Medical data standardization processing method and system based on multi-source data

Abstract

The application relates to the technical field of data processing, and discloses a medical data standardization processing method and system based on multi-source data. The method comprises the following steps: collecting multi-source medical original data and preprocessing to obtain a standardized data set, extracting time sequence metadata, performing dynamic time warping matching to obtain a time corresponding sequence, calibrating time sequence deviation to generate a unified time sequence data set; multi-modal fusion is performed to generate a fusion feature vector, the inter-modal connection strength is calculated, and the weight is optimized to obtain an enhanced feature completeness index; the high-dimensional vector is dimensionally reduced and clustered according to the index, a physiological state topology graph is constructed to generate a patient overall feature representation; and the patient overall feature representation is input into a feedforward neural network to process time sequence and modal collaborative clues, so that a complication risk assessment result is obtained. The application realizes accurate alignment of multi-source data time sequences, strengthens modal collaboration, and improves the accuracy and reliability of complication risk assessment.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for standardizing medical data processing based on multi-source data. Background Technology

[0002] Currently, the healthcare field is witnessing the large-scale accumulation of multi-source data. Electronic medical records, imaging examinations, laboratory tests, wearable device monitoring, and patient self-reports are becoming the core basis for clinical diagnosis and disease prediction. At the same time, the widespread adoption of smart chips enables wearable devices and portable medical testing instruments to collect physiological indicators at high frequencies, further enriching the sources and dimensions of medical data and posing higher technical requirements for the integration and processing of multi-source medical data.

[0003] Current mainstream medical data standardization technologies mostly target a single data source, structuring text for electronic medical records and standardizing numerical units for laboratory data. While some solutions attempt to integrate multi-source data, they only address superficial numerical range mapping and format conversion, failing to perform precise temporal alignment to address differences in recording time and acquisition granularity. Furthermore, the independent processing of each modality ignores the inherent relationships between text descriptions, instrument monitoring values, and image features, resulting in standardized datasets with significant temporal biases and fragmented modal information. Subsequent analyses based on this data are prone to information loss and distorted feature representations. This makes it difficult for downstream medical tasks, such as risk prediction and assisted diagnosis, to achieve the accuracy and robustness required for clinical application. Therefore, a standardized processing method capable of deeply integrating multi-source heterogeneous medical data across spatiotemporal and semantic dimensions is urgently needed.

[0004] Existing technologies cannot solve the problem of integrating multi-source heterogeneous medical data in terms of recording time, collection granularity, and expression form, and there is a problem that they cannot achieve effective correspondence and collaborative representation of medical data of different modalities. Summary of the Invention

[0005] This invention provides a method and system for standardizing medical data based on multi-source data, in order to solve the problem that existing technologies cannot achieve effective correspondence and collaborative representation of medical data of different modalities.

[0006] Firstly, in order to solve the above-mentioned technical problems, the present invention provides a method for standardizing medical data based on multi-source data, comprising: Collect raw medical data from multiple sources, and perform format-unified preprocessing on the raw medical data from multiple sources to obtain a standardized multi-source dataset; Extract the time-series metadata from the standardized multi-source dataset, match the time-series metadata through dynamic time warping to obtain the relative time offset, and establish a reference time axis based on the relative time offset to obtain the time-corresponding sequence; The time-series deviation of the time-corresponding sequence is detected, and the time-corresponding sequence with deviation is interpolated and filled to obtain a reconstructed monitoring sequence. The sampling rate of the reconstructed monitoring sequence is calibrated to obtain a unified time-series dataset. Based on the unified time series dataset, multimodal data fusion is performed to generate a multimodal embedding matrix. Textual and numerical features related to time series and content are extracted from the multimodal embedding matrix to obtain a set of associated features. The set of associated features is then weighted and integrated to obtain a fused feature vector. Calculate the intermodal relationship strength of the fused feature vector. If the intermodal relationship strength is lower than a preset strength threshold, iteratively optimize the initial fusion weights of the fused feature vector to obtain an updated fused feature vector. Construct an expression form collaboration matrix based on the updated fused feature vector to obtain the enhanced feature completeness index. A low-dimensional embedding space is constructed based on the enhanced feature integrity index. The high-dimensional vectors in the updated fused feature vectors are mapped to the low-dimensional embedding space to obtain low-dimensional representation vectors. Cluster analysis is performed on the low-dimensional representation vectors to generate an overall feature representation of the patient. The patient's overall characteristics are used as prediction input, and the temporal relationships and modal synergistic cues of the data are processed by a pre-trained feedforward neural network to obtain the risk assessment results of medical complications.

[0007] Secondly, the present invention provides a medical data standardization processing system based on multi-source data, comprising: The data acquisition module is used to collect raw medical data from multiple sources and perform format-unified preprocessing on the raw medical data from multiple sources to obtain a standardized multi-source dataset. The time series matching module is used to extract the time series metadata of the standardized multi-source dataset, match the time series metadata through dynamic time warping to obtain the relative time offset, and establish a reference time axis based on the relative time offset to obtain the time corresponding sequence. The time series calibration module is used to detect the time series deviation of the time-corresponding sequence, interpolate and fill the time-corresponding sequence with deviation to obtain the reconstructed monitoring sequence, and perform sampling rate calibration on the reconstructed monitoring sequence to obtain a unified time series dataset. The modality fusion module is used to perform multimodal fusion of data based on the unified time series dataset, generate a multimodal embedding matrix, extract time-series and content-related text features and numerical features from the multimodal embedding matrix to obtain a set of associated features, and weight and integrate the set of associated features to obtain a fused feature vector. The weight enhancement module is used to calculate the intermodal connection strength of the fused feature vector. If the intermodal connection strength is lower than a preset strength threshold, the initial fusion weight of the fused feature vector is iteratively optimized to obtain an updated fused feature vector. Based on the updated fused feature vector, an expression form collaboration matrix is ​​constructed to obtain the enhanced feature completeness index. The dimensionality reduction and clustering module is used to construct a low-dimensional embedding space based on the enhanced feature integrity index, map the high-dimensional vector in the updated fused feature vector to the low-dimensional embedding space to obtain a low-dimensional representation vector, and perform cluster analysis on the low-dimensional representation vector to generate an overall feature representation of the patient. The results output module is used to take the patient's overall characteristics as prediction input, process the temporal relationships and modal synergistic cues of the data through a pre-trained feedforward neural network, and obtain the risk assessment results of medical complications.

[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention obtains a standardized multi-source dataset by performing a unified preprocessing operation such as integrity verification, timestamp synchronization and semantic standardization coding on multi-source medical raw data. Then, it extracts time-series metadata and obtains the corresponding time sequence through dynamic time warping and matching. For time-series deviations, it performs interpolation filling and sampling rate calibration to generate a unified time-series dataset. From data preprocessing to time-series calibration, a time dimension warping scheme is formed for the whole process, which unifies the recording time of multi-source heterogeneous medical data, aligns the collection granularity, greatly eliminates information loss and analysis deviation caused by time-series deviations, realizes accurate alignment of multi-source heterogeneous medical data in the time dimension, and improves the time-series consistency of medical data.

[0009] (2) Based on a unified time series dataset, this invention performs multimodal fusion to generate a multimodal embedding matrix, extracts the associated feature set and integrates it into a fusion feature vector, optimizes the fusion weights that do not reach the threshold by calculating the intermodal connection strength, and constructs an expression form synergy matrix to obtain an enhanced feature completeness index, which strengthens the intrinsic synergy between multimodal data, fully explores the association features between medical monitoring numerical data and doctor consultation natural language data, and makes the fused feature vector more realistically reflect the patient's physiological state, thus improving the feature expression completeness of multi-source medical data.

