A multi-modal cardiovascular proxy phenotype unified representation method with medical semantic constraints

Abstract

The application provides a multi-modal cardiovascular proxy phenotype unified representation method, system and electronic equipment with medical semantic constraints. A multi-modal data object containing object identification, device parameters, time attributes, quality states and data bloodlines is established; quality control, scale calibration, time mapping and medical semantic analysis are performed on each mode; the analysis results are mapped to microcirculation function, circulation load, autonomic nervous-electrophysiology, oxygenation state, vascular structure, data quality, time evolution and individual baseline dimension object; a medical semantic constraint network is established to perform consistency check, conflict identification and credibility correction; a unified representation object containing medical semantics, quality, credibility, time, baseline, version and evidence coordinates is generated, and incremental update is performed on new data.

Description

Technical Field

[0001] This invention relates to the fields of biomedical informatics, medical knowledge representation, multimodal medical data processing, medical artificial intelligence, and digital healthcare, and particularly to a unified representation method, system, and electronic device for multimodal cardiovascular surrogate phenotypes with medical semantic constraints.

[0002] This invention addresses cardiovascular-related medical data originating from different medical acquisition devices, sampling frequencies, data formats, and time points. It establishes a computable data object, a unified medical semantic label, a medical dimension object, a medical semantic constraint network, and a unified representation object of proxy phenotypes. Through quality and credibility propagation, incremental updates, and version management, the unified representation object can serve as a standardized basic representation layer between different downstream medical artificial intelligence systems. Background Technology

[0003] With the development of portable testing devices, wearable sensors, medical imaging equipment, and telemedicine platforms, a variety of cardiovascular-related medical data can be obtained in clinical, physical examination, primary care screening, and home monitoring scenarios, including fingertip microcirculation images or videos, electrocardiogram signals, pulse wave signals, blood pressure data, blood oxygen data, and retinal images. These data reflect different levels of physiological states, such as peripheral microcirculation perfusion, cardiac electrical activity, autonomic nervous system regulation, circulatory pressure load, tissue oxygenation, and vascular structure. They are characterized by relatively low acquisition costs, repeatability, and suitability for continuous follow-up.

[0004] Different modalities of medical data exhibit significant differences in data format, sampling frequency, time scale, dimensions, device output format, quality assessment methods, and medical meaning. Fingertip microcirculation data are typically represented as images, videos, or perfusion sequences; electrocardiograms and pulse waves are presented as high-frequency time series; blood pressure can be a single measurement, a home measurement, or a dynamic monitoring sequence; blood oxygen saturation can be a continuous waveform or event statistics; and retinal images are structural images with a lower sampling frequency. The lack of a unified object association and time mapping method between different data types easily leads to incorrect correspondences between personnel, tasks, devices, and time windows.

[0005] Existing multimodal medical data processing solutions typically establish separate input formats, feature names, fusion models, and output structures for a specific task. For example, ECG models are used for rhythm recognition, blood pressure and pulse wave models for circulatory load assessment, and medical imaging models for structural abnormality detection. Even when different models use relevant data from the same subject, their internal representations are often not directly interchangeable, requiring each downstream task to repeatedly perform data access, preprocessing, feature extraction, and representation learning.

[0006] Some techniques employ feature concatenation, latent vectors, or neural network embedding for multimodal fusion. However, these representations typically depend on the specific model structure, training data, and target task, and the vector positions lack stable and reusable medical meaning across systems. When the model structure or task changes, the original representation is difficult to transfer directly; when results are abnormal, it is also difficult to determine whether the abnormality originates from microcirculation, circulatory load, autonomic nervous system, electrophysiology, oxygenation, vascular structure, or data quality.

[0007] Typical fusion vectors usually only store numerical values, often neglecting to retain information such as source modality, device parameters, acquisition time, data quality, reliability, individual baselines, constraint conflicts, and historical versions. This results in data of varying quality being processed downstream under the same conditions, and new data may overwrite old results or trigger duplicate calculations of all historical data, hindering traceability, incremental maintenance, and cross-system sharing in long-term monitoring.

[0008] Different sources may yield contradictory results for the same medical condition. For example, abnormal blood oxygen saturation may be accompanied by hypoperfusion or probe detachment, blood pressure may rise significantly in the short term but heart rate and other circulatory parameters may remain stable, or structural imaging and short-term functional data may fall within different effective time windows. Without constraints on physiological correlation, temporal continuity, and quality reliability propagation, the system may easily interpret acquisition anomalies or temporal mismatches as changes in the actual medical condition.

[0009] Furthermore, existing medical terminology standards or data exchange formats mainly address field naming and communication format issues, and typically cannot organize multimodal raw data, preprocessed data, medical semantics, medical dimensions, quality, reliability, individual baselines, conflict states, and version relationships into a unified proxy table representation object that can be directly invoked by multiple computing systems.

[0010] Therefore, there is a need for a unified representation technology that can establish object associations and data lineages for multimodal cardiovascular medical data, perform quality control, equipment calibration, time mapping and medical semantic parsing, construct medical dimension objects with fixed medical meanings, and use medical semantic constraint networks for conflict identification and credibility propagation. This technology should also support local updates of new data, version chain maintenance and unified interface output, thereby reducing redundant calculations and improving data consistency across devices, time and applications. Summary of the Invention

[0011] I. Purpose of the Invention The purpose of this invention is to provide a unified representation method, system, and electronic device for multimodal cardiovascular surrogate phenotypes with medical semantic constraints, in order to solve the problems in the prior art such as unstable association of multimodal medical data objects, difficulty in transmitting data quality and time status, inconsistent medical meaning, lack of interpretability of fusion representation, and the need to repeatedly process all historical data for new data.

[0012] This invention transforms heterogeneous medical data into a unified representation object that includes medical semantics, dimensional expression, quality, credibility, time, individual baseline, version, and source evidence by establishing standardized multimodal medical data objects, unified medical semantic labels, medical dimensional objects, and a medical semantic constraint network. Through local dimensional updates and version chain maintenance, the unified representation object can be reused by different downstream systems without having to repeatedly parse all the original data.

[0013] II. Technical Solution To achieve the above objectives, the present invention provides a unified representation method for multimodal cardiovascular surrogate phenotypes with medical semantic constraints, including steps S101 to S107.

