Multi-modal data-driven old heart failure patient re-admission risk prediction method

Abstract

The invention discloses a multi-modal data-driven old heart failure patient re-admission risk prediction method, which comprises the following steps: collecting clinical diagnosis and treatment, physiological monitoring, image examination and lifestyle related multi-source data to construct a data set, and screening and calibrating risk factors through a heart failure risk factor weight learning algorithm; extracting dynamic vulnerability characteristics of the time sequence data by using an old-age heart failure time sequence vulnerability evaluator, and analyzing a cross-organ risk transmission mechanism by using an organ level risk transmission analysis model; inputting the two types of analysis results into an intelligent heart failure risk hierarchical management platform, determining evaluation dimensions, mapping feature data, distributing index weights, and constructing a multi-dimensional risk evaluation matrix; and finally, based on the matrix, risk values are calculated through a re-admission risk prediction model and are graded. According to the method, accurate risk prediction is realized through multi-modal data fusion and multi-technology cooperation, scientific support is provided for clinical diagnosis and treatment management, and the method is suitable for a re-admission risk assessment scene of the senile heart failure patient.

Description

Technical Field

[0001] This invention relates to the field of risk prediction technology for heart failure patients, and in particular to a multimodal data-driven method for predicting readmission risk in elderly heart failure patients. Background Technology

[0002] Heart failure, a prevalent chronic cardiovascular disease among the elderly, is characterized by its insidious progression, strong correlation with physiological functions, and significant individual variability. Readmissions not only exacerbate the physical and mental burden on patients but also significantly increase the consumption of medical resources. With the aging population, the number of elderly heart failure patients continues to expand, making accurate prediction of readmission risk a crucial requirement for optimizing medical resource allocation and improving the efficiency of diagnosis and treatment management. Current clinical practice often relies on single-dimensional data or traditional assessment tools for risk assessment, making it difficult to integrate multi-source heterogeneous data from clinical diagnosis and treatment, physiological monitoring, imaging examinations, and lifestyle factors. This fails to comprehensively capture the dynamic changes in patients' conditions and their correlation with organ function. Therefore, there is an urgent need to construct a multimodal data-driven integrated risk prediction system to achieve accurate assessment and stratified management of readmission risk for elderly heart failure patients.

[0003] Existing technologies have significant limitations in predicting readmission risk in elderly patients with heart failure. On the one hand, existing methods often lack deep integration of multimodal data and precise screening of labeled risk factors, frequently using a single type of data in isolation for prediction. This ignores the synergistic effects of multiple dimensions such as clinical indicators, physiological state, imaging features, and lifestyle habits, resulting in insufficient comprehensiveness and accuracy in risk assessment. On the other hand, existing technologies fail to fully consider the temporal changes in physiological function and the risk transmission mechanisms between organs in elderly patients. They cannot dynamically capture the evolution of patient vulnerability across different time dimensions, nor can they analyze the risk transmission pathways between cardiac and related organ dysfunction. This makes it difficult for prediction models to reflect the dynamics and complexity of disease progression and to meet the actual clinical need for accurate prediction of readmission risk. Summary of the Invention

[0004] To overcome the shortcomings and deficiencies of existing technologies, this invention provides a multimodal data-driven method for predicting readmission risk in elderly patients with heart failure.

[0005] The technical solution adopted in this invention is a multimodal data-driven method for predicting readmission risk in elderly patients with heart failure, comprising the following steps: S1, collecting clinical diagnosis and treatment data, physiological indicator monitoring data, imaging examination data, and lifestyle-related data of elderly patients with heart failure to construct a multimodal raw dataset; S2, performing correlation analysis and importance ranking on various risk factors in the multimodal raw dataset using a heart failure risk factor weight learning algorithm to select a set of calibrated factors strongly associated with readmission risk; S3, using an elderly heart failure temporal vulnerability assessor to dynamically extract features from the temporal data in the calibrated factor set. S4. Capture the vulnerability characteristics of patients' physiological state and disease changes at different time dimensions; S5. Use an organ-level risk propagation analysis model to model the cross-organ risk transmission path of the functional data of the heart and related organs involved in the calibration factor set, and analyze the risk diffusion mechanism of organ dysfunction; S6. Input the temporal vulnerability characteristics and organ-level risk propagation analysis results into the intelligent heart failure risk stratification management platform, and construct a multi-dimensional risk assessment matrix in combination with the preset risk assessment dimensions; S7. Calculate the readmission risk value of patients based on the risk assessment matrix through the readmission risk prediction model for elderly heart failure patients, and complete the risk level classification.

[0006] Furthermore, the expression for the heart failure risk factor weight learning algorithm is: ,in, This represents the weight of the j-th risk factor in the i-th data category. This is the correlation adjustment coefficient. Let be the observed value of the j-th risk factor in the i-th data category, and y be the readmission risk label variable. Let be the influence coefficient of the k-th correlation factor. Let be the correlation coefficient between the j-th risk factor and the k-th risk factor. The total number of data categories, The number of risk factors in each data category. The entropy penalty coefficient is... Let the information entropy of the j-th risk factor be , This represents the observed value of the nth risk factor in the m-th data category. The covariance between the re-admission risk label variable y and the re-admission risk label variable y.

[0007] Furthermore, the expression for the temporal vulnerability assessor for heart failure in the elderly is: ,in, Let t be the patient's temporal vulnerability value. This is the time-series integral adjustment coefficient. Let be the physiological state characteristic function at time s. for The stability coefficient of the disease at any given time. for Factors influencing medical intervention at any given time The weighting coefficients are the second derivatives. is the vulnerability amplification factor at time s.

[0008] Furthermore, the expression for the organ-level risk propagation analysis model is: ,in, Let be the risk value of the o-th organ at time t. Let be the transmission efficiency from the p-th organ to the q-th risk transmission pathway. for The risk value of the p-th organ at time p. Let be the response coefficient of the o-th organ at time t to the q-th conduction pathway. This is the path gradient adjustment factor. Let t be the risk propagation function from the p-th organ to the o-th organ. Regarding conduction parameters gradient, Let be the response coefficient of the r-th organ at time t to the q-th conduction pathway.

[0009] Furthermore, the risk assessment matrix construction expression of the intelligent heart failure risk stratification management platform is as follows: ,in, A risk assessment matrix with L rows and K columns. The weight coefficient of the k-th indicator in the i-th evaluation dimension. This represents the temporal vulnerability feature value corresponding to the k-th indicator of the i-th evaluation dimension. Let L be the organ-level risk value corresponding to the k-th indicator of the i-th assessment dimension, and L be the total number of assessment dimensions. The number of indicators under each evaluation dimension.

[0010] Furthermore, the expression for the readmission risk prediction model for elderly heart failure patients is as follows: ,in, This is the risk value for readmission. It is the Sigmoid activation function. , This is the model weight matrix. , The bias term is ReLU, which is a linear rectified function. The risk amplification factor at time t, Let T be the total risk at the organ level at time t, where T is the time span of the time series data.

[0011] Further, S3 includes the following sub-steps: S31, the time-series data in the calibration factor set are arranged in an ordered manner according to timestamps, and multiple consecutive time windows are divided. Each time window corresponds to a fixed time interval, and the statistical characteristics and trend characteristics of risk factors within each time window are extracted; S32, the trend change rate is calculated based on the characteristic differences between adjacent time windows, and a vulnerability threshold range is set in combination with the progression pattern of heart failure, and abnormal time windows exceeding the threshold range are initially screened out; S33, the time-series data within the abnormal time windows are analyzed in a fine-grained manner to capture the mutation points and gradual trends of physiological indicators, and a two-dimensional vulnerability feature matrix with time dimension and indicator dimension is constructed; S34, the missing data is completed by time-series interpolation, and the feature matrix is ​​dimensionally normalized to form a standardized time-series vulnerability feature vector for subsequent risk assessment.