[0010] (3) Based on the enhanced feature completeness index, the present invention maps the fused feature vector to the low-dimensional embedding space to obtain the low-dimensional representation vector. After cluster analysis, a physiological state topology map is constructed and the overall feature representation of the patient is generated. The redundancy of the high-dimensional fused features is eliminated, and the physiological state is intuitively depicted. At the same time, relying on the temporal modeling network and causal mask, the temporal and modal synergistic rules are accurately captured, ensuring the completeness and accuracy of the feature representation and avoiding the evaluation bias caused by the confusion of temporal relationships. Thus, the risk assessment of complications is based on complete and accurate patient features, which improves the accuracy and reliability of the results of the risk assessment of medical complications. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the medical data standardization processing method based on multi-source data provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the medical data standardization processing system based on multi-source data provided in the second embodiment of the present invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] Reference Figure 1 The first embodiment of the present invention provides a method for standardizing medical data based on multi-source data, including the following steps: S11, Collect raw medical data from multiple sources, and perform format-unified preprocessing on the raw medical data from multiple sources to obtain a standardized multi-source dataset; S12, extract the time-series metadata of the standardized multi-source dataset, match the time-series metadata through dynamic time warping to obtain the relative time offset, and establish a reference time axis based on the relative time offset to obtain the time-corresponding sequence; S13, detect the time-series deviation of the time-corresponding sequence, interpolate and fill the time-corresponding sequence with deviation to obtain the reconstructed monitoring sequence, and calibrate the sampling rate of the reconstructed monitoring sequence to obtain a unified time-series dataset; S14. Perform multimodal data fusion based on the unified time series dataset to generate a multimodal embedding matrix. Extract time-series and content-related text features and numerical features from the multimodal embedding matrix to obtain a set of associated features. Weighted integration of the set of associated features yields a fused feature vector. S15, calculate the intermodal connection strength of the fused feature vector. If the intermodal connection strength is lower than a preset strength threshold, iteratively optimize the initial fusion weight of the fused feature vector to obtain an updated fused feature vector. Construct an expression form collaboration matrix based on the updated fused feature vector to obtain the enhanced feature completeness index. S16. Construct a low-dimensional embedding space based on the enhanced feature integrity index, map the high-dimensional vector in the updated fused feature vector to the low-dimensional embedding space to obtain a low-dimensional representation vector, perform cluster analysis on the low-dimensional representation vector, and generate an overall feature representation of the patient. S17, the patient's overall characteristics are used as prediction input, and the temporal relationships and modal synergistic cues of the data are processed by a pre-trained feedforward neural network to obtain the risk assessment results of medical complications.

[0014] In step S11, it is necessary to collect multi-source medical raw data, and perform format unification preprocessing on the multi-source medical raw data to obtain a standardized multi-source dataset, including: Collect raw medical data from multiple sources, perform integrity verification on the raw medical data from multiple sources, and generate a clean data stream; Diagnostic terms and numerical matrices are extracted from the clean data stream and timestamp synchronization is performed to construct a time-aligned multimodal data sequence; The diagnostic and treatment terms in the multimodal data sequence are mapped to a pre-stored medical knowledge graph to generate semantically standardized codes. The semantically standardized codes are then concatenated with the numerical matrix to form a fusion feature base vector. The fused feature base vectors are then subjected to structured encapsulation to obtain a standardized multi-source dataset.

[0015] It should be noted that the multi-source medical raw data includes five core data categories: electronic medical records, imaging examinations, laboratory tests, wearable device monitoring, and patient self-reports. These serve as the basic data sources for standardized processing. Data collection covers objective monitoring and subjective description data throughout the entire clinical diagnosis and treatment process. The integrity verification threshold is that the integrity of the required fields is greater than or equal to 90%. This threshold is set based on the minimum data quality requirements for effective clinical analysis. Field integrity is calculated by dividing the actual number of required fields by the preset number of required fields. This threshold can be flexibly adjusted within the range of 80%-95% according to the accuracy requirements of clinical application scenarios. Data below the threshold is prioritized for completion by associating with the patient's unique identifier. Data that cannot be completed is isolated.

[0016] The pre-stored medical knowledge graph adopts a three-layer architecture: ontology layer, data layer, and association layer. The ontology layer contains core concepts such as diseases, symptoms, medical terms, and physiological indicators, as well as the hierarchical relationships between these concepts. The data layer includes official standardized medical codes such as ICD-10 disease codes, SNOMED CT medical terminology sets, and LOINC test item codes. The association layer stores clinical semantic relationships between concepts, such as causality, association, and subordination. The data sources for this knowledge graph are medical coding standards published by the National Health Commission, clinical practice guidelines from core medical journals, and real clinical practice data from tertiary hospitals. The update mechanism is a quarterly incremental update plus an annual comprehensive update, synchronizing with the latest medical coding standards and clinical practice consensus.

[0017] The time precision of timestamp synchronization is uniformly set to the minute level. This precision is based on the granularity of routine medical data collection and recording, and can be adjusted to the second level according to the high-frequency collection needs of wearable devices. The dimension of the fusion feature base vector is jointly determined by the number of semantic features of diagnostic and treatment terms and the number of index features of the numerical matrix. The structured encapsulation adopts the HL7 FHIR dedicated medical data format, while being compatible with the general CSV data format, ensuring the compatibility and transmissibility of data between different medical information systems.

[0018] In this step, firstly, raw medical data from multiple sources are collected through the medical data acquisition interface from the hospital information system, laboratory information system, medical image archiving and communication system, wearable device data backend, and patient self-report terminal. The data is then grouped according to the patient's unique identifier to form a raw dataset. Subsequently, integrity verification is performed on the raw dataset. By matching the preset required fields of each data entry, the field completeness is calculated. For data below the threshold, an association completion process is initiated. The related data in the raw dataset is retrieved using the patient's unique identifier to complete the fields. If completion fails, the data is marked as incomplete and isolated. Finally, data that meets the integrity requirements is retained to generate a clean data stream.

[0019] Furthermore, core diagnostic and treatment terms from electronic medical records and patient self-reports are extracted from the clean data stream using natural language processing technology. Numerical physiological indicators from laboratory tests and wearable device monitoring are also extracted and a numerical matrix is ​​constructed. Simultaneously, the original timestamps of all data are extracted, and timestamps from different data sources are uniformly converted to Beijing timestamps using timestamp calibration and calibrated to the set time precision. Based on the calibrated timestamps, the diagnostic and treatment terms and the numerical matrix are matched in terms of time dimension to construct a time-aligned multimodal data sequence.

[0020] Furthermore, time-aligned multimodal data sequences are input into a pre-stored medical knowledge graph. Precise terminology matching is used to match diagnostic and treatment terms with standardized medical concepts in the knowledge graph, generating corresponding semantically standardized codes. Vector concatenation is then used to concatenate the vector form of the semantically standardized codes with the numerical matrix, forming a fusion feature base vector. During the concatenation process, the dimensions of the semantic and numerical features are ensured to be independent and non-overlapping. Finally, data structuring encapsulation is used to encapsulate the fusion feature base vector, adding identifiers such as data headers, unique patient identifiers, data collection time, and data source according to a preset medical data format. After encapsulation, a standardized multi-source dataset is obtained, achieving triple standardization in data format, semantics, and timestamps.

[0021] For example, multi-source raw medical data of a hypertensive patient is collected, including electronic medical records from the hospital information system, cardiac MRI imaging reports from the medical imaging archive and communication system, blood routine and biochemical test data from the laboratory information system, heart rate and blood pressure monitoring data from wearable devices, and the patient's self-reported symptoms, diet, and exercise. Integrity checks are performed on the raw data. Ten mandatory fields are preset, including patient identification, collection time, data type, and core indicators. If a wearable device data entry is missing the device model field, it is directly retained if the field completeness meets the threshold requirement of 90%. If a patient's self-reported data entry is missing the symptom onset time field, it is retained after being supplemented by linking to the electronic medical record. Diagnostic terms such as "grade 3 hypertension" and "left ventricular hypertrophy" are extracted from the clean data stream. Indicators such as blood pressure 140 / 90 mmHg and heart rate 85 beats / minute are extracted to construct a numerical matrix. All data timestamps are uniformly calibrated to minute-level Beijing timestamps to construct a time-aligned multimodal data sequence. The sequence is mapped to a medical knowledge graph. "Grade 3 hypertension" is matched to generate ICD-10 code I10.3, and "left ventricular hypertrophy" is matched to generate ICD-10 code I51.7. The semantically standardized codes are concatenated with the numerical matrix to form a fusion feature base vector. Finally, the standardized multi-source dataset is encapsulated in HL7 FHIR format.

