Medical multi-source cross-platform data real-time monitoring and intelligent risk early warning method

Abstract

The present application relates to medical information processing technology field, specifically for medical multi-source cross-platform data real-time monitoring and intelligent risk early warning method, including the following steps: through multiple medical institutions, the patient's vital signs and medical terminology text are converted into embedded vectors, mutual information calculation is used to screen the semantic contribution degree dimension to generate an optimized medical terminology embedding set, the coordinate axis of the embedded vector of different platforms is rotated to realize cross-platform alignment, time synchronization and smoothing processing are performed to obtain cross-platform joint monitoring data flow, the intelligent risk discrimination model is inputted, the risk score is generated based on mutual information matching calculation, and real-time risk early warning state information is generated, in the present application, the multi-source fusion accuracy and coordination are improved by optimizing the medical terminology vector and cross-platform alignment, the monitoring intelligence is enhanced by deeply analyzing the vital signs and terminology semantics, the data sharing application effect is improved, the health risk score is dynamically calculated based on real-time synchronization and intelligent discrimination, and the early warning precision response speed is improved.

Description

Technical Field

[0001] This invention relates to the field of medical information processing technology, and in particular to a method for real-time monitoring and intelligent risk warning of multi-source cross-platform medical data. Background Technology

[0002] The field of medical information processing technology encompasses the collection, processing, storage, analysis, and application of medical data. This technology area includes various techniques and equipment used to improve the efficiency and accuracy of the healthcare industry, particularly in disease diagnosis, patient monitoring, and treatment plan design. Core aspects include standardized processing of medical data, cross-platform data sharing and integration, real-time monitoring technology, intelligent risk warning systems, and the security protection of medical information. With the continuous development of information technology, the field of medical information processing not only focuses on the basic management of medical data but also emphasizes improving the accuracy and real-time performance of clinical decision support systems through intelligent analysis.

[0003] Traditional medical multi-source, cross-platform data real-time monitoring and intelligent risk warning methods refer to collecting diverse data from different medical devices, sensors, and systems for real-time monitoring and analysis, and implementing intelligent risk warnings based on this data. This method relies on data integration technology to effectively aggregate data from different sources and utilizes big data analytics and machine learning techniques to monitor patient status and predict potential risks in real time. Traditional methods primarily employ rule-based monitoring systems and simple threshold alarm mechanisms, lacking intelligent analysis and adaptive warnings for complex disease patterns.

[0004] Existing technologies rely on rule-based monitoring systems and simple threshold alarm mechanisms, which struggle to handle complex disease patterns and individual differences, resulting in a lack of intelligent adaptability in early warning systems. Limited data integration capabilities prevent the full exploitation of potential risks from multi-source data, particularly the effective fusion and analysis of cross-platform data. Traditional methods fail to fully utilize the deep semantics of medical terminology, leading to poor data matching and integration across different platforms. This results in insufficient real-time performance and accuracy of monitoring systems, hindering their ability to effectively support personalized clinical decision-making, and increasing the risk of misjudgments or delays, thus impacting the timeliness and accuracy of early warning systems. Summary of the Invention

[0005] To address the shortcomings of existing technologies that rely on rule-based monitoring systems and simple threshold alarm mechanisms, which struggle to handle complex disease patterns and individual differences, resulting in a lack of intelligent adaptability in early warning systems; limited data integration capabilities that fail to fully exploit potential risks in multi-source data, especially the effective fusion and analysis of cross-platform data; and the inability of traditional methods to fully utilize the deep semantics of medical terminology, leading to poor data matching and integration across different platforms, insufficient real-time performance and accuracy of monitoring systems, inability to effectively support personalized clinical decision-making, and a high risk of misjudgment or delays, thus affecting the timeliness and accuracy of early warning systems, this invention provides a method for real-time monitoring and intelligent risk early warning of multi-source cross-platform medical data.