[0012] Further, S4 includes the following sub-steps: S41, classifying the organ function data involved in the calibration factor set, clarifying the functional indicator system of the heart and related organs such as the lungs, kidneys, and liver, and establishing organ-indicator mapping relationships; S42, based on the correlation between anatomical structure and physiological function, constructing a directed graph model of risk transmission between organs, defining nodes as organs, edges as risk transmission paths, and assigning an initial transmission probability to each edge; S43, calibrating the transmission probability through organ dysfunction data in multimodal data, analyzing the changes in the intensity and direction of risk transmission between organs under different pathological states, and updating the edge weights of the directed graph model; S44, based on the updated directed graph model, using a path search algorithm to traverse all possible risk propagation paths, calculating the risk propagation intensity of each path, and analyzing the risk diffusion mechanism of organ dysfunction.

[0013] Further, S5 includes the following sub-steps: S51, clarifying the core dimensions of risk assessment, including the degree of temporal vulnerability, the degree of organ function impairment, the scope of risk propagation, the treatment response effect, the basic health status and lifestyle adaptability, and setting corresponding assessment indicators for each dimension; S52, classifying and mapping the temporal vulnerability characteristics and organ-level risk propagation analysis results according to the assessment dimensions, and converting different types of feature data into assessment values ​​of a unified dimension; S53, determining the weight allocation of each dimension and indicator based on the differences in the importance of each assessment dimension through a heart failure risk factor weight learning algorithm, ensuring that the weights match the correlation with readmission risk; S54, constructing a multi-dimensional risk assessment matrix based on the weight allocation and assessment values, where each element in the matrix corresponds to the comprehensive risk quantification result of the set dimension and set indicators.

[0014] A multimodal data-driven method for predicting readmission risk in elderly patients with heart failure is proposed. This method is implemented through several units, including: a multimodal data integration and acquisition unit, which receives multi-source data from clinical diagnosis and treatment, physiological monitoring, imaging examinations, and lifestyle-related data, and forms a structured multimodal dataset through data format conversion and association mapping to provide data support for subsequent analysis; a risk factor weight dynamic calculation unit, connected to the multimodal data integration and acquisition unit, which performs factor correlation analysis and importance ranking on the dataset using a heart failure risk factor weight learning algorithm, and outputs a set of calibrated risk factors; and a temporal vulnerability feature extraction unit, connected to the risk factor weight dynamic calculation unit, which extracts temporal data from the calibrated factors using an elderly heart failure temporal vulnerability assessor. The system performs dynamic feature capture and vulnerability quantification to generate a temporal vulnerability feature vector; the organ-level risk propagation analysis unit, connected to the risk factor weight dynamic calculation unit, constructs cross-organ risk transmission paths through the organ-level risk propagation analysis model and outputs organ-level risk propagation results; the multi-dimensional risk matrix construction unit, connected to the temporal vulnerability feature extraction unit and the organ-level risk propagation analysis unit respectively, integrates the two types of feature results and combines weight allocation to construct a risk assessment matrix; the readmission risk prediction and stratification unit, connected to the multi-dimensional risk matrix construction unit, calculates risk values ​​and completes level classification through the readmission risk prediction model for elderly heart failure patients, and transmits the results to the intelligent heart failure risk stratification management platform for storage and display.

[0015] Beneficial Effects: This invention proposes a multimodal data-driven method for predicting readmission risk in elderly patients with heart failure. It comprehensively collects multi-source data, including clinical diagnosis and treatment, physiological monitoring, imaging examinations, and lifestyle data, avoiding the limitations of single data dimensions. A specialized risk factor weighting learning mechanism is then used to accurately screen key influencing factors, solving the problems of insufficient multimodal data fusion and inaccurate factor identification in existing technologies, ensuring the comprehensiveness and relevance of risk assessment. A temporal vulnerability assessor dynamically captures the temporal changes in the patient's physiological state and condition. Simultaneously, an organ-level risk propagation analysis model analyzes the risk transmission mechanism between the heart and related organs, overcoming the shortcomings of existing technologies in reflecting the dynamic evolution of the condition and cross-organ risk diffusion, making risk assessment more closely aligned with the complexity of the condition in elderly patients with heart failure. By constructing a multi-dimensional risk assessment matrix through an intelligent heart failure risk stratification management platform, and combining it with a specialized readmission risk prediction model to complete risk value calculation and level classification, a complete technical chain is formed from data collection, feature extraction, mechanism analysis to risk prediction. This enables accurate prediction and stratified management of readmission risk, providing reliable support for optimizing clinical treatment plans and rationally allocating medical resources, and effectively improving the quality of medical management for elderly heart failure patients. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method steps of the present invention; Figure 2 This is a diagram showing the unit composition for implementing the method of the present invention. Detailed Implementation

[0017] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] like Figure 1 As shown, a multimodal data-driven method for predicting readmission risk in elderly patients with heart failure includes the following steps: S1: Collect clinical diagnosis and treatment data, physiological indicator monitoring data, imaging examination data and lifestyle-related data of elderly patients with heart failure to construct a multimodal raw dataset; Specifically, step S1 involves systematically collecting multi-dimensional data from elderly patients with heart failure and constructing a standardized raw dataset. The dataset must cover patients aged 60 and above diagnosed with heart failure, with a sample size of at least 500 cases to ensure data representativeness. Clinical data includes the patient's past medical history, disease duration, medication records, surgical history, outpatient follow-up records, and inpatient medical documents. Medication records must clearly specify the type, dosage, frequency of use, and adjustment time points for each medication. Disease duration must be accurate to the day. Physiological monitoring data includes heart rate, blood pressure, blood oxygen saturation, cardiac function classification, electrolyte levels, liver and kidney function indicators, and myocardial enzyme profiles. Monitoring frequency is set at once every 4 hours during hospitalization and twice a week after discharge, with continuous collection for at least 3 months to ensure the continuity of time-series data. Imaging data includes echocardiography, chest CT, and electrocardiogram. Echocardiography requires extraction of core parameters such as left ventricular ejection fraction and left ventricular end-diastolic diameter. Imaging data must retain its original format and diagnostic report. Lifestyle-related data were collected through structured questionnaires, including smoking history, alcohol consumption history, exercise frequency, dietary habits, sleep duration, and psychological status scores. A questionnaire return rate of over 95% was required. After data collection, the data was processed using a standardized format, establishing a unique patient identifier and mapping it to various data types. Invalid and duplicate data were removed, ultimately forming a multimodal raw dataset encompassing four major categories: clinical, physiological, imaging, and lifestyle data. This dataset provides comprehensive and high-quality data support for subsequent risk factor analysis.