[0022] In step S12, it is necessary to extract the time-series metadata of the standardized multi-source dataset, obtain the relative time offset by dynamic time warping matching of the time-series metadata, and establish a reference time axis based on the relative time offset to obtain the time-corresponding sequence, including: The timestamps and collection granularity information of each data entry are extracted from the standardized multi-source dataset to obtain time-series metadata. Based on the time-series metadata, continuous medical monitoring data and discrete static medical data are separated to obtain continuous time series and discrete time nodes; The continuous time series and the discrete time nodes are matched by dynamic time warping, and the relative time offset between each modal data is calculated. A reference time axis with a unified sampling frequency is established based on the relative time offset, and the continuous time series and the discrete time nodes are mapped to the reference time axis to obtain a time-corresponding sequence.

[0023] It should be noted that time-series metadata is a set of core time attributes for each data entry in a standardized multi-source dataset, containing three types of information: timestamp, collection granularity, and data collection device type. This forms the basis for subsequent data separation and time-series matching. The threshold for distinguishing between continuous medical monitoring data and discrete static medical data is that data with a collection granularity of 5 minutes or less is considered continuous, and data with a collection granularity greater than 1 hour is considered discrete. This threshold is set based on the characteristics of medical data collection. Continuous data mainly consists of wearable device monitoring data, while discrete data mainly consists of electronic medical records, imaging examinations, and laboratory test data. The threshold can be adjusted according to the actual collection capabilities of the collection devices. Dynamic time warping is implemented using classic dynamic programming, and distance calculation is performed using... Using Euclidean distance, the adjustment window size is set to 10% of the continuous time series length. This window size is set according to the time fluctuation characteristics of medical time series data and can be adjusted within the range of 5%-15% to limit the maximum time offset of adjustment and avoid time series distortion caused by over-adjustment. The relative time offset is the average time difference between the continuous time series and the discrete time nodes after matching, and the calculation unit is minutes. The uniform sampling frequency of the reference time axis is determined according to the average acquisition granularity of continuous medical monitoring data. Common acquisition frequencies such as 1 minute, 5 minutes, and 10 minutes are preferred to ensure that the interpolation error during the mapping of continuous time series is minimized. The time range of the reference time axis covers the earliest and latest acquisition times of the standardized multi-source dataset.

[0024] In this step, firstly, by traversing all data entries of the standardized multi-source dataset, the timestamp, collection granularity, and data collection device type information of each data entry are extracted. The data is then grouped according to the unique patient identifier to form time-series metadata for each patient. Subsequently, based on the collection granularity information in the time-series metadata, data separation is used to divide the data into continuous medical monitoring data and discrete static medical data. The continuous medical monitoring data is sorted in ascending order according to the timestamp to construct a continuous time series. This series is a sequence structure of a one-dimensional time axis corresponding to multi-dimensional physiological indicators. For the discrete static medical data, key timestamps are extracted, and duplicate timestamps are removed to form discrete time nodes.

[0025] Furthermore, dynamic time warping matching is performed on continuous time series and discrete time nodes. First, a dynamic programming matrix is ​​initialized, where the rows of the matrix correspond to the time points of the continuous time series, the columns correspond to the discrete time nodes, and the matrix elements are the Euclidean distances of the corresponding time points. Then, the minimum cumulative distance of the matrix is ​​calculated according to the recursive rules of dynamic programming. During the recursion process, the constraint of the warping window is followed to avoid time matching across windows. The optimal matching relationship between the continuous time series and the discrete time nodes is determined based on the path of the minimum cumulative distance. Based on this matching relationship, the time difference between each modality is calculated, and the relative time offset is obtained by averaging the time differences.

[0026] Furthermore, based on the relative time offset and the average acquisition granularity of the continuous time series, a uniform sampling frequency for the reference time axis is determined. The earliest acquisition time of the standardized multi-source dataset is used as the start time, and the latest acquisition time is used as the end time. A reference time axis with equal time intervals is generated according to the uniform sampling frequency. Finally, time mapping is used to interpolate or aggregate the continuous time series according to the sampling frequency of the reference time axis. The discrete time nodes are calibrated according to the relative time offset and mapped to the nearest time point of the reference time axis. After completing the time mapping of all modal data, the corresponding time sequence is obtained.

[0027] For example, time-series metadata is extracted from a standardized multi-source dataset of a diabetic patient. The granularity of blood glucose and heart rate monitoring data from wearable devices is 5 minutes, the granularity of electronic medical record diagnostic records is 1 day, and the granularity of chest CT reports and blood routine test results is 1 time, with the collection time being 8:30 AM. Based on the collection granularity, the 5-minute granularity monitoring data is separated into continuous medical monitoring data, constructing a continuous time series. Diagnostic records, imaging reports, and test results are separated into discrete static medical data, with 8:30 AM and 00:00 AM of the current day extracted as discrete time nodes. Dynamic time warping is used to match the continuous time series and discrete time nodes, setting the warping window to 10% of the continuous series length. The best match between the continuous series at 8:20, 8:25, and 8:30 and the discrete node at 8:30 is calculated, with a mean time difference of -5 minutes, representing the relative time offset. Based on the 5-minute acquisition granularity, the sampling frequency of the reference time axis is determined to be 5 minutes, covering the period from 00:00 to 23:55 of the day. The continuous time series is directly mapped to the reference time axis, and the discrete node 8:30 is calibrated to 8:25 and then mapped to the corresponding time point to obtain the time-corresponding sequence.

[0028] In step S13, it is necessary to detect the time-series deviation of the time-corresponding sequence, interpolate and fill the time-corresponding sequence with deviation to obtain a reconstructed monitoring sequence, and perform sampling rate calibration on the reconstructed monitoring sequence to obtain a unified time-series dataset, including: The temporal consistency of the time-corresponding sequences is detected, and missing time points are identified from the time-corresponding sequences with deviations. The missing time points are numerically estimated using interpolation to obtain the filling time points, and the continuous medical monitoring data is reconstructed in its entirety to generate a reconstructed monitoring sequence. Analyze the data fluctuation amplitude and frequency of the reconstructed monitoring sequence to determine the appropriate target sampling frequency; Perform time-series synchronization calibration on the reconstructed monitoring sequence according to the target sampling frequency to obtain a unified time-series dataset.

[0029] It should be noted that the temporal consistency test uses time point continuity testing. The deviation judgment standard is that there is no data at three or more consecutive time points on the baseline time axis. This standard is set according to the basic requirements of medical time series data continuity analysis and can be adjusted within the range of 1-5 time points according to the actual data collection situation. Missing time points refer to time points on the baseline time axis where there is no corresponding data. The interpolation method includes two types: linear interpolation and trend-based smoothing interpolation. Linear interpolation is suitable for time intervals where physiological indicators change slowly, while smoothing interpolation is implemented using the moving average method with a window size of 5 time points, which is suitable for time intervals where physiological indicators fluctuate greatly. The numerical error threshold for interpolation is 5%, which is set according to the conventional measurement error range of medical physiological indicators. When the interpolation result error exceeds the threshold, it is re-estimated by combining the physiological indicator change trend of adjacent time points.

[0030] The data fluctuation amplitude is quantified by the coefficient of variation, which is the ratio of the standard deviation to the mean. The fluctuation frequency is quantified by the number of extreme value changes per unit time. In one possible embodiment, the rules for determining the target sampling frequency can be as follows: if the coefficient of variation is less than or equal to 20% and the fluctuation frequency is less than or equal to 1 time / hour, use 15 minutes / time; if the coefficient of variation is greater than 20% but less than or equal to 50% or the fluctuation frequency is greater than 1 time / hour but less than or equal to 3 times / hour, use 5 minutes / time; and if the coefficient of variation is greater than 50% or the fluctuation frequency is greater than 3 times / hour, use 1 minute / time.

[0031] In this step, the baseline timeline of the corresponding time sequence is traversed by time point continuity detection. The corresponding data is checked point by point, and the number of consecutive time points without data is counted. If the deviation judgment criteria are met, it is determined that there is a time sequence deviation in the corresponding time sequence. At the same time, all time points without corresponding data on the baseline timeline are identified and marked as missing time points.