[0014] Step S101: Acquire multimodal cardiovascular-related medical data of the subjects, establish subject identifiers, data collection task identifiers, and modal data objects; perform integrity checks, quality checks, unit and scale standardization, equipment calibration, and unified time axis mapping on each modality to generate standardized multimodal medical data objects with quality scores, validity masks, unified time windows, and data lineage.

[0015] Step S102: For each standardized multimodal medical data object, call the corresponding medical semantic parser to convert the raw data, preprocessed data or phenotypic parameters into unified medical semantic labels, and record the label code, semantic category, semantic state or standardized value, source modality, applicable time interval, label credibility and parsing rule version for each label.

[0016] Step S103: Based on the versioned medical dimension dictionary, the unified medical semantic labels are mapped to the microcirculation function dimension, circulatory load dimension, autonomic nervous-electrophysiological dimension, oxygenation status dimension, vascular structure dimension, data quality dimension, temporal evolution dimension, and individual baseline dimension; the same parameters of different devices or different modalities are scaled uniformly, and the source evidence, quality, credibility, time, baseline offset, and missing status are retained for each medical dimension.

[0017] Step S104: Using medical dimension objects as nodes and physiological association, temporal continuity and quality reliability propagation relationships as edges, establish a medical semantic constraint network. Calculate the degree of conflict and generate constraint verification results based on the applicable conditions, allowable range and constraint strength of the relationship edges.

[0018] Step S105: Calculate the tag credibility, medical dimension credibility, and overall object credibility level by level based on source quality, semantic parsing credibility, time synchronization degree, source coverage, and conflict degree. When the conflict exceeds the threshold, reduce the corresponding credibility or generate a pending review status.

[0019] Step S106: Encode the set of dimension representations constrained by medical semantics into a unified representation space, generate a proxy table unified representation object containing object header, medical semantics and medical dimension subjects, quality and credibility fields, time and baseline fields, constraint and conflict fields, version fields and data lineage fields, and output it through a unified interface.

[0020] Step S107: When new medical data arrives, locate the associated objects and affected medical dimensions, update only the affected medical dimensions and related quality, credibility, source evidence and constraint subgraphs, generate a new version with parent version association, and maintain the collection time order, version generation order and historical evolution trajectory.

[0021] This invention also provides a unified representation system for multimodal cardiovascular surrogate phenotypes with medical semantic constraints, including a data object creation module, a quality and time processing module, a medical semantic parsing module, a medical dimension expression module, a medical semantic constraint module, a credibility propagation module, a unified representation space construction module, a surrogate phenotype object management module, an incremental update and versioning module, and a unified interface module. The modules exchange standardized multimodal medical data objects, unified medical semantic tags, medical dimension objects, constraint verification results, and unified surrogate phenotype representation objects.

[0022] The present invention also provides an electronic device, including a processor, a memory, and a communication module. The memory stores a computer program, medical semantic parsing rules, a unified semantic dictionary, a medical dimension dictionary, device mapping and calibration rules, and medical semantic constraint rules. When the computer program is executed by the processor, it implements the above-mentioned methods.

[0023] III. Beneficial Effects (1) By establishing a multimodal medical data object that is interconnected with subjects, collection tasks, modalities, devices, time and data lineage, different data forms and different device outputs can establish a definite correspondence within the same object system, and can distinguish different states such as not collected, unusable quality and incomparable time.

[0024] (2) By performing quality inspection, equipment calibration, unit and scale unification and time mapping on different modalities, the data entering the unified representation space has comparable dimensions and effective time range, reducing the impact of equipment differences and time mismatch on subsequent calculations.

[0025] (3) By versioning the unified semantic dictionary and medical dimension dictionary, the device private fields and model outputs are converted into semantic labels and medical dimension objects with stable medical meanings, thus avoiding the unified representation from degenerating into ordinary latent vectors that lack medical meaning.

[0026] (4) By using the medical semantic constraint network, the consistency of the relationship between microcirculation function, circulatory load, autonomic nervous system-electrophysiology, oxygenation status, vascular structure, data quality, time evolution and individual baseline can be verified, and multi-source conflicts can be identified and the cause and source evidence of the conflict can be preserved.

[0027] (5) By establishing a hierarchical propagation mechanism from modal quality to label credibility, medical dimension credibility and overall object credibility, downstream systems can call data according to quality and credibility conditions, avoiding the indiscriminate use of low-quality data and high-quality data.

[0028] (6) By setting medical semantics, medical dimensions, quality, credibility, time, individual baseline, version and source evidence coordinates, data from the same subject at different times, on different devices and in different scenarios can form a unified representation object that is structurally consistent, computable and traceable.

[0029] (7) By adding data-driven local dimension updates and constraint subgraph verification, it is not necessary to reprocess all historical raw data corresponding to the unaffected medical dimensions, thus reducing the overhead of repeated calculations and storage access in continuous monitoring scenarios.

[0030] (8) Through version chain, late data insertion, version rollback, individual baseline version and interface mode version management, the unified representation object can support long-term follow-up, cross-platform sharing and compatible calls of downstream systems with different versions. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating the overall process of the unified representation method for multimodal cardiovascular surrogate phenotypes with medical semantic constraints according to the present invention. Figure 2 This is a schematic diagram illustrating the establishment, quality processing, scale standardization, and time mapping of multimodal medical data objects in this invention; Figure 3 This is a schematic diagram of the medical semantic parser and unified medical semantic tag generation framework of the present invention; Figure 4 This is a schematic diagram of the construction and structure of the medical dimension object in this invention; Figure 5 This is a schematic diagram illustrating the medical semantic constraint relationships, conflict handling, and credibility propagation of the present invention; Figure 6 This is a schematic diagram of the unified representation space, proxy table, unified representation object, and unified interface structure of the present invention; Figure 7 This invention adds a flowchart for medical data-driven incremental updates, version management, and historical trajectory maintenance. Figure 8 This is a structural diagram of the multimodal cardiovascular surrogate phenotype unified representation system of the present invention.