[0019] S2, through the heart failure risk factor weight learning algorithm, performs correlation analysis and importance ranking on various risk factors in the multimodal raw dataset, and selects a set of calibrated factors that are strongly associated with readmission risk. Specifically, step S2 uses a heart failure risk factor weight learning algorithm to perform in-depth analysis of all potential risk factors in the multimodal raw dataset. First, the total number of risk factors in the dataset is determined, including at least 30 clinical diagnosis and treatment risk factors, at least 25 physiological indicators, at least 15 imaging features, and at least 10 lifestyle-related risk factors. The first step of the algorithm implementation is risk factor correlation analysis, which calculates the correlation strength between each factor and the readmission risk outcome. The correlation strength calculation requires selecting appropriate statistical methods based on the factor data type to ensure the accuracy of the analysis results. The second step is importance ranking. Based on the correlation strength analysis results, a multi-dimensional scoring mechanism is used to rank all risk factors. The scoring dimensions include correlation significance, data reliability, and clinical relevance. The weight allocation for each dimension is determined through expert review, with correlation significance accounting for at least 40% of the weight. During the factor selection process, a selection threshold was set, and the top 30% of risk factors in terms of importance scores were included in the factor set. Simultaneously, it was ensured that the factor set included core indicators from all four data categories to avoid an excessively high proportion of data from any single category. Specifically, the proportion of clinical diagnosis and treatment factors and physiological indicators should each be no less than 25%, and the proportion of imaging features and lifestyle factors should each be no less than 15%. After selection, the validity of the factor set was validated using a validation sample size of 20% of the total sample. The screening effect was assessed by calculating the predictive contribution of the factor set; the contribution should reach at least 70%. If this requirement was not met, the selection threshold was adjusted, and the selection was repeated. The final result was a set of factor sets that were strongly correlated with readmission risk, had reliable data quality, and comprehensive coverage, laying the foundation for subsequent feature extraction and model construction.

[0020] S3 utilizes an elderly heart failure temporal vulnerability assessor to dynamically extract features from temporal data in a calibrated factor set, capturing the vulnerability characteristics of patients' physiological state and disease changes at different time dimensions. Specifically, step S3 utilizes an elderly heart failure temporal vulnerability assessor to extract dynamic features from the time-series data in the calibrated factor set. The process begins by screening the time-series data in the calibrated factor set, identifying 15-20 core time-series factors such as heart rate, blood pressure, blood oxygen saturation, and cardiac function indicators. Then, time windows are divided using a sliding window method. The window size is set to 7 days, and the sliding step is 1 day. The continuously collected time-series data is divided into multiple consecutive and partially overlapping time windows, ensuring that each window includes sufficient monitoring data points, with no fewer than 12 data points per window. For each time window, statistical and dynamic features are extracted. Statistical features include mean, standard deviation, median, and extreme values, while dynamic features include rate of change, trend, and fluctuation amplitude. The rate of change requires calculating the absolute and relative changes between adjacent data points. After feature extraction, the evolution of features within different time windows is analyzed, taking into account the pathophysiological characteristics of heart failure in the elderly. This captures short-term fluctuations and long-term trends in the patient's physiological state, identifying vulnerability signals before disease deterioration, such as increased heart rate fluctuations and decreased blood pressure stability. Simultaneously, the extracted vulnerability features are standardized, unifying feature dimensions and constructing a two-dimensional vulnerability feature matrix including time and feature dimensions. The number of rows corresponds to the number of time windows, and the number of columns corresponds to the number of extracted feature types, with a total of no less than 50 feature types. This process accurately captures vulnerability features of changes in the patient's physiological state and disease condition across different time dimensions, providing dynamic and fine-grained feature support for subsequent risk assessment and improving the timeliness and sensitivity of risk prediction. S4 uses an organ-level risk propagation analysis model to model the cross-organ risk transmission path of functional data of the heart and related organs involved in the calibration factor set, and analyzes the risk diffusion mechanism of organ dysfunction. Specifically, step S4 employs an organ-level risk propagation analysis model to systematically model the functional data of the heart and related organs involved in the calibration factor set. First, the scope of core organs is defined, including five key organs: heart, lungs, kidneys, liver, and brain. Each organ corresponds to no fewer than eight data indicators, such as left ventricular ejection fraction for the heart, partial pressure of oxygen in the lungs, and glomerular filtration rate for the kidneys. The first step in model construction is the initialization of cross-organ risk transmission pathways. Based on the correlation between anatomical structure and physiological function, direct and indirect correlation pathways between organs are identified, clarifying the transmission relationship between the heart as a core organ and other organs. No fewer than 20 core transmission pathways are constructed, with each pathway defining its starting organ, ending organ, and intermediate transmission links. The second step is pathway modeling, using a directed graph model to formally describe the transmission pathways. Organs are used as nodes, and transmission pathways are directed edges. Each edge is assigned an initial transmission probability, set based on physiological mechanism research results. The initial transmission probability from the heart to the lungs and kidneys is no less than 0.6. The third step is model calibration. Using organ dysfunction data from multimodal datasets, iterative optimization algorithms are employed to adjust the conduction probability and intensity parameters of the pathways, with at least 100 iterations to ensure consistency between the model output and clinical realities. The fourth step is risk diffusion mechanism analysis. By simulating the risk propagation process under different organ dysfunction scenarios, the risk propagation efficiency and impact range of each pathway are calculated, identifying key conduction nodes and high-risk propagation pathways. For example, the risk propagation efficiency of the pathway from cardiac dysfunction to the lungs needs to be quantified. Through this process, the risk diffusion mechanism of organ dysfunction is clearly analyzed, providing in-depth organ-level analysis for a comprehensive assessment of patient readmission risk.

[0021] S5 inputs the temporal vulnerability characteristics and organ-level risk propagation analysis results into the intelligent heart failure risk stratification management platform, and constructs a multi-dimensional risk assessment matrix by combining the preset risk assessment dimensions; Specifically, step S5 integrates the temporal vulnerability characteristics and organ-level risk propagation analysis results into the intelligent heart failure risk stratification management platform to construct a multi-dimensional risk assessment matrix. First, the platform needs to pre-configure risk assessment dimensions. Combining the core influencing factors of readmission risk in elderly heart failure patients, six primary assessment dimensions are clearly defined: degree of temporal vulnerability, degree of organ function impairment, scope of risk propagation, treatment response effect, basic health status, and lifestyle suitability. Each primary dimension has 5-8 secondary assessment indicators, such as cardiac function impairment score and pulmonary function impairment score under the organ function impairment dimension. Second, feature result mapping is performed. The temporal vulnerability characteristics extracted in step S3 and the organ-level risk propagation results obtained in step S4 are classified and matched according to the definition of the assessment indicators. Different types and scales of feature data are converted into assessment values ​​within a unified range, with the assessment value range set from 0 to 10 points, where a higher score indicates a higher risk. Subsequently, the weight allocation was determined. The weights of each primary dimension and secondary indicator were calculated using a heart failure risk factor weight learning algorithm. Within the primary dimensions, the weights of temporal vulnerability and organ function impairment were each no less than 20%, while other dimensions had a weight allocation of no less than 10%. The weights of secondary indicators were determined based on their importance within the dimension, ensuring that the weight allocation conformed to the logic of clinical risk assessment. Finally, a multi-dimensional risk assessment matrix was constructed. The rows of the matrix corresponded to the primary assessment dimensions, and the columns to the secondary assessment indicators. Matrix elements were the product of the corresponding weight and assessment value, forming a standardized risk assessment matrix of approximately 6 rows and 35 columns. The matrix data underwent normalization to ensure consistent value ranges for matrix elements, providing structured and quantifiable input data for subsequent risk value calculations.

[0022] S6. Based on the risk assessment matrix, the readmission risk value of elderly heart failure patients is calculated using a readmission risk prediction model to complete the risk level classification.

