Multi-modal health data feature extraction method and system

Abstract

The invention provides a multi-modal health data feature extraction method and system, and relates to the technical field of medical data processing, and the method comprises the steps: obtaining a multi-modal health data set which comprises time sequence physiological data, image data and text data; performing random noise processing on the time sequence physiological data to extract blood pressure fluctuation characteristics; based on the blood pressure fluctuation characteristics, performing structural noise processing on the image data to extract morphological characteristics of the fundus blood vessels; in the process of extracting the morphological features of the fundus blood vessels, dynamically positioning three feature anchor points on a fundus image corresponding to the image data; the anatomical positions of the three characteristic anchor points are respectively an optic disc edge upper quadrant reference point, a macular region bitamporal trunk blood vessel bifurcation core point and an arteriovenous cross indentation center point; and constructing a dynamic response area based on the three feature anchor points. The cardiovascular event risk assessment precision can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical data processing, in particular to a multi-modal health data feature extraction method and system. BACKGROUND

[0002] In the field of cardiovascular disease risk assessment, multi-modal health data includes time-series physiological data, image data and text data, and its comprehensive analysis is the key to improving the accuracy of assessment. With the development of medical monitoring, imaging and electronic health record technology, multi-source heterogeneous health data has been gradually realized. How to effectively extract relevant features and support risk level judgment has become the core research direction in this field.

[0003] The existing technology for feature extraction of health data mostly adopts the mode of independent processing and simple fusion. Usually, isolated feature extraction is first performed on the three types of data, such as extracting fluctuation features from blood pressure signals, extracting blood vessel shape features from fundus images, and extracting drug-related features from electronic health records. Then, feature splicing or simple weighting is used to complete fusion, and then risk assessment is carried out.

[0004] These methods have the following technical defects: the extraction of features of each modality is isolated, the inherent dynamic correlation between modalities is not fully utilized, such as the physiological correlation between blood pressure fluctuation and fundus blood vessel shape, and the synergistic correlation between drug behavior and blood pressure and blood vessel state, and there is a lack of targeted dynamic calibration of single-modality features. The fundus blood vessel shape changes with the blood pressure fluctuation, and the drug behavior also affects the blood vessel state through the blood pressure, but the existing technology does not combine these correlations for calibration or accurate mining when extracting related features, resulting in insufficient accuracy of single-modality features. This isolated extraction and lack of correlation calibration mode fails to exploit the complementarity and correlation between modalities, reducing the effectiveness of fusion, and ultimately making it difficult to meet the clinical needs of cardiovascular event risk level assessment accuracy. SUMMARY

[0005] The technical problem to be solved by the present application is to provide a multi-modal health data feature extraction method and system, which can fully exploit the complementarity and correlation of data, and improve the accuracy of features and the effectiveness of fusion.

[0006] To solve the above technical problems, the technical solutions of the present application are as follows: In a first aspect, a multi-modal health data feature extraction method is provided, the method comprising: obtaining a multi-modal health data set, the multi-modal health data set comprising time-series physiological data, image data and text data; performing random noise processing on the time-series physiological data to extract blood pressure fluctuation features; Based on the blood pressure fluctuation characteristics, the image data is processed for structural noise to extract the fundus vascular morphology characteristics; in the process of extracting the fundus vascular morphology characteristics, three feature anchor points are dynamically positioned on the fundus image corresponding to the image data; the anatomical positions of the three feature anchor points are the upper quadrant reference point on the optic disc edge, the core point of the main trunk blood vessel bifurcation on the temporal side of the macular area, and the center point of the arteriovenous crossing pressure trace; a dynamic response region is constructed based on the three feature anchor points; The dynamic response region is divided into structural units to form a plurality of analysis units; a feature compensation coefficient is formed according to the deformation characteristics of the analysis unit; the three quantitative indicators of the fundus vascular morphology characteristics are calibrated based on the feature compensation coefficient to obtain the calibrated fundus vascular morphology characteristics; the blood pressure fluctuation characteristics and the calibrated fundus vascular morphology characteristics are fused to generate a first correlation feature; Based on the first correlation feature, the text data is processed for semantic noise to extract the medication adherence characteristics; the medication adherence characteristics and the first correlation feature are fused to generate a second correlation feature; The second correlation feature is processed for multi-modal fusion to obtain the cardiovascular event risk level.

[0007] Further, a multi-modal health data set is obtained, which includes time series physiological data, image data and text data, including: The time series physiological signals from the medical monitoring equipment are collected, the image data generated by the fundus imaging equipment is obtained, and the related text information is extracted from the electronic health record system; The time series physiological signals are processed for signal quality evaluation to obtain the time series physiological signals after quality evaluation; The image data is processed for image quality verification to obtain the image data after quality verification; The text information is processed for structured preprocessing to obtain the preprocessed text information; The time series physiological signals after quality evaluation, the image data after quality verification, and the preprocessed text information are standardized and integrated to obtain the standardized and integrated data; the standardized and integrated data is organized into a multi-modal health data set in a unified format.

[0008] Further, the time series physiological data is processed for random noise to extract the blood pressure fluctuation characteristics, including: The blood pressure signals in the time series physiological data are processed for segmentation to obtain a plurality of blood pressure signal segments; Random noise interference is applied to each blood pressure signal segment to generate a noise-enhanced blood pressure signal segment; The time domain feature analysis is performed on each noise-enhanced blood pressure signal segment to calculate the time domain feature vector; the frequency domain feature analysis is performed on each noise-enhanced blood pressure signal segment to extract the frequency domain feature vector; The time domain feature vector is fused with the frequency domain feature vector to form a fused feature vector; Based on the fused feature vector, a feature selection calculation is performed to screen out blood pressure fluctuation features with the highest discriminability; the blood pressure fluctuation features are normalized to form a standardized blood pressure fluctuation feature set.

[0009] Further, based on the blood pressure fluctuation features, structural noise processing is performed on the image data to extract fundus vessel morphological features; in the process of extracting the fundus vessel morphological features, three feature anchor points are dynamically positioned on the fundus image corresponding to the image data; the anatomical positions of the three feature anchor points are the upper quadrant reference point on the edge of the optic disc, the core point of the main trunk vessel bifurcation on the temporal side of the macular region, and the center point of the arteriovenous crossing indentation; based on the three feature anchor points, a dynamic response region is constructed, including: According to the standardized blood pressure fluctuation feature set, the injection parameters of the structural noise are determined; Based on the injection parameters, structural noise processing is performed on the fundus image in the image data to obtain a processed fundus image; The vessel network topology structure is segmented from the processed fundus image; The three feature anchor points are dynamically positioned on the vessel network topology structure, and the anatomical positions of the three feature anchor points include the upper quadrant reference point on the edge of the optic disc, the core point of the main trunk vessel bifurcation on the temporal side of the macular region, and the center point of the arteriovenous crossing indentation; Based on the three feature anchor points, a feature response map is constructed, i.e. a dynamic response region is constructed; The vessel morphological feature parameters are extracted from the feature response map, and the vessel morphological feature parameters are optimized to form the final fundus vessel morphological features.

[0010] Further, the dynamic response region is divided into structural units to form a plurality of analysis units; a feature compensation coefficient is formed according to the deformation characteristics of the analysis units; based on the feature compensation coefficient, the three quantitative indexes of the fundus vessel morphological features are calibrated to obtain calibrated fundus vessel morphological features; the blood pressure fluctuation features and the calibrated fundus vessel morphological features are fused to generate first correlation features, including: The three feature anchor points in the feature response map are taken as generation elements to construct a plane point set; based on the generation elements, the dynamic response region is spatially divided by adopting a plane point set Voronoi diagram construction calculation method to obtain a plurality of Voronoi units; Each Voronoi unit is defined as an analysis unit, and the spatial geometric features of each analysis unit are calculated with the corresponding generation element as the center; Correlate the spatial geometric features with the time-series changes of the blood pressure fluctuation features, calculate the deformation parameters of each analysis unit relative to the generating element; based on the deformation parameters of all analysis units relative to the respective generating elements, construct a deformation feature matrix; Perform eigenvalue decomposition on the deformation feature matrix to extract the main deformation components; according to the correlation of the main deformation components with the spatial distribution of the generating elements, calculate the feature compensation coefficients; Through the feature compensation coefficients, dynamically calibrate the three quantitative indicators of the fundus vascular morphological features, and obtain the calibrated fundus vascular morphological features; Fuse the calibrated fundus vascular morphological features with the standardized blood pressure fluctuation features in multiple scales, based on the spatial distribution characteristics of the generating elements, and through feature importance evaluation, generate a first correlation feature containing dynamic correlation information of blood vessels and blood pressure.

[0011] Further, based on the first correlation feature, perform semantic noise processing on the text data to extract medication adherence features; fuse the medication adherence features with the first correlation feature to generate a second correlation feature, including: Based on the first correlation feature, determine the intensity parameter of the semantic noise processing; According to the intensity parameter, perform semantic disturbance processing on the medication records in the text data to extract medication frequency features and medication regularity features from the disturbed medication records; Based on the medication frequency features and the medication regularity features, construct a medication behavior mode vector; Perform correlation degree analysis on the medication behavior mode vector and the first correlation feature to obtain a correlation degree analysis result; according to the correlation degree analysis result, calculate a feature fusion weight coefficient; Based on the feature fusion weight coefficient, perform weighted fusion on the medication behavior mode vector and the first correlation feature, and through feature dimension reduction processing, generate a second correlation feature containing correlation information of blood vessel morphology, blood pressure fluctuation and medication behavior.