[0006] To achieve the above objectives, this invention employs a method for real-time monitoring and intelligent risk warning of multi-source cross-platform medical data, comprising the following steps:

[0007] S1: Obtain patient vital signs and medical terminology texts from multiple medical institutions, convert the texts into corresponding medical terminology embedding vectors, and combine them with patient vital signs to generate a preliminary fused medical dataset.

[0008] S2: Call the aforementioned preliminary fused medical dataset, use the mutual information calculation method to analyze the semantic information contribution of the medical term embedding vectors, filter the dimensions whose contribution meets the contribution threshold, and label the heterogeneous dimensions between different platforms to generate an optimized medical term embedding set.

[0009] S3: Based on the optimized medical terminology embedding set, rotate and adjust the coordinate axes of the embedding vectors for different platforms, align the cross-platform embedding space by calculating the coordinate rotation angle, and output the aligned embedding vector data structure.

[0010] S4: Based on the aligned embedded vector data structure and the patient's vital signs, perform time synchronization processing, smooth the vital sign parameter data, and obtain a cross-platform joint monitoring data stream;

[0011] S5: Invoke the cross-platform joint monitoring data stream input to the intelligent risk discrimination model, compare vital signs with preset vital sign thresholds at each time step, calculate the associated risk score based on the statistical deviation, and generate real-time risk warning status information.

[0012] As a further aspect of the present invention, the preliminary fused medical dataset includes a unified data identifier, a feature integration domain, and a parameter association box; the optimized medical terminology embedding set includes a semantic filtering domain, a weight distribution set, and a structural annotation layer; the aligned embedding vector data structure includes a coordinate correspondence table, a spatial reference box, and an alignment transformation domain; the cross-platform joint monitoring data stream includes a temporal continuous band, a parameter fusion channel, and a joint observation trajectory; and the real-time risk warning status information includes a risk indication level, a threshold comparison domain, and a warning trigger signal.

[0013] As a further aspect of the present invention, the specific steps of S1 are as follows:

[0014] S101: Obtain patient vital sign data frames from multiple medical institutions, perform interval discretization calculations on heart rate, blood pressure, and body temperature values, and perform serialization rearrangement based on the vital sign collection timestamps. Then, aggregate and normalize the rearranged values ​​according to the time index to generate a vital sign sequence.

[0015] S102: Call the vital signs sequence, obtain medical terminology text through a medical institution, extract medical terminology units, perform mapping calculations between the extracted character sequence and the medical terminology embedding vector, and perform alignment calculations between the mapping result and the vital signs sequence to obtain the terminology embedding alignment sequence.

[0016] S103: Based on the term embedding alignment sequence, call the normalized index in the vital signs sequence, perform weighted aggregation calculation on the vectors of the two sequences at the same time index position, and integrate the aggregated vectors in chronological order to establish a preliminary fused medical dataset.

[0017] As a further aspect of the present invention, the specific steps of S2 are as follows:

[0018] S201: Based on the medical term embedding vectors in the preliminary fused medical dataset, for each embedding vector's dimension and term semantic annotation value, the mutual information calculation method is used to quantize the probability distribution of the dimension and annotation value, perform joint distribution quantization and difference measurement, and generate a semantic contribution matrix.

[0019] S202: Call the semantic contribution matrix, compare the contribution value of each dimension with the contribution threshold, filter the dimensions whose contribution value is greater than the contribution threshold, and aggregate the corresponding indices to obtain the semantic contribution dimension set;

[0020] S203: Based on the semantic contribution dimension set, call the medical term embedding vectors in the preliminary fusion medical dataset, perform consistency marking, mark the differences in the values ​​of embedding vectors from different platforms on the same dimension, bind the platform identifier and dimension index, and finally obtain the optimized medical term embedding set.

[0021] As a further aspect of the present invention, the contribution threshold is determined based on the statistical characteristics, data distribution patterns, and the requirements of the tasks involved in the preliminary fusion of medical terminology embedding vectors in the medical dataset.