[0023] Specifically, step S6, based on a multi-dimensional risk assessment matrix, calculates and classifies risk values ​​using a readmission risk prediction model for elderly heart failure patients. The model is deployed on an intelligent heart failure risk stratification management platform, with a response time of no more than 3 seconds per case. During risk value calculation, the elements of the multi-dimensional risk assessment matrix are first used as model input. The model performs feature fusion and nonlinear transformation on the input data through multi-layered calculations. The fusion process must fully utilize the correlation information of indicators across various dimensions, highlighting the impact of key risk factors. The risk value range is set from 0 to 1; the closer the value is to 1, the higher the readmission risk. After calculation, the effectiveness of the risk value needs to be validated using indicators such as confusion matrix and ROC curve, ensuring that the model's prediction accuracy is no less than 85% and the AUC value is no less than 0.8. The risk grading uses the quartile method, dividing risk values ​​into four levels: low risk (0-0.25), low-to-medium risk (0.25-0.5), medium-to-high risk (0.5-0.75), and high risk (0.75-1.0). Each risk level requires a clearly defined framework of clinical intervention recommendations. After grading, the risk values, risk levels, and corresponding indicator contribution analysis results are stored in the platform database. Simultaneously, a standardized risk assessment report is generated. This report must include core content such as patient basic information, risk scores for each dimension, risk level, and high-risk factor alerts, providing clinicians with clear and intuitive decision-making support. This enables precise quantification and scientific stratification of readmission risk in elderly heart failure patients, supporting the development of personalized treatment plans and the optimal allocation of medical resources.

[0024] Preferably, the expression for the heart failure risk factor weight learning algorithm is: ,in, This represents the weight of the j-th risk factor in the i-th data category. This is the correlation adjustment coefficient. Let be the observed value of the j-th risk factor in the i-th data category, and y be the readmission risk label variable. Let be the influence coefficient of the k-th correlation factor. Let be the correlation coefficient between the j-th risk factor and the k-th risk factor. The total number of data categories, The number of risk factors in each data category. The entropy penalty coefficient is... Let the information entropy of the j-th risk factor be , This represents the observed value of the nth risk factor in the m-th data category. The covariance between the re-admission risk label variable y and the re-admission risk label variable y.

[0025] Specifically, the heart failure risk factor weight learning algorithm is used to quantify the impact of various risk factors in multimodal data on the readmission risk of elderly heart failure patients. The implementation process first clarifies the parameter value range to ensure the accuracy and rationality of the calculation. The correlation adjustment coefficient is set to a range of 0.7-0.9, and the entropy penalty coefficient is set to a range of 0.1-0.3. This range has been validated through extensive clinical data and can balance the impact of factor correlation and data dispersion on weight calculation. During algorithm implementation, the observed risk factor values ​​for each type of data in the multimodal dataset are first obtained, clarifying the total number of data categories and the number of risk factors in each category. The data categories correspond to four main categories: clinical diagnosis and treatment, physiological monitoring, imaging examinations, and lifestyle. The number of risk factors in each category is determined based on the actual collected data and is not less than a preset threshold. Then, the covariance between each risk factor and the readmission risk label variable is calculated to reflect the strength of their linear association. Simultaneously, the correlation coefficient between the risk factor and other related factors is calculated. The interaction between related factors is adjusted using the correlation factor influence coefficient, which ranges from 0.05 to 0.2. Next, the information entropy of each risk factor is calculated to measure its data dispersion. An entropy penalty coefficient is used to suppress the weights of factors with excessive dispersion and insufficient reliability. Finally, the weight of each risk factor is obtained through a comprehensive calculation of the numerator and denominator. The numerator integrates positive influencing factors such as covariance and interactions between related factors, while the denominator includes the sum of the covariances of all factors and the entropy penalty term. This ensures that the weight calculation highlights strongly correlated factors while avoiding the excessive dominance of a single factor. The final output weights can be directly used to calibrate risk factor screening, providing accurate basis for factor importance in subsequent feature extraction.

[0026] Preferably, the expression for the temporal vulnerability assessor for heart failure in the elderly is: ,in, Let t be the patient's temporal vulnerability value. This is the time-series integral adjustment coefficient. Let be the physiological state characteristic function at time s. for The stability coefficient of the disease at any given time. for Factors influencing medical intervention at any given time The weighting coefficients are the second derivatives. is the vulnerability amplification factor at time s.

[0027] Specifically, the temporal vulnerability assessor for elderly heart failure is used to dynamically capture the vulnerability of patients' physiological states at different time points. During implementation, the temporal integral adjustment coefficient ranges from 0.6 to 0.8, and the second derivative weighting coefficient ranges from 0.4 to 0.6. This parameter setting effectively balances the contributions of temporal integral characteristics and abrupt change characteristics to vulnerability assessment. When the assessor runs, the time span is first determined, covering the monitoring period from 3 months before admission to 6 months after discharge. The physiological state characteristic function value is obtained at each time point. This function is constructed by linearly combining key temporal factors such as heart rate, blood pressure, and cardiac function indicators. Then, the first derivative of the physiological state characteristic function is calculated to reflect the instantaneous rate of change in the physiological state, while the second derivative is calculated to capture fluctuations in the rate of change, i.e., the abrupt change trend of physiological indicators. The disease stability coefficient ranges from 0.3 to 0.5, and the medical intervention impact factor is dynamically assigned based on adjustments to patient medication and changes in treatment measures, ranging from 0.2 to 0.8, with higher values ​​indicating higher treatment intensity. The vulnerability amplification factor ranges from 1.1 to 1.5, used to enhance vulnerability characteristics at key time points. The algorithm integrates dynamic changes over a time span through integral operations, adjusts the weights of different time periods using an exponential function term to highlight the impact of recent physiological changes on vulnerability, and captures key abrupt changes through the maximum value of the second derivative. Finally, it outputs the temporal vulnerability value at each time point, which directly reflects the degree of instability of the patient's condition at a specific time point, providing dynamic and continuous temporal feature support for risk assessment.

[0028] Preferably, the expression for the organ-level risk propagation analysis model is: ,in, Let be the risk value of the o-th organ at time t. Let be the transmission efficiency from the p-th organ to the q-th risk transmission pathway. for The risk value of the p-th organ at time p. Let be the response coefficient of the o-th organ at time t to the q-th conduction pathway. This is the path gradient adjustment factor. Let t be the risk propagation function from the p-th organ to the o-th organ. Regarding conduction parameters gradient, Let be the response coefficient of the r-th organ at time t to the q-th conduction pathway.

[0029] Specifically, the organ-level risk transmission analysis model is used to analyze the risk transmission patterns between the heart and related organs. In implementation, the transmission efficiency of risk transmission pathways between organs ranges from 0.3 to 0.9, assigned based on the closeness of physiological connections between organs. The transmission efficiency between the heart and directly related organs such as the lungs and kidneys is no less than 0.7, and the transmission efficiency between indirectly related organs is no more than 0.5. The pathway gradient adjustment factor ranges from 0.08 to 0.15 to ensure that the changes in the risk transmission gradient conform to physiological and pathological patterns. During model implementation, the types of organs participating in the risk transmission analysis are first identified, including core organs such as the heart, lungs, kidneys, liver, and brain. The initial risk value for each organ at different time points is determined, based on the quantification of the abnormality level of organ functional indicators. Subsequently, a directed graph model of risk transmission between organs is constructed, clarifying the number and direction of transmission pathways. Each organ corresponds to no less than 3 and no more than 8 transmission pathways. The response coefficient of each organ to different transmission pathways at each time point is calculated, ranging from 0.1 to 0.9. The response coefficient is dynamically adjusted according to the organ's functional state; the higher the degree of functional abnormality, the larger the response coefficient. The gradient change of risk propagation between organs is calculated using a risk propagation function to reflect the sensitivity of the conduction pathway. The model integrates the risk contributions of all precursor organs and all conduction pathways through a double summation operation, combined with the normalization of the response coefficient and the gradient adjustment term, to obtain the risk value of the target organ at the current time point. This calculation process needs to be iterative, covering the entire monitoring period, and finally, the risk diffusion path and intensity of different organ dysfunctions are analyzed, providing in-depth organ-level analysis results for a comprehensive assessment of the patient's overall risk.