[0012] Further, perform multi-modal fusion processing on the second correlation feature to obtain a cardiovascular event risk level, including: Perform feature standardization processing on the second correlation feature to obtain a standardized multi-modal feature set; Based on the feature importance evaluation logic, calculate the weight coefficients of the features in the multi-modal feature set; According to the weight coefficients, perform weighted fusion processing on the multi-modal feature set to obtain a weighted fused feature; Through a multi-level feature extraction network, extract deep correlation features from the weighted fused feature; Based on the deep correlation features, calculate the risk probability of a cardiovascular event by using a risk assessment classification mechanism; According to the risk probability, in combination with a preset risk threshold, a final cardiovascular event risk level is determined.

[0013] In a second aspect, a multi-modal health data feature extraction system comprises: An acquisition module is configured to acquire a multi-modal health data set, the multi-modal health data set comprising time-series physiological data, image data, and text data. A processing module is configured to perform random noise processing on the time-series physiological data to extract blood pressure fluctuation features. A positioning module is configured to perform structural noise processing on the image data based on the blood pressure fluctuation features to extract fundus blood vessel morphology features; during the extraction of the fundus blood vessel morphology features, three feature anchor points are dynamically positioned on a fundus image corresponding to the image data; the anatomical positions of the three feature anchor points are, respectively, an upper quadrant reference point on the edge of the optic disc, a core point of a main blood vessel bifurcation on the temporal side of the macular region, and a center point of a dynamic-venous intersection pressure trace; a dynamic response region is constructed based on the three feature anchor points. A calibration module is configured to divide the dynamic response region into structural units to form a plurality of analysis units; a feature compensation coefficient is formed according to the deformation characteristics of the analysis units; three quantitative indicators of the fundus blood vessel morphology features are calibrated based on the feature compensation coefficient to obtain calibrated fundus blood vessel morphology features; the blood pressure fluctuation features and the calibrated fundus blood vessel morphology features are fused to generate first associated features. A fusion module is configured to extract medication adherence features by performing semantic noise processing on the text data based on the first associated features; the medication adherence features and the first associated features are fused to generate second associated features; and the cardiovascular event risk level is obtained by performing multi-modal fusion processing on the second associated features. one or more processors; a storage device configured to store one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the method.

[0014] In a fourth aspect, a computer-readable storage medium stores a program, which, when executed by a processor, implements the method.

[0015] The above-mentioned scheme of the present application at least has the following beneficial effects: Because the feature extraction scheme based on the step-by-step driving of the inter-modal dynamic correlation is adopted, the blood pressure fluctuation features are extracted from the time series physiological data by random noise processing, and then the image data is subjected to structural noise processing based on the features, the three feature anchor points of the quadrant reference point on the optic disc edge, the core point of the main temporal trunk blood vessel bifurcation in the macular area and the center point of the arteriovenous intersection indentation are dynamically positioned to construct a dynamic response area, the feature compensation coefficient is obtained through the analysis unit division and deformation characteristic analysis, the eye fundus blood vessel morphological features are calibrated, and the first correlation features are fused with the blood pressure fluctuation features, then the medication adherence features are extracted based on the first correlation features and the semantic noise processing of the text data is completed, and finally the risk level is obtained through the multi-modal fusion processing, so the core technical problems of the single-modal feature accuracy and the low fusion effectiveness caused by the isolation of the multi-modal feature extraction and the lack of inter-modal correlation calibration in the prior art are overcome, and the complementarity and correlation between the multi-modal data are effectively mined, the accuracy of feature extraction and the effectiveness of feature fusion are effectively improved, and finally the precision and reliability of the cardiovascular event risk level evaluation are improved, and the clinical application requirements are better met. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a flowchart of a multi-modal health data feature extraction method provided by an embodiment of the present application.

[0017] Figure 2 is a schematic diagram of a multi-modal health data feature extraction system provided by an embodiment of the present application. DETAILED DESCRIPTION

[0018] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided so that the present disclosure can be more thoroughly understood and the scope of the present disclosure can be accurately conveyed to those skilled in the art.

[0019] As Figure 1 shown, an embodiment of the present application proposes a multi-modal health data feature extraction method, which comprises the following steps: Step 1, acquiring a multi-modal health data set, the multi-modal health data set containing time series physiological data, image data and text data; Step 2, random noise processing is performed on the time series physiological data to extract blood pressure fluctuation features; Step 3, based on the blood pressure fluctuation characteristics, the image data is processed for structural noise to extract the fundus vascular morphology characteristics; in the process of extracting the fundus vascular morphology characteristics, three feature anchor points are dynamically positioned on the fundus image corresponding to the image data; the anatomical positions of the three feature anchor points are the upper quadrant reference point on the optic disc edge, the core point of the main trunk blood vessel bifurcation on the temporal side of the macular area, and the center point of the arteriovenous crossing pressure trace; a dynamic response area is constructed based on the three feature anchor points; Step 4, the dynamic response area is divided into structural units to form a plurality of analysis units; a feature compensation coefficient is formed according to the deformation characteristics of the analysis unit; based on the feature compensation coefficient, the three quantitative indexes of the fundus vascular morphology characteristics are calibrated to obtain the calibrated fundus vascular morphology characteristics; the blood pressure fluctuation characteristics and the calibrated fundus vascular morphology characteristics are fused to generate a first associated feature; Step 5, based on the first associated feature, the text data is processed for semantic noise to extract the medication adherence characteristics; the medication adherence characteristics and the first associated feature are fused to generate a second associated feature; Step 6, the second associated feature is processed for multi-modal fusion to obtain a cardiovascular event risk level.

[0020] In the embodiments of the present application, the beneficial effects are reflected in the pertinence of feature extraction, the relevance of association and the reliability of evaluation results. By taking the dynamic association between modalities as the core driving, the blood pressure fluctuation characteristics are first extracted from the time sequence physiological data, then the image data is processed in pertinence based on this, the feature anchor points of three specific anatomical positions are accurately positioned and a dynamic response area is constructed, the compensation coefficient is generated by analyzing the deformation characteristics of the blood pressure fluctuation time sequence change analysis unit, the accurate calibration of the fundus vascular morphology characteristics is realized, and the single modality feature is more in line with the physiological association law. The text data is processed based on the first associated feature obtained by the previous fusion, and the medication adherence characteristics extracted can form a deep echo with the physiological and image characteristics, rather than existing in isolation. Through step-by-step association, calibration and fusion of various modal characteristics, the complementary value of multi-source data is fully released, the accuracy of feature extraction and the effectiveness of fusion are greatly improved, and finally the output cardiovascular event risk level is more in line with the needs of clinical actual scene, providing more reliable technical support for disease risk assessment.

[0021] In a preferred embodiment of the present application, the above step 1 can include: Step 1.1, collecting time-series physiological signals from medical monitoring devices, acquiring image data generated by fundus imaging devices, and extracting relevant text information from electronic health record systems, specifically including: based on the comprehensive demand of multi-source data for cardiovascular event risk assessment, selecting medical monitoring devices with continuous monitoring function, and continuously collecting time-series physiological signals such as systolic blood pressure, diastolic blood pressure, heart rate, and blood oxygen saturation of the monitoring object at a sampling frequency of every 5 minutes; using a fundus camera with a resolution not less than 2000x2000 pixels, shooting the fundus of both eyes of the monitoring object according to the standard fundus photography specification, ensuring that the image data can clearly show the key anatomical structures such as optic disc, macular area, and arteriovenous vessels; according to the unique identification number of the monitoring object, searching and extracting the medication name, medication frequency, medication dose, diagnosis time, diagnosis result, and symptom description within the past 12 months from the hospital's unified electronic health record system, realizing the comprehensive coverage of three types of core data: time-series physiological signals, image data, and text information, and providing complete data support for the correlation extraction of multi-modal features.

[0022] Step 1.2, signal quality assessment of time-series physiological signals, obtaining quality-assessed time-series physiological signals, specifically including: using a fixed time length segmentation method combined with statistical analysis techniques to assess the quality of the collected time-series physiological signals segment by segment, with each segment length set to 30 seconds; calculating three core indicators for each segment: signal-to-noise ratio, data integrity, and abnormal value proportion. The signal-to-noise ratio is calculated by the ratio of signal effective amplitude to noise amplitude, the data integrity is measured by the proportion of valid data points to total data points in the segment, and the abnormal value is determined by statistical analysis within the normal physiological range. Set the qualified standard as signal-to-noise ratio not less than 30dB, data integrity not less than 95%, and abnormal value proportion not more than 3%, retain all signal segments that meet the standard, mark and remove the segments that do not meet the standard, avoid low-quality physiological signals interfering with the extraction of blood pressure fluctuation features, and ensure the reliability of time-series physiological data from the source.

[0023] Step 1.3, image quality verification of image data, forming quality-verified image data, specifically including: evaluating image clarity by calculating the gray level change amplitude of blood vessel edges, higher change amplitude indicating clearer image, and detecting contrast by analyzing the distribution range of image gray values to ensure that the gray difference between blood vessels and surrounding tissues meets the recognition requirements; for common interference in image data, identify whether there are motion artifacts, uneven lighting artifacts by identifying irregular light and dark areas, blurred trailing images, etc. in the image; at the same time, verify whether the optic disc contour is complete, the macular area positioning is accurate, and the arteriovenous vessels are continuous, to ensure the visibility of key anatomical structures, and refer to the clinical fundus image quality evaluation standard to select image data that meets all indicators.