[0032] The connecting lines in the attached diagram only represent data transfer, semantic mapping, constraint propagation, version association, or interface call relationships. Modules can be merged, split, configured in parallel, or distributed across different computing nodes depending on deployment conditions. Detailed Implementation

[0033] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments. The embodiments are used to illustrate possible implementations of the present invention and are not intended to limit the scope of protection. Those skilled in the art can adjust the fields, thresholds, rules, number of medical dimensions, constraint forms, and deployment locations according to data modalities, device capabilities, application scenarios, and storage architectures.

[0034] The "unified representation object of surrogate phenotype" as referred to in this invention refers to a standardized, computable data object formed from multimodal medical data after object association, quality and time processing, medical semantic parsing, medical dimension expression, and medical semantic constraints. This object describes the cardiovascular-related medical status of the subject within a specific time or time interval and carries information on source, quality, reliability, individual baseline, conflict, and version; the representation object itself does not directly provide a disease diagnosis conclusion.

[0035] The "unified representation space" referred to in this invention is not limited to a two-dimensional or three-dimensional geometric space in a physical sense, but rather is a computable representation system composed of multiple coordinates with definite medical meanings, fixed field definitions, and version relationships. Each coordinate can be implemented using numerical values, vectors, levels, categories, graph nodes, key-value fields, or combinations thereof.

[0036] The "medical semantic constraints" referred to in this invention include verification or restriction relationships established based on human physiological relationships, medical knowledge rules, temporal continuity, data quality, and source consistency. Medical semantic constraints can be used as hard constraints to exclude states that do not meet the conditions, or as soft constraints to calculate the degree of conflict, correct credibility, or generate a sign to be reviewed.

[0037] Example 1: Multimodal Medical Data Objects, Quality Scoring, and Time Mapping like Figure 1 and Figure 2 As shown, the system first establishes a master index for the subjects. The master index can use internal patient identifiers from medical institutions, research subject codes, or desensitized unique identifiers. A collection task identifier is generated for each collection task, and a modality data object identifier is generated for each modality within the same task to prevent erroneous association of data from different personnel, at different times, or in different batches.

[0038] In outpatient settings, the system can receive fingertip microcirculation video, single-lead or twelve-lead ECG, pulse wave, clinic blood pressure or ambulatory blood pressure segments, blood oxygenation sequence, and optional retinal images. Retinal images can be color fundus photographs, laser-scanned fundus images, optical coherence tomography (OCT) angiography, or other images that support vascular structural dimensions.

[0039] Each modal data object records at least the data modality, file location, device manufacturer and model, sampling rate, resolution, unit, start and end times of data acquisition, time zone, checksum, integrity status, synchronization status, and quality status. The data object includes a raw data layer, a preprocessed data layer, and a phenotypic parameter layer, and records the reference relationships between the raw file, processing results, and parameter results through a data lineage field.

[0040] For continuous physiological signals, the system converts the device's internal time to a unified reference time axis. When multiple devices cannot achieve hardware synchronization, software alignment can be performed based on the acquired event markers, ECG R peaks, pulse peaks, manual trigger markers, or network timestamps. The system saves the offset, allowable error, and effective time window of each modality relative to the reference time axis.

[0041] Modal quality scores are obtained by weighting multiple quality indicators for the corresponding modality. Finger-tip microcirculation quality indicators may include sharpness, motion artifacts, analyzable area proportion, perfusion visibility, and blood flow tracking stability; ECG or pulse wave quality indicators may include contact status, baseline drift, saturation, motion artifacts, waveform periodicity, and effective duration; blood pressure quality indicators may include effective reading proportion, posture, cuff status, and aberrant jumps; blood oxygen quality indicators may include hypoperfusion, probe detachment, aberrant fluctuations, and data continuity; retinal image quality indicators may include exposure, focus, optic disc or macular visibility, and vascular segmentability. The overall quality score for the m-th modality is calculated according to formula (1) in claim 2.

[0042] Among them, s m,r β represents the r-th normalized quality sub-index of the m-th mode; m,r R represents the corresponding non-negative quality weight; m q represents the number of quality sub-indices contained in the m-th mode; m Let represent the overall quality score of the m-th mode, satisfying 0 ≤ q. m ≤1. The quality sub-indicators include one or more of the following: effective data ratio, signal-to-noise ratio, motion or contact artifacts, device status, data continuity, and time synchronization status.

[0043] The system distinguishes between physical missing data, unavailability of quality data, and incomparability of time data: an "uncollected" state is generated when a certain modality is not collected; a "quality unqualified" state is generated when data exists but its quality does not meet the minimum requirements; and a "time mismatch" state is generated when data quality is acceptable but the acquisition time exceeds the synchronization tolerance. All of these states are stored as data object attributes and validity masks.

[0044] Example 2: Medical Semantic Parsing and Unified Tag Generation like Figure 3 As shown, the medical semantic parsing module establishes corresponding parsers for each modality. The parsers can employ threshold rules, statistical rules, knowledge graph mapping, feature mapping models, machine learning models, or combinations thereof. Private fields from different devices and local terms from different institutions are converted into unified label codes through a terminology mapping table.

[0045] For fingertip microcirculation data, the system can obtain the visible area of ​​microvessels, vascular density, perfusion area, red blood cell movement trajectory, flow velocity distribution, perfusion uniformity, pulsation state and pre- and post-stimulation response, and generate unified medical semantic tags such as reduced perfusion, uneven perfusion, unstable blood flow or weakened reactivity.

[0046] For ECG data, the system identifies heart rate, RR interval, rhythm type, QRS duration, QT or corrected QT interval, heart rate variability, and abnormal rhythm events; for pulse waves, the system identifies pulse peak, rise time, reflected wave, waveform stability, and compliance-related parameters; and generates semantic labels for heart rate, rhythm, electrical conduction, autonomic regulation, and peripheral circulation.

[0047] For blood pressure data, the system distinguishes between single blood pressure readings, home blood pressure readings, and ambulatory blood pressure monitoring, generating semantic labels for systolic blood pressure, diastolic blood pressure, pulse pressure, mean arterial pressure, diurnal variability, nocturnal decrease, morning peak, and circulatory load. For blood oxygenation data, the system generates semantic labels for mean oxygenation, minimum oxygenation, duration of hypoxia, deoxygenation events, oxygenation fluctuations, and recovery status.