[0022] The selection of dimensions with contribution values ​​greater than the contribution threshold involves comparing the contribution value of each dimension with the contribution threshold one by one. If the contribution value is greater than the threshold, the dimension index is included in the semantic contribution dimension set.

[0023] As a further aspect of the present invention, the specific steps of S3 are as follows:

[0024] S301: Based on the optimized medical term embedding set, select cross-platform semantically consistent medical term embedding vectors, obtain the vector representation in the embedding space, calculate the angle difference between vectors, extract rotation parameters, and generate a rotation parameter array.

[0025] S302: Call the rotation parameter array and the optimized medical term embedding set, perform angle correction based on each vector, adjust the vector value sequence according to the radian value in the rotation parameter array, and aggregate the rotated vectors to obtain the rotation vector matrix;

[0026] S303: Call the rotation vector matrix to perform dimension reconstruction on the vector sequence, integrate adjacent sequences according to the position alignment rules, establish the aligned vector set structure framework, and perform convergence processing on the vectors to obtain the aligned embedded vector data structure.

[0027] As a further aspect of the present invention, the specific steps of S4 are as follows:

[0028] S401: Based on the aligned embedded vector data structure and the timestamp sequence in the patient's vital signs, perform a matching judgment operation to obtain the time interval difference. Compare the difference with a preset synchronization threshold. If the difference is lower than the threshold, mark the synchronization time point and generate a time synchronization index matrix.

[0029] S402: Call the time synchronization index matrix, traverse the timestamp positions, obtain the corresponding vital sign data frames, calculate the difference between adjacent parameters, and compare it with the preset smoothing benchmark value. When the difference is lower than the benchmark value, perform a weighted operation, aggregate multiple parameter data, and obtain a smoothing parameter vector set.

[0030] S403: Based on the smoothing parameter vector set, perform cross-source aggregation operation on the parameter vector, execute alignment embedding vector data structure mapping calculation, perform concatenation and merging of the mapped data with the original data, perform data frame splicing, and obtain cross-platform joint monitoring data stream.

[0031] As a further aspect of the present invention, the preset synchronization threshold is set based on the sampling period of the timestamp sequence, the accuracy of the acquisition channel, the time step span, and the synchronization link calibration error, and the synchronization limit is set according to the distinguishable interval between adjacent timestamp markers.

[0032] The preset smoothing baseline value is obtained by statistical fitting based on the acquisition noise range, equipment accuracy, parameter fluctuation range and change gradient of vital sign parameters, through the amplitude distribution of the difference between adjacent parameters.

[0033] As a further aspect of the present invention, the specific steps of S5 are as follows:

[0034] S501: Based on the cross-platform joint monitoring data stream, the intelligent risk discrimination model compares the heart rate, blood pressure and body temperature values ​​in vital signs with the corresponding vital sign thresholds at each time step, performs judgment calculations based on the difference magnitude, and generates deviation matrix values.

[0035] S502: Call the deviation matrix value, calculate the similarity between the deviation sequence and the corresponding vital sign threshold sequence based on the distance method, and perform normalization processing based on the similarity result to obtain the associated risk parameter set;

[0036] S503: Based on the associated risk parameter set, determine the magnitude relationship between the risk parameters and the risk judgment benchmark value, filter out risk parameters whose magnitude is greater than the risk judgment benchmark value, aggregate the parameter index and encode it, and generate real-time risk warning status information.

[0037] As a further aspect of the present invention, the vital sign threshold is obtained by collecting the original health data of the target monitoring object and performing statistical analysis. The mean and standard deviation of the collected heart rate data, blood pressure data and body temperature data are calculated respectively, and the interval formed by adding and subtracting the standard deviation multiple of the mean is used as the vital sign threshold range of the corresponding vital signs.

[0038] The risk discrimination benchmark value is determined by performing quantile analysis on the deviation matrix values, extracting samples whose values ​​in the set of associated risk parameters are located in a set percentile interval, and calculating the sample mean as the risk discrimination benchmark value.