[0024] Step 1.4, structured preprocessing of text information to obtain preprocessed text information, specifically including: first, unstructured text is divided into independent lexical units, and then through the identification of medical field special vocabulary, the key information fields such as drug name, drug time, drug dose, diagnosis date, diagnosis result, and symptom type are accurately extracted; adopt the preset format rule to unify the field content, for example, convert the drug time to the standard format of year, month, day, hour, and minute, and mark the drug dose according to the unified unit of milligram and milliliter; by comparing the authoritative medical term library, correct the wrong words and ambiguous expressions in the text, and filter out the redundant content such as administrative notes and irrelevant consultation records which are irrelevant to health feature extraction, finally convert the scattered and disordered text information into structured text data which is structured, field-specific and can be directly used for feature analysis.

[0025] Step 1.5, standardizing and integrating the quality-evaluated time-series physiological signals, the quality-verified image data and the preprocessed text information to obtain standardized and integrated data; organizing the standardized and integrated data into a unified format of multi-modal health data set, specifically including: formulating a unified multi-modal data format specification, time-series physiological signals are stored in CSV format, fields include monitoring time stamp, systolic pressure, diastolic pressure, heart rate, blood oxygen saturation and other core parameters; image data adopts DICOM standard format, with shooting time, equipment parameters, monitoring object identifier and other metadata; text data adopts JSON format, with key information fields as keys and corresponding contents as values; taking the unique identification number of the monitoring object and the monitoring time stamp as the core key field, the quality-evaluated time-series physiological signals, the quality-verified image data and the preprocessed text information are field-aligned and associated matched; constructing an efficient data index system to quickly retrieve corresponding data through key fields, ensuring that the three types of data of the same monitoring object under the same time dimension can be accurately corresponded; finally, these multi-source data which have been unified in format and associated matched are organized into a multi-modal health data set which is structured, associated and efficient in retrieval.

[0026] In the embodiments of the present application, by comprehensively collecting the time sequence physiological signals of the medical monitoring device, the image data of the fundus imaging device and the text information in the electronic health record, the comprehensive coverage of multi-source health data is realized, which provides a rich data basis for subsequent multi-modal feature extraction. The signal quality assessment, image quality verification and text structured preprocessing are carried out for different types of data, which can accurately eliminate invalid and interference data, regularize unstructured information, improve data quality from the source, and avoid the influence of poor data on the analysis results. And by standardizing and integrating the processed data of various types into a unified format, the format difference and compatibility problem of multi-source heterogeneous data are successfully eliminated, so that different modal data can efficiently and cooperatively participate in the subsequent feature extraction process, providing standardized and consistent data support for feature calibration and fusion based on inter-modal correlation, and further ensuring the smoothness of the entire feature extraction process and the accuracy of the final risk assessment result.

[0027] In a preferred embodiment of the present application, step 2 can include: Step 2.1, the blood pressure signal in the time sequence physiological data is segmented and processed to obtain a plurality of blood pressure signal segments, specifically including: referring to the short physiological period of about 2 minutes of sympathetic nerve regulation of blood pressure in clinic, determining to use a 2-minute fixed time window to segment and process the blood pressure signal in the time sequence physiological data; a non-overlapping interval is set between the windows to avoid repeated analysis of the same data point, which not only ensures that each signal segment can independently reflect the blood pressure dynamic change in a specific period, but also prevents the fragmentation of features caused by too short window or the masking of key fluctuation details caused by too long window; the continuous blood pressure data is sequentially cut from front to back in time sequence to form a plurality of blood pressure signal segments with consistent length, and for the last remaining data less than 2 minutes, the mean data of the previous adjacent signal segment is filled to complete 2 minutes, while the filled part is marked to distinguish the original data, ensuring the integrity of the data without affecting the authenticity of the features; to adapt to the time dimension correlation with image data and text data, the starting and ending time stamps of each signal segment are recorded to seconds, and a time index is established.

[0028] Step 2.2, a random noise interference is applied to each blood pressure signal segment to generate a noise-enhanced blood pressure signal segment, specifically including: first, the amplitude range and standard deviation of the original blood pressure signal are counted, the clinical physiological range of 90 to 220 mmHg for systolic pressure and 60 to 140 mmHg for diastolic pressure is taken as a reference to determine the intensity boundary of the random noise; the noise amplitude is controlled between 3% to 5% of the maximum amplitude of the original blood pressure signal, wherein the upper limit of the systolic pressure noise amplitude is set to 5% and the diastolic pressure is set to 3%, which fits the characteristics that the natural fluctuation of systolic pressure is more obvious in the clinic; a normal distribution random noise sequence with a mean of 0 and a standard deviation of 10% of the original blood pressure signal standard deviation is generated to ensure that the noise distribution matches the fluctuation characteristics of the signal itself and avoids conflicts between noise and signal rules; the noise sequence is superimposed on the corresponding data points of each blood pressure signal segment point by point, and the signal value after superimposition is verified point by point whether it is still within the clinical physiological blood pressure range, if it is out of range, the corresponding noise amplitude is adjusted to the signal compliance, so as to simulate common actual scenarios such as poor electrode contact of the device, slight electromagnetic interference of the environment, and patient limb micro-motion in clinical monitoring, and generate noise-enhanced blood pressure signal segments.

[0029] Step 2.3, time domain feature analysis is performed on each noise-enhanced blood pressure signal segment to calculate a time domain feature vector; frequency domain feature analysis is performed on each noise-enhanced blood pressure signal segment to extract a frequency domain feature vector, specifically including: time domain feature analysis is performed on each noise-enhanced blood pressure signal segment, and indicators are calculated around five core dimensions of blood pressure fluctuation, including level, dispersion, extreme value, speed and frequency; the level dimension calculates the mean of systolic and diastolic pressure, the dispersion dimension calculates the variance and standard deviation, the extreme value dimension extracts the peak value of systolic pressure, the valley value of diastolic pressure and the range of the two, the speed dimension calculates the time for the peak value to rise from the valley value and the time for the valley value to fall from the peak value, and the frequency dimension counts the number of fluctuations within 2 minutes when the difference between adjacent peak values exceeds 5 mmHg and the average peak interval of adjacent fluctuations, and these 10 indicators are combined in the logical order of level, dispersion, extreme value, speed and frequency to form a time domain feature vector; the time domain blood pressure signal is converted into a frequency domain signal through mathematical transformation, focusing on two key frequency bands related to cardiovascular regulation, 0.04 to 0.15 Hz is the sympathetic nerve activity regulation frequency band, and 0.15 to 0.4 Hz is the respiratory rhythm associated frequency band, the frequency peak, energy value and energy proportion of total energy of each frequency band are calculated, and the main frequency bandwidth, i.e. the main frequency plus or minus 0.02 Hz and the total energy value of the full frequency band are extracted, and these 8 indicators are integrated in the order of frequency band energy, energy proportion and frequency parameter to form a frequency domain feature vector, which comprehensively captures the numerical change and frequency distribution characteristics of blood pressure fluctuation.

[0030] Step 2.4, the time domain feature vector and the frequency domain feature vector are fused at the feature level to form a fused feature vector, specifically including: first, the time domain feature vector and the frequency domain feature vector are respectively subjected to standardization processing, taking the mean value of the feature in all signal segments as the benchmark, subtracting the mean value from the feature value and then dividing by the standard deviation, so that the processed features have a mean value of 0 and a standard deviation of 1, completely eliminating the fusion bias caused by different units and orders of magnitude of the two types of vectors; assigning weights based on the correlation between the features and the core law of blood pressure fluctuation, determining the correlation strength by statistical analysis of the difference between each feature in known blood pressure abnormal samples and normal samples, the more significant the difference, the greater the weight, while limiting the sum of the weights of the time domain feature vector and the frequency domain feature vector to 1, ensuring that the two types of features occupy an equal proportion in the fusion; aligning the dimensions of the vectors after weight adjustment, corresponding the features with physiological correlation one by one, such as the number of fluctuations in the time domain and the main frequency in the frequency domain, the rise time in the time domain and the main frequency bandwidth in the frequency domain, integrating by element-by-element addition to form a fused feature vector containing 18 indicators, realizing the complementary correlation of time domain numerical change law and frequency domain frequency characteristics, and highlighting the comprehensive characteristics of blood pressure fluctuation.

[0031] Step 2.5, based on the fused feature vector, the most discriminant blood pressure fluctuation features are selected by feature selection calculation; the blood pressure fluctuation features are normalized to form a standardized blood pressure fluctuation feature set, specifically including: selecting four types of sample groups including normal blood pressure, prehypertension, stage 1 to 2 hypertension and history of previous cardiovascular events, calculating the ratio of the variance between the four types of sample groups and the within-group variance for each feature to quantify the discriminability of the feature, the larger the ratio, the stronger the feature's ability to identify different health states; sorting the features from high to low according to the discriminability, testing a 20% to 40% feature retention ratio, and finding that a 30% ratio has both stable discriminability and maximum redundancy reduction to reduce computational complexity, so the top 30% features are retained; using linear transformation method for normalization processing, first, the maximum value, minimum value and 95% quantile value of each feature after screening are calculated, extreme feature values exceeding the 95% quantile value are replaced with the 95% quantile value to avoid distortion of the normalization result by extreme values, and then the feature values are mapped to a unified numerical interval of 0 to 1 by subtracting the minimum value from the feature value and dividing by the difference between the maximum value and the minimum value; finally, a standardized blood pressure fluctuation feature set is formed to ensure that each feature has a unified basis for comparison.