[0032] Furthermore, the changing trends of physiological indicators at adjacent time points with data are analyzed, and corresponding interpolation methods are selected based on the changing trends. If the physiological indicators at adjacent time points change smoothly, linear interpolation is used to estimate the values ​​of physiological indicators at the missing time points based on the linear relationship between the two points. If the physiological indicators at adjacent time points fluctuate significantly, trend-based smoothing interpolation is used to calculate the values ​​of physiological indicators at the missing time points using the moving average method. The numerical error between the interpolation result and the theoretical trend is calculated. If the error meets the threshold requirement, the estimated value is used as the filler value to generate the filler time point. If the error exceeds the threshold, the window size of the moving average method is adjusted and interpolation is repeated until the error meets the requirements. The filler time points are then integrated with the original time points to reconstruct the medical monitoring data and generate a reconstructed monitoring sequence without missing time points. Furthermore, all values ​​of each physiological indicator in the reconstructed monitoring sequence are extracted, the coefficient of variation and fluctuation frequency of each physiological indicator are calculated, the dynamic change characteristics of the physiological indicators are quantified, and a suitable target sampling frequency is matched for the reconstructed monitoring sequence based on the preset target sampling frequency and the calculation results of the coefficient of variation and fluctuation frequency, so as to ensure that the sampling frequency can completely capture the dynamic change characteristics of the physiological indicators. Finally, resampling is used to perform time-series synchronization calibration on the reconstructed monitoring sequence. If the original sampling frequency of the reconstructed monitoring sequence is higher than the target sampling frequency, the original data is downsampled by data mean aggregation, and the mean within the time window is taken as the value under the target sampling frequency. If the original sampling frequency is lower than the target sampling frequency, upsampling is performed by interpolation to supplement the missing values ​​under the target sampling frequency. At the same time, discrete data such as imaging examinations and laboratory tests are remapped to the calibrated time axis, and finally a unified time-series dataset is obtained.

[0033] For example, a temporal consistency test was performed on the time-series data of a patient with coronary heart disease. The baseline time axis sampling frequency was 5 minutes. The test found that there was no heart rate or blood pressure monitoring data at 5 time points within the interval of 14:00-14:20 on the same day. This indicated a temporal discrepancy, and these time points were marked as missing time points. The physiological indicators at adjacent time points of 13:55 and 14:25 showed significant fluctuations. A moving average method using 5 time point windows was used for smoothing interpolation to estimate the values ​​of the missing time points. The interpolation error was 3%, meeting the threshold requirement. Filled time points were generated, and the monitoring sequence was reconstructed. The coefficient of variation for heart rate in the reconstructed sequence was calculated to be 35%, with a fluctuation frequency of 2 times / hour; the coefficient of variation for blood pressure was 28%, with a fluctuation frequency of 1.5 times / hour. A sampling frequency of 5 minutes / time was matched according to the rules, consistent with the original sampling frequency. The reconstructed sequence values ​​were directly retained, and the discrete data were remapped to obtain a unified time-series dataset.

[0034] In step S14, multimodal data fusion needs to be performed based on the unified time-series dataset to generate a multimodal embedding matrix. Textual and numerical features related to time series and content are extracted from the multimodal embedding matrix to obtain a set of associated features. The set of associated features is then weighted and integrated to obtain a fused feature vector, including: Medical monitoring numerical data and doctor consultation natural language data are extracted from the unified time series dataset. The doctor consultation natural language data is vectorized to obtain consultation text vectors. The consultation text vector is fused with the medical monitoring numerical data to generate a multimodal embedding matrix; Textual and numerical features with clinical semantic relevance within the same time window are selected from the multimodal embedding matrix to obtain a set of associated features; An initial fusion weight is assigned to the associated feature set, and the associated feature set is weighted and integrated based on the initial fusion weight to obtain a fusion feature vector.

[0035] It should be noted that the medical monitoring numerical data refers to numerical physiological indicators such as wearable device monitoring and laboratory tests in a unified time-series dataset, while the doctor's consultation natural language data refers to unstructured text data such as electronic medical records, consultation dialogues in patient self-reports, and symptom descriptions.

[0036] In one possible implementation, vectorization can be achieved using a pre-trained BERT model in the medical field. The pre-training corpus consists of professional text data such as medical consultation records, clinical guidelines, and medical literature. This corpus originates from publicly available medical text resources and compliant clinical data, including nationally published clinical treatment guidelines, core medical journal articles, publicly available medical consultation datasets, and anonymized real consultation records and electronic medical records from top-tier hospitals. All corpora undergo compliance review and anonymization to ensure no patient privacy is compromised. Model training employs a two-stage approach: unsupervised pre-training followed by supervised fine-tuning. During pre-training, the model learns medical semantic features through Masked Language Modeling (MLM) and Next Sentence Prediction (NSP) tasks. During fine-tuning, labeled medical text semantic matching datasets are used to optimize model parameters to meet the vectorization requirements of consultation texts. The model adopts the basic BERT architecture, with specific parameters of 768 hidden layer dimensions, 12 attention heads, 12 encoder layers, and a vocabulary size adapted to medical terminology. The training tasks are medical text mask prediction and medical semantic next-sentence prediction, using the Adam optimizer with an initial learning rate of 2e-5 and a batch size of 32. Iterative training is performed until the loss on the validation set stabilizes and converges. The training data requires preprocessing, including word segmentation, stop word removal, and format standardization, to ensure the consistency and validity of the input data.

[0037] The multimodal embedding matrix has the dimension of the number of time points multiplied by the sum of the dimension of the consultation text vector and the dimension of the medical monitoring numerical data features. The fusion method uses feature concatenation to ensure that text features and numerical features are independently distributed in the matrix. The criteria for determining temporal and content association are that they fall within the same 30-minute time window and that the symptoms described by the text features and the physiological indicators of the numerical features have a clinical correlation. The clinical correlation is determined based on the conceptual associations in a pre-stored medical knowledge graph. The fusion weights range from 0 to 1, and the sum of the fusion weights of all features is 1. The weight values ​​are determined based on both the clinical importance of the features and the reliability of the data. Clinical importance is determined based on… The correlation between features and diseases in the pre-stored medical knowledge graph is quantified, and data reliability is quantified based on the data collection accuracy and missing rate. The specific calculation method is that the fusion weight is equal to the sum of the clinical importance score multiplied by 0.6 and the data reliability score multiplied by 0.4, and then divided by the sum of the calculated values ​​of all features. The quantification range of both the clinical importance score and the data reliability score is 0-1. The weight coefficients of 0.6 and 0.4 in this calculation method are set according to the proportion of indicator importance and data reliability in clinical analysis, and the coefficient values ​​can be adjusted according to the actual application scenario. The dimension of the fusion feature vector is consistent with the number of features in the associated feature set, and the integration of features is achieved through weighted summation.

[0038] In this step, medical monitoring numerical data and doctor consultation natural language data are first separated from the unified time-series dataset through feature extraction. The medical monitoring numerical data is normalized to map the values ​​to the 0-1 range, eliminating the dimensional differences between different indicators. The doctor consultation natural language data is preprocessed by performing word segmentation, stop word removal, and part-of-speech tagging to remove meaningless words and retain core diagnostic terms and symptom descriptions. Then, the preprocessed doctor consultation natural language data is input into a BERT model pre-trained in the medical field. The model's encoding layer converts the text into a dense vector representation, resulting in a consultation text vector that fully captures the semantic information and contextual features of the text.

[0039] Furthermore, feature concatenation is employed to concatenate the consultation text vector with the normalized medical monitoring numerical data along the feature dimension. Using the time points of the unified time series dataset as rows and the concatenated features as columns, a multimodal embedding matrix is ​​constructed, with each row of the matrix corresponding to a multimodal fusion feature at a given time point. Subsequently, based on preset time series and content association judgment criteria, associated features are extracted from the multimodal embedding matrix. First, the time axis is divided into 30-minute time windows. Within each time window, text features and numerical features are matched through feature matching. Based on a pre-stored medical knowledge graph, it is determined whether there is a clinical association between the two. Text features and numerical features that simultaneously meet the time series and content association requirements are selected to form an associated feature set.

[0040] Furthermore, each feature in the associated feature set is scored for clinical importance and data reliability. The fusion weight of each feature is calculated according to the fusion weight calculation method to ensure that the sum of all weights is 1. A weighted summation is used to multiply each feature in the associated feature set by its corresponding fusion weight and then sum them to obtain a fusion feature vector. This vector integrates the core associated features of multimodal data.