[0030] Preferably, the risk assessment matrix construction expression of the intelligent heart failure risk stratification management platform is as follows: ,in, A risk assessment matrix with L rows and K columns. The weight coefficient of the k-th indicator in the i-th evaluation dimension. This represents the temporal vulnerability feature value corresponding to the k-th indicator of the i-th evaluation dimension. Let L be the organ-level risk value corresponding to the k-th indicator of the i-th assessment dimension, and L be the total number of assessment dimensions. The number of indicators under each evaluation dimension.

[0031] Specifically, the intelligent heart failure risk stratification management platform's risk assessment matrix integrates multi-dimensional features and quantifies risk assessment indicators. In implementation, the total number of assessment dimensions is set to six, corresponding to temporal vulnerability, organ function impairment, risk propagation scope, treatment response effectiveness, basic health status, and lifestyle adaptability. Each assessment dimension has 5-8 indicators, ensuring comprehensiveness and targeted indicators. The weight coefficient of the k-th indicator in each assessment dimension is calculated using a heart failure risk factor weight learning algorithm, ranging from 0.05 to 0.3, with a total weight coefficient of 1. The average weight coefficient for indicators corresponding to temporal vulnerability and organ function impairment is no less than 0.2, highlighting the importance of core assessment dimensions. During matrix construction, the temporal vulnerability feature values ​​output by the temporal vulnerability feature extraction unit and the organ-level risk values ​​output by the organ-level risk propagation analysis unit are first classified and matched according to the correspondence between assessment dimensions and indicators, ensuring that each indicator has a corresponding feature value and risk value. Subsequently, the weight coefficient of the k-th indicator in each assessment dimension, the corresponding time-series vulnerability eigenvalue, and the organ-level risk value are multiplied to obtain the comprehensive risk quantification result for that indicator. Following the rule of rows for assessment dimensions and columns for indicators, the comprehensive risk quantification results of all indicators are arranged in an ordered manner to form an L-row, K-column risk assessment matrix. Each element in the matrix is ​​a quantified value between 0 and 10, with larger values ​​indicating higher risk for that dimension and indicator. During the construction process, it is crucial to ensure consistent data format and dimension matching to avoid matrix construction anomalies due to data heterogeneity. The resulting multi-dimensional risk assessment matrix can systematically integrate risk information from multiple aspects, including time series and organ-level data, providing structured and quantifiable input data for subsequent risk prediction models, ensuring the comprehensiveness and systematic nature of the risk assessment.

[0032] Preferably, the expression for the readmission risk prediction model for elderly heart failure patients is: ,in, This is the risk value for readmission. It is the Sigmoid activation function. , This is the model weight matrix. , The bias term is ReLU, which is a linear rectified function. The risk amplification factor at time t, Let T be the total risk at the organ level at time t, where T is the time span of the time series data.

[0033] Specifically, the readmission risk prediction model for elderly heart failure patients is used to ultimately calculate the readmission risk value. During implementation, the model weight matrix is ​​obtained through iterative optimization using training data. The dimensions of the weight matrix are determined based on the number of columns in the risk assessment matrix and the number of hidden layer nodes. The number of hidden layer nodes is set to 64-128 to ensure sufficient model fitting ability. The bias term ranges from -0.5 to 0.5 and is used to adjust the baseline value of the model output. The time span T is determined based on the time series data collection period, typically 90-180 days, covering the key diagnosis, treatment, and rehabilitation stages of the patient. The risk amplification factor at time t ranges from 1.0 to 1.3 and is dynamically adjusted over time, with a value no lower than 1.2 within one month after discharge to highlight the importance of short-term risk. During model implementation, the multi-dimensional risk assessment matrix is ​​first input into the model. Feature extraction and nonlinear transformation are performed through linear transformation and the ReLU activation function. The ReLU function effectively enhances the model's ability to express complex features and suppresses the gradient vanishing problem. Subsequently, a second linear transformation and a sigmoid activation function are used to map the output to the 0-1 interval, yielding a preliminary risk value. Simultaneously, the product of the temporal vulnerability value at each time point and the sum of the overall organ-level risks is calculated. After adjustment with a risk amplification factor, this product is multiplied again to obtain a temporal risk correction term. This correction term strengthens the influence of dynamic risk characteristics on the final prediction result. Finally, the preliminary risk value is multiplied by the temporal risk correction term to obtain the final readmission risk value, which ranges from 0 to 1, with values ​​closer to 1 indicating a higher readmission risk. During model implementation, parameters need to be continuously optimized using validation datasets to ensure a prediction accuracy of no less than 85%. The final output risk value can be directly used for risk level classification, providing accurate risk quantification for clinical diagnosis and treatment decisions.

[0034] Preferably, step S3 includes the following sub-steps: S31, arranging the time-series data in the calibration factor set in an ordered manner according to timestamps, dividing it into multiple consecutive time windows, each time window corresponding to a fixed time interval, and extracting the statistical characteristics and trend characteristics of risk factors within each time window; S32, calculating the trend change rate based on the characteristic differences between adjacent time windows, setting a vulnerability threshold range in conjunction with the progression pattern of heart failure, and initially screening out abnormal time windows that exceed the threshold range; S33, performing fine-grained analysis on the time-series data within the abnormal time windows, capturing the mutation points and gradual trends of physiological indicators, and constructing a two-dimensional vulnerability feature matrix with time and indicator dimensions; S34, completing the missing data through time-series interpolation, performing dimensional regularization on the feature matrix, forming a standardized time-series vulnerability feature vector for subsequent risk assessment.

[0035] Specifically, step S3 includes four sub-steps. In S31, the time-series data in the calibration factor set are first sorted in ascending order by timestamp to ensure the temporal continuity of the data. Then, a sliding window method is used to divide the time window, with the window duration set to 7 days and the sliding step size to 1 day. This parameter setting balances time resolution and computational efficiency. Each window must include at least 12 valid data points. The extracted statistical features include basic indicators such as mean, standard deviation, median, and extreme values. The trend features include linear trend slope, fluctuation amplitude, and rate of change between adjacent data points, comprehensively capturing the static and dynamic attributes of the data within the window. In S32, the rate of change of the trend is calculated based on the difference of the same feature between adjacent windows. Referring to the clinical patterns of the progression of heart failure in the elderly, a vulnerability threshold range is set through expert consultation. The upper and lower limits of the range are determined based on three times the standard deviation of historical data. Windows exceeding this range are judged as abnormal time windows, initially focusing on high-risk time segments. S33 employs a fine-grained analytical approach for abnormal time windows, refining the time granularity to the hourly level. Trend fitting and mutation detection algorithms capture abrupt changes and gradual trends in physiological indicators, with a sensitivity of 0.85 for mutation point detection to ensure no key changes are missed. A two-dimensional vulnerability feature matrix is ​​then constructed, with rows corresponding to time segments and columns corresponding to extracted feature types, totaling no fewer than 50 feature types. S34 uses linear interpolation to complete missing data, ensuring the missing data proportion does not exceed 10% of the total data in the window. The feature matrix is ​​standardized, mapping all feature values ​​to the 0-1 range. Principal component analysis is used for dimensionality normalization, retaining principal components with a cumulative contribution rate of no less than 85%, ultimately forming a standardized temporal vulnerability feature vector. This provides structured, high-quality temporal feature support for subsequent risk assessment.

