Heart Failure Readmission Risk Assessment System Based on Cardiac Imaging Feature Analysis

Abstract

This invention discloses a heart failure readmission risk assessment system based on cardiac imaging feature analysis, belonging to the field of cardiac imaging analysis technology. By analyzing speckle texture features in ultrasound images and combining them with respiratory phase and image quality information, a dynamic and regular quantitative index is generated, thereby achieving accurate risk assessment. It can identify potentially high-risk patients in advance, even those with normal or mildly abnormal ejection fractions, and effectively addresses the impact of respiratory phase changes and image quality interference on assessment results. By weighted fusion of image quality and physiological state information, the stability and accuracy of the assessment results are improved, making it particularly suitable for primary healthcare institutions. This invention provides a more objective and accurate risk assessment, offering a scientific basis for doctors to develop personalized treatment plans, thereby significantly improving the early identification and treatment outcomes of heart failure patients.

Description

Technical Field

[0001] This invention relates to the field of cardiac image analysis technology, specifically a heart failure readmission risk assessment system based on cardiac image feature analysis. Background Technology

[0002] Heart failure is a serious disease caused by impaired cardiac function, and its early diagnosis and readmission risk prediction are crucial for treatment effectiveness and patient prognosis. Traditional methods for assessing heart failure mainly rely on macroscopic indicators such as ejection fraction and heart chamber size, and typically employ imaging techniques such as echocardiography to evaluate cardiac function. While these traditional methods can effectively reflect the gross structural and functional status of the heart, in some patients with normal or mildly abnormal ejection fractions, traditional indicators are less sensitive in capturing potential microscopic myocardial changes, especially at the level of myocardial microstructure.

[0003] In recent years, with the development of image processing technology, methods for assessing myocardial function based on speckle texture features in ultrasound images have gradually emerged. Speckle texture reflects the scattering of ultrasound waves in myocardial tissue. Studies have shown that changes in myocardial microstructure lead to changes in texture features, providing a new approach for the early detection of heart failure. However, existing techniques often overlook the dynamic changes in the patient's respiratory status during the examination. The different phases of respiration affect the quality of ultrasound images and the position of the heart, which may lead to inconsistencies and decreased accuracy in the analysis results. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a heart failure readmission risk assessment system based on cardiac imaging feature analysis.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows:

[0006] In a first aspect, the present invention discloses a heart failure readmission risk assessment system based on cardiac imaging feature analysis, comprising:

[0007] The multi-source data acquisition module is used to simultaneously acquire standard apical section echocardiographic sequence data and real-time respiratory status data of the target patient;

[0008] The respiratory phase segmentation module is used to segment the echocardiogram sequence data according to the respiratory phase based on the real-time respiratory state data, so as to obtain several sets of image segment data with consistent respiratory phases.

[0009] The image segment processing module performs the following processing on each image segment data:

[0010] Extract image patch data of the left ventricular endocardial surface region from each frame of the image fragment data, and calculate speckle texture features based on the image patch data;

[0011] Arrange the speckle texture features of each frame of the image in chronological order to form a texture feature sequence of the image fragment data;

[0012] The texture feature sequence is segmented and aligned according to the cardiac cycle, and the consistency measure of texture features at the same phase point in multiple cardiac cycles is calculated to obtain a dynamic regularity quantification index.

[0013] The comprehensive quantification module is used to normalize the dynamic regularity quantification index based on the respiratory phase type associated with the image segment data, and to perform weighted fusion of the normalized dynamic regularity quantification index with the image quality confidence associated with the image segment data to generate a comprehensive regularity quantification index.

[0014] The risk assessment output module compares the comprehensive regularity quantitative indicators with preset risk assessment thresholds and outputs the heart failure readmission risk assessment result.

[0015] Secondly, this invention discloses a method for assessing the risk of readmission for heart failure based on cardiac imaging feature analysis, comprising the following steps:

[0016] Simultaneously acquire standard apical section echocardiographic sequence data and real-time respiratory status data of the target patient;

[0017] Based on the real-time respiratory status data, the echocardiogram sequence data is segmented according to the respiratory phase to obtain several sets of image segment data with consistent respiratory phases.

[0018] For each image segment data, perform the following processing:

[0019] Extract image patch data of the left ventricular endocardial surface region from each frame of the image fragment data, and calculate speckle texture features based on the image patch data;

[0020] Arrange the speckle texture features of each frame of the image in chronological order to form a texture feature sequence of the image fragment data;

[0021] The texture feature sequence is segmented and aligned according to the cardiac cycle, and the consistency measure of texture features at the same phase point in multiple cardiac cycles is calculated to obtain a dynamic regularity quantification index.

[0022] Based on the respiratory phase type associated with the image segment data, the dynamic regularity quantitative index is normalized by physiological state, and the normalized dynamic regularity quantitative index is weighted and fused with the image quality confidence associated with the image segment data to generate a comprehensive regularity quantitative index.

[0023] The comprehensive regularity quantitative indicators are compared with the preset risk assessment thresholds to output the risk assessment results of heart failure readmission.

[0024] Part Three, this invention discloses a computer storage medium storing computer-executable instructions, which, when executed, implement the heart failure readmission risk assessment system based on cardiac imaging feature analysis as disclosed in Part One.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0026] 1. By incorporating respiratory phase information, this invention addresses the problem of respiratory status interfering with image analysis results in existing technologies. By segmenting and normalizing echocardiogram sequences based on respiratory phase, the comparability of myocardial texture features under different physiological states is improved, thereby enhancing the sensitivity of readmission risk prediction for heart failure patients. This is particularly beneficial for patients with normal or mildly abnormal ejection fractions, enabling the early identification of potentially high-risk patients.

[0027] 2. This invention solves the error problem caused by image quality differences in traditional echocardiographic analysis by weighted fusion of image quality confidence and dynamic regularity quantification indicators. Through the weighted fusion method, this invention ensures that the impact of image quality on the assessment results is minimized, significantly improving the reliability of the assessment results, and providing consistent and objective risk assessment results for different patients and under different conditions. Attached Figure Description

[0028] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. In the drawings, the same reference numerals are used to refer to the same parts. Wherein:

[0029] Figure 1 This is a system architecture diagram of the present invention;

[0030] Figure 2 This is a data timing diagram of the present invention;

[0031] Figure 3 This is a flowchart of the method of the present invention. Detailed Implementation

[0032] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0033] Application Overview

[0034] This invention relates to a heart failure readmission risk assessment system based on cardiac imaging feature analysis. It aims to assess the readmission risk of heart failure patients by combining echocardiographic data and real-time respiratory status data. The core technology of this method lies in analyzing the speckle texture features reflecting microscopic changes in the myocardium in ultrasound images and combining this with respiratory phase and image quality information to generate dynamic and regular quantitative indicators, thereby achieving accurate risk assessment. This technology can improve the risk identification ability of heart failure patients with normal or mildly abnormal ejection fractions, especially in patient groups where traditional assessment methods cannot provide effective risk indications, providing a low-cost and efficient auxiliary diagnostic tool.

[0035] Currently, traditional methods for assessing heart failure primarily rely on macroscopic indicators such as ejection fraction and cardiac chamber size. While these methods can effectively reflect gross functional and structural changes in the heart, their sensitivity and accuracy are limited for patients with normal or mildly abnormal ejection fractions, making it difficult to identify potential early heart failure. Furthermore, existing techniques often overlook the impact of respiratory phase on ultrasound image quality and changes in cardiac position, leading to inconsistent assessment results due to image quality fluctuations, thus affecting the accuracy of risk assessment. Therefore, how to integrate respiratory status and image quality information within the existing ultrasound image analysis framework to eliminate physiological changes and image quality interference has become a pressing technical problem to be solved.

[0036] This invention aims to solve the aforementioned technical problems and effectively achieves its objectives through the following technical means: First, it simultaneously acquires standard apical echocardiographic sequence data and real-time respiratory status data of the target patient; then, based on respiratory phase information, it segments the echocardiographic sequence data according to respiratory phase to obtain image segment data with consistent respiratory phases; next, it extracts speckle texture features from each frame of the image and arranges them in chronological order to form a texture feature sequence; subsequently, it segments and aligns these texture feature sequences according to the cardiac cycle, calculates the consistency measure of texture features, and obtains a dynamic regularity quantitative index; finally, it performs weighted fusion and physiological state normalization based on image quality confidence to generate a comprehensive regularity quantitative index, compares it with a preset risk threshold, and outputs the heart failure readmission risk assessment result.

[0037] By implementing this invention, the accuracy and stability of heart failure readmission risk assessment can be significantly improved, especially in patients with normal or mildly abnormal ejection fractions, enabling early detection of potential risks. This method effectively eliminates the influence of fluctuations in respiratory status and image quality, improving the consistency and reliability of assessment results. Simultaneously, this invention provides primary healthcare institutions with a low-cost, efficient solution, reducing technical barriers and enabling the widespread application of heart failure risk assessment tools in more clinical settings. By comprehensively considering dynamic patterns, image quality, and physiological status, this invention provides physicians with more objective and accurate heart failure readmission risk assessment results, assisting them in developing personalized treatment and follow-up plans, thereby improving patient treatment outcomes and quality of life.

[0038] like Figure 1 , Figure 2 As shown, this application discloses a heart failure readmission risk assessment system based on cardiac imaging feature analysis, which includes the following modules: multi-source data acquisition module, respiratory phase segmentation module, image segment processing module, comprehensive quantification module, and risk assessment output module. Each module will be described in detail below.

[0039] The multi-source data acquisition module simultaneously acquires two types of heterogeneous time-series data: standard apical section echocardiographic sequence data of the target patient and real-time respiratory status data of the target patient.