[0041] For example, medical monitoring data such as blood glucose, heart rate, and blood oxygen saturation from 30 to 120 minutes postprandially were extracted from a unified time-series dataset of patients with abnormal postprandial blood glucose. Natural language data from doctor's consultations, including patient descriptions such as "frequent chest tightness and palpitations two hours after meals, accompanied by hand tremors, sweating, and fatigue," were also extracted. The numerical data were normalized, and the text data was segmented and stop-word removed to obtain core vocabulary. This vocabulary was then input into a medical BERT model to obtain a 768-dimensional consultation text vector. The consultation text vector was concatenated with the 3-dimensional numerical data, constructing a multimodal embedding matrix with four time points as rows. After dividing the data into 30-minute time windows, a clinical association was found between "palpitations," "sweating," and "fatigue" and elevated heart rate and blood glucose within the 60-120 minute postprandial window. These features were selected to form an associated feature set. The clinical importance and data reliability of the features within the set were scored, and the fusion weights of each feature were calculated using a specific method. The weighted sum was then used to obtain the fused feature vector.

[0042] In step S15, the intermodal relationship strength of the fused feature vector needs to be calculated. If the intermodal relationship strength is lower than a preset strength threshold, the initial fusion weights of the fused feature vector are iteratively optimized to obtain an updated fused feature vector. Based on the updated fused feature vector, an expression form of the collaboration matrix is ​​constructed to obtain the enhanced feature completeness index, including: The cosine similarity is calculated for the text modality feature vector and the numerical modality feature vector in the fused feature vector. The modality spacing is calculated based on the cosine similarity, and the modality spacing is mapped to the intermodal connection strength. If the intermodal connection strength is lower than a preset strength threshold, the initial fusion weights are iteratively optimized using the gradient descent method until the intermodal connection strength is not lower than the preset strength threshold or the maximum number of iterations is reached, and the updated fusion weights are obtained. An updated fusion feature vector is generated by weighted integration based on the updated fusion weights. The updated fusion feature vector is then mapped to a numerical range based on a preset physiological index clinical normal range to obtain a mapped value. Calculate the mapping deviation between the mapped value and the median value of the preset clinical normal range, and construct an expression form collaboration matrix with the number of features as rows and the time window as columns; The ratio of the number of non-zero mapping deviation elements to the total number of elements in the cooperative matrix of the expressed form is used as the feature coverage rate, and the feature coverage rate is used as the feature completeness index after enhancement.

[0043] It should be noted that the cosine similarity ranges from -1 to 1, with the value closer to 1 indicating a higher similarity between the two modal features. The modal distance is 1 minus the cosine similarity, ranging from 0 to 2. The intermodal connection strength uses a linear mapping to map the modal distance to the 0-1 range, with the value closer to 1 indicating a higher connection strength between the modal features. The preset strength threshold is set based on the synergy requirements of medical multimodal data fusion and has been verified through a large number of clinical multimodal data fusion experiments. An strength threshold in the range of 0.5-0.7 can achieve experimental results with clear clinical semantic association of multimodal features and stable fusion effect. In this embodiment, it is set to 0.6 for reference and can be dynamically adjusted according to the actual application scenario.

[0044] The fusion weight optimization is achieved using the gradient descent method, with the optimization objective being to maximize the intermodal connection strength. The learning rate is set to 0.01, and the maximum number of iterations is set to 100. These learning rate and iteration number are set based on the convergence speed of the weight optimization and can be adjusted according to the convergence of the optimization process. The optimized updated fusion weights still meet the basic requirements of a value range of 0-1 and the sum of all weights being 1. The target range for numerical range mapping is the clinical normal range of preset physiological indicators. After mapping, the feature values ​​are unified to the 0-1 interval. The mapping deviation is the absolute difference between the mapped value and 0.5, where 0.5 is the median value of the clinical normal range. The value range of the mapping deviation is 0 to 0.5. The smaller the value, the closer the feature value is to the clinical normal range. The preset clinical normal range of physiological indicators is derived from national clinical testing standards and internationally recognized medical norms. It is determined based on the clinical statistical research results of different physiological indicators, the distribution law of physiological parameters in healthy populations, and clinical practice consensus to ensure the authority and applicability of the range.

[0045] The dimension of the expression form co-coordinated matrix is ​​the number of features in the fused feature vector multiplied by the number of time windows in the unified time series dataset. The matrix elements are the mapping deviations of each feature within each time window. The feature coverage rate is the ratio of the number of non-zero mapping deviation elements in the expression form co-coordinated matrix to the total number of elements in the matrix. The value ranges from 0 to 1. The closer the value is to 1, the higher the feature completeness and co-coordinatedness. The enhanced feature completeness index is the specific value of the feature coverage rate.

[0046] In this step, the text modal feature vector and the numerical modal feature vector are first separated from the fused feature vector. The cosine similarity between the two vectors is calculated. The modal distance is obtained according to the modal distance calculation method. Then, the modal distance is converted into the intermodal connection strength in the interval of 0 to 1 through linear mapping, thus completing the quantitative evaluation of the intermodal connection strength.

[0047] Furthermore, the calculated intermodal connection strength is compared with a preset strength threshold. If the connection strength is greater than or equal to the threshold, it indicates that the intermodal synergy meets the requirements of clinical analysis, and the original fusion weights are directly retained. If the connection strength is less than the preset strength threshold, it indicates that the intermodal synergy is insufficient, and the fusion weight optimization process is immediately initiated. With maximizing the intermodal connection strength as the core optimization objective, the gradient descent method is used to iteratively optimize the initial fusion weights. In each iteration, the gradient of the intermodal connection strength is calculated, and the value of the fusion weights is adjusted according to the learning rate. At the same time, the value range and summation constraints of the weights are strictly guaranteed. After iterating to a preset number of times, the updated fusion weights are obtained.

[0048] Furthermore, based on the corresponding clinical normal ranges of each feature, the feature values ​​of the fused feature vector under updated fusion weights are mapped to a range of 0-1. The absolute difference between each feature mapping value and 0.5 is calculated to obtain the mapping bias. Using the number of features in the fused feature set as rows and the number of time windows in the unified time series dataset as columns, the mapping bias of each feature within each time window is sequentially filled into a matrix to construct an expression form coordination matrix. This matrix intuitively reflects the degree of coordination of multimodal features in numerical expression form. Finally, the number of non-zero mapping bias elements in the expression form coordination matrix and the total number of elements in the matrix are counted, and the ratio of the two is calculated to obtain the feature coverage rate. This feature coverage rate is used as the feature completeness index after enhancement. This index quantitatively reflects the feature completeness and modal coordination of the fused feature vector after weight optimization.

[0049] For example, text and numerical modal feature vectors are separated from the fused feature vector of a patient with hypertension and diabetes. The calculated cosine similarity is 0.56, the modal distance is 0.44, and the mapped intermodal connection strength is 0.78, which is higher than the threshold of 0.6. Therefore, the original fusion weights are directly retained. For a critically ill patient, the calculated cosine similarity of the fused feature vector is 0.04, and the intermodal connection strength is 0.52, which is lower than the threshold. Weight optimization is then initiated. With the goal of maximizing the intermodal connection strength, a learning rate of 0.01 and 100 iterations are set. Gradient descent is used for optimization. The original text modal weights were 0.55 and numerical modal weights were 0.45. After optimization, these are updated to 0.42 for text and 0.58 for numerical, and the connection strength is recalculated, increasing it to 0.76. Numerical range mapping was performed based on the clinical normal ranges of heart rate, blood pressure, and blood glucose. The mapping deviation of each feature was calculated. A 5×4 co-occurrence matrix was constructed with 5 features and 4 time windows. 19 non-zero mapping deviation elements were counted, and there were a total of 20 elements. The feature coverage rate was 0.95, which was used as the feature integrity index after enhancement.

[0050] In step S16, a low-dimensional embedding space needs to be constructed based on the enhanced feature integrity index. The high-dimensional vectors in the updated fused feature vectors are mapped to the low-dimensional embedding space to obtain a low-dimensional representation vector. Cluster analysis is then performed on the low-dimensional representation vector to generate a holistic patient feature representation, including: A low-dimensional embedding space is constructed based on the enhanced feature integrity index. The high-dimensional vector in the updated fused feature vector is mapped to the low-dimensional embedding space to obtain the low-dimensional representation vector. The low-dimensional representation vector is segmented by a sliding window to identify physiological state segments, and the physiological state segments are divided into physiological state clusters by cluster analysis. A state transition matrix is ​​constructed using the physiological state clusters as nodes, and a physiological state topology graph is generated based on the state transition matrix. The physiological state topology is analyzed to obtain a global structural complexity index, which is then used as a core feature to generate a representation of the patient's overall characteristics.