[0036] Preferably, step S4 includes the following sub-steps: S41, classifying the organ function data involved in the calibration factor set, clarifying the functional indicator system of the heart and related organs such as the lungs, kidneys, and liver, and establishing an organ-indicator mapping relationship; S42, constructing a directed graph model of risk transmission between organs based on the correlation between anatomical structure and physiological function, defining nodes as organs, edges as risk transmission paths, and assigning an initial transmission probability to each edge; S43, calibrating the transmission probability through organ dysfunction data in multimodal data, analyzing the changes in the intensity and direction of risk transmission between organs under different pathological states, and updating the edge weights of the directed graph model; S44, based on the updated directed graph model, using a path search algorithm to traverse all possible risk propagation paths, calculating the risk propagation intensity of each path, and analyzing the risk diffusion mechanism of organ dysfunction.

[0037] Specifically, step S4 includes four sub-steps. S41 first classifies and organizes the organ function data involved in the calibration factor set, identifying the five core related organs: heart, lungs, kidneys, liver, and brain. Each organ corresponds to no fewer than eight functional indicators, establishing a precise mapping relationship between organs and functional indicators to ensure the accuracy and completeness of data classification, providing a clear data foundation for subsequent modeling. S42, based on the principle of the correlation between human anatomical structure and physiological function, constructs a directed graph model of risk transmission between organs. Organs are used as nodes in the graph, and risk transmission paths are used as directed edges. No fewer than 20 core transmission paths are defined. Initial transmission probabilities are set according to the closeness of the physiological correlation between organs; the initial transmission probability from the heart to the lungs and kidneys is set to 0.7, and the initial transmission probability to the liver and brain is set to 0.5, consistent with the understanding of clinical physiological mechanisms. S43 utilizes organ dysfunction data from multimodal datasets and employs an iterative optimization algorithm to calibrate the transmission probability. The number of iterations is set to 100, and the convergence threshold is set to 0.001. By analyzing the intensity changes and directional shifts of risk transmission under different pathological states, the edge weights of the directed graph model are dynamically updated to adapt the model to the individual patient's condition characteristics. S44, based on the updated directed graph model, uses a depth-first search algorithm to traverse all possible risk propagation paths. The maximum path search depth is set to 4 layers to balance computational efficiency and path coverage integrity. The risk propagation intensity of each path is calculated by weighting the product of edge weights and the transmission efficiency coefficient. Ultimately, the main risk diffusion paths, key transmission nodes, and intensity distribution of organ dysfunction are identified, providing organ-level in-depth analysis results for comprehensive risk assessment.

[0038] Preferably, step S5 includes the following sub-steps: S51, clarifying the core dimensions of risk assessment, including the degree of temporal vulnerability, the degree of organ function impairment, the scope of risk propagation, the treatment response effect, the basic health status and lifestyle adaptability, and setting corresponding assessment indicators for each dimension; S52, classifying and mapping the temporal vulnerability characteristics and organ-level risk propagation analysis results according to the assessment dimensions, and converting different types of feature data into assessment values ​​of a unified dimension; S53, determining the weight allocation of each dimension and indicator based on the differences in the importance of each assessment dimension through a heart failure risk factor weight learning algorithm, ensuring that the weights match the correlation with readmission risk; S54, constructing a multi-dimensional risk assessment matrix based on the weight allocation and assessment values, where each element in the matrix corresponds to the comprehensive risk quantification result of the set dimension and set indicators.

[0039] Specifically, step S5 includes four sub-steps. S51 clarifies six core risk assessment dimensions: temporal vulnerability level, organ function impairment level, risk propagation scope, treatment response effect, basic health status, and lifestyle adaptability. Each dimension sets 5-8 assessment indicators. The selection of indicators combines clinical experience with core risk prediction needs to ensure comprehensive assessment dimensions and highly targeted indicators, providing a clear structural framework for matrix construction. S52 accurately classifies and maps the temporal vulnerability features extracted in step S3 and the organ-level risk propagation results obtained in step S4 according to the definitions of assessment dimensions and indicators. A linear transformation method is used to uniformly convert feature data of different dimensions and ranges into assessment values ​​of 0-10. The conversion process references the historical maximum and minimum values ​​of the data to ensure the comparability and rationality of the assessment values. S53 calculates the weights of each assessment dimension and indicator based on the heart failure risk factor weight learning algorithm. The dimensional weights are determined using the analytic hierarchy process (AHP). The weights for temporal vulnerability and organ function impairment are both set to 0.2, the weights for risk propagation and treatment response are both set to 0.15, and the weights for basic health status and lifestyle fit are both set to 0.1. Indicator weights are allocated within their respective dimensions based on their importance, ensuring that the weight allocation conforms to the logic of clinical risk assessment. S54, based on the determined weight allocation and the transformed assessment values, constructs a multi-dimensional risk assessment matrix of approximately 6 rows and 35 columns. Each element in the matrix is ​​the product of the corresponding dimension indicator's weight and the assessment value. The matrix is ​​then normalized, standardizing all element values ​​to the 0-1 range, ultimately forming a structured and quantifiable risk assessment matrix, providing comprehensive and systematic input data support for subsequent risk prediction models.

[0040] The heart failure risk factor weight learning algorithm is an algorithm in this invention that quantifies the impact of various risk factors in multimodal data on the readmission risk of elderly heart failure patients. It clarifies the importance and priority of different risk factors through multi-dimensional correlation analysis and quantitative calculations. The implementation process first integrates multi-source data from four categories: clinical diagnosis and treatment, physiological monitoring, imaging examinations, and lifestyle, identifying no fewer than 80 potential risk factors. Then, it calculates the covariance between each factor and the readmission risk label to quantify the linear correlation strength. Simultaneously, it calculates the correlation coefficient between the factor and other related factors, adjusting for interactions using a correlation factor influence coefficient of 0.05-0.2. Information entropy is introduced to measure data dispersion, and an entropy penalty coefficient of 0.1-0.3 is used to suppress the weights of low-reliability factors. Finally, standardized weights are output through comprehensive calculations of the numerator and denominator. The numerator integrates positive factors such as covariance and the interaction of related factors, while the denominator includes the sum of the covariances of all factors and an entropy penalty term. This algorithm accurately selects the top 30% of the labeling factors by importance, eliminates redundant information, and ensures that subsequent steps focus on core influencing factors. It solves the problems of chaotic and ambiguous risk factors in multimodal data. By scientifically quantifying weights, it provides a precise basis for feature extraction and model construction, making risk prediction more targeted, avoiding the neglect of labeling factors or interference from invalid factors, ensuring the scientific nature of the prediction system, and improving the overall prediction accuracy.

[0041] The temporal vulnerability assessor for heart failure in the elderly is a tool that dynamically captures the vulnerability characteristics of a patient's physiological state over time and quantifies the degree of instability of the condition. It outputs quantitative vulnerability values ​​at different time points through temporal data analysis and multi-parameter calculation. The implementation requires first determining a monitoring period covering 3 months before admission to 6 months after discharge, screening 15-20 core time-series factors, dividing time segments using a sliding window method with a window size of 7 days and a sliding step of 1 day, extracting features such as mean, standard deviation, and rate of change within each window, calculating the first derivative of the physiological state characteristic function to reflect the instantaneous rate of change, and the second derivative to capture mutation trends, adjusting the time-series weights by combining a disease stability coefficient of 0.3-0.5 and a medical intervention impact factor of 0.2-0.8, integrating long-term trends through a time-series integral adjustment coefficient of 0.6-0.8, and strengthening mutation characteristics with a second derivative weight coefficient of 0.4-0.6 and a vulnerability amplification factor of 1.1-1.5, and finally outputting the vulnerability value at each time point, dynamically presenting the evolution of the patient's condition over time and capturing vulnerability signals before the condition deteriorates. This assessment tool breaks through the limitations of traditional static assessments, accurately reflects the temporal characteristics of the course of heart failure in the elderly, provides continuous and dynamic feature support for risk prediction, makes risk assessment more in line with the actual changes in the patient's condition, improves the timeliness and sensitivity of prediction, and provides key evidence for early intervention.