[0032] In the embodiment of the present application, by segmenting the blood pressure signal in the time sequence physiological data, independent signal segments focusing on the blood pressure change rule in different time periods are obtained, providing accurate units for subsequent detailed processing; random noise interference is applied to each signal segment, enhancing the robustness of the blood pressure signal features and making them more consistent with the characteristics of clinical actual data; through time domain and frequency domain feature analysis, the blood pressure fluctuation characteristics are comprehensively captured from two dimensions of numerical change and frequency distribution, avoiding feature omission; the time domain and frequency domain feature vectors are fused at the feature level, integrating multi-dimensional information to form a more representative fused feature vector; core discriminant features are selected through feature selection and normalized processing, eliminating redundant information and eliminating dimension differences, and finally forming a standardized blood pressure fluctuation feature set, which not only reduces the computational complexity, but also provides reliable and standardized core basis for image data processing and multi-modal feature fusion.

[0033] In a preferred embodiment of the present application, step 3 can include: Step 3.1, according to the standardized blood pressure fluctuation feature set, determining the injection parameters of structural noise, specifically including: in-depth analysis of the core indicators in the standardized blood pressure fluctuation feature set, including the fluctuation amplitude standard deviation of systolic and diastolic blood pressure, the main frequency distribution interval, the extreme value change range within 2 minutes, etc.; establishing a quantitative mapping relationship between these indicators and the structural noise injection parameters, the structural noise intensity parameter is increased by 0.5% for every 5mmHg increase in systolic blood pressure fluctuation amplitude standard deviation, and the noise spatial density parameter is increased by 1 point per square millimeter when the fluctuation main frequency is in the 0.15 to 0.4Hz interval; referring to the correlation data of blood pressure fluctuation amplitude and micro deformation of fundus blood vessels in clinical research, setting the noise intensity reference range to 2% of the original image gray value dynamic range, and increasing to 4% when the fluctuation amplitude exceeds 1.5 times the clinical average, the noise shape adopts irregular point distribution simulating the fine wrinkles of blood vessel wall, the gray influence range of a single noise point is controlled within 3*3 pixels, and finally the intensity, density and morphology three-dimensional injection parameter combination dynamically matched with the blood pressure fluctuation feature is formed.

[0034] Step 3.2: Based on the injection parameters, structural noise processing is performed on the fundus image in the image data to obtain the processed fundus image. Specifically, this includes: based on the determined injection parameters, the vascular region and background region of the fundus image are first distinguished by the vascular contour mask obtained in the previous preprocessing; structural noise with the lower limit of the parameters is applied to the vascular region, and noise with the upper limit of the parameters is applied to the background region to avoid excessive interference with vascular edge recognition; noise application adopts a pixel-by-pixel random triggering mechanism, and the triggering probability is negatively correlated with the vascular density of the region, with the triggering probability reduced by 30% in densely vascular areas; the gray value change of each pixel is checked in real time to ensure that the gray value after processing is maintained within the reasonable range of 0 to 255 in clinical images, and the gray value change of a single pixel does not exceed 10% of the original value; after processing, the structural clarity score of the optic disc and macula is calculated, with a full score of 10 points, and a score of not less than 8 points is considered qualified, retaining the processed fundus image with complete key anatomical structures and continuously distinguishable vascular contours.

[0035] Step 3.3 involves segmenting the vascular network topology from the processed fundus image. Specifically, this includes: enhancing the contrast of the processed fundus image; using local histogram equalization to divide the image into 16×16 sub-regions; adjusting the grayscale distribution within each sub-region to increase the grayscale difference between the blood vessels and surrounding tissues to 1.2 to 1.5 times the original value; segmenting the blood vessels using a region growing method, selecting points with typical grayscale values ​​between 50 and 100 and at least 3 pixels in their 8-neighborhood with a grayscale difference of less than 10 as seed points; limiting the grayscale difference between adjacent pixels to no more than 15 during the growing process; stopping when no new pixels are added after 3 consecutive iterations; performing morphological processing on the preliminary segmentation results, using 3×3 circular structuring elements for one expansion to fill small breaks, followed by one erosion to remove isolated noise points, preserving the continuity of the main trunk and the integrity of the branches of the blood vessels; finally, extracting the coordinates of the bifurcation and intersection nodes of the blood vessels, recording the length and orientation angle of the blood vessel segments between nodes, and constructing a complete vascular network topology including node type and edge attributes.

[0036] Step 3.4: Dynamically locate three feature anchor points on the vascular network topology. The anatomical locations of these three feature anchor points include the upper quadrant reference point of the optic disc edge, the core point of the bifurcation of the main temporal vessels in the macula, and the center point of the arteriovenous crossing impression. Specifically, this involves: identifying the optic disc region by combining the circular morphological features with the grayscale distribution (the grayscale value of the optic disc is usually 20 to 30 lower than that of the surrounding tissue); outlining the optic disc edge using edge detection; establishing a polar coordinate system with the optic disc center as the origin, where 0° represents the horizontal temporal side; dividing the system into four quadrants clockwise, with the upper quadrant ranging from 90° to 180°; selecting the point furthest from the optic disc center and located on the main vessel course within this range as the upper quadrant reference point of the optic disc edge; and shifting the reference point temporally from the optic disc center. The approximate range of the macular region is determined by measuring 3 to 4 mm and offsetting downwards by 0.5 mm. Within this region, the area with the lowest gray value is identified as the fovea of ​​the macula. The main blood vessels with a diameter greater than 100 μm in the temporal direction are traced, and the location where the blood vessel first bifurcates with a bifurcation angle greater than 60° is found. The bifurcation apex is taken as the core point of the bifurcation of the main blood vessel in the temporal direction of the macular region. Based on the gray value of the blood vessels, arteries are usually 10 to 15 brighter than veins. In terms of diameter, the diameter of an artery is about 2 / 3 that of a vein. Arteries and veins are distinguished, and the intersection nodes are traversed to identify the location where the vein is compressed by the artery and the diameter narrows by more than 20%. The center point of the narrowest point is taken as the center point of the arteriovenous intersection impression. After localization, the coordinates of each point are fine-tuned in combination with the extreme value period of blood pressure fluctuation characteristics to ensure consistency with the correlation of blood pressure fluctuations.

[0037] Step 3.5: Construct a feature response map based on three feature anchor points, i.e., construct a dynamic response region. Specifically, this includes: using the three feature anchor points as core references, calculating the straight-line distance L1 between the upper quadrant reference point of the optic disc edge and the core point of the bifurcation of the temporal main blood vessel in the macula, the distance L2 between the bifurcation core point and the center point of the arteriovenous crossing impression, and the distance L3 between the center point of the impression and the optic disc reference point, using the triangle formed by these three sides as the base region; based on the maximum range ΔP in the blood pressure fluctuation characteristics, i.e., the difference between the peak systolic blood pressure and the trough diastolic blood pressure, expanding the base region outward by an expansion distance = 0.01mm / mmHg × ΔP, the larger the range, the wider the expansion range, ensuring coverage of vascular areas significantly affected by blood pressure fluctuations; within the expanded region, retaining all vascular branches with a diameter greater than 50μm that are directly or indirectly connected to the three anchor points, and removing terminal vessels with a diameter less than 30μm and non-vascular background tissue, forming a closed polygon boundary by connecting the endpoints of the outermost vessels, constructing a feature response map, i.e., a dynamic response region, with clear boundaries and focusing on key vascular structures.

[0038] Step 3.6: Extract vascular morphology feature parameters from the feature response map and optimize these parameters to form the final fundus vascular morphology features. Specifically, this includes: extracting core vascular morphology parameters from the dynamic response region, including the average diameter of the main trunk vessels (obtained by averaging three uniformly distributed points), the angle between branch vessels and the main trunk (calculated using two vessel orientation vectors), vessel tortuosity (the ratio of the actual vessel length to the straight-line distance between the start and end points), the number of vessel branches per unit area (vessel density), and the vessel wall thickness (the average distance from the vessel edge to the center). The extracted parameters were screened for outliers, referencing clinically normal ranges such as arterial diameter 100 to 200 μm and venous diameter 150 to 250 μm, and values ​​exceeding the range by ±30% were removed. The parameter sequence was smoothed using a moving average method with 5 consecutive measurement points to reduce fluctuations caused by noise. The correlation coefficient between each parameter and the standardized blood pressure fluctuation characteristic set index was calculated, and parameters with an absolute correlation coefficient greater than 0.3 were retained, while redundant indicators with weak correlation were removed. Finally, a fundus vascular morphology feature that accurately reflects the correlation between blood pressure fluctuations and fundus vascular morphology was formed.

[0039] In this embodiment of the invention, through the collaborative design of each step, the dynamic correlation between image data and blood pressure fluctuation features is effectively realized, significantly improving the accuracy and specificity of fundus vascular morphology features. Structural noise injection parameters are determined based on standardized blood pressure fluctuation features, enabling noise processing to accurately simulate the physiological impact of blood pressure fluctuations on fundus vessels, avoiding interference from indiscriminate noise in feature extraction. Parameter-based processing of fundus images enhances the adaptability of features to real clinical scenarios while preserving the integrity of key anatomical structures. Segmenting the vascular network topology provides a clear structural basis for feature anchor point localization, ensuring the fundamental reliability of vascular morphology analysis. Dynamically locating feature anchor points at three specific anatomical locations accurately captures key vascular regions closely related to blood pressure fluctuations, strengthening the specificity of the physiological correlation. Constructing a dynamic response region focuses on vascular structures significantly affected by blood pressure fluctuations, avoiding interference from irrelevant regions. Extracting and optimizing vascular morphology parameters, eliminating redundant information, and retaining features strongly correlated with blood pressure fluctuations, the resulting fundus vascular morphology features truly reflect the dynamic correlation between the two, providing high-quality image feature support for multimodal feature fusion and risk assessment.