[0051] It should be noted that, in one possible embodiment, the principal components of principal component analysis are selected based on the cumulative variance contribution rate, with a threshold of 90%. This threshold is set according to the information retention requirements of feature dimensionality reduction and can be adjusted within the range of 85%-95%. This ensures that the low-dimensional representation vector after dimensionality reduction retains more than 90% of the effective information of the high-dimensional fused feature vector. The dimension of the low-dimensional representation vector is usually 5-8 dimensions, with the specific dimension determined according to the cumulative variance contribution rate. The window size of the sliding window is 10 time points of the unified time series dataset, and the step size is 5 time points. This window size and step size are set according to the time persistence characteristics of medical physiological states and can be adjusted according to the rate of change of physiological indicators. The physiological state segment is a subset of the low-dimensional representation vector obtained after the sliding window is segmented, and each subset corresponds to a continuous time interval.

[0052] Cluster analysis can be implemented using the K-means clustering algorithm. The number of clusters K is determined using the elbow rule. By calculating the sum of squares of errors under different K values, the K value with the slowest decreasing trend of error is selected as the optimal number of clusters. Typically, K is 3-5. Physiological state clusters are sets of vectors with similar physiological characteristics obtained after clustering, and each cluster corresponds to a specific physiological state. The state transition matrix is ​​a square matrix with the same dimension as the number of physiological state clusters. The matrix elements are the probabilities of transitioning from one physiological state cluster to another between adjacent physiological state segments. The transition probabilities are calculated by dividing the number of inter-cluster transitions by the total number of transitions. The physiological state topology graph is a directed weighted graph. Nodes are physiological state clusters, and the node size is proportional to the frequency of cluster occurrence. Directed edges represent the transition relationships between physiological state clusters, and the thickness of the edges is proportional to the transition probability.

[0053] The global structural complexity index includes three core indicators: average out-degree, clustering coefficient, and network diameter. The average out-degree reflects the diversity of physiological state transitions, the clustering coefficient reflects the local correlation of physiological states, and the network diameter reflects the span of physiological state transitions. The quantification range of all three indicators is 0-1, achieved through linear normalization. The overall patient characteristics are represented as the feature vector of the global structural complexity index, with a dimension of 3, which comprehensively reflects the dynamic evolution characteristics and overall complexity of the patient's physiological state.

[0054] In this step, the completeness of the fused feature vector is first determined based on the enhanced feature completeness index. If the enhanced feature completeness index is greater than or equal to the preset index threshold, it indicates that the completeness of the fused feature vector is high, and principal component analysis is directly performed on the high-dimensional vector in the fused feature vector. If the index is less than the preset index threshold, feature completion is performed on the fused feature vector before the projection operation is performed. Principal component analysis is used to reduce the dimensionality of the high-dimensional fused feature vector. First, the high-dimensional vector is decentered, the covariance matrix is ​​calculated, and the eigenvalues ​​and eigenvectors of the covariance matrix are solved. The eigenvalues ​​are sorted from largest to smallest, and eigenvectors are selected sequentially until the cumulative variance contribution rate reaches the set threshold. The high-dimensional vector is then projected onto the low-dimensional embedding space formed by the selected eigenvectors to obtain the low-dimensional representation vector.

[0055] Furthermore, a sliding window is used to segment the low-dimensional representation vector. According to the preset window size and step size, the continuous low-dimensional representation vector is divided into multiple vector subsets. Each subset corresponds to the physiological characteristics of a time interval and is identified as a physiological state segment. The low-dimensional representation vectors of all physiological state segments are normalized to eliminate dimensional differences. Then, the K-means clustering algorithm is used to perform cluster analysis on the normalized physiological state segments. First, K initial cluster centers are randomly selected, and the Euclidean distance between each physiological state segment and each cluster center is calculated. The segment is assigned to the nearest cluster center, and the mean of each cluster is recalculated as the new cluster center. The above process is repeated until the cluster centers no longer change or the maximum number of iterations is reached. The optimal number of clusters is determined by the elbow rule, and the physiological state segments are divided into multiple physiological state clusters. Then, the number of inter-cluster transitions between adjacent physiological state segments is counted, and the transition probability of each cluster transitioning to other clusters is calculated. The transition probabilities are filled into a matrix with the physiological state clusters as rows and columns to construct a state transition matrix.

[0056] Furthermore, a physiological state topology graph is generated based on the state transition matrix. Physiological state clusters are used as nodes, and the node size is adjusted according to the frequency of cluster occurrence, with larger nodes appearing more frequently. The transition relationships between clusters are used as directed edges, and the thickness of the edges is adjusted according to the transition probability, with thicker edges appearing more frequently. Finally, the physiological state topology graph is analyzed, and three global structural complexity indicators—average out-degree, clustering coefficient, and network diameter—are calculated. Linear normalization is used to map the values ​​of the three indicators to the 0-1 interval. The normalized three indicators are used as core features to construct a 3-dimensional feature vector, which represents the overall characteristics of the patient.

[0057] It is worth noting that the preset completeness threshold is the minimum effective completeness standard for dimensionality reduction and clustering after feature fusion of multimodal medical data. The numerical value is based on extensive clinical multimodal data fusion experiments. When the enhanced feature completeness index is greater than or equal to 0.8, the modal synergy, information coverage, and consistency of clinical representation of the fused features meet the accuracy requirements for subsequent dimensionality reduction, clustering, and topology graph construction. Below 0.8, problems such as feature loss, modal fragmentation, and state recognition distortion are likely to occur. This threshold can be flexibly adjusted within the range of 0.75-0.85 according to the needs of clinical scenarios.

[0058] For example, the enhanced feature integrity index of a diabetic patient is 0.85. Principal component analysis was directly performed on the high-dimensional vector of the fused feature vector to reduce dimensionality. The first 6 feature vectors were selected, with a cumulative variance contribution rate of 92%, resulting in a 6-dimensional low-dimensional representation vector. A sliding window with a window size of 10 time points and a step size of 5 time points was used to divide the data into 12 physiological state segments. After normalization, K-means clustering was used, and the optimal K value was determined to be 4 using the elbow rule. Four physiological state clusters were obtained: "Stable Recovery Cluster", "Mild Sympathetic Activation Cluster", "Significant Abnormal Stress Cluster", and "Delayed Recovery Cluster". The "Significant Abnormal Stress Cluster" contains 5 consecutive segments. The number of transitions between clusters was counted and the transition probability was calculated. A 4×4 state transition matrix was constructed. A physiological state topology map was generated based on the matrix, and a unidirectional strong transition path from the "Stable Recovery Cluster" to the "Significant Abnormal Stress Cluster" could be observed. The average out-degree was calculated to be 0.85, the clustering coefficient was 0.68, and the network diameter was 4. After normalization, the values ​​were 0.85, 0.68, and 0.8, respectively. A 3D representation of the overall patient characteristics was constructed, reflecting the characteristics of patients' postprandial physiological state being prone to abnormal stress and delayed recovery.

[0059] In step S17, the overall patient feature representation is used as the prediction input. A pre-trained feedforward neural network processes the temporal relationships and modal co-occurrences of the data to obtain the medical complication risk assessment result, including: A complication prediction input matrix is ​​constructed based on the overall patient feature representation. A causal masking operation is applied to the complication prediction input matrix to establish the temporal sequence of data. Modal collaboration cues are extracted based on the temporal sequence. The modal collaborative cues are input into a pre-trained feedforward neural network with temporal modeling capabilities for feature extraction to obtain hidden layer feature vectors; The hidden layer feature vector is mapped to the output layer and the classification probability is calculated. The medical complication risk assessment result is output based on the classification probability.