[0042] The organ-level risk propagation analysis model is a model for analyzing the risk transmission patterns between the heart and related organs and reconstructing the complexity of disease progression. Through directed graph modeling and iterative computation, it quantifies the cross-organ risk transmission paths and intensity. Its implementation first identifies five core organs: heart, lungs, kidneys, liver, and brain. Each organ corresponds to at least eight functional indicators, and at least 20 core conduction pathways are constructed. Initial conduction probabilities are set based on anatomy and physiological function, with the initial conduction probability between the heart and lungs / kidneys not less than 0.7, and between indirectly related organs not more than 0.5. Subsequently, using organ dysfunction data, the conduction efficiency is optimized and calibrated through more than 100 iterations (values ​​ranging from 0.3 to 0.9). A pathway gradient adjustment factor of 0.08 to 0.15 is used to adjust conduction sensitivity, and the response coefficients of each organ to different pathways are calculated (0.1 to 0.9). A double summation operation integrates the risk contributions of all precursor organs and conduction pathways to obtain the risk value of the target organ at each time point, clearly presenting the risk diffusion paths, key conduction nodes, and intensity distribution of organ dysfunction. This model overcomes the shortcomings of traditional methods that isolate and assess the function of a single organ, fully reconstructs the pathophysiological mechanism of the interaction between multiple organs in elderly patients with heart failure, provides in-depth analysis at the organ level for risk assessment, makes risk prediction more comprehensive and in line with the nature of the disease, and provides scientific support for the clinical development of multi-organ synergistic intervention plans.

[0043] The intelligent heart failure risk stratification management platform is an integrated carrier that integrates multi-source data, coordinates various technologies, and realizes risk assessment and stratification. It constructs a closed-loop management system from data processing to risk output through the collaborative operation of multiple modules. Its implementation requires first building a data integration module to receive multi-source heterogeneous data and complete format standardization and correlation mapping. Then, it connects to a heart failure risk factor weight learning algorithm to complete factor screening, extracts dynamic features through a temporal vulnerability assessor, calls an organ-level risk propagation analysis model to analyze cross-organ risk mechanisms, and then constructs an assessment system including 6 primary dimensions and approximately 35 secondary indicators. Dimension weights are assigned through algorithms (the weights for temporal vulnerability and organ function impairment dimensions are no less than 0.2), converting the feature data into a unified assessment value of 0-10. A multi-dimensional risk assessment matrix is ​​constructed and normalized to the 0-1 interval. Finally, it connects to a readmission risk prediction model to calculate the risk value (0-1), classifies it into four risk levels using the quartile method, and coordinates the output results of each technical link to achieve accurate risk quantification, level classification, and result display. This platform breaks down information barriers between various technical aspects, forming a systematic and standardized risk prediction process. It solves the problems of fragmentation and non-standard processes in traditional prediction methods, significantly improving the efficiency and reliability of risk prediction. It provides clinicians with clear and intuitive decision-making basis, supports the formulation of personalized treatment plans and the optimal allocation of medical resources, and effectively improves the quality of medical management for elderly patients with heart failure.

[0044] like Figure 2As shown, a multimodal data-driven method for predicting readmission risk in elderly patients with heart failure is presented. This method is implemented through different units, including: a multimodal data integration and acquisition unit, which receives multi-source data from clinical diagnosis and treatment, physiological monitoring, imaging examinations, and lifestyle-related data, and forms a structured multimodal dataset through data format conversion and association mapping to provide data support for subsequent analysis; a risk factor weight dynamic calculation unit, connected to the multimodal data integration and acquisition unit, which performs factor correlation analysis and importance ranking on the dataset using a heart failure risk factor weight learning algorithm, and outputs a set of calibrated risk factors; and a temporal vulnerability feature extraction unit, connected to the risk factor weight dynamic calculation unit, which extracts temporal vulnerability features from the calibrated factors using an elderly heart failure temporal vulnerability assessor. The system performs dynamic feature capture and vulnerability quantification on sequential data to generate temporal vulnerability feature vectors. An organ-level risk propagation analysis unit, connected to a risk factor weight dynamic calculation unit, constructs cross-organ risk transmission paths through an organ-level risk propagation analysis model and outputs organ-level risk propagation results. A multi-dimensional risk matrix construction unit, connected to both the temporal vulnerability feature extraction unit and the organ-level risk propagation analysis unit, integrates the two types of feature results and combines weight allocation to construct a risk assessment matrix. A readmission risk prediction and stratification unit, connected to the multi-dimensional risk matrix construction unit, calculates risk values ​​and completes grading through a readmission risk prediction model for elderly heart failure patients, transmitting the results to an intelligent heart failure risk stratification management platform for storage and display.

[0045] A multimodal data-driven method for predicting readmission risk in elderly heart failure patients has been developed. This method comprehensively collects heterogeneous data from multiple sources, including clinical diagnosis and treatment, physiological monitoring, imaging examinations, and lifestyle data, avoiding the information bias caused by single data dimensions. Furthermore, a specialized risk factor weight learning mechanism accurately identifies key factors strongly correlated with readmission risk, addressing the shortcomings of existing technologies in integrating multimodal data and identifying ambiguous influencing factors. This makes risk assessment more comprehensive and targeted. Simultaneously, a multi-dimensional assessment system is constructed based on an intelligent heart failure risk stratification management platform, forming a closed-loop management system from data collection to risk stratification, significantly improving the systematic and standardized nature of the prediction process.

[0046] This method utilizes a temporal vulnerability assessor to deeply explore the evolutionary characteristics of patients' physiological states and disease conditions across different time dimensions, dynamically presenting the changing trends of disease vulnerability and ensuring that risk assessment aligns with the temporal characteristics of heart failure in the elderly. Addressing the issue of existing technologies neglecting inter-organ risk correlations, it employs an organ-level risk propagation analysis model to analyze the risk transmission pathways and diffusion mechanisms between the heart and related organs, fully reconstructing the complexity of disease progression and overcoming the shortcomings of traditional methods that isolate and assess single-organ function. By integrating various feature results through a multi-dimensional risk assessment matrix and combining them with a specialized prediction model, it achieves accurate calculation and classification of risk values, significantly improving the accuracy and reliability of readmission risk prediction and providing scientific support for clinical diagnosis and management.

[0047] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," "link," and "fix" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0048] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-modal data-driven method for predicting re-hospitalization risk of elderly heart failure patients, characterized in that, Comprise the following steps: S1, collect the clinical diagnosis and treatment data, physiological index monitoring data, image examination data and lifestyle related data of elderly heart failure patients, and construct a multi-modal original data set; S2, analyze the correlation and importance of each risk factor in the multi-modal original data set by a heart failure risk factor weight learning algorithm, and screen out a set of calibration factors that are strongly associated with rehospitalization risk; S3, use the elderly heart failure time series vulnerability evaluator to extract dynamic features from the time series data in the calibration factor set, and capture the vulnerability features of the patient's physiological state and disease changes at different time dimensions; S4, use the organ-level risk propagation analysis model to model the cross-organ risk transmission path of the heart and related organ function related data in the calibration factor set, and analyze the risk diffusion mechanism of organ dysfunction; S5, input the time series vulnerability features and organ-level risk propagation analysis results into the intelligent heart failure risk stratification management platform, and construct a multi-dimensional risk assessment matrix combined with the preset risk assessment dimensions; S6, based on the risk assessment matrix, calculate the rehospitalization risk value of the elderly heart failure patient through the heart failure patient rehospitalization risk prediction model, and complete the risk grade division.