[0039] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0040] This invention achieves efficient multi-source data fusion by optimizing medical terminology embedding vectors and adjusting cross-platform data alignment, improving coordination between different platforms and the accuracy of data integration. By utilizing deep analysis of the semantic information of patient vital signs and medical terms, the intelligence level of data monitoring is enhanced, significantly improving the sharing and application of cross-platform data. Based on real-time synchronous processing and intelligent risk assessment, the risk score of the patient's health status is dynamically calculated, avoiding the limitations of static thresholds, improving the accuracy and response speed of the early warning system, and ensuring the timeliness and reliability of risk warnings in complex clinical environments. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the accompanying drawings without creative effort.

[0042] Figure 1This is a schematic diagram of the steps of the present invention;

[0043] Figure 2 This is a detailed schematic diagram of S1 of the present invention;

[0044] Figure 3 This is a detailed schematic diagram of S2 of the present invention;

[0045] Figure 4 This is a detailed schematic diagram of S3 of the present invention;

[0046] Figure 5 This is a detailed schematic diagram of S4 of the present invention;

[0047] Figure 6 This is a detailed schematic diagram of S5 of the present invention. Detailed Implementation

[0048] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0049] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0050] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0051] In this embodiment of the invention, sometimes the subscript such as W1 is written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0052] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0053] Please see Figure 1 This invention provides a method for real-time monitoring and intelligent risk warning of multi-source cross-platform medical data, comprising the following steps:

[0054] S1: Obtain patient vital signs and medical terminology texts from multiple medical institutions, convert the texts into corresponding medical terminology embedding vectors, and combine them with patient vital signs to generate a preliminary fused medical dataset.

[0055] S2: Call the preliminary fused medical dataset, use the mutual information calculation method to analyze the semantic information contribution of medical term embedding vectors, filter the dimensions whose contribution meets the contribution threshold, and label the heterogeneous dimensions between different platforms to generate an optimized medical term embedding set.

[0056] S3: Based on the optimized medical terminology embedding set, rotate and adjust the coordinate axes of the embedding vectors for different platforms, align the cross-platform embedding space by calculating the coordinate rotation angle, and output the aligned embedding vector data structure.

[0057] S4: Based on the aligned embedded vector data structure and patient vital signs, perform time synchronization processing, smooth the vital sign parameter data, and obtain a cross-platform joint monitoring data stream;

[0058] S5: Call the cross-platform joint monitoring data stream input to the intelligent risk discrimination model, compare vital signs with preset vital sign thresholds at each time step, calculate the associated risk score based on statistical deviation, and generate real-time risk warning status information.

[0059] The initial fusion of medical datasets includes unified data identification, feature integration domain, and parameter association box; the optimized medical terminology embedding set includes semantic filtering domain, weight distribution set, and structural annotation layer; the aligned embedding vector data structure includes coordinate correspondence table, spatial reference box, and alignment transformation domain; the cross-platform joint monitoring data stream includes temporal continuous band, parameter fusion channel, and joint observation trajectory; and the real-time risk warning status information includes risk indication level, threshold comparison domain, and warning trigger signal.

[0060] Please see Figure 2 The specific steps of S1 are as follows:

[0061] S101: Obtain patient vital sign data frames from multiple medical institutions, perform interval discretization calculations on heart rate, blood pressure, and body temperature values, and perform serialization rearrangement based on the vital sign collection timestamps. Then, aggregate and normalize the rearranged values ​​according to the time index to generate a vital sign sequence.