[0040] In a preferred embodiment of the present invention, step 4 above may include: Step 4.1: Using the three feature anchor points in the feature response map as generators, construct a planar point set; based on the generators, spatially partition the dynamic response region using a planar point set Voronoi diagram construction calculation method to obtain multiple Voronoi units. Specifically, this includes: accurately recording the two-dimensional coordinates of the three feature anchor points based on the pixel coordinate system of the fundus image, with coordinate values ​​accurate to a single pixel, forming a planar point set containing three core generators; using the polygonal boundary of the dynamic response region as the spatial range, partitioning all pixels within the region using the planar point set Voronoi diagram construction logic; firstly, traversing each pixel within the region and calculating the Euclidean distance from that pixel to the three generators. The distance is obtained by taking the square root of the sum of the squares of the differences in pixel coordinates. For each pixel, the distances to the three generators are compared, and the pixel is assigned to the region corresponding to the nearest generator. If the distances from a pixel to two generators are equal, the pixel is assigned to the region of the adjacent pixel that has already been assigned to, ensuring continuous boundaries. All pixels belonging to the same generator naturally form a closed polygonal unit. The boundaries of these units are formed by connecting points that are equidistant from two generators, ultimately forming multiple Voronoi units that cover the entire dynamic response region, with non-overlapping and complete boundaries. Each unit is centered on the corresponding feature anchor point and completely wraps the vascular structure around the anchor point that is affected by blood pressure fluctuations.

[0041] Step 4.2 defines each Voronoi unit as an analysis unit. Centered on the corresponding generator, the spatial geometric features of each analysis unit are calculated. Specifically, this includes: defining each Voronoi unit as an analysis unit, and using the coordinates of the corresponding generator as the core reference, calculating multiple spatial geometric features; when calculating the area, only the number of effective pixels belonging to the vascular network within the unit is counted, excluding background pixels to ensure the area reflects the actual vascular distribution range; when calculating the perimeter, pixel-by-pixel tracing is performed along the unit boundary, recording the connection count of adjacent boundary pixels, and accumulating the perimeter value reflecting the shape complexity of the unit by the pixel side length of each continuous boundary segment; the centroid coordinates are determined by averaging the x-axis and y-axis coordinates of all effective vascular pixels within the unit, clearly showing the offset direction of the centroid relative to the generator; the shortest bounding rectangle is determined by fitting the farthest point of the unit edge, and the ratio of the long side to the short side of the rectangle is calculated to reflect the stretching degree of the unit; the straight-line distance from the centroid to the generator is calculated using coordinate differences to quantify the spatial positional relationship between the two. All features revolve around the generator, accurately reflecting the spatial attributes of the analysis unit relative to the core anchor point.

[0042] Step 4.3 involves performing a correlation analysis between the spatial geometric features and the temporal changes of blood pressure fluctuation features, calculating the deformation parameters of each analysis unit relative to its generator; based on the deformation parameters of all analysis units relative to their respective generators, a deformation feature matrix is ​​constructed, specifically including: extracting time-series data matching the time range of the dynamic response region from the standardized blood pressure fluctuation feature set, dividing the time nodes at 2-minute intervals, with each node containing four indicators: systolic blood pressure fluctuation amplitude, diastolic blood pressure fluctuation amplitude, dominant frequency value, and frequency of extreme values, forming a complete blood pressure fluctuation time series; and then analyzing the spatial geometric features (area, perimeter, centroid offset, etc.) of each analysis unit according to... Geometric feature sequences at the same time points are compiled and compared with blood pressure fluctuation time series on a time-by-time basis. The similarity of the changing trends of the two within the same time period is calculated. The higher the trend synchronization, the stronger the correlation. Based on the correlation results, deformation parameters are quantified: the area expansion rate is the ratio of the current area to the baseline area when blood pressure is stable, the centroid offset is the difference between the x-axis and y-axis of the centroid coordinates relative to the generator, and the perimeter change amplitude is the percentage of the difference between the current perimeter and the baseline perimeter. All analysis units are arranged in order as matrix rows, and each deformation parameter is arranged as matrix columns to construct a deformation feature matrix with complete data and corresponding rows and columns. The matrix element values ​​intuitively reflect the degree of deformation of each unit.

[0043] Step 4.4 involves eigenvalue decomposition of the deformation feature matrix to extract the main deformation components. Based on the correlation between the main deformation components and the spatial distribution of the generator elements, feature compensation coefficients are calculated. Specifically, this includes: preprocessing the deformation feature matrix by scaling each parameter using the mean and standard deviation across all cells to eliminate differences in the magnitude of different deformation parameters and ensure the matrix data is on the same analytical scale; calculating the covariance matrix of the processed matrix; and obtaining eigenvalues ​​and eigenvectors by decomposing the covariance matrix. Larger eigenvalues ​​indicate a more pronounced deformation pattern reflected by the corresponding eigenvectors. The cumulative contribution of the eigenvalues ​​is then selected. The top three eigenvectors with a rate of over 85% are selected as the main deformation components, which can cover most of the deformation information. The numerical distribution of each main deformation component in the analysis units corresponding to the three generators is analyzed, and the variance ratio of the component in each unit is calculated. The higher the variance ratio, the stronger the influence of the component on the unit. This determines the spatial correlation weight between the component and the generator. The correlation weight and the corresponding eigenvalue are normalized and weighted to obtain the feature compensation coefficient of each generator. The coefficient value is controlled between 0.8 and 1.2 to ensure the calibration effectiveness and conform to the physiological range of vascular deformation.

[0044] Step 4.5: Dynamically calibrate the three quantitative indicators of fundus vascular morphology using feature compensation coefficients to obtain the calibrated fundus vascular morphology characteristics. Specifically, this includes: identifying the three core quantitative indicators of fundus vascular morphology characteristics: the average diameter of the main vessel trunk (take the average of three evenly distributed measurement points); the average angle between branch vessels and the main trunk (take the average angle of all first-order branches); and the average vessel tortuosity (take the average tortuosity of the main vessel). Based on the physiological functions of the three generators, establish the correspondence between the compensation coefficients and the indicators: the coefficient at the upper quadrant reference point of the optic disc edge corresponds to the average diameter of the main vessel trunk, as the main vessel trunk in the optic disc region is most directly affected by blood pressure. The coefficient of the core point of the bifurcation of the main blood vessel in the temporal region of the macula corresponds to the average angle of the branches, as the bifurcation angle is sensitive to blood pressure fluctuations; the coefficient of the center point of the arteriovenous crossing impression corresponds to the average tortuosity of the blood vessel, as the tortuosity of the blood vessel at the crossing is easily affected by blood pressure; when the compensation coefficient is greater than 1, the corresponding index is adjusted upward according to the coefficient ratio to match the state of vasodilation when blood pressure rises; when the coefficient is less than 1, the index is adjusted downward according to the coefficient ratio to match the state of vasoconstriction when blood pressure falls; after calibration, each index is checked to see if it is within the clinical normal range, such as arterial diameter of 100 to 200 μm, to ensure that the data is true and reliable, and finally the calibrated fundus vascular morphology characteristics that are dynamically matched with blood pressure fluctuations are obtained.

[0045] Step 4.6 involves multi-scale fusion of the calibrated fundus vascular morphology features and standardized blood pressure fluctuation features. Based on the spatial distribution characteristics of the generators, a first correlation feature containing dynamic information about the relationship between blood vessels and blood pressure is generated through feature importance assessment. Specifically, this includes: using a multi-scale fusion strategy to divide the calibrated fundus vascular morphology features into fine-scale and coarse-scale features: fine-scale features include local details such as the diameter of branch vessels less than 100 μm and the angle of minute branches less than 30°, reflecting the immediate response of blood vessels to short-term blood pressure fluctuations; coarse-scale features include the overall density of the vascular network, the length of blood vessels per unit area, and the main... The orientation and angle range of blood vessels relative to the horizontal line, among other global attributes, reflect the adaptive changes of blood vessels to long-term blood pressure trends. Standardized blood pressure fluctuation characteristics are broken down by time scale: short-term fluctuation characteristics from 2 minutes to 1 hour (e.g., high-frequency fluctuation amplitude corresponding to fine-scale vascular morphology), and long-term trend characteristics from 1 hour to 12 hours (e.g., average blood pressure change rate corresponding to coarse-scale vascular morphology). The importance of features is assessed based on the clinical significance of the generators: the optic disc, as the origin of blood vessels, receives a 20% increase in feature weight; the macula, closely associated with visual function, receives a 15% increase in weight; and the arteriovenous crossing area, a high-risk area for lesions, receives a 15% increase in weight. Vascular morphology features and blood pressure fluctuation characteristics at different scales are integrated with a weighted average across dimensions. The correlation between each pair of features is calculated, retaining features with a high correlation coefficient greater than 0.3, and removing duplicate or weakly correlated information. Finally, a first associated feature containing dynamic correlation information between vascular morphology and blood pressure fluctuation is generated.