[0040] The standard apical echocardiographic sequence data of the target patient is, in form, a continuous digital video frame sequence. Essentially, it represents the temporal variation of a two-dimensional grayscale acoustic projection of the heart, a moving organ. The pixel grayscale values ​​in each frame directly characterize the distribution of scattered intensity after the interaction of ultrasound waves with the myocardial microstructures (such as muscle fibers and collagen networks). This data is acquired because changes in the functional state, arrangement, and interstitial components of cardiomyocytes directly affect their acoustic properties, leaving detectable imprints in the dynamic speckle pattern. This data carries the most fundamental spatial and temporal information required to assess cardiac function, serving as the raw material for all subsequent analyses.

[0041] Real-time respiratory status data of the target patient is typically a waveform signal that varies over time (such as chest and abdominal movement displacement, airflow, or a proxy signal derived from the image). Real-time respiratory status data directly characterizes the dynamic process of periodic changes in intrathoracic pressure, i.e., the respiratory cycle. Respiratory motion is one of the most significant physiological sources of interference affecting cardiac imaging: during inspiration, intrathoracic negative pressure increases, venous return increases, cardiac preload increases, and the ventricles tend to fill; the opposite occurs during expiration. This periodic load change leads to overall cardiac displacement, rotation, and morphological alterations, which can superimpose with the motion generated by the heart's own contraction and relaxation, severely obscuring the analysis of the heart's own micro-motion patterns. Therefore, real-time respiratory status data provides an accurate time reference for subsequent steps to identify and isolate interference from the respiratory phase, ensuring that the analysis focuses on the manifestations under the heart's inherent rhythm.

[0042] The two types of data mentioned above are acquired synchronously, which ensures that each frame of ultrasound image can be assigned a clear respiratory phase label, so that the image data can be processed in the correct physiological context and avoids parsing errors caused by time domain mismatch.

[0043] The respiratory phase segmentation module is used to segment echocardiographic sequence data according to respiratory phase. This module receives echocardiographic sequence data and real-time respiratory status data acquired by the multi-source data acquisition module. First, it identifies key phase inflection points (such as the onset and end of inspiration, the onset and end of expiration) from the waveform of the real-time respiratory status data, thereby dividing the continuous respiratory cycle into phase windows with clear physiological significance (such as early inspiration, the inspiratory plateau, and early expiration). Then, based on the timestamps of the image frames in the echocardiographic sequence data, it categorizes them into their respective respiratory phase windows, finally obtaining image segment data with consistent respiratory phase.

[0044] The significance of image segment data with consistent respiratory phase is that all image frames within each segment are acquired from a short time window when the heart is in a similar state of thoracic pressure and load. This is essentially a state clustering of the image data based on external physiological conditions.

[0045] The heart is a dynamic system modulated by respiration. Directly analyzing long sequences containing multiple respiratory cycles is equivalent to mixing observational data from different operating states, resulting in low signal-to-noise ratios and an inability to distinguish whether changes originate from the heart's own function or respiratory modulation. The purpose of segmentation is to transform non-stationary long-time-series signals into multiple quasi-steady-state subsequences that are approximately stationary under physiological conditions within short time windows.

[0046] Subsequent processing treats each image segment data that is consistent with the breathing time as an analysis unit. The cardiac motion variation within it mainly comes from its own rhythm, while the systemic variation caused by breathing has been limited to the greatest extent, creating the preconditions for subsequent extraction of microscopic features that characterize the heart's own rhythm.

[0047] The image segment processing module performs depth feature extraction and regularity calculation on each stable image segment. The specific data flow relationship is as follows: image segment data → image block data → speckle texture features → texture feature sequence → dynamic regularity quantification index.

[0048] The specific processing procedure of this module is as follows:

[0049] First, image patch data of the left ventricular endocardial surface region is extracted from each frame of the image segment data using image processing techniques (such as edge detection and region growing). This image patch data directly represents the innermost layer of myocardial tissue in the ventricular wall. The reason for choosing the left ventricular endocardial surface region is that the subendocardial myocardial layer is rich in longitudinally arranged muscle fibers, which contribute the most to the longitudinal contractile function of the heart. It is also the terminal area of ​​coronary artery perfusion and is most sensitive to pathological processes such as ischemia and fibrosis. If the entire heart chamber or the entire myocardium is analyzed, a large amount of irrelevant information (such as intracardiac blood and epicardial fat) will be introduced during the analysis process. This irrelevant information will dilute the key signals and lead to distortion of the analysis results. By selecting the left ventricular endocardial surface region, the analysis process is focused on the most functionally core and pathologically vulnerable tissue, so that the micromotor features reflecting the functional state of this specific tissue can be extracted more purely in subsequent steps.

[0050] Subsequently, speckle texture features were calculated from image patch data of the left ventricular endocardial surface region. These features were a set of values ​​calculated using texture analysis algorithms (such as Local Binary Pattern (LBP) and derived indices of the Gray-Level Co-occurrence Matrix (GLCM). The speckle in ultrasound images is not noise, but rather an interference pattern formed by the interaction of coherent sound waves with subwavelength-scale tissue microstructures. Therefore, speckle texture features essentially characterize the spatial statistical properties of the myocardial tissue microstructure (such as muscle bundle arrangement and interstitial components). For example, homogeneous tissue produces fine, regular speckles; while fibrotic or disordered tissue results in coarse, unevenly contrasted speckles.

[0051] Next, the speckle texture features of each frame are arranged chronologically to form a texture feature sequence of the image segment data. This texture feature sequence characterizes how the microstructural properties of the subendocardial myocardium dynamically evolve with each heartbeat within a segment lasting tens of seconds. For example, it describes the continuous change in myocardial tissue during a cardiac cycle, from a relatively relaxed state during diastole (corresponding to a certain texture pattern) to a state of tight arrangement, shortening, and thickening of muscle fibers during systole (corresponding to a change in texture pattern). Sequencing the speckle texture features aims to obtain the dynamic pattern of speckle texture features changing with the cardiac cycle. The texture feature sequence contains rich functional information about myocardial contraction coordination and diastolic relaxation rates, providing raw materials for subsequent rhythmic analysis.

[0052] Next, the boundaries of cardiac cycles in the texture feature sequence are identified (usually based on the R wave of a synchronized electrocardiogram or the main cycle estimated from the sequence itself). Using these boundary points as a reference, the continuous time series is cut into independent cardiac cycle segments, and interpolation resampling and other techniques are used to strictly align the cycle segments of different lengths on the time axis so that the same physiological moment in different cycles (such as 50ms after the R wave) has the same index position on the data sequence.

[0053] Finally, based on the texture feature sequence, a dynamic regularity quantification index is obtained by calculating the consistency measure of texture features at the same phase point across multiple cardiac cycles, thus quantifying the repeatability of myocardial micromotion. For each identical physiological phase point on the aligned data from multiple cycles, the dispersion (e.g., standard deviation, variance, or cross-correlation coefficient) between texture feature vectors at that point across all cycles is calculated. This consistency measure directly and quantitatively characterizes the repeatability and stability of myocardial micromotion patterns across multiple consecutive heartbeats. High consistency means the heart functions like a sophisticated instrument, highly reproducing the same microscopic deformation patterns with each beat; low consistency suggests subtle, inconsistent, or disordered contraction and relaxation processes of the myocardium, a sensitive marker of early cardiac dysfunction. A heart with adequate systolic function but decreased regularity may have a worse long-term prognosis than a heart with good regularity and equivalent function. Therefore, the dynamic regularity quantification index can serve as a new dimension for assessing cardiac mechanical stability and coordination, and its physiological significance complements volumetric indices such as ejection fraction.

[0054] The comprehensive quantification module integrates preliminary indicators from all respiratory phase segments, performs two-level correction and fusion based on respiratory phase type and image quality confidence, and generates a stable comprehensive regularity quantification indicator.

[0055] The respiratory phase type is associated with the image segment data and characterizes the specific load state of the heart at the time the image segment data was acquired (such as "end-expiration" or "mid-inspiratory").

[0056] Because the cardiac load varies at different respiratory phases, even for the same heart, the theoretical expected values ​​of its dynamic regularity quantitative indicators may differ. For example, increased cardiac filling at the end of inspiration may physiologically increase the dynamic regularity quantitative indicators. Therefore, respiratory phase type is used to normalize the dynamic regularity quantitative indicators to eliminate this systematic bias caused by differences in physiological state. Specifically, this can be done by querying a predefined or learned correction coefficient table or function based on the respiratory phase type data, correcting the original indicators to an equivalent value under a standardized reference physiological state (such as quiet end-expiration). This makes indicators from different respiratory phases eligible for direct comparison and fusion.

[0057] Image quality confidence data is also associated with image segment data. This data is a quantitative assessment of the imaging quality of each image segment (such as signal-to-noise ratio, edge sharpness, and degree of motion artifacts), characterizing the reliability of the image segment data as evidence.

[0058] Because the image quality of different segments varies due to random factors such as acoustic window and patient cooperation, noise in low-quality segments would contaminate the final result if all segments were treated equally. Therefore, a fusion weight is assigned to each normalized dynamic regularity quantification index based on the image quality confidence data, with higher image quality confidence data receiving a larger weight. Then, a weighted average or similar operation is performed.

[0059] The comprehensive regularity quantification index is the final output after dual optimization through physiological state correction and quality-weighted fusion. It characterizes the overall level of myocardial micromotor regularity observed throughout the observation period, after removing respiratory interference, correcting for physiological baselines, and suppressing random image noise. The comprehensive regularity quantification index has stronger robustness, stability, and physiological comparability, making it a reliable basis for the final risk assessment.