[0048] For retinal images, the system can identify arteriovenous diameter, arteriovenous ratio, vascular density, fractal dimension, tortuosity, bifurcation structure, and microvascular lesions. When retinal images are used as an enhanced modality, their absence does not prevent the construction of other medical dimension objects. The credibility of the r-th medical semantic tag of the m-th modality is calculated according to formula (2) in claim 3.

[0049] Where cm,r is the credibility of the r-th medical semantic tag in the m-th modality; qm is the quality score of the corresponding modality; pm,r is the parsing confidence of the medical semantic parser for the tag; and hm,r is the time validity factor of the tag within a unified time window.

[0050] Example 3: Construction and Scale Unification of Medical Dimensional Objects like Figure 4 As shown, the medical dimension expression module maps unified medical semantic tags into microcirculation function dimension, circulatory load dimension, autonomic nervous system-electrophysiology dimension, oxygenation state dimension, vascular structure dimension, data quality dimension, temporal evolution dimension, and individual baseline dimension based on a versioned medical dimension dictionary. A medical dimension can be supported by a single modality or by multiple modalities.

[0051] For similar parameters generated by different devices, unit conversion and device calibration are performed first. Parameters that can be directly converted are converted using unit conversion and linear or nonlinear calibration; parameters that cannot be directly converted are expressed using range mapping, distributed calibration, semantic hierarchy, or standardized intervals. Both the device mapping table and calibration function record version numbers.

[0052] Each medical dimension object includes at least a dimension identifier, dimension value or dimension vector, source modality set, source evidence set, quality status, dimension credibility, time index, individual baseline offset, constraint label, dimension dictionary version, and missing mask. When multiple sources support the same medical dimension, the system retains the contributions and evidence citations of each source. The medical dimension vector and individual baseline offset are calculated according to formulas (3) and (4) in claim 4, respectively.

[0053] Where dk is the vector of the k-th medical dimension; Mk is the number of candidate sources supporting the k-th medical dimension; zk,m is the normalized source vector after the m-th source is mapped to the k-th medical dimension; ak,m is the source availability mask; wk,m is the non-negative source weight; and ε is a positive number to avoid zero denominators. The system simultaneously saves the contribution identifier and evidence citation of each source, and the individual baseline offset is calculated using the following formula:

[0054] Example 4: Medical semantic constraints, degree of conflict, and credibility propagation like Figure 5 As shown, the medical semantic constraint module uses each medical dimension object as a node and physiological associations, temporal relationships, and quality propagation relationships as relation edges. Each relation edge records the relation type, direction, applicable conditions, allowed range, constraint strength, and rule version.

[0055] Coupling constraints are set between the microcirculation function dimension and the circulatory load dimension; regulatory constraints are set between the circulatory load dimension and the autonomic nervous system-electrophysiology dimension; oxygen supply-perfusion constraints are set between the oxygenation status dimension and the microcirculation function dimension; structure-function constraints are set between the vascular structure dimension and the microcirculation function dimension; continuity constraints are set between the time evolution dimension and the individual baseline dimension; and the data quality dimension acts on other medical dimensions through credibility propagation relationships.

[0056] The above constraints do not require different medical dimensions to maintain a fixed proportion, but rather specify a reasonable combination or allowable range under a given time scale, quality conditions, and application scenario. The low-frequency vascular structure dimension and the high-frequency microcirculation function dimension can use different recent update times to avoid misinterpreting short-term functional fluctuations as structural changes. The edge conflict value and the degree of conflict of medical dimensions are calculated according to formulas (5) and (6) in claim 5, respectively.

[0057] Where χe is the edge conflict value of the e-th semantic relation edge; ge is the relation function; Ωe is the allowed interval; βe is the constraint strength; Dist(·) represents the normalized distance from the output of the relation function to the allowed interval; E(k) represents the set of semantic relation edges connected to the k-th medical dimension; ηk,e represents the non-negative aggregation weight of the corresponding edge.

[0058] When the degree of conflict is below the first conflict threshold, the original dimensional state is retained; when the degree of conflict is between the first and second conflict thresholds, a mild conflict label is generated and the credibility is reduced; when the degree of conflict exceeds the second conflict threshold, a verification label is generated, and evidence from each source is retained without forcing a unique state to be output. The credibility of the conflict-corrected medical dimension and the overall credibility of the proxy phenotype unified representation object are calculated according to formulas (7) and (8) in claim 6, respectively.

[0059] Where ck0 is the basic credibility of the k-th medical dimension; ck is the credibility of the medical dimension after conflict correction; λk is the conflict penalty coefficient; χk is the conflict degree of the k-th medical dimension; C is the overall credibility of the proxy phenotype unified representation object; ρk is the non-negative importance weight of the k-th medical dimension; and K is the number of medical dimensions participating in the overall credibility calculation.

[0060] Example 5: Unified Representation Space and Proxy Tables for Unified Object Representation like Figure 6 As shown, the unified representation space construction module receives a set of medical dimension representations that have been constrained and verified, and generates medical semantic coordinates, medical dimension coordinates, quality coordinates, credibility coordinates, time coordinates, individual baseline coordinates, version coordinates, and source evidence coordinates.

[0061] Medical semantic coordinates record the unified medical semantic category and status; medical dimension coordinates record the values, vectors, or levels of each medical dimension; quality coordinates record source availability, quality scores, and artifact status; credibility coordinates record label-level, dimension-level, and object-level credibility; time coordinates record the effective time window and time sequence; individual baseline coordinates record the current state's offset relative to the baseline; version coordinates record object generation, updates, and parent-child version relationships; and source evidence coordinates record data lineage and supporting evidence.

[0062] A unified representation space can be implemented using relational data structures, graph data structures, vector or tensor data structures, key-value or document data structures, and combinations thereof. In the vector implementation, each vector position is fixed by a versioned medical dimension dictionary, and the training model cannot arbitrarily change its medical meaning; when adding a new dimension, a new dictionary version is generated and the old and new mappings are saved.

[0063] The proxy table uniformly represents objects including an object header, medical semantics and medical dimensions, quality and reliability fields, time and individual baseline fields, constraint and conflict fields, version fields, and data lineage fields. Objects can be generated according to collection tasks, time points, continuous monitoring windows, or clinical events.

