteristic parameters are used as negative index weights, the atrioventricular coupling characteristic parameters are used as positive index weights, and the clinical correction factors are combined and weighted to output a single comprehensive risk score ranging from 0 to 100.

8. The real-time cardiac risk assessment and early warning system according to claim 7, characterized in that, The tiered alerts are executed based on a set three-tier threshold logic: A green indicator suggesting good coupling is displayed when the value is below the first threshold. A yellow indicator is displayed between the first and second thresholds, prompting a detailed inspection. A red flashing warning is displayed if the value exceeds the second threshold, indicating a high likelihood of heart failure.

9. The real-time cardiac risk assessment and early warning system according to claim 1, characterized in that, The system has a built-in self-verification mechanism. The maximum strain peak value extracted by the biomechanical index automated calculation module during the reserve period must be equal to the sum of the strain difference during the pipeline period and the strain value during the contraction period, so as to be used for background physical verification of the algorithm results.

10. The real-time cardiac risk assessment and early warning system according to any one of claims 1 to 9, characterized in that, The edge-side terminal is wirelessly connected to a smart ultrasound probe to monitor and quantify cardiac diastolic dysfunction and remodeling status in real time before the traditional ejection fraction declines.

Images