[0060] The risk assessment output module compares comprehensive regular quantitative indicators with preset risk assessment thresholds and outputs the risk assessment results for heart failure readmission.

[0061] By comparing comprehensive, regular quantitative indicators with risk assessment thresholds (which may involve multi-level judgments of multiple thresholds), the system outputs a discrete risk level (e.g., "low risk," "medium risk," "high risk") or a continuous risk probability. The heart failure readmission risk assessment results characterize the likelihood of a current patient being readmitted due to heart failure within a specific future timeframe. This provides clinicians with objective, quantitative decision support information, which can be used to identify high-risk patients, optimize follow-up strategies, and guide individualized treatment, thereby potentially reducing readmission rates. The aforementioned risk assessment threshold is a preset value, an objective cutoff point determined through standardized retrospective clinical cohort studies.

[0062] This application further discloses the process of normalizing physiological states for dynamic regularity quantitative indicators, as follows:

[0063] A1. Refined classification of respiratory phase types;

[0064] The respiratory phase types of image segment data are refined and classified into expiratory-dominant phase, inspiratory-dominant phase, or respiratory transition phase.

[0065] The respiratory phase type is a structured classification label determined by the algorithm. The determination is based not only on the instantaneous amplitude of the respiratory signal, but also on the temporal position, duration, and trend of the signal change of the image segment throughout the respiratory cycle.

[0066] The cardiac preload and afterload (i.e., filling pressure and ejection resistance) are not constant throughout inspiration or expiration. The expiratory-dominant phase encompasses the period from peak expiratory pressure to end-expiratory phase, characterized by a positive increase in intrathoracic pressure and a relative decrease in venous return; the inspiratory-dominant phase is the opposite; and the respiratory transition phase captures the brief transition between inspiration and expiration or between respiration, during which the rate of change in intrathoracic pressure is greatest, and its instantaneous mechanical effects on the heart may be the most complex. This classification method is used in subsequent analyses to address the problem that traditional methods cannot distinguish the differentiated impacts of different phases of the respiratory cycle on cardiac load.

[0067] A2. Establish a quantitative mapping relationship between respiration and cardiac load;

[0068] The dynamic regularity quantification index of image segment data is input into the pre-trained phase-load regression model; the phase-load regression model takes the respiratory phase type and proportion as input and outputs the estimated value of cardiac load status.

[0069] A phase-load regression model is a mathematical function constructed using machine learning (such as linear regression, support vector regression, or neural networks) or based on statistical analysis of population physiological data. Its inputs are the aforementioned respiratory phase type data, along with optional quantitative features (such as the proportion of a specific phase type in the segment, signal intensity). Its output is an estimate of cardiac load status. This estimate is a continuous scalar or vector designed to quantitatively characterize the expected relative level of the overall load borne by the heart (especially the left ventricle) at a specific respiratory phase. For example, the model might be trained to output a higher load estimate (corresponding to higher preload) during the "inspiratory-dominant phase" and a lower load estimate during the "expiratory-dominant phase."

[0070] Directly dependent dynamic regularity indicators are affected by workload status: regularity indicators may perform better due to the Frank-Starling mechanism when workload is moderately increased; however, they may decline due to decompensation when workload is abnormally high. The role of the phase-load regression model is to provide a baseline workload prediction for each segment, independent of the patient's current cardiac function and based solely on their respiratory status. By estimating using the phase-load regression model, the observed indicator values ​​are correlated with a predicted workload status based on pure respiratory status, providing a benchmark for subsequent targeted compensation.

[0071] A3. Perform reverse compensation calculation;

[0072] The estimated cardiac load status output by the phase-load regression model is dynamically corrected against the standard load status reference value, i.e., reverse compensation calculation.

[0073] The estimated cardiac workload is compared to a pre-defined standard workload reference. The standard workload reference typically corresponds to an ideal, stable assessment state, such as the quiet end-expiratory phase, where cardiac workload is relatively stable and reproducible, and is often used as a standard window for echocardiography. The resulting difference quantifies the degree and direction of deviation of the current segment's workload from the standard state.

[0074] Based on this difference, the original dynamic regularity quantitative indicators are mathematically compensated and adjusted. For example, if the phase-load regression model estimates the current state as high load (the difference is positive), and high load may theoretically inflate the regularity indicators, the compensation algorithm will lower the dynamic regularity quantitative indicators according to certain rules (possibly non-linearly) to offset the inflated portion caused by the load, bringing them back to the indicator level that the heart should exhibit under the assumption of a standard load state.

[0075] This step is the essential operation for normalizing physiological states, generating a set of dynamic, regular quantitative indicators after normalization. All segments of these dynamic, regular quantitative indicators have been mathematically projected or corrected to the same comparable physiological baseline (standard load state). This fundamentally solves the systematic bias that arises when directly comparing indicators collected at different respiratory phases due to varying physiological conditions. After this processing, the differences in indicators from different segments such as end-expiration and inspiration will more purely reflect the differences in cardiac function itself, rather than the influence of respiratory-induced load changes. This significantly improves the fairness and accuracy of subsequent fusion to generate comprehensive, regular quantitative indicators, ensuring that the final risk assessment is based on a more stable measure of intrinsic cardiac function.

[0076] This application also further discloses the specific process of calculating reverse compensation using dynamic regularity quantitative indicators, as follows:

[0077] A31. Using the difference between the estimated cardiac load state and the standard load state reference value as input, the corresponding compensation coefficient is obtained by using a piecewise compensation function.

[0078] The model-predicted physiological state (estimated cardiac workload) is quantitatively compared with an ideal, stable assessment benchmark (standard workload reference, typically corresponding to quiet end-expiration) to obtain a difference value. This difference value characterizes the deviation of the current image segment's cardiac workload state relative to the standard assessment conditions. A positive difference value indicates that the estimated workload is higher than the reference benchmark (e.g., during inspiration), while a negative value indicates that it is lower than the benchmark (e.g., during deep expiration). This difference value is used as the sole input to the subsequent compensation function, thus transforming the complex physiological state differences into a clear driving signal that can be processed by a mathematical function.

[0079] A32. The overall structure and interval division of the piecewise compensation function;

[0080] The absolute value of the difference is taken to quantify the degree of deviation. Based on the magnitude of the absolute value of the difference, three difference intervals are divided and different compensation strategies are applied to each interval.

[0081] Based on the absolute value of the difference, from low to high, the three difference intervals are: the first difference interval, the second difference interval, and the third difference interval; the following is a separate introduction to each of the three difference intervals:

[0082] First Difference Interval: This interval corresponds to situations where the absolute value of the difference is relatively small, typically encompassing physiological fluctuations or mild pathological conditions. Within the first difference interval, the piecewise compensation function is configured to output a first compensation coefficient that exhibits a first linear relationship with the difference value. This means that the compensation coefficient changes proportionally with the difference value, and the compensation behavior is smooth and predictable.

[0083] This functional relationship simulates the linear compensatory capacity of the heart within its physiological range or a range of mild abnormalities. Within this range, cardiac function responds relatively linearly to changes in preload (volume load) according to the Frank-Starling law. Therefore, the compensation for dynamic regularity quantitative indicators also adopts a linear relationship, aiming to smoothly and accurately eliminate the effects of these mild load changes.

[0084] Second Difference Interval: When the absolute value of the difference increases and enters the second difference interval, the piecewise compensation function switches to the second nonlinear relationship; as the absolute value of the difference increases, the rate of change of the second compensation coefficient decreases. In other words, the second compensation coefficient still increases with the increase of the difference value, but the rate of increase gradually slows down.

[0085] The piecewise compensation function in the second difference interval directly simulates the nonlinear characteristics of the heart's transition from the compensated to the decompensated phase. When the workload increases significantly, the heart may gradually approach the limit of its functional reserve. At this point, the deterioration of its mechanical motion (reflected in microscopic regularity) may intensify, but it also means that the heart's own regulatory capacity decreases, and its response to changes in external workload becomes sluggish and unstable. The use of nonlinear compensation with a decreasing rate of change is precisely to match this physiological reality: for large workload deviations, the compensation behavior should become more cautious and conservative. Overcompensation (assuming that linearity continues to increase rapidly) may incorrectly correct those functional decline signals that are truly caused by decompensation. The design of the second difference interval effectively prevents miscorrection caused by mechanical linear compensation under obvious pathological conditions, protecting the integrity of the real pathological signals.

[0086] Third difference interval: When the absolute value of the difference is very large, the piecewise compensation function no longer pursues a precise functional relationship with the difference value, but instead limits the output third compensation coefficient to a preset saturation compensation value.

[0087] The piecewise compensation function in the third difference interval establishes an upper or lower limit for the compensation amount. This firstly prevents absurdly large, numerically unstable compensation amounts from being generated when extreme abnormal data (which may be due to signal acquisition errors, severe arrhythmias, or abnormal model estimations) are input, thus avoiding overflow or distortion of the final indicator. Secondly, physiologically speaking, it also implies that there is uncertainty in the understanding of the impact of extreme load deviations, therefore adopting the most conservative and limited compensation strategy to minimize potential risks.

[0088] A33. Compensation execution and indicator generation, completing mathematical corrections;

[0089] By using the first, second, or third compensation coefficients output by the piecewise compensation function, the dynamic regularity quantitative index is corrected to generate a dynamic regularity quantitative index after normalization of physiological state.