[0064] The unified representation object of this invention does not directly limit cardiovascular magnetic resonance surrogate indicators, disease categories, or examination recommendations. Downstream systems can read the corresponding medical dimensions and credibility to perform specific tasks, but the downstream output does not change the basic coordinate definition, thereby achieving separation between the representation layer and the application layer.

[0065] Example 6: Unified Interface and Pattern Version Management The proxy phenotypic object management module provides creation, reading, updating, retrieval, comparison, and archiving operations. Retrieval criteria can include subject, time range, medical semantic tags, medical dimension status, confidence threshold, and version status. Comparison operations can compare changes in medical dimensions between different versions of the same subject, or compare objects formed from different sources within the same time window.

[0066] The unified interface can be implemented using application programming interfaces (APIs), messaging interfaces, database views, or file exchange. Required output fields include at least the object identifier, medical dimension representation, medical dimension credibility, overall credibility, time index, and version index. Depending on the access permissions, further output can include the source modality, quality status, conflict identifier, resolution rule version, and historical reference relationships.

[0067] Both the field structure and the medical dimension dictionary are managed using schema versions. When a new field is added or a medical dimension is expanded, the interface returns the new schema version and saves the field mapping between the old and new schemas; the old version calls the system to read the original fields according to the old schema, while the new version calls the system to read the expanded fields.

[0068] In one downstream call example, the longitudinal monitoring system reads consecutive versions of the microcirculation function dimension, circulatory load dimension, and oxygenation status dimension, as well as the reliability of each version, to form a quality-weighted time trajectory. This eliminates the need for the system to re-access and parse the original microcirculation video, ECG files, and blood oxygenation sequences.

[0069] Example 7: Adding incremental data updates, version management, and historical tracking like Figure 7 As shown, when new medical data arrives, the incremental update and versioning module searches for related objects based on the subject identifier, collection task identifier, collection time, and modality identifier. If no object exists, an initial version is created; if an object exists, it is determined whether the new data is used to update the current time window, create a new time point object, add a source to an existing medical dimension, or create a new medical dimension.

[0070] After the newly added data is object-based and undergoes medical semantic parsing, the affected medical dimensions are located based on unified medical semantic labels. For example, newly added ambulatory blood pressure data typically updates the circulatory load dimension, data quality dimension, and time evolution dimension; newly added electrocardiogram data updates the autonomic nervous system-electrophysiology dimension; and newly added retinal images update the vascular structure dimension, without recalculating medical dimensions unrelated to the newly added data.

[0071] After a local dimension update, only the constraint subgraph containing the affected medical dimension is re-executed. The system adjusts the confidence of the relevant dimension and the overall confidence of the object based on the quality of the new source, conflict status, and source coverage. The values, sources, and constraint results of unaffected dimensions remain unchanged.

[0072] Each update generates a new version, recording the parent version, update time, updated data object, updated dimension, value before update, value after update, change in confidence level, and constraint results. Old versions remain read-only and are not overwritten by new versions.

[0073] The system maintains both the data collection time order and the version generation order. Late data collection times can be inserted into the middle of historical data, while its version generation time is later than the current version. If an association error or device time error is detected, the system can roll back the current valid version pointer or generate a corrected version, while retaining the original version and the reason for the operation.

[0074] Individual baselines can be selected from a single high-quality, stable version, or a reference range can be formed by multiple stable versions. When a baseline changes, a new baseline version is generated while the old baseline is retained, allowing historical comparisons to be reproduced according to the baseline used at the time.

[0075] Example 8: Cross-device and longitudinal follow-up applications This embodiment uses continuous data from the same subject in two scenarios: outpatient and home. In the outpatient setting, data on fingertip microcirculation, electrocardiogram (ECG), blood pressure, blood oxygen saturation, and retinal imaging are collected; at home, wearable ECG or pulse wave data, home blood pressure, and blood oxygen saturation are collected. The system establishes separate outpatient and home data objects and links them using the subject's master index.

[0076] Outpatient retinal images are used to establish vascular structure dimensions. Due to the low frequency of retinal image acquisition, subsequent home versions can continue to reference the most recent valid vascular structure dimensions, and record the most recent update time and validity period in the time attribute; its reliability can decay according to preset rules based on time intervals.

[0077] Different brands of blood pressure monitors output different fields and units. The system uses a device mapping table to convert them into a unified cyclic load dimension input. Different ECG devices have different sampling rates. During the semantic parsing stage, waveform analysis is performed according to the actual sampling rate, but the final output includes heart rate, rhythm, and autonomic nerve labels with the same encoding rules.

[0078] When a home blood oxygen saturation level is abnormally low and is accompanied by low perfusion or probe detachment markers, the data quality dimension reduces the reliability of that segment. The oxygen supply-perfusion constraint will not turn this single low-quality anomaly into a high-reliability hypoxia state, but will instead preserve the anomaly value, the cause of the quality issue, and the marker indicating that it needs to be retested.

[0079] When multiple consecutive high-quality time windows show an increase in the circulatory load dimension and a decrease in the microcirculatory function dimension, the unified representation object records the directional consistency, source evidence, and credibility of the two medical dimensions. Downstream longitudinal systems can perform trend analysis based on this, but trend judgments and diagnostic conclusions are not essential components of the unified representation layer of this invention.

[0080] Example 9: Modular Implementation of the System like Figure 8 As shown, the data object creation module connects to the medical acquisition equipment or data access platform to receive data, generate object identifiers, record device and time information, and establish data lineage. The quality and time processing module performs integrity checks, quality checks, unit and scale standardization, device calibration, and unified timeline mapping.

[0081] The medical semantic parsing module calls the corresponding modality parser to generate unified medical semantic tags. The medical dimension expression module organizes the tags and parameters into medical dimension objects according to the medical dimension dictionary. The medical semantic constraint module reads the constraint rule base and performs consistency checks and conflict degree calculations on the medical dimensions; the credibility propagation module calculates the overall credibility of the tags, medical dimensions, and objects.

[0082] The unified representation space construction module generates a unified representation space based on the coordinate pattern. The proxy table object management module is responsible for object generation, storage, retrieval, comparison, and output. The incremental update and versioning module listens for new data events and performs local updates, constraint subgraph verification, and version generation for the affected medical dimensions. The unified interface module outputs the required fields based on the caller's identity, permissions, and pattern version.