[0060] It should be noted that the dimension of the complication prediction input matrix is ​​1 multiplied by the sum of the dimension of the patient's overall features and the dimension of the statistical features of the physiological state clusters. The statistical features of the physiological state clusters include the occurrence ratio of each cluster and the mean of the center vector. This matrix is ​​a structured numerical matrix and serves as the input basis for the feedforward neural network. The causal masking operation is implemented using a lower triangular mask matrix. The dimension of the mask matrix is ​​consistent with the time dimension of the complication prediction input matrix. The lower triangular elements in the mask matrix are 1, and the upper triangular elements are 0. After multiplying with the complication prediction input matrix, future time information is masked, retaining only the current and previous time information, and strictly establishing the temporal order of the data. The extraction of modal collaborative cues is implemented using a self-attention mechanism, with the number of attention heads set to 8. This number is set according to the requirements of multimodal feature association analysis and can be adjusted within the range of 4-16.

[0061] The feedforward neural network is a three-layer fully connected network structure. The input layer dimension is consistent with the dimension of the modal collaboration cue. The first hidden layer has a dimension of 128, the second hidden layer has a dimension of 64, and the output layer has a dimension of 3, corresponding to the three risk levels of complications: low risk, medium risk, and high risk. The ReLU activation function is used between each layer. The model training requires a medical dataset labeled with the occurrence of complications as samples (including the patient's overall feature representation, temporal relationship, and modal collaboration cue labels). After normalization and outlier removal preprocessing, the training data is trained using the Adam optimizer with an initial learning rate of 1e-4 and the cross-entropy loss function as the optimization objective. The training is iteratively trained until the classification accuracy on the validation set is stable. An early stopping mechanism is used to avoid overfitting and ensure the model's generalization ability.

[0062] The classification probability is calculated using the softmax activation function, mapping the output layer values ​​to the 0-1 range, with the sum of all class probabilities equal to 1. The classification probability threshold is set as follows: a high-risk probability greater than or equal to 0.5 is considered high-risk; a high-risk probability greater than or equal to 0.3 but less than 0.5 is considered medium-risk; and a high-risk probability less than 0.3 is considered low-risk. This threshold is set according to the grading standards for clinical complication risk assessment and can be adjusted based on the severity of the disease and clinical intervention needs. The medical complication risk assessment results are classified as low, medium, or high risk, and the classification probability for each level is output, providing a quantitative reference for clinical decision-making.

[0063] In this step, the overall characteristics of the patient are taken as the core. Combined with the occurrence ratio of each physiological state cluster, the mean of the central vector and other statistical features, the complication prediction input matrix is ​​constructed by feature splicing. The matrix is ​​then normalized to map the values ​​to the 0-1 interval to ensure the consistency of the dimensions of the input data.

[0064] Furthermore, a lower triangular mask matrix with the same time dimension as the complication prediction input matrix is ​​constructed. The mask matrix is ​​multiplied by the complication prediction input matrix, and a causal masking operation is applied to mask information from future times, retaining only the feature data from the current and previous times. This strictly establishes the temporal sequence of the data and ensures the causal nature of the analysis. Then, a self-attention mechanism is used to perform feature analysis on the masked complication prediction input matrix, calculating the attention weights between features at different time steps and different modalities. By weighted summation, feature combinations with strong temporal and content correlations are highlighted, and modal collaboration cues that can reflect the collaborative patterns of multimodal data are extracted.

[0065] Furthermore, modal collaborative cues are input into the feedforward neural network. First, the feature vector is passed through the input layer to the first hidden layer, where it undergoes a nonlinear transformation using the ReLU activation function to achieve preliminary feature extraction, resulting in a 128-dimensional feature vector. This vector is then passed to the second hidden layer, where the ReLU activation function further extracts deeper features, yielding a 64-dimensional hidden layer feature vector that captures the core features of complication risk. Finally, the hidden layer feature vector is passed to the output layer, where the softmax activation function maps the output layer values ​​to classification probabilities in the 0-1 range. The sum of the classification probabilities for the three risk levels is 1. Based on a preset classification probability threshold, the classification probabilities are determined to establish the level of medical complication risk assessment, and the specific classification probability for each level is output, completing the entire risk assessment process.

[0066] For example, a 1×32 complication prediction input matrix is ​​constructed by combining the 3D overall patient feature representation of a diabetic patient with the occurrence ratio of four physiological state clusters and the mean of the center vector. After normalization, a 32×32 lower triangular mask matrix is ​​constructed, and a causal masking operation is applied to block future information. A self-attention mechanism with eight attention heads is used to extract modal collaboration cues, and it is found that the sequential collaboration mode weight between abnormal heart rate and sweating signals reaches 0.67. The modal collaboration cues are input into a three-layer fully connected feedforward neural network, and after two ReLU activation transformations, a 64-dimensional hidden layer feature vector is obtained. After mapping to the output layer, the low-risk probability (0.12), medium-risk probability (0.31), and high-risk probability (0.57) are calculated using the softmax function. Based on the threshold, the patient is judged to be high-risk, indicating a higher probability of postprandial hypoglycemic coma, cardiovascular events, and other complications.

[0067] In summary, this invention discloses a medical data standardization processing method based on multi-source data, which can improve the accuracy and reliability of complication risk assessment.

[0068] Reference Figure 2 The second embodiment of the present invention provides a medical data standardization processing system based on multi-source data, comprising: The data acquisition module is used to collect raw medical data from multiple sources and perform format-unified preprocessing on the raw medical data from multiple sources to obtain a standardized multi-source dataset. The time series matching module is used to extract time series metadata from the standardized multi-source dataset, obtain a relative time offset by dynamic time warping matching of the time series metadata, and establish a reference time axis based on the relative time offset to obtain a time-corresponding sequence. The time series calibration module is used to detect the time series deviation of the time-corresponding sequence, interpolate and fill the time-corresponding sequence with deviation to obtain the reconstructed monitoring sequence, and perform sampling rate calibration on the reconstructed monitoring sequence to obtain a unified time series dataset. The modality fusion module is used to perform multimodal fusion of data based on the unified time series dataset, generate a multimodal embedding matrix, extract time-series and content-related text features and numerical features from the multimodal embedding matrix to obtain a set of associated features, and weight and integrate the set of associated features to obtain a fused feature vector. The weight enhancement module is used to calculate the intermodal connection strength of the fused feature vector. If the intermodal connection strength is lower than a preset strength threshold, the initial fusion weight of the fused feature vector is iteratively optimized to obtain an updated fused feature vector. Based on the updated fused feature vector, an expression form collaboration matrix is ​​constructed to obtain the enhanced feature completeness index. The dimensionality reduction and clustering module is used to construct a low-dimensional embedding space based on the enhanced feature integrity index, map the high-dimensional vector in the updated fused feature vector to the low-dimensional embedding space to obtain a low-dimensional representation vector, and perform cluster analysis on the low-dimensional representation vector to generate an overall feature representation of the patient. The results output module is used to take the patient's overall characteristics as prediction input, process the temporal relationships and modal synergistic cues of the data through a pre-trained feedforward neural network, and obtain the risk assessment results of medical complications.