[0090] The correction operation for dynamic regularity quantitative indicators using compensation coefficients is usually a multiplication operation, for example:

[0091] ;in, It is a dynamic and regular quantitative indicator after normalization of physiological state. It is a dynamic and regular quantitative indicator. It is the first compensation coefficient, the second compensation coefficient, or the third compensation coefficient. It could be a value slightly greater than, equal to, or less than 1, corresponding to raising, maintaining, or lowering the original indicator, respectively.

[0092] Through this step, each raw index from different respiratory phase segments is individually adjusted based on its own associated compensation coefficient, estimated by the physiological model and determined by a piecewise function. The output dynamic regularity quantitative index after normalization of physiological state has mathematically eliminated the systematic bias introduced by changes in load due to respiration to the greatest extent possible, allowing all indexes to be compared from a unified starting line of standard physiological load.

[0093] The specific mathematical implementation of the piecewise compensation function mentioned above is as follows:

[0094] Input is This refers to the difference between the estimated cardiac workload and the standard workload reference value. The output of the piecewise compensation function is the compensation coefficient. .

[0095] Set a threshold for the absolute values ​​of the two differences: , (in ), which correspond to the boundaries of the first, second, and third difference intervals, respectively.

[0096] The first difference interval, i.e. In this range, the compensation coefficient and difference value It exhibits a linear relationship, aiming to achieve smooth and predictable correction.

[0097] At this point, the first compensation coefficient Calculated using the following linear function:

[0098] ;

[0099] in, These are signed difference values ​​(positive indicates high load, negative indicates low load).

[0100] It is a preset linear gain coefficient ( This determines the sensitivity of the compensation coefficient to load differences.

[0101] 1 is the baseline value, meaning that when the difference value When the load is at standard load, the compensation factor is 1 and no correction is made.

[0102] The first linear relationship directly simulates the approximate linear response of the heart within physiological fluctuations or mild abnormal load ranges. According to the Frank-Starling mechanism, within a certain range, cardiac output (or related functional indicators) is linearly positively correlated with preload (volume load). Therefore, when a slight deviation from the standard state of load is detected, linear compensation is used to inversely correct this effect, which conforms to basic physiological principles. This achieves smooth, controllable, and predictable correction for physiological disturbances or early mild pathological disturbances. A preset linear gain coefficient is used. This can be determined by regression analysis on data from healthy individuals or patients with clearly defined stages, so that the compensated indicators can more stably reflect the intrinsic function of the heart under mild load changes, providing clean baseline data for subsequent fusion analysis. The value depends on the statistical analysis of a large amount of training data, for example, The range is set to be several times the standard deviation of the estimated load values ​​of healthy volunteers under calm breathing conditions, to ensure that the range covers the major physiological fluctuations.

[0103] The second difference interval, i.e. Second compensation coefficient Calculated using a nonlinear function, this function must satisfy the following condition: When it increases, The rate of change decreases.

[0104] A typical implementation that is both engineering-practice- and physiologically-interpretable is to use a saturated growth model, such as a fractional function or a modified exponential function of the following form;

[0105] Fractional functions:

[0106] ;

[0107] in, , All parameters are preset and take positive values. This function is for... The derivative (rate of change) with Increases and decreases.

[0108] Modified exponential function:

[0109] ;

[0110] in, , All parameters are preset and have positive values. This is a symbolic function. This function... The increase asymptotically tends towards Its rate of change (derivative) also follows It increases but then decreases exponentially.

[0111] The second nonlinear relationship aims to simulate the decompensation trend of cardiac function under moderate to severe abnormal load. When the load deviation is significant, the heart's self-regulation ability decreases, and its functional indicators become sluggish and unstable in response to further load changes (nonlinearity, saturation effect). The decreasing rate of change in the above functional form is precisely to correspond to this reduced response sensitivity. The second nonlinear relationship provides a prudent and robust compensation strategy. Compared with continuous linear amplification compensation, the decreasing rate of change ensures that the compensation behavior is not overly aggressive when faced with large load differences. This effectively prevents the erroneous compensation of the real functional decline caused by cardiac decompensation itself, thereby protecting the integrity of the pathological signal. For example, the load difference value of a heart with poor regularity due to severe volume overload... The value is large, but the nonlinear compensation function will only apply a limited correction (tending towards the saturation value). or This is to avoid generating a false, corrected-looking indicator.

[0112] parameter , or , and interval boundaries The value of can be determined and optimized through machine learning training on clinical datasets. The training objective can be to ensure that the compensated indicator, after excluding cases clearly dominated by load differences, has the strongest independent association with the gold standard measurement of cardiac function or clinical endpoint events. This ensures that the compensation function is not only mathematically sound but also clinically effective in prediction.

[0113] In a specific system, it is necessary to ensure that the function is at the interval boundary ( To ensure a continuous or smooth transition, so as to avoid abrupt changes in the compensation coefficient. For example, by setting parameters to make .

[0114] The third difference interval, namely When the absolute value of the input cardiac load status difference value Exceeding the upper limit of the second difference interval When entering the third difference interval, the compensation function will ignore The specific value is constant, and a preset saturation compensation value is output. The main function of the saturation compensation value is to act as an absolute output limiter for the compensation system.

[0115] The preset saturation compensation value is determined based on the algorithm's robustness requirements and a safety boundary derived from comprehensive statistical analysis of clinical data. The design principles, determination methods, and technical effects of this saturation compensation value are fully explained below.

[0116] Algorithm robustness requirement: From the perspective of signal processing and system control, this is a standard strategy to prevent overshoot and ensure numerical stability. It ensures that under any extreme input, the system output (compensation coefficient) is limited to a safe, pre-verified range, thereby avoiding non-physiological extreme values ​​(such as close to zero or abnormally large values) in the normalized index due to uncontrolled compensation coefficients, which could lead to serious distortion of risk assessment results or system errors.

[0117] Saturation compensation value The preset is usually based on a combination of one or more of the following methods:

[0118] 1. Determination of statistical extreme values ​​based on the distribution of clinical data:

[0119] In a dataset containing a large number of healthy individuals and patients with varying degrees of heart failure, the nonlinear function is used to calculate the difference between all samples within the second interval of variation. The resulting second compensation coefficients. Then, the high percentile (e.g., the 99th or 99.5th percentile) of the distribution of these second compensation coefficients is taken as... Candidate values.

[0120] This means It covers over 99% of the compensation needs reasonably derived from interpretable load differences in the training data. The remaining less than 1% of extreme cases are considered anomalies or model extrapolation risk areas, and the system uniformly uses this statistically reasonable upper limit to cap these cases.

[0121] To ensure safety, it may rise slightly, such as ,in For a small safety margin.

[0122] 2. Boundary determination based on consensus among physiological or clinical experts:

[0123] Collaborate with domain experts (cardiac physiologists, heart failure specialists) to determine a reasonable upper limit for the compensation coefficient based on physiological knowledge. For example, expert consensus might suggest that, based on respiratory phase load differences, the maximum correction to myocardial micro-regularity indicators should not exceed a certain percentage of the original indicator. (For example ), then, can be set (For positive compensation) and (For negative compensation, a separate lower limit saturation value needs to be set, or the absolute value should be taken.)

[0124] 3. Smooth connection constraint at the boundary of the function with the second difference interval:

[0125] To ensure the overall smoothness of the function, avoid... A jump occurs at that point. The setting needs to satisfy the boundary continuity condition, that is:

[0126] ;

[0127] This condition can be used for solving or verification. The constraint equations. When the nonlinear function in the second difference interval... Once the form is determined, It is a computable value, and setting it directly as the saturation value is a natural choice.

[0128] Preset saturation compensation value Once the above scheme is determined, it fundamentally eliminates the possibility of extreme abnormal estimates in the respiratory load model caused by sensor failure, severe signal noise, or sudden cardiac arrhythmias in patients (such as premature beats). This compensates for the possibility of the system outputting absurd results. The system output is clamped within a credible range.

[0129] This application also discloses the weighted fusion process of dynamic regularity quantitative indicators, as follows:

[0130] B1. Perform structured decomposition on image quality confidence data;

[0131] The image quality confidence data is deconstructed into three independent sub-features with clear physical meaning: signal-to-noise ratio feature value, boundary sharpness feature value, and artifact perturbation feature value.

[0132] Signal-to-noise ratio (SNR) characteristic value: obtained by calculating the gray-level statistics (such as the ratio of mean to standard deviation) of the target myocardial region and the anechoic area (such as the heart chamber) in the same frame in the image block data; the SNR characteristic value directly characterizes the most basic signal purity of the ultrasound image, reflecting the ultrasound penetration, whether the gain setting is appropriate, and whether there is electronic noise.

[0133] Boundary sharpness feature value: obtained by calculating the statistical characteristics (such as average or maximum value) of the gradient magnitude of pixels near the endocardial boundary (such as using the Sobel operator); the boundary sharpness feature value directly characterizes the clarity and recognizability of the endocardial boundary, which is closely related to the resolution of the ultrasound probe, the direction of the sound beam, and the characteristics of the tissue interface.

[0134] Artifact perturbation features: These are quantified by analyzing unexpected changes in image block data between consecutive frames, such as calculating the variance of the time series of block matching errors or detecting spatial noise in a specific frequency band. Artifact perturbation features directly characterize the instability and distortion of image content caused by residual respiratory motion, slight patient movement, or multiple ultrasound reflections.

[0135] The aforementioned decomposition of image quality confidence data addresses the issues of a single quality assessment dimension and information loss. It refines quality assessment into three orthogonal dimensions: signal strength (signal-to-noise ratio), geometric positioning accuracy (boundary sharpness), and temporal series stability (artifact perturbation). This structured decomposition forms the basis for subsequent refined and differentiated weight allocation, enabling the system to distinguish between a noisy but clearly defined image and a well-signaled but motion-blurred image, thus making a more reasonable contribution assessment.