2. The multi-modal data-driven rehospitalization risk prediction method for elderly heart failure patients according to claim 1, characterized in that, The expression for the heart failure risk factor weight learning algorithm is: ,in, This represents the weight of the j-th risk factor in the i-th data category. This is the correlation adjustment coefficient. Let be the observed value of the j-th risk factor in the i-th data category, and y be the readmission risk label variable. Let be the influence coefficient of the k-th correlation factor. Let be the correlation coefficient between the j-th risk factor and the k-th risk factor. The total number of data categories, The number of risk factors in each data category. The entropy penalty coefficient is... Let the information entropy of the j-th risk factor be , This represents the observed value of the nth risk factor in the m-th data category. The covariance between the re-admission risk label variable y and the re-admission risk label variable y.

3. The multi-modal data-driven rehospitalization risk prediction method for elderly heart failure patients according to claim 1, characterized in that, The expression of the old heart failure timing vulnerability evaluator is: wherein, is the timing vulnerability value of the patient at time t, is the timing integral adjustment coefficient, is the physiological state characteristic function at time s, is is the disease stability coefficient at time t, is is the medical intervention influence factor at time t, is the second derivative weight coefficient, is the vulnerability amplification factor at time s.

4. The multi-modal data-driven rehospitalization risk prediction method for elderly heart failure patients according to claim 1, characterized in that, The expression for the organ-level risk propagation analysis model is: ,in, Let be the risk value of the o-th organ at time t. Let be the transmission efficiency from the p-th organ to the q-th risk transmission pathway. for The risk value of the p-th organ at time p. Let be the response coefficient of the o-th organ at time t to the q-th conduction pathway. This is the path gradient adjustment factor. Let t be the risk propagation function from the p-th organ to the o-th organ. Regarding conduction parameters gradient, Let be the response coefficient of the r-th organ at time t to the q-th conduction pathway.

5. The multi-modal data-driven rehospitalization risk prediction method for elderly heart failure patients according to claim 1, wherein, The risk assessment matrix construction expression of the intelligent heart failure risk stratification management platform is: Wherein, is a risk assessment matrix of L rows and K columns, is a weight coefficient of the kth index of the Ith evaluation dimension, is a time sequence vulnerability characteristic value corresponding to the kth index of the Ith evaluation dimension, is an organ-level risk value corresponding to the kth index of the Ith evaluation dimension, L is the total number of evaluation dimensions, is the number of indexes under each evaluation dimension.

6. The multi-modal data-driven rehospitalization risk prediction method for elderly heart failure patients according to claim 1, characterized in that, The expression of the rehospitalization risk prediction model for the elderly heart failure patients is: wherein, is a rehospitalization risk value, is a Sigmoid activation function, , is a model weight matrix, , is a bias term, and ReLU is a linear rectification function, is a risk amplification coefficient at time t, is a total sum of overall organ-level risks at time t, and T is a time span of time series data.

7. The multi-modal data-driven rehospitalization risk prediction method for elderly heart failure patients according to claim 1, wherein, S3 comprises the following steps: S31, arrange the time series data in the calibration factor set in order according to the time stamp, divide a plurality of continuous time windows, each time window corresponds to a fixed time interval, and extract the statistical features and change trend features of the risk factors in each time window; S32, calculate the trend change rate based on the feature difference of adjacent time windows, set the vulnerability threshold interval according to the progression rule of heart failure disease, and preliminarily screen out abnormal time windows that exceed the threshold interval; S33, analyze the time series data in the abnormal time window in detail, capture the mutation point and gradual change trend of the physiological index, and construct a two-dimensional vulnerability feature matrix of time dimension and index dimension; S34, complete the missing data by time series interpolation, regularize the feature matrix, and form a standardized time series vulnerability feature vector for subsequent risk assessment.

8. The multi-modal data-driven rehospitalization risk prediction method for elderly heart failure patients according to claim 1, wherein, S4 comprises the following steps: S41, classify the organ function data in the calibration factor set, and clearly define the function index system of the heart and the lungs, kidneys, and liver associated organs, and establish the organ-index mapping relationship; S42, based on the correlation between anatomical structure and physiological function, construct a directed graph model of inter-organ risk transmission, define the nodes as organs and the edges as risk transmission paths, and assign an initial transmission probability to each edge; S43, calibrate the transmission probability through the organ function abnormal data in the multi-modal data, analyze the transmission intensity and direction change of risk between organs under different pathological states, and update the edge weight of the directed graph model; S44, based on the updated directed graph model, use a path search algorithm to traverse all possible risk propagation paths, calculate the risk propagation intensity of each path, and analyze the risk diffusion mechanism of organ dysfunction.

9. The multi-modal data-driven rehospitalization risk prediction method for elderly heart failure patients according to claim 1, wherein, The S5 comprises the following steps: S51, determining the core dimensions of risk assessment, including the time sequence vulnerability degree, the organ function damage degree, the risk propagation range, the treatment response effect, the basic health condition and the lifestyle adaptation degree, and setting corresponding evaluation indexes for each dimension; S52, classifying and mapping the time sequence vulnerability characteristics and the organ level risk propagation analysis results according to the evaluation dimensions, and converting different types of characteristic data into evaluation values of uniform dimensions; S53, determining the weight distribution of each dimension and index according to the differences in the importance of each evaluation dimension through a heart failure risk factor weight learning algorithm, and ensuring that the weight is matched with the rehospitalization risk relevance; S54, constructing a multi-dimensional risk assessment matrix based on the weight distribution and the evaluation values, and setting the comprehensive risk quantification results of each dimension and index in each element of the matrix.

10. The multi-modal data-driven rehospitalization risk prediction method for elderly heart failure patients according to any one of claims 1-9, characterized in that, The method is realized by different units, including: a multi-modal data integration and acquisition unit for receiving multi-source data associated with clinical diagnosis and treatment, physiological monitoring, image examination and lifestyle, forming a structured multi-modal data set through data format conversion and association mapping, and providing data support for subsequent analysis; a risk factor weight dynamic calculation unit connected with the multi-modal data integration and acquisition unit, performing factor relevance analysis and importance sorting on the data set through a heart failure risk factor weight learning algorithm, and outputting a calibrated risk factor set; a time sequence vulnerability feature extraction unit connected with the risk factor weight dynamic calculation unit, performing dynamic feature capture and vulnerability quantification on the time sequence data in the calibrated factor through an elderly heart failure time sequence vulnerability evaluator, and generating a time sequence vulnerability feature vector; an organ level risk propagation analysis unit connected with the risk factor weight dynamic calculation unit, constructing a cross-organ risk transmission path through an organ level risk propagation analysis model, and outputting an organ level risk propagation result; a multi-dimensional risk matrix construction unit connected with the time sequence vulnerability feature extraction unit and the organ level risk propagation analysis unit, integrating the two types of feature results and constructing a risk assessment matrix combined with the weight distribution; a rehospitalization risk prediction and stratification unit connected with the multi-dimensional risk matrix construction unit, calculating the risk value and completing the grade division through an elderly heart failure patient rehospitalization risk prediction model, and transmitting the results to an intelligent heart failure risk stratification management platform for storage and display.

Images