[0083] The modules described above can all be deployed on the medical institution's local server, or the data object creation and preliminary quality processing can be deployed on the data acquisition terminal, while the medical semantic parsing, constraint processing, and unified representation space construction can be deployed in the cloud. For scenarios with high privacy or network requirements, only the unified medical semantic tags and medical dimension objects can be uploaded without uploading the original data.

[0084] The system can be configured with a rule management unit, a version management unit, and an audit log unit. The rule management unit maintains medical semantic parsing rules, a unified semantic dictionary, a medical dimension dictionary, constraint relationships, and a device mapping table; the version management unit maintains rule versions, object versions, and interface mode versions; and the audit log unit records data access, rule changes, medical dimension updates, interface calls, and version rollbacks.

[0085] Example 10: Electronic device and computer program implementation The method of this invention can be executed by an electronic device. The electronic device includes at least one processor, a memory, a communication module, and a data bus. The memory stores a computer program, medical semantic parsing rules, a unified semantic dictionary, a medical dimension dictionary, medical semantic constraint rules, device mapping and calibration rules, and a unified representation object of proxy phenotypes.

[0086] When the processor executes a computer program, it completes the following tasks: receiving multimodal medical data, associating data objects, processing quality and time, parsing medical semantics, expressing medical dimensions, verifying medical semantic constraints, propagating credibility, constructing a unified representation space, generating proxy phenotypic unified representation objects, incremental updates, version management, and interface output.

[0087] The communication module is used to exchange multimodal medical data objects, medical dimension objects, and proxy phenotypic unified representation objects with medical acquisition equipment, hospital information systems, telemedicine platforms, home monitoring platforms, cloud servers, or downstream application systems.

[0088] Electronic devices can be servers, workstations, personal computers, mobile terminals, edge computing devices, portable medical devices, or distributed systems composed of multiple computing nodes. Different processing steps can be executed on different nodes, as long as standardized multimodal medical data objects, unified medical semantic tags, medical dimension objects, or proxy phenotypic unified representation objects can be transferred between nodes.

[0089] Computer programs can be stored on disks, solid-state drives, optical storage media, removable storage media, or other computer-readable storage media. The thresholds, scoring methods, number of coordinates, field names, and module deployment locations mentioned above are merely examples and do not alter the core technical concept of this invention: forming a traceable and incrementally updatable unified representation object through medical semantic constraints.

[0090] It should be noted that the technical features of the above embodiments can be combined with each other where there is no conflict. Any equivalent substitutions, simple transformations, or combinations made based on this specification and the accompanying drawings are implementations that can be covered by the technical solutions of this invention.

Claims

1. A unified representation method for multimodal cardiovascular surrogate phenotypes with medical semantic constraints, characterized in that, Includes the following steps: Step S101: Acquire multimodal cardiovascular-related medical data of the subjects, establish subject identifiers, data collection task identifiers, and modal data objects; perform integrity checks, quality checks, unit and scale standardization, and time mapping on each modal data to obtain standardized multimodal medical data objects with modal identifiers, device identifiers, sampling parameters, unified time windows, quality scores, validity masks, and data lineage relationships; Step S102: For each standardized multimodal medical data object, call the corresponding medical semantic parser to convert the original data, preprocessed data or phenotypic parameters of each modality into unified medical semantic labels. The medical semantic labels include label encoding, semantic category, semantic state or standardized value, source modality, applicable time interval, label credibility and parsing rule version. Step S103: Based on the preset medical dimension dictionary, the unified medical semantic label is mapped to multiple medical dimension objects. Unit conversion, device calibration, range mapping or distribution calibration are performed on the same type of parameters from different devices or different modalities. Dimension values ​​or dimension vectors, source evidence sets, data quality attributes, dimension credibility, time index, individual baseline offset, constraint labels and missing masks are associated with each medical dimension object. Step S104: Using each medical dimension object as a node and the physiological association, temporal continuity, and quality reliability propagation relationship between medical dimensions as relation edges, establish a medical semantic constraint network; based on the applicable conditions, allowable range, and constraint strength corresponding to each relation edge, perform consistency verification and conflict degree calculation on the medical dimension objects to obtain constraint verification results and conflict status. Step S105: Based on the source quality, semantic parsing credibility, time synchronization degree, source coverage and conflict degree of each medical dimension object, calculate the semantic tag credibility, medical dimension credibility and representation object credibility level by level; when the conflict degree exceeds the preset threshold, generate a conflict identifier and correct the corresponding medical dimension credibility to obtain a set of dimension representations constrained by medical semantics. Step S106: Encode the dimensional representation set into the unified representation space of the multimodal cardiovascular surrogate phenotype to generate a unified representation object of the surrogate phenotype; the unified representation space includes medical semantic coordinates, medical dimensional coordinates, quality coordinates, confidence coordinates, time coordinates, individual baseline coordinates, version coordinates and source evidence coordinates; the unified representation object of the surrogate phenotype includes object identifier, data fields corresponding to each coordinate, conflict status and data lineage information; Step S107: When new medical data is received, locate the associated proxy phenotypic unified representation object and the affected medical dimension based on the subject identifier, collection task identifier, time identifier and modality identifier of the new medical data. Only update the affected medical dimension and its associated quality attribute, credibility attribute, source evidence and constraint verification results, generate a new version proxy phenotypic unified representation object with parent version association, and output the current effective version, historical version index and update difference through the unified representation interface.