[0136] B2. Two-stage decision-making using a dynamic weighting function;

[0137] B21. Admission determination based on signal-to-noise ratio characteristic values;

[0138] First, the signal-to-noise ratio (SNR) feature value of each image segment is compared with a preset threshold. Only when the SNR feature value is greater than the preset threshold is the image segment data enabled and allowed to proceed to the next processing step; otherwise, the segment is directly excluded.

[0139] The signal-to-noise ratio (SNR) is a fundamental prerequisite for the reliability of image information. If the underlying signal is severely contaminated by noise, then no matter how clear its boundaries appear or how many artifacts there are, any texture feature sequences and their regularity indicators extracted from it have lost their credible physical meaning. By using this hard threshold judgment, the system fundamentally eliminates the possibility of extremely low-quality data contaminating the fusion results, ensuring that the datasets participating in the fusion have a minimum level of information integrity.

[0140] B22. Fine-tuning the weights based on boundary sharpness eigenvalues ​​and artifact perturbation eigenvalues, and calculating the initial weight base;

[0141] For image segments that pass the admission criteria, the function calculates an initial weight base by combining their boundary sharpness feature values ​​and artifact perturbation feature values ​​using a weighted combination formula. This weighted combination formula typically assigns positive weights to boundary sharpness (the sharper the image, the higher the weight contribution) and negative weights to artifact perturbation (the larger the perturbation, the lower the weight contribution), and may include an interaction term between the two.

[0142] This step involves fine-tuning within the qualified data; even with a satisfactory signal-to-noise ratio, the usability of different segments still varies. A segment with sharp boundaries and no artifacts is an ideal specimen for analyzing myocardial micromotor and should be given high weight. Conversely, a segment with a satisfactory signal-to-noise ratio but blurry boundaries or obvious artifacts may provide information with geometric errors or local distortions, and its weight should be reduced. The initial weighting is precisely the first quantitative assessment of this information ideality.

[0143] B23. The final fusion weight value is obtained by using a nonlinear decay function;

[0144] The initial weight base is fed into a nonlinear decay function to generate the final fused weight values. The key behavior of the nonlinear decay function is defined by a preset weight threshold:

[0145] High-quality retention zone: When the initial weight base is greater than the preset weight threshold, the output value of the nonlinear decay function approaches the input value; this means that for high-quality segments, their weights are basically retained, encouraging them to contribute information fully.

[0146] Low-quality acceleration suppression region: When the initial weight base is less than or equal to a preset weight threshold, the output value of the nonlinear decay function decays at an accelerated rate as the input decreases. This means that for segments with critical or below-critical quality, their weights will be reduced significantly by a rate exceeding linear proportion.

[0147] The above process simulates the nonlinear judgment of human experts on the credibility of initial weighting as evidence:

[0148] When the initial weighted base, representing the quality of evidence, slowly decreases from excellent to good, the system's trust in it decreases relatively slowly; however, when the quality of evidence further falls into the range of barely usable or poor, the system's trust in it drops sharply. The accelerated decay mechanism mathematizes this intuition. First, this mechanism achieves strong suppression of low-to-medium quality data, ensuring that the final fusion result is dominated by high-quality fragments, greatly reducing the risk of overall index fluctuations due to poor image quality in some parts. Second, compared to simple linear weighting, this non-linear suppression provides stronger denoising capabilities, making the system more tolerant of image quality fluctuations and enhancing its applicability in complex real-world clinical acquisition environments.

[0149] B3. Weighted average generates a comprehensive index;

[0150] Using the fusion weight value calculated for each enabled segment, a weighted average is performed on the corresponding dynamic regularity quantitative indicators that have been normalized by physiological state to obtain the final comprehensive regularity quantitative indicator.

[0151] Weighted averaging is a standard mathematical operation for information fusion, and its weights are the fusion weight values ​​output by the dynamic allocation function mentioned above.

[0152] Following the aforementioned quality assessment and weighting process, this step is performed on a purified and balanced set of evidence. High-quality, high-confidence segments dominate the final composite index; the influence of low-quality segments is effectively limited; and invalid segments are completely excluded. The resulting comprehensive and regular quantitative index is a highly robust cardiac function assessment result that encapsulates the essence of high-quality imaging information and resists interference from various image acquisition defects to the greatest extent. This provides a reliable data foundation for making accurate risk assessments.

[0153] The calculation process for the initial weight base mentioned in step B22 is further disclosed in this application, as follows:

[0154] B221, Eigenvalue normalization processing;

[0155] The original boundary sharpness eigenvalues ​​and artifact perturbation eigenvalues ​​are normalized to obtain normalized boundary sharpness values ​​and normalized artifact perturbation values. Normalization is a mathematical transformation process, typically achieved through linear scaling, mapping to the cumulative distribution function based on historical data distribution, etc. It maps two original eigenvalues ​​that may have different dimensions, numerical ranges, or even inconsistent distribution characteristics to a unified, dimensionless numerical interval (e.g., ...). or Furthermore, it is ensured that the larger the mapped value, the better the image quality in that dimension.

[0156] B222. Calculate the geometric mean as the basic weighting factor;

[0157] Calculate the geometric mean of the normalized boundary sharpness value and the normalized artifact perturbation value, and use it as the basic weighting factor. .

[0158] Compared to the arithmetic mean, the geometric mean has a core mathematical characteristic: it is sensitive to low-value anomalies. Specifically, the baseline weighting factor will only be high when both the normalized boundary sharpness value and the normalized artifact perturbation value are high; if either of these values ​​is low, it will significantly lower the baseline weighting factor result. This logic simulates the "weakest link" principle in cardiac ultrasound image analysis: the overall reliability of an image is determined by its weakest link. An image with clear boundaries but severe artifacts, or an image with almost no artifacts but blurred boundaries, both have serious deficiencies in diagnostic value.

[0159] Geometric mean as the basic weighting factor This approach mandates that both key quality dimensions must be excellent in tandem to achieve a high baseline score. This reflects the reliability of true clinical information better than the arithmetic mean (where a high score in one dimension can partially compensate for a low score in another), effectively preventing "unbalanced" images from receiving undue high weight.

[0160] B223. Calculate the mass unevenness;

[0161] While calculating the basic weighting factors, the quality imbalance, i.e., the absolute value of the difference between the normalized boundary sharpness value and the normalized artifact perturbation value, was calculated in parallel.

[0162] The mass unevenness is a simple scalar quantity whose range is within the range of Between these values, the balance of the quality structure is assessed.

[0163] The purpose of introducing the quality imbalance is that, even with the same geometric mean (i.e. the same absolute level), different quality structures may imply different risks.

[0164] For example:

[0165] Scenario 1: The normalized boundary sharpness value and the normalized artifact perturbation value are both 0.9;

[0166] Scenario 2: The normalized boundary sharpness value is 0.99, and the normalized artifact perturbation value is 0.81.

[0167] The geometric mean of Scenario 1 and Scenario 2 are similar, but the former's quality structure is obviously more ideal and reliable.

[0168] Quality Imbalance is specifically designed to capture and quantify the degree of this imbalance. A larger quality imbalance value indicates a greater mismatch in the quality performance of the two dimensions, suggesting that the image may have a single but significant defect (such as artifacts or blur). This defect pattern can be more misleading than a case where both dimensions are moderate but balanced.

[0169] B224. Calculate the equilibrium adjustment factor using the equilibrium adjustment function;

[0170] The equilibrium adjustment factor is calculated based on the monotonically decreasing equilibrium adjustment function.

[0171] The equilibrium adjustment function is explicitly configured as follows:

[0172] When the quality is uneven When the value is 0, the balance adjustment factor This is the maximum value (usually 1, indicating no penalty);

[0173] When the quality is uneven When increased, the balance adjustment factor Non-linear decrease.

[0174] The function can take the form of:

[0175] Inverse decay: ;

[0176] Exponential decay: ;

[0177] in, , All of these are preset coefficients, and all of them have values ​​greater than 0.

[0178] The above process is a direct response to and penalty mechanism for quality imbalance. It means that as the degree of imbalance increases from slight to significant, the penalty for the weights increases exponentially. This simulates expert judgment: slight imbalances are tolerable, but the reliability of images with severe imbalances drops sharply.

[0179] Balance adjustment factor Its function is as a penalty coefficient: multiplying the base weight factor by If the image quality is balanced ( Small, If the image quality is close to 1, the basic weighting factor is largely preserved; if the image quality is severely uneven ( big, If the value is significantly less than 1, the base weight factor will be significantly discounted. This mechanism ensures that the final weight is highly sensitive to the balance of the quality structure, making balanced, high-quality images take precedence over imbalanced, pseudo-high-quality images.

[0180] B225, Initial weight base for synthesis;

[0181] Finally, the basic weighting factors With balance adjustment factor Multiply to obtain the initial weight base. :

[0182] .

[0183] Basic weighting factor This represents the potential for quality contribution under the ideal equilibrium assumption, while the equilibrium adjustment factor... This represents the discount that needs to be applied due to actual imbalances. Multiplying the two together produces a comprehensive weight value that requires both a high absolute level and a high degree of balance.

[0184] An image must simultaneously possess high edge sharpness and low artifact perturbation (to ensure...) Only with a high value can a high base weight be obtained.

[0185] At the same time, these two dimensions must also be balanced (ensuring) The value is small, therefore Only by having a high value can one avoid being severely punished.

[0186] Only image segments that simultaneously meet the criteria of "high quality" and "balanced quality" can obtain the highest initial weight base. This design allows the weights to more finely distinguish images of different quality modes, thus enabling the most comprehensive and reliable image data to play a dominant role in subsequent fusion.