[0046] In this embodiment of the invention, through the progressive and coordinated linkage of each step, the problem of isolated extraction of vascular morphology and blood pressure fluctuation features and lack of dynamic correlation in the prior art is effectively solved, and the accuracy and correlation of the features are significantly improved. Using three feature anchor points as generators, a Voronoi diagram construction method is employed to spatially divide the dynamic response region, accurately focusing on key vascular areas affected by blood pressure fluctuations and providing a highly targeted unit carrier for analysis. The spatial geometric features of each analysis unit are calculated, laying a quantitative foundation for deformation analysis. By correlating spatial geometric features with temporal changes in blood pressure fluctuations, a deformation feature matrix is ​​constructed, achieving a deep binding between vascular spatial morphology and dynamic blood pressure changes. The main deformation components are extracted and feature compensation coefficients are calculated, providing an effective basis for vascular morphology feature calibration. The core quantitative indicators of fundus vascular morphology are dynamically calibrated using compensation coefficients, ensuring that vascular morphology features truly reflect the impact of blood pressure fluctuations and avoiding biases caused by static extraction. Finally, through multi-scale fusion and feature importance assessment, the calibrated vascular morphology features and standardized blood pressure fluctuation features are integrated to generate the first associated feature, comprehensively capturing the dynamic correlation information between blood vessels and blood pressure, providing a high-quality, strongly correlated core foundation for further multimodal feature fusion.

[0047] In a preferred embodiment of the present invention, step 5 above may include: Step 5.1: Based on the first association feature, determine the intensity parameter of semantic noise processing. Specifically, this includes: extracting core indicators from the first association feature that reflect the dynamic correlation between vascular morphology and blood pressure fluctuations, including the synchronization rate between the change amplitude of the main vascular diameter and the fluctuation amplitude of systolic blood pressure, the correlation between the trend of vascular tortuosity change and the time of occurrence of blood pressure extreme values, and the matching degree between the change of branch angle and the change of blood pressure dominant frequency; normalizing these indicators and mapping them to the range of 0 to 1 by (indicator value - minimum indicator value) / (maximum indicator value - minimum indicator value); setting the baseline range of the semantic noise processing intensity parameter to 0.1 to 0.3. When the mean of the three core indicators exceeds 0.7, the intensity parameter is increased to 0.3; when the mean is lower than 0.3, it is decreased to 0.1; when the mean is between 0.3 and 0.7, it is adjusted according to the linear relationship of (mean - 0.3) × 0.5 + 0.1 to ensure that the intensity parameter and the first association feature are dynamically adapted to each other.

[0048] Step 5.2: Based on the intensity parameters, perform semantic perturbation processing on the medication records in the text data to extract medication frequency features and medication pattern features from the perturbated medication records. Specifically, this includes: performing semantic perturbation processing on the medication records in the text data based on the determined intensity parameters; the medication records contain information such as the generic name of the drug, brand name, single dose, daily frequency of administration, specific administration time, start and end dates of administration, and dose adjustment records; the perturbation only targets non-critical information, such as replacing twice daily with twice a day, replacing after breakfast with after breakfast, and perturbing consecutive... Continued use is replaced with continuous use, and the replacement ratio is strictly equal to the intensity parameter value; key information such as drug name, dosage value, start and end dates remain unchanged to avoid changing the core meaning of the record; extract medication frequency characteristics from the perturbed record: count the total number of times each type of drug is taken each week, calculate the proportion of each drug's frequency to the total weekly frequency, and record the percentage of days taken within the week; extract medication pattern characteristics: count the longest consecutive days of medication, calculate the standard deviation of the interval between two adjacent doses, and record the number of dose adjustments per month to reflect the stability and adjustment trend of medication behavior.

[0049] Step 5.3: Based on the medication frequency characteristics and medication pattern characteristics, construct a medication behavior pattern vector. This includes: quantifying the extracted medication frequency characteristics and medication pattern characteristics. In the medication frequency characteristics, the total number of times taken per week is recorded according to the actual count; the frequency percentage of a single drug is the number of times that drug is taken per week divided by the total number of times per week; and the percentage of medication days taken per week is the actual number of medication days taken per week divided by 7. In the medication pattern characteristics, the longest consecutive medication days is the maximum number of days taken without interruption in the record; the standard deviation of the medication interval is obtained by calculating the standard deviation of the time difference between all two adjacent medications in hours; and the number of dose adjustments per month is the total number of dose changes in the current month, with any period less than a month converted to the actual number of days. Arrange these quantified features in a fixed order: first, arrange the medication frequency characteristics, such as total number of times, single drug percentage, and medication day percentage; then arrange the medication pattern characteristics, such as consecutive days, standard deviation of the interval, and number of adjustments, forming a medication behavior pattern vector containing 6 indicators. Each element in the vector corresponds to a specific quantified value of one feature.

[0050] Step 5.4: Perform a correlation analysis between the medication behavior pattern vector and the first associated feature to obtain the correlation analysis results. Based on the correlation analysis results, calculate the feature fusion weight coefficients, specifically including: aligning the medication behavior pattern vector and the first associated feature in the time dimension, using a 7-day sliding window to ensure that the data of both cover the same time period within each window; calculating the correlation between each dimension of the medication behavior pattern vector and the corresponding dimension of the first associated feature within each window, using the Pearson correlation coefficient as a measure. The larger the absolute value of the correlation coefficient, the stronger the association between the two in that dimension; assigning feature fusion weight coefficients based on the correlation analysis results, assigning higher weights to features with an absolute correlation coefficient greater than 0.4, medium weights to those between 0.2 and 0.4, and lower weights to those less than 0.2; the total weight coefficient is 1, with the weight ratio of the medication behavior pattern vector controlled between 30% and 40%, and the weight ratio of the first associated feature controlled between 60% and 70%. The specific ratio is fine-tuned based on the average of the overall correlation between the two. The higher the correlation, the closer the weight ratio of the medication behavior pattern vector is to 40%.

[0051] Step 5.5: Based on the feature fusion weight coefficient, the medication behavior pattern vector and the first associated feature are weighted and fused. Through feature dimensionality reduction, a second associated feature containing information on the correlation between vascular morphology, blood pressure fluctuation, and medication behavior is generated. Specifically, this includes: weighting and integrating the medication behavior pattern vector and the first associated feature dimension by dimension according to the feature fusion weight coefficient. The fusion value for each dimension is the value of the corresponding dimension of the medication behavior pattern vector multiplied by its weight, plus the value of the corresponding dimension of the first associated feature multiplied by its weight; performing dimensionality reduction on the integrated preliminary fused feature, using principal component analysis to calculate the contribution rate of each principal component, and retaining the top principal components with a cumulative contribution rate of 90%; during the dimensionality reduction process, core features are screened through loading coefficients, prioritizing the retention of components with an absolute value of loading coefficient greater than 0.5. These components typically reflect the interaction between vascular morphology changes and medication adjustments, and the correlation between blood pressure fluctuation trends and medication frequency; the final generated second associated feature dimension is controlled between 40% and 50% of the preliminary fused feature, which reduces data redundancy while fully preserving the cross-correlation information among vascular morphology, blood pressure fluctuation, and medication behavior.

[0052] In this embodiment of the invention, through the collaborative design of each step, deep fusion of text data and the first associated feature is achieved, effectively improving the comprehensiveness and relevance of the feature. The intensity parameter of semantic noise processing is determined based on the first associated feature to ensure that semantic perturbation conforms to the correlation characteristics between vascular morphology and blood pressure fluctuations, avoiding indiscriminate perturbation that could damage the authenticity of the feature. Semantic perturbation processing of medication records is performed according to the intensity parameter, enhancing the adaptability of medication features to actual clinical scenarios and accurately extracting the core features of medication frequency and patterns. A medication behavior pattern vector is constructed, transforming abstract medication behavior into a quantifiable feature form, providing a standardized data foundation for multimodal fusion. Correlation analysis clarifies the correlation strength between medication behavior and vascular morphology and blood pressure fluctuation features, and scientifically reasonable fusion weight coefficients are assigned accordingly, making the fusion process more targeted. Finally, a second associated feature is generated through weighted fusion and feature dimensionality reduction, effectively integrating the correlation information of vascular morphology, blood pressure fluctuations, and medication behavior. This solves the problem of isolated multimodal features in existing technologies, providing more comprehensive, accurate, and clinically valuable feature support for cardiovascular event risk assessment.

[0053] In a preferred embodiment of the present invention, step 6 above may include: Step 6.1 involves standardizing the second association features to obtain a standardized multimodal feature set. This includes: collecting second association feature data from all clinical samples, covering vascular morphology change indicators, blood pressure fluctuation trend parameters, medication behavior patterns, and cross-correlation information among these three features; classifying features by dimension, statistically analyzing all sample feature values ​​under each dimension, and calculating the mean and standard deviation for that dimension; performing standardization transformation on each feature value of each sample by subtracting the mean of the corresponding dimension from the feature value and then dividing by the standard deviation of the corresponding dimension, so that the overall distribution of each feature after processing satisfies a mean of 0 and a standard deviation of 1; considering the possible random measurement errors in clinical data, extreme values ​​exceeding the mean plus or minus 3 times the standard deviation are truncated and replaced with boundary values ​​of the mean plus or minus 3 times the standard deviation to avoid extreme values ​​disrupting the overall distribution pattern of the features; after standardization, all features are within the same numerical range, eliminating fusion bias caused by dimensional differences, and forming a standardized multimodal feature set that can be directly used for subsequent analysis.