[0069] It should be noted that the medical data standardization processing system based on multi-source data provided in this embodiment of the invention is used to execute all the process steps of the medical data standardization processing method based on multi-source data in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0070] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0071] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1.A method for medical data standardization processing based on multi-source data, characterized in that, include: Collect raw medical data from multiple sources, and perform format-unified preprocessing on the raw medical data from multiple sources to obtain a standardized multi-source dataset; Extract the time-series metadata from the standardized multi-source dataset, match the time-series metadata through dynamic time warping to obtain the relative time offset, and establish a reference time axis based on the relative time offset to obtain the time-corresponding sequence; The time-series deviation of the time-corresponding sequence is detected, and the time-corresponding sequence with deviation is interpolated and filled to obtain a reconstructed monitoring sequence. The sampling rate of the reconstructed monitoring sequence is calibrated to obtain a unified time-series dataset. Based on the unified time series dataset, multimodal data fusion is performed to generate a multimodal embedding matrix. Textual and numerical features related to time series and content are extracted from the multimodal embedding matrix to obtain a set of associated features. The set of associated features is then weighted and integrated to obtain a fused feature vector. Calculate the intermodal relationship strength of the fused feature vector. If the intermodal relationship strength is lower than a preset strength threshold, iteratively optimize the initial fusion weights of the fused feature vector to obtain an updated fused feature vector. Construct an expression form collaboration matrix based on the updated fused feature vector to obtain the enhanced feature completeness index. A low-dimensional embedding space is constructed based on the enhanced feature integrity index. The high-dimensional vectors in the updated fused feature vectors are mapped to the low-dimensional embedding space to obtain low-dimensional representation vectors. Cluster analysis is performed on the low-dimensional representation vectors to generate an overall feature representation of the patient. The patient's overall characteristics are used as prediction input, and the temporal relationships and modal synergistic cues of the data are processed by a pre-trained feedforward neural network to obtain the risk assessment results of medical complications. 2.The method of claim 1, wherein, The process involves collecting raw medical data from multiple sources, performing format-unified preprocessing on the raw medical data to obtain a standardized multi-source dataset, including: Collect raw medical data from multiple sources, perform integrity verification on the raw medical data from multiple sources, and generate a clean data stream; Diagnostic terms and numerical matrices are extracted from the clean data stream and timestamp synchronization is performed to construct a time-aligned multimodal data sequence; The diagnostic and treatment terms in the multimodal data sequence are mapped to a pre-stored medical knowledge graph to generate semantically standardized codes. The semantically standardized codes are then concatenated with the numerical matrix to form a fusion feature base vector. The fused feature base vectors are then subjected to structured encapsulation to obtain a standardized multi-source dataset. 3.The method of claim 1, wherein, The process of extracting time-series metadata from the standardized multi-source dataset, matching the time-series metadata using dynamic time warping to obtain a relative time offset, and establishing a baseline time axis based on the relative time offset to obtain a time-corresponding sequence includes: The timestamps and collection granularity information of each data entry are extracted from the standardized multi-source dataset to obtain time-series metadata. Based on the time-series metadata, continuous medical monitoring data and discrete static medical data are separated to obtain continuous time series and discrete time nodes; The continuous time series and the discrete time nodes are matched by dynamic time warping, and the relative time offset between each modal data is calculated. A reference time axis with a unified sampling frequency is established based on the relative time offset, and the continuous time series and the discrete time nodes are mapped to the reference time axis to obtain a time-corresponding sequence. 4.The method of claim 3, wherein, The process involves detecting temporal deviations in the time-corresponding sequences, interpolating and padding the time-corresponding sequences with deviations to obtain reconstructed monitoring sequences, and calibrating the sampling rate of the reconstructed monitoring sequences to obtain a unified time-series dataset, including: The temporal consistency of the time-corresponding sequences is detected, and missing time points are identified from the time-corresponding sequences with deviations. The missing time points are numerically estimated using interpolation to obtain the filling time points, and the continuous medical monitoring data is reconstructed in its entirety to generate a reconstructed monitoring sequence. Analyze the data fluctuation amplitude and frequency of the reconstructed monitoring sequence to determine the appropriate target sampling frequency; Perform time-series synchronization calibration on the reconstructed monitoring sequence according to the target sampling frequency to obtain a unified time-series dataset. 5.The method of claim 1, wherein, The process involves performing multimodal data fusion based on the unified time-series dataset to generate a multimodal embedding matrix. Textual and numerical features related to time series and content are extracted from the multimodal embedding matrix to obtain a set of associated features. The associated feature set is then weighted and integrated to obtain a fused feature vector, including: Medical monitoring numerical data and doctor consultation natural language data are extracted from the unified time series dataset. The doctor consultation natural language data is vectorized to obtain consultation text vectors. The consultation text vector is fused with the medical monitoring numerical data to generate a multimodal embedding matrix; Textual and numerical features with clinical semantic relevance within the same time window are selected from the multimodal embedding matrix to obtain a set of associated features; An initial fusion weight is assigned to the associated feature set, and the associated feature set is weighted and integrated based on the initial fusion weight to obtain a fusion feature vector. 6.The method of claim 5, wherein, The calculation of the intermodal relationship strength of the fused feature vector, if the intermodal relationship strength is lower than a preset strength threshold, iteratively optimizes the initial fusion weights of the fused feature vector to obtain an updated fused feature vector, and constructs an expression form collaboration matrix based on the updated fused feature vector to obtain an enhanced feature completeness index, including: The cosine similarity is calculated for the text modality feature vector and the numerical modality feature vector in the fused feature vector. The modality spacing is calculated based on the cosine similarity, and the modality spacing is mapped to the intermodal connection strength. If the intermodal connection strength is lower than a preset strength threshold, the initial fusion weights are iteratively optimized using the gradient descent method until the intermodal connection strength is not lower than the preset strength threshold or the maximum number of iterations is reached, and the updated fusion weights are obtained. An updated fusion feature vector is generated by weighted integration based on the updated fusion weights. The updated fusion feature vector is then mapped to a numerical range based on a preset physiological index clinical normal range to obtain a mapped value. Calculate the mapping deviation between the mapped value and the median value of the preset clinical normal range, and construct an expression form collaboration matrix with the number of features as rows and the time window as columns; The ratio of the number of non-zero mapping deviation elements to the total number of elements in the cooperative matrix of the expressed form is used as the feature coverage rate, and the feature coverage rate is used as the feature completeness index after enhancement. 7.The method of claim 1, wherein, The process involves constructing a low-dimensional embedding space based on the enhanced feature completeness index, mapping the high-dimensional vectors in the updated fused feature vectors to the low-dimensional embedding space to obtain low-dimensional representation vectors, and performing cluster analysis on the low-dimensional representation vectors to generate an overall patient feature representation, including: A low-dimensional embedding space is constructed based on the enhanced feature integrity index. The high-dimensional vector in the updated fused feature vector is mapped to the low-dimensional embedding space to obtain the low-dimensional representation vector. The low-dimensional representation vector is segmented by a sliding window to identify physiological state segments, and the physiological state segments are divided into physiological state clusters by cluster analysis. A state transition matrix is ​​constructed using the physiological state clusters as nodes, and a physiological state topology graph is generated based on the state transition matrix. The physiological state topology is analyzed to obtain a global structural complexity index, which is then used as a core feature to generate a representation of the patient's overall characteristics. 8.The method of claim 1, wherein, The process of using the patient's overall characteristics as prediction input, and processing the temporal relationships and modal co-occurrences of the data through a pre-trained feedforward neural network to obtain the risk assessment results for medical complications includes: A complication prediction input matrix is ​​constructed based on the overall patient feature representation. A causal masking operation is applied to the complication prediction input matrix to establish the temporal sequence of data. Modal collaboration cues are extracted based on the temporal sequence. The modal collaborative cues are input into a pre-trained feedforward neural network with temporal modeling capabilities for feature extraction to obtain hidden layer feature vectors; The hidden layer feature vector is mapped to the output layer and the classification probability is calculated. The medical complication risk assessment result is output based on the classification probability. 9.A medical data standardization processing system based on multi-source data, characterized in that, include: The data acquisition module is used to collect raw medical data from multiple sources and perform format-unified preprocessing on the raw medical data from multiple sources to obtain a standardized multi-source dataset. The time series matching module is used to extract the time series metadata of the standardized multi-source dataset, match the time series metadata through dynamic time warping to obtain the relative time offset, and establish a reference time axis based on the relative time offset to obtain the time corresponding sequence. The time series calibration module is used to detect the time series deviation of the time-corresponding sequence, interpolate and fill the time-corresponding sequence with deviation to obtain the reconstructed monitoring sequence, and perform sampling rate calibration on the reconstructed monitoring sequence to obtain a unified time series dataset. The modality fusion module is used to perform multimodal fusion of data based on the unified time series dataset, generate a multimodal embedding matrix, extract time-series and content-related text features and numerical features from the multimodal embedding matrix to obtain a set of associated features, and weight and integrate the set of associated features to obtain a fused feature vector. The weight enhancement module is used to calculate the intermodal connection strength of the fused feature vector. If the intermodal connection strength is lower than a preset strength threshold, the initial fusion weight of the fused feature vector is iteratively optimized to obtain an updated fused feature vector. Based on the updated fused feature vector, an expression form collaboration matrix is ​​constructed to obtain the enhanced feature completeness index. The dimensionality reduction and clustering module is used to construct a low-dimensional embedding space based on the enhanced feature integrity index, map the high-dimensional vector in the updated fused feature vector to the low-dimensional embedding space to obtain a low-dimensional representation vector, and perform cluster analysis on the low-dimensional representation vector to generate an overall feature representation of the patient. The results output module is used to take the patient's overall characteristics as prediction input, process the temporal relationships and modal synergistic cues of the data through a pre-trained feedforward neural network, and obtain the risk assessment results of medical complications.

Images