[0187] Regarding the nonlinear decay function mentioned in step B23, this application also provides further disclosure, as follows:

[0188] The nonlinear decay function uses the initial weight base This is the input, and its value can be between 0 and 1.

[0189] when Greater than the preset weight threshold At that time, the fusion weight value The formula for determining the value of is:

[0190] ; It is a decay rate parameter that is greater than zero.

[0191] The above formula is based on a preset weight threshold. It consists of a constant term and a series of correction terms.

[0192] Correction items It has the following characteristics:

[0193] when rigidity greater than At that time, that is It is a small positive number.

[0194] Due to the exponential term At this point, the entire formula is:

[0195] The formula's output value is very close to the input value. .

[0196] along with As it continues to increase, the exponential term begins to decrease, causing the growth rate of the correction term to slow down, resulting in... Although still following It will increase, but will gradually fall slightly below... And the gap increases It increases in size and expands slightly.

[0197] decay rate parameter The intensity of this discount was controlled: The smaller the value, the slower the exponent decays, and the better the output. The closer ; The larger the value, the more significant it is for values ​​significantly above the threshold. The weights of these segments will be more significantly suppressed to prevent a very small number of extremely high-weight segments from excessively dominating the fusion result, thus enhancing the balance of the system.

[0198] when Less than or equal to the preset weight threshold At that time, the fusion weight value The formula for determining the value of is:

[0199] ; It is a parameter representing the acceleration attenuation intensity that is greater than zero.

[0200] At this point, the formula is composed of the input. With exponential decay factor The product of these two terms, the core of which lies in the exponent. .

[0201] Regarding the above formula Taking the derivative, we can prove that in Within the interval, its derivative is positive, but its second derivative is negative (when...). (When it is large enough). This means... Follow The rate of decrease (the absolute value of the first derivative) is itself accelerating, meaning there is an accelerating decay effect.

[0202] when from When the point begins to fall (e.g.) ), due to the square term Very small ( The exponent term is close to 1, so The decline was gradual.

[0203] along with Further reduce (e.g.) ), Become exponent term Significantly smaller than ,lead to Not only because of the multiplier As it shrinks and decreases, it suffers a double blow due to the sharp decrease in the exponential term, causing the rate of decline to increase rapidly.

[0204] Accelerated decay intensity parameters Controlling the severity of inhibition: The larger the value, the faster the exponential decay; for anything below the threshold... The weighting base is drastically reduced, and the final fusion weights are drastically compressed. This ensures that the contribution of segments that barely meet the quality standards or are substandard to the final comprehensive index is suppressed to an extremely low level, thereby significantly improving the purity and robustness of the fusion results.

[0205] threshold Attenuation rate parameter and accelerated decay intensity parameters These are all preset parameters, which can be determined by optimization on the training dataset. The optimization objective can be to maximize the predictive association strength between comprehensive regularity quantitative indicators and clinical endpoints (such as readmission events), or to minimize the variation of indicators in repeated measures.

[0206] To ensure function smoothness, it is usually done in The condition requires that the two piecewise expressions have equal values. This imposes a constraint: This condition is automatically met.

[0207] This application further discloses the process for acquiring image segment data, as follows:

[0208] E1. Filtering, feature extraction and phase inflection point identification of respiratory signals;

[0209] Using real-time respiratory data as the operational object, the goal of the operation is to identify phase transition points with clear physiological significance within each respiratory cycle.

[0210] First, real-time respiratory data is filtered and features are extracted. Real-time respiratory data is typically a continuous waveform signal (such as chest and abdominal movement, respiratory airflow, or a respiratory proxy signal derived from image processing). First, high-frequency noise and baseline drift are removed using digital filtering (such as bandpass filtering). Then, feature extraction and event detection are performed according to preset recognition rules. These preset recognition rules are based on the morphological features of the respiratory waveform, for example:

[0211] Identify local maxima as the peak of inspiratory activity or the start of expiratory activity.

[0212] Identify local minimum points as the end of expiration or the beginning of inspiration.

[0213] By calculating the first derivative (slope) of the signal, the zero-crossing point where the slope changes from positive to negative or from negative to positive can be identified. The zero-crossing point may correspond to the instant when the direction of respiratory flow changes.

[0214] These identified key time points are called phase inflection points, which divide the continuous respiratory waveform into segments with clear physiological significance in the time domain.

[0215] Phase inflection points are objective markers of state transitions within the respiratory cycle. Identifying these points is the basis for subsequent staging of the respiratory cycle. Transforming subjective and vague descriptions of inspiration or expiration into a series of repeatable timestamps provides temporal anchors for establishing the mapping relationship between image frames and respiratory phases.

[0216] E2. Divide the respiratory phase window and define the phase type;

[0217] Based on the phase inflection points identified in the previous step, the continuous, cyclical respiratory cycle is divided into discrete, standardized respiratory phase windows.

[0218] A breathing phase window is defined as the interval between two consecutive specific inflection points.

[0219] For example, the period between the onset of inspiration and the peak of inspiration can be defined as the inspiratory acceleration phase; the period between the peak of inspiration and the onset of expiration can be defined as the inspiratory plateau phase or the end of inspiration. The phase type defined for each window is a structured label that contains not only information about inhalation or exhalation, but also the relative position and physiological significance within that cycle.

[0220] This step constructs a structured, discrete physiological state time coordinate system, enabling precise modeling of different load change phases within the respiratory cycle. For example, although early and late expiration both belong to expiration, intrathoracic pressure and cardiac load may differ. This division allows the system to more accurately correlate specific respiratory phases with expected cardiac load states, providing direct input data for subsequent physiological state normalization processing based on respiratory phase type.

[0221] E3, ultrasound image frames are mapped to the respiratory phase window;

[0222] When acquiring sequential images, echocardiography equipment records an acquisition timestamp for each frame of the echocardiogram sequence data. Simultaneously, each respiratory phase window has its start and end time boundaries (defined by inflection points). By comparing the timestamps of the image frames with the time boundaries of each phase window, each image frame is mapped to a unique corresponding respiratory phase window.

[0223] E4. Merging and fragment generation of image frame sequences;

[0224] Iterate through all respiratory phase windows, and for each window, collect all ultrasound image frames mapped to that window. If these image frames are temporally consecutive in the original sequence, they are merged sequentially and packaged into a single image segment data. Simultaneously, this segment is associated with the phase type of the respiratory phase window it belongs to, serving as its metadata.

[0225] Each frame within a generated image segment shares the exact same respiratory phase type. This means that the respiratory modulation force exerted on the heart is relatively stable and consistent over the several seconds of the segment. This provides a quasi-steady-state analysis environment for subsequent image segment processing modules, ensuring that observed cardiac motion variations within the segment are primarily attributed to the heart's own rhythm rather than respiratory interference, significantly improving the signal-to-noise ratio. By generating image segment data, preliminary decoupling of respiratory motion interference from the heart's inherent motion is achieved at the data preprocessing level, laying a solid foundation for subsequent exploration of the pure myocardial micromotor regularity.

[0226] This application further discloses the process for obtaining speckle texture features, specifically as follows:

[0227] F1, multi-scale spatial filtering processing;

[0228] First, multi-scale spatial filtering is performed on the image patch data to extract the filter response maps at at least two different spatial scales.

[0229] Multi-scale spatial filtering refers to using a set of filters with different spatial support domains (such as Gaussian kernels, Gaussian-Laplacian kernels, or Gabor filter banks of different sizes) to perform convolution operations on the same image patch data. Each filter produces a filtered response map, whose pixel values ​​represent the response intensity of the original image within a specific spatial frequency band defined by that filter. For example, a small-scale Difference of Gaussians (DoG) filter is sensitive to fine, high-frequency speckle grains; while a large-scale smoothing filter suppresses high-frequency details and highlights the macroscopic trends of texture.

[0230] The formation of ultrasonic speckle is the result of the coherent superposition of numerous sub-resolution scatterers within the sound beam. The particle size and distribution exhibited are closely related to the point spread function (resolution) of the ultrasonic system and the spatial distribution density of scatterers within the tissue. The purpose of multi-scale filtering is to decouple the mixed speckle signals at frequencies.

[0231] Small-scale response: Primarily captures subtle speckle fluctuations close to the system resolution limit, generated by the microstructures of cardiomyocytes themselves and their adjacent interstitial structures (such as capillaries and fine fibrosis). Texture changes at this scale may be highly sensitive to cell swelling, edema, or early fibrosis.

[0232] Large-scale response: This primarily reflects the gradual intensity variation pattern of speckle particles spanning several dimensions, resulting from more macroscopic structures such as muscle bundles, interfascicular connective tissue, or sheet-like fibrotic regions. This scale characteristic may be more specific for myocardial remodeling and scar formation.

[0233] By extracting responses at multiple scales in parallel, texture components that may contain different physiological and pathological information are effectively separated at the source of feature extraction, laying the foundation for the subsequent construction of information-rich feature descriptors.

[0234] F2. Calculation of local statistical distribution characteristics;

[0235] For each filtered response map, the statistical distribution characteristics of the local region are calculated. The statistical distribution characteristics include at least variance characteristics and entropy characteristics.

[0236] The aforementioned local region refers to sliding a fixed-size window (such as 5x5 or 7x7 pixels) on the filter response map, and performing calculations independently at each window position, rather than calculating global statistics for the entire image patch.

[0237] Variance characteristics of local regions: Calculate the variance of the gray values ​​of all pixels within the window to quantify the fluctuation range (contrast) of the response intensity in that local region. In speckle images, a large variance indicates that the acoustic impedance of the scatterers (tissue microstructures) in that region changes drastically or is unevenly arranged.