[0062] By establishing encrypted communication links with servers of multiple medical institutions, a distributed database interface is invoked to read data from the target patient within a specified time period. The raw monitoring data stream contains values ​​for four dimensions: heart rate, systolic blood pressure, diastolic blood pressure, and body temperature. First, interval discretization is performed on the data for each dimension to define the effective heart rate interval. systolic blood pressure effective range Effective range of diastolic blood pressure and the effective range of body temperature The collected data frames are iterated through, and values ​​falling outside the above-mentioned range are identified as noise and removed. For values ​​within the range, they are processed according to the collection timestamp. The sequential execution of serialization and rearrangement integrates the chaotic asynchronous data stream into a monotonically increasing time series table, and sets a time aggregation window. The rearranged data stream is segmented into 60-second intervals. The arithmetic mean of vital signs within each time window is extracted as the representative value for that time index. Then, normalization is performed on the aggregated values, and a normalization baseline upper limit is set. and lower limit The maximum and minimum values ​​of the above valid intervals are respectively obtained by calling the formula:

[0063] ;

[0064] The normalized index for the calculation dimension, taking the average heart rate of 85 bpm actually collected by the patient as an example, is calculated by substituting the parameters. Similarly, if the systolic blood pressure is 130 mmHg, the normalized result is... ,body temperature The normalization result is The calculations for each time node are performed sequentially, and the normalized values ​​are indexed by time. Combined into vector form:

[0065] ;

[0066] The final result is a time-aligned sequence of vital signs.

[0067] S102: Call the vital signs sequence, obtain the medical terminology text from the medical institution, extract the medical terminology units, perform mapping calculations between the extracted character sequence and the medical terminology embedding vector, and perform alignment calculations between the mapping result and the vital signs sequence to obtain the terminology embedding alignment sequence.

[0068] The system retrieves the generated vital sign sequence, reads the time span it covers, and sends a query command to the hospital's electronic medical record system. It then retrieves the chief complaint records, ward round records, and diagnostic logs entered by doctors within that time span. The system uses a natural language processing interface to segment the obtained text, removes stop words, and extracts medical terminology units. A pre-built medical terminology vector database, generated using a Word2Vec model with pre-trained word vector dimensions, is provided. Using the dimension of matching vital sign vectors, the extracted lexical units are traversed and their corresponding word vectors are retrieved from the database, as shown in Table 1. If the word "palpitation" is extracted from the text, its corresponding vector is extracted. If "high heat" is extracted, then extract the vector. The extracted word vectors are bound to the timestamps generated by the text, and time alignment calculations are performed to convert the text timestamps... Time index mapped to vital sign sequence Above, the judgment condition is This means that the text time falls within half the width of the center point of the aggregation window. If multiple terms exist within the same window, vector addition is performed on the vectors of the multiple terms and the average is taken. For example, if "palpitation" and "chest tightness" appear within the same minute (assuming the vector of chest tightness is...), then... Then the merged term embedding vector is:

[0069] ;

[0070] For time windows without text records, the term embedding vector is set to zero, and matching is completed point by point to obtain a term embedding aligned sequence that is completely consistent with the time axis of the vital signs sequence.

[0071] Table 1: Examples of Vector Mapping for Medical Terminology

[0072]

[0073] As shown in Table 1, different medical terms are mapped to four-dimensional vectors with specific physiological correlation weights based on the pre-training results, which are then used for subsequent fusion calculations.

[0074] S103: Based on the term embedding aligned sequence, the normalized index in the vital sign sequence is called, and the vectors at the same time index position of the two sequences are weighted and aggregated. The aggregated vectors are then integrated in chronological order to establish a preliminary fused medical dataset.