[0054] Step 6.2, based on the feature importance assessment logic, calculate the weight coefficients of each feature in the multimodal feature set. Specifically, this includes: assessing the feature importance of the standardized multimodal feature set by first dividing the data into a training part and a validation part in a 7:3 ratio. The training part is used to calculate the feature contribution, and the validation part is used to verify the stability of the results. Construct multiple decision tree structures, each generated by recursively splitting nodes, with a maximum depth limit of 10 layers to avoid excessive focus on details. When splitting at each node, calculate the contribution of each feature to reducing node impurity. Impurity is measured using the Gini index, specifically the Gini index of the node before splitting minus the weighted sum of the Gini indices of the two child nodes after splitting. Accumulate the total contribution of each feature in all decision trees, and then normalize by dividing the total contribution by the sum of the total contributions of all features to obtain the weight coefficient of each feature. Features with higher weight coefficients usually include the interaction features between changes in arteriovenous cross impressions and systolic blood pressure fluctuations, and the correlation features between long-term medication adherence and changes in the diameter of the main blood vessel, etc. These features have a more critical impact on cardiovascular event risk assessment.

[0055] Step 6.3: Based on the weighting coefficients, perform weighted fusion processing on the multimodal feature set to obtain the weighted fused features. Specifically, this includes: multiplying each feature value in the standardized multimodal feature set by its corresponding weighting coefficient to obtain the weighted feature value of each feature, ensuring that features with high importance occupy a larger proportion in the fusion result; accumulating all weighted feature values ​​dimension by dimension to form a comprehensive feature vector, i.e., the weighted fused features; during the fusion process, strictly preserving the clinical correlation logic between each feature and cardiovascular risk, for example, superimposing the weighted values ​​of high-frequency blood pressure fluctuation features and abnormal vascular branch angle features to highlight the cumulative effect of their combined effect on vascular damage; at the same time, controlling the contribution ratio of each feature through weight normalization to avoid a single feature, such as extreme blood pressure values, from excessively dominating the fusion result, ensuring that the fused features can comprehensively integrate multimodal information and accurately focus on core risk factors.

[0056] Step 6.4 involves extracting deep-level correlated features from the weighted fused features using a multi-level feature extraction network. Specifically, this includes constructing a multi-level feature extraction network containing an input layer, three hidden layers, and an output layer. The input layer dimension is consistent with the weighted fused feature dimension, while the output layer dimension is set to 32 dimensions to balance information retention and computational efficiency. The first hidden layer has 128 neurons and uses the ReLU activation function for non-linear transformation, primarily extracting basic correlated features, such as the correspondence between medication frequency and mean blood pressure, and the synchronous change pattern between blood vessel diameter and peak systolic blood pressure. The second hidden layer has 64 neurons and introduces a dropout mechanism with a dropout rate of 0.2, randomly and temporarily discarding some neurons to prevent overfitting due to excessive reliance on local features, while further compressing the feature dimension to capture... Intermediate correlation patterns are established, such as the lagged correlation between long-term regular medication and improvement in vascular tortuosity, and the dynamic influence of blood pressure fluctuation amplitude and vascular branch density. The third hidden layer is configured with 32 neurons, focusing on core risk correlations and extracting deep correlation features, such as the synergistic changes between abnormal blood pressure diurnal rhythm and macular vascular lesions after medication adjustment, and the interactive influence between aggravated arteriovenous cross-entropy and insufficient antihypertensive drug dosage. The network training uses the cross-entropy loss function to measure prediction error, and uses the Adam optimizer to dynamically adjust the weights of each layer. The learning rate is initially set to 0.001 and decays by 10% every 10 iterations. Iterative training stops when the loss function does not decrease for 5 consecutive cycles on the validation set. Finally, 32-dimensional deep correlation features are obtained from the output layer. These features can reflect the potential complex correlations between multimodal data that are difficult to observe directly.

[0057] Step 6.5, based on deep correlation features, calculates the risk probability of cardiovascular events using a risk assessment classification mechanism. This includes: first, organizing over 100,000 clinical follow-up data points, dividing the data into positive and negative groups based on whether cardiovascular events will occur within the next 5 years; both groups contain dimensional information corresponding to the deep correlation features; dividing each dimension of the deep correlation features into intervals, for example, vascular morphology-related features are divided into mild, moderate, and severe intervals based on the degree of abnormality; blood pressure fluctuation features are divided into stable, moderate, and severe intervals based on amplitude; and medication behavior features are divided into good, average, and poor intervals based on adherence; traversing all interval combinations of feature dimensions, counting the number of samples for each combination in the positive and negative groups, and calculating the event occurrence rate for that combination, i.e., the number of samples in the positive group divided by the total number of samples in that combination, as the basic risk probability for that combination. For combinations with fewer than 50 samples, the mean probability of adjacent similar combinations is used to supplement the risk probability, ensuring that each combination has a reliable probability reference. A mapping table between feature combinations and risk probabilities is established, matching the deep correlation features of the current sample to the corresponding interval combinations, and extracting the base risk probability from the mapping table. The extracted base risk probability is calibrated based on the overall cardiovascular event incidence rate of the full follow-up data. If the combination probability is higher than the overall incidence rate, it is appropriately lowered; if it is lower, it is appropriately higher, ensuring that the probability distribution conforms to the actual clinical incidence pattern. The probability discrimination of high-risk combinations is emphasized. For example, the combination of severe vascular abnormalities, drastic blood pressure fluctuations, and poor medication adherence has a significantly higher calibrated risk probability than the combination of single-dimensional abnormalities. Finally, the accurate quantification of cardiovascular event risk probability for the current sample is obtained, with the probability value controlled between 0 and 1.

[0058] Step 6.6: Based on the risk probability and a preset risk threshold, determine the final cardiovascular event risk level. Specifically, this includes: referencing cardiovascular disease prevention and treatment guidelines from various state cardiology societies and the National Heart Association, and combining incidence data from large-scale population cohort studies in China, setting three risk thresholds: a low-risk threshold of 0.3 and a medium-risk threshold of 0.7. That is, a risk probability less than 0.3 is considered low risk, between 0.3 and 0.7 is considered medium risk, and greater than 0.7 is considered high risk. For special populations, such as elderly people over 65 years old with diabetes and other high-risk groups, the low-risk threshold is lowered to 0.25, and the medium-risk threshold is lowered to 0.65 to more sensitively identify potential risks. A risk level indicates a low likelihood of cardiovascular events, and it is recommended to have an annual fundus vascular examination, dynamic blood pressure monitoring, and medication adherence assessment. A medium-risk level indicates the presence of clear risk factors, and it is recommended to have a follow-up examination every 3 to 6 months, adjust the dosage of antihypertensive or lipid-lowering drugs, and strengthen lifestyle interventions such as limiting sodium intake and engaging in 150 minutes of moderate-intensity exercise per week. A high-risk level indicates a high risk of developing an event in the short term, requiring immediate comprehensive examinations such as coronary CT and carotid ultrasound to develop a personalized intervention plan, such as strengthening antihypertensive treatment to control systolic blood pressure below 130 mmHg and assessing the necessity of vascular interventional treatment, ultimately forming a risk level result that is intuitive and has clear clinical operational guidance.

[0059] In this embodiment of the invention, a comprehensive design encompassing standardization, weighted fusion, deep extraction, and risk grading achieves efficient integration of multimodal features and accurate risk assessment, effectively enhancing the reliability and practicality of cardiovascular event risk judgment. Standardization of the second associated feature eliminates dimensional differences between different modal features, laying a unified and comparable foundation for fusion and assessment. Calculating the weight coefficients of each feature based on a feature importance assessment algorithm accurately identifies core features that play a crucial role in risk assessment, avoiding redundant feature interference. Weighted fusion based on the weight coefficients strengthens the contribution of key information, achieving effective integration of multimodal features. Deep associated features are mined through a multi-level feature extraction network, overcoming the limitations of surface features and capturing the potential complex correlations between vascular morphology, blood pressure fluctuations, and medication behavior. A risk assessment classification mechanism is used to calculate risk probabilities, transforming deep associated features into quantifiable risk indicators, improving the objectivity of the assessment. Combining preset risk thresholds to determine the final risk level makes the assessment results more intuitive and aligned with clinical application scenarios, providing an intuitive, comprehensive, and practical decision-making basis for early warning and intervention of cardiovascular events.