[0238] Entropy characteristics of local regions: Calculate the Shannon entropy of the grayscale value distribution within the window to quantify the randomness and disorder of the grayscale pattern in that local region. High entropy values ​​mean that pixel intensity changes lack discernible patterns and the texture is complex; low entropy values ​​mean that the intensity distribution is more regular and more predictable.

[0239] Traditional texture analysis often calculates global statistics for the entire region of interest, which averages out spatial heterogeneity information. Myocardial lesions (such as ischemia and fibrosis) are often focal and unevenly distributed. Local window statistics can spatially locate and quantify this heterogeneity, preserving the statistical "peaks" or "troughs" that may be present in the lesion area, which is crucial for the detection of early and focal lesions.

[0240] The variance characteristics of local regions are directly related to the mean scattering intensity and microstructural contrast of the tissue. For example, cellular edema may lead to an overall change in scattering intensity, and the difference in acoustic impedance between fibrotic tissue and normal myocardium can increase local contrast. The entropy characteristics of local regions are closely related to the structural orderliness of the tissue. Healthy, regularly arranged myocardial fibers may produce relatively uniform, highly periodic speckle patterns, resulting in lower local entropy values; while structurally disordered, anisotropic diseased tissues will produce more random, disordered speckle patterns, leading to higher local entropy values.

[0241] By using the two sets of complementary local statistics mentioned above, the speckle patterns at each scale were quantified in a fine-grained and spatially discriminable manner from the two dimensions of intensity and structure.

[0242] F3, concatenation of multidimensional feature vectors;

[0243] All variance and entropy features calculated from the same image patch data at different spatial scales are concatenated in a determined scale order to form a multidimensional feature vector, which serves as the final speckle texture feature.

[0244] Suppose we use M scales, each scale yielding one variance eigenvalue and one entropy eigenvalue. Arranging all 2M eigenvalues ​​in a predetermined order (e.g., scale 1 variance, scale 1 entropy, scale 2 variance, scale 2 entropy… scale M variance, scale M entropy) forms a 2M-dimensional feature vector. If more scales are used or more local statistical moments (such as skewness and kurtosis) are calculated, the dimension of the feature vector increases accordingly.

[0245] This feature vector systematically integrates multi-scale and multi-attribute information, describing the contrast and order of tissue in both fine and coarse structures. For example, an early fibrosis may show little change in entropy at a small scale, but a decrease in variance at a large scale (because the texture of the fibrotic region becomes homogeneous). This multi-dimensional combination makes the features more capable of representing and distinguishing different types of pathological changes.

[0246] This application further discloses the process for obtaining dynamic regularity quantitative indicators, specifically including:

[0247] G1, the cycle markers that identify the start and end points of the cardiac cycle;

[0248] Identify cycle markers for the start and end points of each cardiac cycle in the texture feature sequence;

[0249] The aforementioned texture feature sequence is a multidimensional vector sequence that varies over time. Periodic markers typically correspond to pivotal moments in electromechanical activity, most commonly the time points corresponding to the R-wave on an electrocardiogram (ECG), which can be obtained by aligning with synchronously acquired ECG signals. If no synchronous ECG is available, the periodic markers can be derived from the sequence itself, for example, by analyzing signals derived from endocardial motion or overall brightness changes to find periodic extrema or by determining the principal period through autocorrelation analysis. These periodic markers are typically defined as the onset of ventricular electroexcitation, serving as physiological triggers for mechanical contraction.

[0250] This step transforms the continuous, absolute timeline into a relative cycle count and intra-cycle phase coordinate system based on the heart's own rhythm, thereby determining which data segments belong to the same heartbeat and performing phase matching between different heartbeats. All subsequent regularity analyses will lose physiological accuracy and comparability.

[0251] G2, perform segmentation of the feature subsequence;

[0252] Based on the identified periodic markers, the long texture feature sequence is segmented into feature subsequences corresponding to multiple consecutive cardiac cycles.

[0253] In terms of operation, with the first Taking the periodic marker as the starting point, the 1st cycle marker... Using the periodic marker point as the endpoint, this segment of multidimensional vector is extracted from the texture feature sequence, thus forming the first... Each feature subsequence contains a complete evolution of speckle texture features in the endocardial region from the onset of electrical excitation to the next electrical excitation within that cardiac cycle. Decomposing the non-stationary long-term data stream into a set of parallelizable, naturally aligned, finite-length data packets (R waves) provides structured data input for subsequent comparisons of the microscopic evolution patterns of different heartbeats within the same physiological framework, and is a prerequisite for performing cross-cycle comparisons.

[0254] G3, Alignment and Resampling of Feature Subsequences;

[0255] The segmented feature subsequences are aligned on the time axis based on their periodic markers, and the aligned feature subsequences are then resampled.

[0256] The starting points (i.e., period markers) of all feature subsequences are set to zero on the time axis to achieve precise alignment of different heartbeats at their starting moments. Due to heart rate variability, the lengths (i.e., RR intervals) of each feature subsequence differ. Therefore, resampling (such as linear interpolation or spline interpolation) is required to map each feature subsequence of varying lengths to a sequence with the same number of intervals. A standard-length sequence of characteristic phase points. The points are evenly or non-uniformly distributed on a standardized cardiac cycle time axis (e.g., from 0% to 100%) according to physiological models.

[0257] Alignment ensures that the same physiological starting state of different heartbeats is compared together;

[0258] Resampling ensures that at any given, standardized percentage of physiological phase (e.g., mid-systolic -30%, early diastolic -70%), there exists a feature vector from each heartbeat available for comparison.

[0259] This step enables the system to compare the texture features of all heartbeats at the same physiological moment of mid-systole, rather than simply comparing the first... Millisecond data (which may be in the middle of systole for heartbeat A, but already in diastole for heartbeat B) lays the temporal foundation for quantifying consistency across cycles.

[0260] G4. Calculation of the consistency metric value of the phase point;

[0261] At each standardized feature phase point, the aggregated dispersion of the texture feature vectors of all aligned and resampled cardiac cycles at that point is calculated and used as a consistency measure for that phase point.

[0262] Aggregate dispersion is a statistical measure used to quantify a set of multidimensional data points (here) Each heartbeat at the corresponding phase point The degree of concentration of these eigenvectors. Common methods include calculating the concentration of these eigenvectors. The trace (total variance) of the covariance matrix of each vector, the standard deviation of the mean Euclidean distance to the centroid or principal component, etc. The smaller the aggregation dispersion, the more similar and concentrated the texture features of all heartbeats at that phase point, that is, the higher the cross-cycle consistency; conversely, it indicates large differences and low consistency.

[0263] G5. Statistical summary of consistency metrics;

[0264] For all The consistency metrics of each characteristic phase point are statistically summarized to obtain the final dynamic regularity quantification index.

[0265] Statistical summary is The consistency metric of each feature phase point (a length of...) The sequence is condensed. Common methods include: calculating its arithmetic mean (reflecting the overall average regularity), taking the minimum value (reflecting the most irregular and disordered period), calculating the proportion of consistent phase points below a certain threshold (reflecting the range of regularity loss), or calculating the area under the curve of the sequence, etc.

[0266] Mean: Provides a general assessment of overall stability.

[0267] Minimum: Sensitive to worst performance and may be more associated with specific events (such as the risk of arrhythmia).

[0268] Low consistency percentage: quantifies the duration of abnormal states over time.

[0269] The resulting dynamic regularity quantitative index is a composite biomarker that integrates information from multiple heartbeat cycles and all physiological phases throughout the entire observation period. It directly characterizes the programmed and repeatable precision with which myocardial tissue performs its periodic mechanical movements, and is a high-level, integrated digital twin of cardiac mechanical function homeostasis and coordination.

[0270] like Figure 3 As shown, this application also discloses a method for assessing the readmission risk of heart failure based on cardiac imaging feature analysis. The method follows the logic of "synchronous data acquisition → physiological interference removal → microscopic feature extraction → rhythm regularity quantification → multi-source index correction and fusion → risk decision-making". The data format gradually transforms from raw images, videos and respiratory signals into texture feature sequences and periodic consistency indicators, and is finally condensed into a robust comprehensive quantitative indicator for risk classification.

[0271] The specific steps are as follows:

[0272] Synchronous data acquisition: Echocardiographic sequences and real-time respiratory status data were acquired simultaneously to lay the foundation for establishing image-physiological time correlation in the future.

[0273] Respiratory phase segmentation: Based on respiratory state data, ultrasound sequences are segmented into multiple image segments with consistent respiratory phases. This isolates the cardiac load and morphological periodic changes caused by respiratory motion, creating conditions for analyzing the steady-state performance of the heart under its own rhythm.

[0274] Image segment processing (microscopic feature extraction and regularity analysis):

[0275] Image patch extraction and speckle texture feature calculation: Focusing on key areas of the subendocardial myocardium, speckle texture features reflecting the microstructure of the tissue are calculated, and image information is transformed into numerical tissue attribute descriptors.

[0276] Texture feature serialization and temporal analysis: The speckle texture features are arranged into a sequence according to time, and the repeatability and stability of myocardial micro-motion in continuous heartbeats are quantified through operations such as cardiac cycle alignment and phase point consistency measurement (calculation of aggregation dispersion), generating dynamic regularity quantitative indicators.

[0277] Comprehensive quantization processing (correction and fusion):

[0278] Physiological state normalization: Based on the respiratory phase type associated with image segments, the dynamic regularity quantitative indicators are corrected to eliminate the assessment bias caused by different respiratory phases (different cardiac load states) and make all segment indicators comparable.