[0075] Alignment sequence based on term embedding With vital signs sequence For each time index Perform weighted aggregation calculations on multimodal data and set fusion weight coefficients. and ,in The weights representing objective vital signs data. Weights representing subjective medical text data are set. , The weighting is based on the premise that the real-time performance and accuracy of objective data collected by instruments are higher than those of subjective descriptions, and a fusion calculation formula is constructed accordingly. Substitute the calculated values ​​into the formula and select the vital sign vector at a certain moment. (Corresponding to heart rate 85, systolic blood pressure 130, etc.) and terminology embedding vector (Corresponding to palpitations and chest tightness), perform element-wise weighted summation to calculate the fusion value of the first dimension (heart rate dimension). Calculate the fusion value of the second dimension (blood pressure dimension). Calculate the third dimension fusion value Calculate the fusion value of the fourth dimension (body temperature dimension). The calculated fusion values ​​from the four dimensions are then repackaged into a fusion vector. The results indicate that, after incorporating patients' subjective descriptions of "palpitations and chest tightness," the risk phenotypic values ​​of the original heart rate and blood pressure were appropriately modified. Following this logic, the entire sequence was traversed chronologically. The vectors are concatenated in order to establish a preliminary fused medical dataset.

[0076] Please see Figure 3 The specific steps of S2 are as follows:

[0077] S201: Based on the medical term embedding vectors in the preliminary fusion medical dataset, for each embedding vector's dimension and term semantic annotation value, the mutual information calculation method is used to quantize the probability distribution of the dimension and annotation value, perform joint distribution quantization and difference measurement, and generate a semantic contribution matrix.

[0078] Based on the constructed preliminary fused medical dataset, the first step is to call the medical terminology embedding vector matrix from the dataset, and then process each feature dimension of the matrix. (For example, the aforementioned heart rate correlation dimension, blood pressure correlation dimension, etc.) and preset semantic annotation values ​​of terms. Perform association analysis, where semantic annotation values The terms are set as binary classification labels, representing the risk level of the term in a clinical emergency setting. 0 represents low-risk concern, and 1 represents high-risk concern. (Selection of dimensions) That is, the data related to heart rate are discretized, and a discretization threshold is set. Iterate through the vector values ​​in this dimension; if the value is greater than 0.4, map it to 1; otherwise, map it to 0, generating discrete variables. Simultaneously, read the corresponding semantic annotation values ​​to generate variables. Call the joint probability distribution calculation module to perform statistics. and The frequency of simultaneous occurrence is used to estimate the joint probability. and respectively count and marginal probability and As shown in Table 2, the mutual information calculation formula is invoked based on the statistical results:

[0079] ;

[0080] In the formula The joint probability of the representation dimension value and the risk label occurring simultaneously. and Represents the marginal probability of independent occurrence. The advantage of this formula, which is based on the natural logarithm, lies in its ability to measure values ​​within known dimensions. In the case of labels The degree of uncertainty reduction, thus quantifying the semantic contribution of this dimension to risk assessment, is calculated by substituting the actual statistical data in Table 2. In this case, the calculation item is ,for In this case, the calculation item is The remaining cross-items have a probability of 0 and are not included in the accumulation, thus yielding the mutual information value for that dimension. The results show that the heart rate association dimension has a very high dependency on the risk label. Following this logic, the mutual information values ​​of the dimensions are calculated. For example, dimension 2 is 0.455, dimension 3 is 0.112, and dimension 4 is 0.045. The calculation results are arranged by dimension index to generate a semantic contribution matrix.

[0081] Table 2: Joint Probability Distribution Statistics of Semantic Dimensions

[0082]

[0083] As shown in Table 2, this table displays the probability distribution sampling data when calculating mutual information for the heart rate correlation dimension. The total number of samples is 500, and the joint probability is obtained based on the sample frequency normalization and used for the weight calculation of the subsequent mutual information formula.

[0084] S202: Call the semantic contribution matrix, compare the contribution value of each dimension with the contribution threshold, filter the dimensions whose contribution value is greater than the contribution threshold, and aggregate the corresponding indices to obtain the semantic contribution dimension set;

[0085] The set of dimensional mutual information values ​​recorded in the semantic contribution matrix is ​​used to retrieve the values. To initiate the thresholding process, first calculate the arithmetic mean of the set. As a basic statistical feature, the operation is performed. Next, calculate the standard deviation. First calculate the variance ,but Construct a contribution threshold calculation formula ,in As the adjustment coefficient, set If the average value is used as the baseline for screening, then This threshold setting process dynamically defines the boundary of effective information based on statistical distribution patterns, avoiding the failure problem of fixed thresholds under different data distributions. It performs numerical comparison and filtering operations, and for dimension 1, it determines... Keep index 1, and for dimension 2, make a decision. Keep index 2, and for dimension 3, make a decision. Remove index 3, and for dimension 4, make a judgment. Remove index 4, aggregate the index values ​​that meet the conditions, and generate a containing index. The set of semantic contribution dimensions.