[0060] like Figure 2 As shown, embodiments of the present invention also provide a multimodal health data feature extraction system, including: The acquisition module is used to acquire multimodal health datasets, which include time-series physiological data, image data, and text data. The processing module is used to process time-series physiological data with random noise in order to extract blood pressure fluctuation features; The localization module is used to process structural noise in image data based on blood pressure fluctuation characteristics in order to extract fundus vascular morphology features. During the extraction of fundus vascular morphology features, three feature anchor points are dynamically located on the fundus image corresponding to the image data. The anatomical locations of the three feature anchor points are the reference point in the upper quadrant of the optic disc edge, the core point of the bifurcation of the main vascular trunk in the temporal region of the macula, and the center point of the arteriovenous crossing impression. A dynamic response region is constructed based on the three feature anchor points. The calibration module is used to divide the dynamic response region into structural units to form several analysis units; a feature compensation coefficient is formed based on the deformation characteristics of the analysis units; three quantitative indicators of fundus vascular morphology are calibrated based on the feature compensation coefficient to obtain calibrated fundus vascular morphology features; the blood pressure fluctuation features are fused with the calibrated fundus vascular morphology features to generate the first associated feature. The fusion module is used to extract medication adherence features based on a first association feature by performing semantic noise processing on the text data; to fuse the medication adherence feature with the first association feature to generate a second association feature; and to perform multimodal fusion processing on the second association feature to obtain the cardiovascular event risk level. The above description is a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method of multi-modal health data feature extraction, characterized in that, The method comprises: acquiring a multi-modal health data set containing time-series physiological data, image data and text data; random noise processing is performed on the time-series physiological data to extract blood pressure fluctuation features; based on the blood pressure fluctuation features, structural noise processing is performed on the image data to extract fundus blood vessel morphology features; in the process of extracting the fundus blood vessel morphology features, three feature anchor points are dynamically positioned on the fundus image corresponding to the image data; the anatomical positions of the three feature anchor points are, respectively, the upper quadrant reference point on the edge of the optic disc, the core point of the main blood vessel bifurcation on the temporal side of the macular area, and the center point of the arteriovenous crossing pressure trace; a dynamic response region is constructed based on the three feature anchor points; the dynamic response region is divided into structural units to form a plurality of analysis units; a feature compensation coefficient is formed according to the deformation characteristics of the analysis units; the three quantitative indicators of the fundus blood vessel morphology features are calibrated based on the feature compensation coefficient to obtain calibrated fundus blood vessel morphology features; the blood pressure fluctuation features and the calibrated fundus blood vessel morphology features are fused to generate first associated features; based on the first associated features, semantic noise processing is performed on the text data to extract medication adherence features; the medication adherence features and the first associated features are fused to generate second associated features; the second associated features are subjected to multi-modal fusion processing to obtain a cardiovascular event risk level.

2. The multi-modal health data feature extraction method of claim 1, wherein, Acquiring a multi-modal health data set containing time-series physiological data, image data and text data, comprising: collecting time-series physiological signals from medical monitoring devices, acquiring image data generated by fundus imaging devices, and extracting relevant text information from electronic health record systems; signal quality assessment is performed on the time-series physiological signals to obtain quality-assessed time-series physiological signals; image quality verification is performed on the image data to form quality-verified image data; structured preprocessing is performed on the text information to obtain preprocessed text information; the quality-assessed time-series physiological signals, the quality-verified image data and the preprocessed text information are standardized and integrated to obtain standardized and integrated data; the standardized and integrated data is organized into a multi-modal health data set in a unified format.

3. The multi-modal health data feature extraction method of claim 2, wherein, Random noise processing is performed on the time-series physiological data to extract blood pressure fluctuation features, comprising: segmenting the blood pressure signals in the time-series physiological data to obtain a plurality of blood pressure signal segments; random noise interference is applied to each blood pressure signal segment to generate noise-enhanced blood pressure signal segments; time-domain feature analysis is performed on each noise-enhanced blood pressure signal segment to calculate a time-domain feature vector; frequency-domain feature analysis is performed on each noise-enhanced blood pressure signal segment to extract a frequency-domain feature vector; feature-level fusion is performed on the time-domain feature vector and the frequency-domain feature vector to form a fused feature vector; based on the fused feature vector, feature selection calculation is performed to screen out the most discriminative blood pressure fluctuation features; the blood pressure fluctuation features are normalized to form a standardized blood pressure fluctuation feature set.

4. The multi-modal health data feature extraction method of claim 3, wherein, Based on the blood pressure fluctuation characteristics, the image data is processed for structural noise to extract the fundus vascular morphology characteristics; in the process of extracting the fundus vascular morphology characteristics, three feature anchor points are dynamically positioned on the fundus image corresponding to the image data; the anatomical positions of the three feature anchor points are the upper quadrant reference point on the edge of the optic disc, the core point of the main trunk blood vessel bifurcation on the temporal side of the macular area, and the center point of the arteriovenous crossing indentation; Based on the three feature anchor points, a dynamic response region is constructed, including: According to the standardized blood pressure fluctuation characteristics set, the injection parameters of the structural noise are determined; Based on the injection parameters, the fundus image in the image data is processed for structural noise to obtain the processed fundus image; The vascular network topology structure is segmented from the processed fundus image; The three feature anchor points are dynamically positioned on the vascular network topology structure, and the anatomical positions of the three feature anchor points include the upper quadrant reference point on the edge of the optic disc, the core point of the main trunk blood vessel bifurcation on the temporal side of the macular area, and the center point of the arteriovenous crossing indentation; Based on the three feature anchor points, a feature response atlas is constructed, that is, a dynamic response region is constructed; The vascular morphology characteristic parameters are extracted from the feature response atlas, and the vascular morphology characteristic parameters are optimized to form the final fundus vascular morphology characteristics.

5. The multi-modal health data feature extraction method of claim 4, wherein, The blood pressure fluctuation characteristics and the calibrated fundus vascular morphology characteristics are fused to generate a first associated feature, including: Taking the three feature anchor points in the feature response atlas as generation elements, a plane point set is constructed; based on the generation elements, the dynamic response region is spatially divided by adopting a plane point set Voronoi diagram construction calculation method to obtain a plurality of Voronoi units; Each Voronoi unit is defined as an analysis unit, and the spatial geometric characteristics of each analysis unit are calculated with the corresponding generation element as the center; The spatial geometric characteristics are associated with the time sequence change of the blood pressure fluctuation characteristics, and the deformation parameters of each analysis unit relative to the generation element are calculated; based on the deformation parameters of all analysis units relative to their respective generation elements, a deformation feature matrix is constructed; The deformation feature matrix is subjected to eigenvalue decomposition to extract the main deformation component; according to the correlation between the main deformation component and the spatial distribution of the generation element, a feature compensation coefficient is calculated; The three quantitative indicators of the fundus vascular morphology characteristics are dynamically calibrated by the feature compensation coefficient to obtain the calibrated fundus vascular morphology characteristics; The calibrated fundus vascular morphology characteristics and the standardized blood pressure fluctuation characteristics are fused at multiple scales, and based on the spatial distribution characteristics of the generation elements, the first associated feature containing the dynamic correlation information of blood vessels and blood pressure is generated through feature importance evaluation.

6. The multi-modal health data feature extraction method of claim 5, wherein, Based on the first associated feature, the text data is processed for semantic noise to extract medication adherence characteristics; The medication adherence characteristics and the first associated feature are fused to generate a second associated feature, including: Based on the first associated feature, the intensity parameter of the semantic noise processing is determined; According to the intensity parameter, the medication records in the text data are processed for semantic disturbance to extract the medication frequency characteristics and the medication regularity characteristics from the disturbed medication records; Based on the medication frequency characteristics and the medication regularity characteristics, a medication behavior mode vector is constructed; The medication behavior pattern vector is associated with the first associated feature for correlation degree analysis to obtain a correlation degree analysis result; and a feature fusion weight coefficient is calculated according to the correlation degree analysis result; Based on the feature fusion weight coefficient, the medication behavior pattern vector and the first associated feature are weighted and fused, and through feature dimension reduction processing, a second associated feature containing blood vessel morphology, blood pressure fluctuation and medication behavior association information is generated.

7. The multi-modal health data feature extraction method of claim 6, wherein, The second associated feature is processed through multi-modal fusion to obtain a cardiovascular event risk level, including: The second associated feature is processed through feature standardization to obtain a standardized multi-modal feature set; Based on the feature importance evaluation logic, a weight coefficient of each feature in the multi-modal feature set is calculated; According to the weight coefficient, the multi-modal feature set is processed through weighted fusion to obtain a weighted fused feature; Through a multi-level feature extraction network, deep layer association features are extracted from the weighted fused feature; Based on the deep layer association features, a risk probability of the cardiovascular event is calculated through a risk assessment classification mechanism; According to the risk probability, in combination with a preset risk threshold, a final cardiovascular event risk level is determined.

8. A multi-modal health data feature extraction system, the system implementing the method of any one of claims 1 to 7, characterized in that, It includes: An acquisition module is configured to acquire a multi-modal health data set, the multi-modal health data set containing time series physiological data, image data and text data; A processing module is configured to process the time series physiological data through random noise processing to extract blood pressure fluctuation features; A positioning module is configured to process the image data through structural noise processing based on the blood pressure fluctuation features to extract fundus blood vessel morphology features; in the process of extracting the fundus blood vessel morphology features, three feature anchor points are dynamically positioned on a fundus image corresponding to the image data; the anatomical positions of the three feature anchor points are, respectively, an upper quadrant reference point on the edge of the optic disc, a core point of a main blood vessel bifurcation on the temporal side of the macular region, and a center point of a dynamic-static vein intersection pressure trace; a dynamic response region is constructed based on the three feature anchor points; A calibration module is configured to divide the dynamic response region into a plurality of analysis units to form a feature compensation coefficient based on the deformation characteristics of the analysis units; The three quantitative indicators of the fundus blood vessel morphology features are calibrated based on the feature compensation coefficient to obtain calibrated fundus blood vessel morphology features; The blood pressure fluctuation features and the calibrated fundus blood vessel morphology features are fused to generate a first associated feature; A fusion module is configured to extract medication adherence features by processing the text data through semantic noise processing based on the first associated feature; The medication adherence features and the first associated feature are fused to generate a second associated feature; The second associated feature is processed through multi-modal fusion to obtain a cardiovascular event risk level.

9. A computing device, comprising: It includes: One or more processors; A storage device is configured to store one or more programs, when the one or more programs are executed by the one or more processors, so that the one or more processors implement the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a program, which is executed by the processor to implement the method of any one of claims 1 to 7.

Images