2. The unified representation method for multimodal cardiovascular surrogate phenotypes with medical semantic constraints according to claim 1, characterized in that, In step S101: The multimodal cardiovascular-related medical data includes at least two of the following: fingertip microcirculation images or videos, electrocardiogram signals, pulse wave signals, blood pressure data, and blood oxygen data, and the retinal images are used as the vascular structure enhancement modality when retinal images are obtained. For each mode, a raw data layer, a preprocessed data layer, and a phenotypic parameter layer are established. The raw data layer stores the original output and verification value of the device. The preprocessed data layer stores the data after noise reduction, artifact marking, unit conversion, device calibration, and time alignment. The phenotypic parameter layer stores the parameter values, units, time ranges, and source references calculated from the corresponding data. The reference relationship between each layer is established through the data lineage field. The acquisition time of each device is converted to a unified reference time axis, and the time offset is determined by one or more of the following: hardware synchronization marker, acquisition event marker, ECG R peak, pulse peak, manual trigger marker or network timestamp, and a unified time window is established within the preset synchronization tolerance. For the m-th mode, the modal quality score is as shown in formula (1): （1） Quality sub-indicators may include the proportion of valid data, signal-to-noise ratio, motion or contact artifacts, equipment status, data continuity, and time synchronization status. Each quality indicator and its weight can be calibrated according to the equipment and acquisition scenario. When any modality is in a state of no acquisition, acquisition failure, insufficient valid segments, unqualified quality, time mismatch, or incompatible device parameters, a corresponding status identifier and validity mask are generated respectively, and the corresponding data is not deleted to replace the status record.

3. The unified representation method for multimodal cardiovascular surrogate phenotypes with medical semantic constraints according to claim 1, characterized in that, In step S102: For fingertip microcirculation data, we analyze the semantics of microcirculation perfusion, microvessel density, perfusion uniformity, blood flow velocity, blood flow stability, vascular pulsation, and microcirculation reactivity. For electrocardiogram (ECG) signals, we analyze the semantics of heart rate, heart rhythm, electrical conduction, autonomic nervous system regulation, and abnormal electrical activity load; for pulse wave signals, we analyze the semantics of waveform periodicity, pulse wave rise time, reflected wave, peripheral perfusion, and vascular compliance. For blood pressure data, we analyze the semantics of systolic blood pressure, diastolic blood pressure, pulse pressure, mean arterial pressure, blood pressure variability, morning peak, nocturnal decline, and circulatory load. We analyze the semantics of mean oxygenation, minimum oxygenation, hypoxia load, deoxygenation events, oxygenation fluctuations, and recovery status based on blood oxygenation data. For retinal images, we analyze the semantics of arteriovenous diameter, arteriovenous ratio, vascular density, fractal dimension, vascular tortuosity, bifurcation structure, and microvascular lesions. The medical semantic parser employs threshold rules, statistical rules, knowledge graph mapping, feature mapping models, machine learning models, or combinations thereof, and maps different device fields and local terms to a versioned unified semantic dictionary. The credibility of the r-th medical semantic tag of the m-th modality is shown in formula (2): （2）。 4. The unified representation method for multimodal cardiovascular surrogate phenotypes with medical semantic constraints according to claim 1, characterized in that, In step S103, the multiple medical dimension objects include: The microcirculation function dimension is used to represent microvascular density, perfusion area, perfusion uniformity, blood flow velocity, blood flow fluctuation, vascular pulsation, and microcirculation responsiveness. The circulatory load dimension is used to represent systolic blood pressure, diastolic blood pressure, pulse pressure, mean arterial pressure, blood pressure variability, morning peak, nocturnal drop, circulatory load, and vascular compliance. The autonomic nervous system-electrophysiological dimension is used to represent heart rate, heart rate variability, rhythm stability, electrical conduction time, and abnormal electrical activity load. Oxygenation status dimension is used to represent mean blood oxygen, minimum blood oxygen, duration of hypoxia, frequency of deoxygenation events, depth of deoxygenation, oxygenation fluctuations, and recovery time. The vascular structure dimension is used to represent retinal vessel diameter, arteriovenous ratio, vessel density, vessel branches, tortuosity, fractal dimension, and structural lesions. The data quality dimension is used to represent the quality score of each modality, the proportion of valid data, the type of artifact, the synchronization status, the integrity status, and the device calibration status. The time evolution dimension is used to represent the acquisition time, sampling period, time interval relative to the baseline, direction of change, and rate of change. Individual baseline dimension is used to represent a dimensional reference value or reference range formed by one or more stable and quality-compliant historical versions; For the k-th medical dimension, the standardized source vector, source availability mask, and source weight of that medical dimension will be fused; the medical dimension vector and individual baseline offset are shown in Equation (3) and Equation (4), respectively: （3） （4）。 5. The unified representation method for multimodal cardiovascular surrogate phenotypes with medical semantic constraints according to claim 1, characterized in that, In step S104, the medical semantic constraint network includes medical dimension nodes, semantic relation edges, constraint type, constraint direction, applicable conditions, allowed range, constraint strength, and rule version; The semantic relation edges include at least: coupling constraints between the microcirculation function dimension and the circulatory load dimension, regulatory constraints between the circulatory load dimension and the autonomic nervous system-electrophysiology dimension, oxygen supply-perfusion constraints between the oxygenation state dimension and the microcirculation function dimension, structural-functional constraints between the vascular structure dimension and the microcirculation function dimension, continuity constraints between the time evolution dimension and the individual baseline dimension, and credibility propagation constraints from the data quality dimension to other medical dimensions. For the e-th semantic relation edge connecting the k-th medical dimension and the l-th medical dimension, the edge conflict value and the conflict degree of the k-th medical dimension are shown in Equations (5) and (6), respectively: （5） （6） When χ_k is greater than the first conflict threshold and less than or equal to the second conflict threshold, a mild conflict flag is generated and the credibility of the medical dimension is reduced; when χ_k is greater than the second conflict threshold, a conflict flag to be reviewed is generated, and evidence from each source is retained without forcibly generating a unique and definite state.

6. The unified representation method for multimodal cardiovascular surrogate phenotypes with medical semantic constraints according to claim 5, characterized in that, In step S105: The basic credibility of the medical dimension, c_k^0, is determined based on the modal quality score supporting the k-th medical dimension, the credibility of medical semantic tags, the source coverage, and the degree of time synchronization. The basic credibility is corrected using the degree of conflict in the k-th medical dimension. The credibility of the medical dimension is shown in formula (7): （7） The overall credibility of the proxy phenotypic unified representation object is calculated based on the credibility of each medical dimension and the importance weight of each dimension, as shown in formula (8): （8） When a medical dimension is missing but other sources support it, c_k is recalculated based on the missing mask and the coverage of the remaining sources. When a medical dimension has no valid source, it is marked as undecidable and the actual measured value is not replaced by a speculative value. When the overall credibility C is greater than or equal to the first credibility threshold, the proxy phenotype unified representation object is marked as a high credibility object; when the overall credibility C is less than the first credibility threshold but greater than or equal to the second credibility threshold, it is marked as a restricted credibility object and the reason for low credibility is output; when the overall credibility C is less than the second credibility threshold, it is marked as an object to be supplemented or to be reviewed.