[0279] Quality-weighted fusion: Based on the confidence level of image segments (such as signal-to-noise ratio and sharpness), the metrics of segments of different quality are weighted differently to suppress the noise influence of low-quality data and generate a robust comprehensive regularity quantitative metric. These two steps together ensure the accuracy, fairness, and stability of the final metric.

[0280] Risk assessment output: By comparing comprehensive regular quantitative indicators with preset thresholds trained based on clinical data, discrete risk levels or continuous risk scores are output, completing the transformation from complex data to clinical decision information.

[0281] Embodiments of this application may also be computer-readable storage media storing computer program instructions that, when executed by a processor, cause the processor to perform the steps described in the above-described section of this specification on the heart failure readmission risk assessment system based on cardiac imaging feature analysis, according to the various embodiments of this application.

[0282] Computer-readable storage media may take the form of any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may, for example, include, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0283] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A heart failure readmission risk assessment system based on cardiac imaging feature analysis, characterized in that: include: The multi-source data acquisition module is used to simultaneously acquire standard apical section echocardiographic sequence data and real-time respiratory status data of the target patient; The respiratory phase segmentation module is used to segment the echocardiogram sequence data according to the respiratory phase based on the real-time respiratory state data, so as to obtain several sets of image segment data with consistent respiratory phases. The image segment processing module performs the following processing on each image segment data: Extract image patch data of the left ventricular endocardial surface region from each frame of the image fragment data, and calculate speckle texture features based on the image patch data; Arrange the speckle texture features of each frame of the image in chronological order to form a texture feature sequence of the image fragment data; The texture feature sequence is segmented and aligned according to the cardiac cycle, and the consistency measure of texture features at the same phase point in multiple cardiac cycles is calculated to obtain a dynamic regularity quantification index. The comprehensive quantification module is used to normalize the dynamic regularity quantification index based on the respiratory phase type associated with the image segment data, and to perform weighted fusion of the normalized dynamic regularity quantification index with the image quality confidence associated with the image segment data to generate a comprehensive regularity quantification index. The risk assessment output module is used to compare the comprehensive regularity quantitative indicators with the preset risk assessment threshold and output the risk assessment result of heart failure readmission.

2. The heart failure readmission risk assessment system based on cardiac imaging feature analysis according to claim 1, characterized in that: The process of normalizing the dynamic regularity quantitative indicators for physiological states includes: Based on the respiratory phase type of the image segment data, it is divided into expiratory-dominant phase, inspiratory-dominant phase, or respiratory transition phase. The dynamic regularity quantification index of image segment data is input into a pre-trained phase-load regression model; the phase-load regression model takes respiratory phase type and proportion as input and outputs an estimated value of cardiac load status. Based on the difference between the estimated cardiac load state and the standard load state reference value, the dynamic regularity quantitative index is calculated in reverse to generate the index after physiological state normalization.

3. The heart failure readmission risk assessment system based on cardiac imaging feature analysis according to claim 2, characterized in that: The specific process for calculating the reverse compensation of the dynamic regularity quantitative index is as follows: The difference between the estimated cardiac load status and the standard load status reference value is input into the piecewise compensation function; The piecewise compensation function is configured as follows: When the absolute value of the difference is within the first difference interval, a first compensation coefficient that has a first linear relationship with the difference is output. When the absolute value of the difference is within the second difference interval, a second compensation coefficient that has a second non-linear relationship with the difference is output; wherein the rate of change of the second compensation coefficient decreases as the absolute value of the difference increases. When the absolute value of the difference is within the third difference interval, the output is limited to the third compensation coefficient within the preset saturation compensation value; The absolute values ​​of the differences corresponding to the first difference interval, the second difference interval, and the third difference interval increase sequentially. Based on the first compensation coefficient, the second compensation coefficient, or the third compensation coefficient, the dynamic regularity quantitative index is corrected to generate a dynamic regularity quantitative index after physiological state normalization.

4. The heart failure readmission risk assessment system based on cardiac imaging feature analysis according to claim 1, characterized in that: The weighted fusion process of the dynamic regularity quantitative indicators includes: The image quality confidence data associated with the image segment data is decomposed into signal-to-noise ratio feature value, boundary sharpness feature value, and artifact perturbation feature value; The signal-to-noise ratio feature value, boundary sharpness feature value, and artifact perturbation feature value are input into the dynamic weight allocation function; The dynamic weight allocation function is configured as follows: For any image segment data, based on the comparison result of its signal-to-noise ratio characteristic value and a preset threshold, it is determined whether to enable the image segment data to participate in fusion; For the enabled image segment data, the initial weight base is calculated using a weighted combination formula based on its boundary sharpness feature value and artifact perturbation feature value; The initial weight base is input into a nonlinear decay function, and the output value of the nonlinear decay function is the fusion weight value of the image segment data. The nonlinear decay function is configured such that when the initial weight base is greater than a preset weight threshold, the output value approaches the initial weight base; when the initial weight base is less than or equal to the preset weight threshold, the output value decays faster as the initial weight base decreases. Using the fusion weight values ​​calculated for image segment data, a weighted average is applied to the dynamic regularity quantitative index after physiological state normalization to generate the comprehensive regularity quantitative index.

5. The heart failure readmission risk assessment system based on cardiac imaging feature analysis according to claim 4, characterized in that: The calculation process for the initial weight base is as follows: The boundary sharpness feature value and the artifact perturbation feature value are normalized respectively to obtain the normalized boundary sharpness value and the normalized artifact perturbation value. Calculate the geometric mean of the normalized boundary sharpness value and the normalized artifact perturbation value, and use it as the basic weighting factor; The absolute value of the difference between the normalized boundary sharpness value and the normalized artifact perturbation value is calculated to obtain the quality imbalance. Based on the aforementioned quality imbalance, the balance adjustment factor is calculated using a monotonically decreasing balance adjustment function. The balance adjustment function is configured such that: when the quality imbalance is zero, the balance adjustment factor is at its maximum value; and when the quality imbalance increases, the balance adjustment factor decreases non-linearly. The initial weight base is obtained by multiplying the basic weight factor by the equilibrium adjustment factor.

6. The heart failure readmission risk assessment system based on cardiac imaging feature analysis according to claim 1, characterized in that: The process of acquiring the image segment data is as follows: The real-time respiratory status data is filtered and features are extracted, and the phase inflection points in each respiratory cycle are identified according to preset recognition rules. Based on the phase inflection point, the respiratory cycle is divided into multiple consecutive respiratory phase windows, and a phase type is defined for each respiratory phase window; Each frame of the echocardiogram sequence data is mapped to the corresponding respiratory phase window according to its acquisition timestamp; For image frame sequences that belong to the same phase window and are temporally continuous, they are merged into one image segment data, and the phase type of the phase window is used as the breathing phase type associated with the image segment data.

7. The heart failure readmission risk assessment system based on cardiac imaging feature analysis according to claim 1, characterized in that: The process of obtaining the speckle texture features is as follows: Multi-scale spatial filtering is performed on the image patch data to extract the filter response maps at at least two different spatial scales. For each spatial scale of the filtered response map, the statistical distribution characteristics of its local region are calculated. The statistical distribution characteristics include at least the variance characteristics and entropy characteristics of the local gray-level distribution. The variance and entropy features obtained from the same image patch data at different spatial scales are concatenated in scale order to form a multidimensional feature vector, which serves as the speckle texture feature of the image patch data.

8. The heart failure readmission risk assessment system based on cardiac imaging feature analysis according to claim 1, characterized in that: The process of obtaining the dynamic regularity quantitative indicators is as follows: Identify cycle markers for the start and end points of each cardiac cycle in the texture feature sequence; Based on the periodic markers, the texture feature sequence is divided into multiple feature subsequences corresponding to consecutive cardiac cycles; The segmented feature subsequences are aligned on the time axis based on their periodic markers, and the aligned feature subsequences are resampled to ensure that each cardiac cycle has the same number of feature phase points. At each of the aforementioned feature phase points, the aggregation dispersion of the texture feature vectors of all aligned cardiac cycles at that point is calculated and used as a consistency measure for that phase point. The consistency metric values ​​of all characteristic phase points are statistically summarized to obtain the dynamic regularity quantitative index.

9. A method for assessing the risk of readmission for heart failure based on cardiac imaging feature analysis, characterized in that: Includes the following steps: Simultaneously acquire standard apical section echocardiographic sequence data and real-time respiratory status data of the target patient; Based on the real-time respiratory status data, the echocardiogram sequence data is segmented according to the respiratory phase to obtain several sets of image segment data with consistent respiratory phases. For each image segment data, perform the following processing: Extract image patch data of the left ventricular endocardial surface region from each frame of the image fragment data, and calculate speckle texture features based on the image patch data; Arrange the speckle texture features of each frame of the image in chronological order to form a texture feature sequence of the image fragment data; The texture feature sequence is segmented and aligned according to the cardiac cycle, and the consistency measure of texture features at the same phase point in multiple cardiac cycles is calculated to obtain a dynamic regularity quantification index. Based on the respiratory phase type associated with the image segment data, the dynamic regularity quantitative index is normalized by physiological state, and the normalized dynamic regularity quantitative index is weighted and fused with the image quality confidence associated with the image segment data to generate a comprehensive regularity quantitative index. The comprehensive regularity quantitative indicators are compared with the preset risk assessment thresholds to output the risk assessment results of heart failure readmission.

10. A computer storage medium storing computer-executable instructions, characterized in that, When the computer-executable instructions are executed, they implement the heart failure readmission risk assessment system based on cardiac imaging feature analysis as described in any one of claims 1-8.

Images