7. The unified representation method for multimodal cardiovascular surrogate phenotypes with medical semantic constraints according to claim 1, characterized in that, In step S106: The proxy phenotype unified representation object includes an object header, a medical semantic and medical dimension subject, a quality and credibility field, a time and individual baseline field, a constraint and conflict field, a version field, and a data lineage field; The object header includes an object identifier, a subject identifier, a data collection task identifier, a generation time, a current version identifier, and a parent version identifier; the medical semantics and medical dimensions include a unified set of medical semantic tags, values ​​or vectors of each medical dimension, and evidence of their sources; the quality and credibility fields include a quality vector, credibility of each medical dimension, and overall credibility; the time and individual baseline fields include an effective time window, an individual baseline index, and a baseline offset; the constraint and conflict fields include constraint verification results, conflict degree, conflict identifier, and processing reason; the version field and data lineage field include a dimension dictionary version, a parsing rule version, a constraint rule version, an object pattern version, and a reference to the source data object; The unified representation space is implemented using relational data structures, graph data structures, vector data structures, tensor data structures, key-value data structures, or combinations thereof, and the medical meaning and field positions of each medical dimension are fixed through a versioned medical dimension dictionary; The unified representation interface outputs at least the object identifier, medical dimension expression, medical dimension credibility, overall credibility, time index, and version index, and selectively outputs the source modality, quality status, conflict identifier, parsing rule version, and historical reference relationship according to the calling permissions; when the object field or medical dimension is expanded, a new interface mode version is generated and the field mapping relationship between the new and old modes is saved.

8. The unified representation method for multimodal cardiovascular surrogate phenotypes with medical semantic constraints according to claim 1, characterized in that, Step S107 includes: The existing surrogate phenotypes of newly added medical data are used to determine the unified representation objects associated with them based on the subject identifier, collection task identifier, collection time, and modality identifier. When the collection time of newly added medical data falls within the current unified time window, update the corresponding object of the current time window; when the collection time of newly added medical data falls outside the current unified time window, create a new time point object; when the newly added medical data comes from a new device but supports existing medical dimensions, add source evidence of the existing medical dimensions after performing device calibration; when the newly added medical data supports medical dimensions that do not yet exist, expand the medical dimension dictionary and object schema version. Based on the unified medical semantic tags of the newly added medical data, the affected medical dimensions are located, and only the dimension values, quality attributes, credibility, baseline offset and source evidence of the affected medical dimensions are recalculated, and only the constraint subgraph containing the affected medical dimensions is re-executed. A new version is generated for each update. The new version records the parent version, update time, update source, update fields, state before update, state after update, change in credibility, and constraint verification results. The old version remains in read-only state. The data collection time order and version generation order are maintained separately to support the insertion of late data, determination of the current valid version, version rollback, generation of corrected versions, updating of individual baseline versions, and retrieval of historical evolution trajectories.

9. A unified representation system for multimodal cardiovascular surrogate phenotypes with medical semantic constraints, characterized in that, include: The data object creation module is used to acquire multimodal cardiovascular-related medical data, create subject identifiers, collection task identifiers and modal data objects, and generate the raw data layer, preprocessed data layer, phenotypic parameter layer and data lineage relationship; The quality and time processing module is used to perform integrity checks, quality checks, unit and scale standardization, equipment calibration, unified time axis mapping, and modal quality scoring, and to generate unified time windows, quality scores, and validity masks. The medical semantic parsing module is used to call the medical semantic parser of the corresponding modality, convert the data of each modality into unified medical semantic tags, and calculate the credibility of the medical semantic tags; The medical dimension expression module is used to generate medical dimension objects based on the medical dimension dictionary, including microcirculation function dimension, circulatory load dimension, autonomic nervous system-electrophysiology dimension, oxygenation status dimension, vascular structure dimension, data quality dimension, time evolution dimension, and individual baseline dimension. The medical semantic constraint module is used to establish a medical semantic constraint network and perform consistency checks, conflict degree calculations, conflict identifier generation, and constraint subgraph updates. The credibility propagation module is used to calculate the credibility of the medical dimension and the overall credibility based on source quality, credibility of medical semantic tags, time synchronization degree, source coverage and conflict degree; The unified representation space construction module is used to construct medical semantic coordinates, medical dimension coordinates, quality coordinates, credibility coordinates, time coordinates, individual baseline coordinates, version coordinates, and source evidence coordinates. The proxy table object management module is used to generate, store, retrieve, compare, and output unified representation objects of proxy tables. The incremental update and version module is used to locate the affected medical dimensions based on newly added medical data, and to perform local updates, local constraint reviews, version generation, version rollback, and maintenance of historical evolution trajectories. The unified interface module is used to output unified representation objects of proxy phenotypes to proxy phenotype generation systems, longitudinal analysis systems, risk assessment systems, medical information processing systems, or clinical support systems according to the interface mode version and calling permissions.

10. An electronic device, characterized in that, Includes processor, memory, and communication module; The memory stores computer programs, medical semantic parsing rules, a unified semantic dictionary, a medical dimension dictionary, device mapping and calibration rules, medical semantic constraint rules, and a unified representation object of proxy phenotypes; When the computer program is executed by the processor, the electronic device performs the unified representation method for multimodal cardiovascular surrogate phenotypes with medical semantic constraints as described in any one of claims 1 to 8; The processor is used to perform multimodal medical data object creation, quality and time processing, medical semantic parsing, medical dimension expression, semantic constraint verification, credibility propagation, unified representation space construction, generation of proxy phenotypic unified representation objects, and incremental updates and version management; The communication module is used to exchange multimodal medical data objects, medical dimension objects, and proxy phenotypic unified representation objects with medical data acquisition equipment, hospital information systems, telemedicine platforms, home monitoring platforms, cloud servers, or downstream application systems.

Images