[0086] S203: Based on the semantic contribution dimension set, call the medical term embedding vectors in the preliminary fused medical dataset, perform consistency marking, mark the differences in the values ​​of embedding vectors from different platforms on the same dimension, bind the platform identifier and dimension index, and finally obtain the optimized medical term embedding set.

[0087] Based on semantic contribution dimension set The feature space requiring optimization was identified, and data sources from different medical platforms were connected, including the internal database of a top-tier hospital (identifier ID: ) and regional medical cloud platform (ID: For the same medical term (such as "tachycardia"), an embedding vector is extracted, denoted as . and Consistency checks are performed only on the selected Dimension 1 and Dimension 2, and a formula for calculating the difference is defined. ,in and Representing the two platforms in terms of dimensions The value on, This formula, based on absolute value calculations, aims to quantify the degree of deviation between different data sources in representing the same semantic meaning on key dimensions. Excessive deviation indicates source noise or inconsistent standards, thus setting a consistency tolerance benchmark. This benchmark value is set based on the allowable error range of the normalized original data. Substituting the actual values, assuming that on dimension 1... , ,but ,determination Mark this dimension as having platform consistency, retain and perform weighted fusion, assuming dimension 2... , ,but ,determination If a platform conflict exists in this dimension, it will be automatically invoked, and a higher confidence level will be called. The platform data covers this dimension value and is bound to the platform identifier. Together with dimension index 2, the final output is an optimized medical terminology embedding set after consistency verification and correction.

[0088] Please see Figure 4 The specific steps of S3 are as follows:

[0089] S301: Based on the optimized medical term embedding set, select the cross-platform semantically consistent medical term embedding vectors, obtain the vector representation in the embedding space, calculate the angle difference between vectors, extract rotation parameters, and generate a rotation parameter array;

[0090] Based on the optimized medical terminology embedding set based on the output, the cross-platform mapping metadata structure corresponding to the dataset is first read to identify semantically consistent medical term pairs across different platforms, such as "elevated heart rate," "high blood pressure," and "decreased blood oxygen." For each semantically consistent term pair, its representation vector in the embedding space is read from platform A and platform B respectively, denoted as […]. The aforementioned vector pairs constitute a set of reference points for cross-platform geometric alignment. To quantify the structural offset between two embedding spaces at the same semantic point, a cross-space cosine angle calculation is performed for each pair of reference vectors to obtain rotation parameters reflecting geometric differences. This process utilizes the standard dot product formula, calling the dot product operation of the vector angle. The advantage of this formula is that it transforms the numerical correlation between two feature sequences in high-dimensional space into an intuitive geometric angle, thereby accurately extracting the rotation parameters that need to be decorrelated or aligned, and setting the sampling length. Substitute into a real-world example and extract... The first 5 normalized values ​​are ,extract The first 5 values ​​are Perform dot product operation ,calculate norm ,calculate norm Substituting into the formula, we get (For ease of explanation, only the first 5 data points are shown here. If calculated based on the full N=100 data set, due to noise accumulation, the cross-platform correlation value commonly seen in actual monitoring is...) ),but Radius, calculated similarly. 47 radians The calculated angle values ​​are filled into the corresponding cells of the angle difference matrix, as shown in Table 3. After traversing the cross-platform semantic term pairs, the angle values ​​are extracted from the matrix as rotation parameters to generate a rotation parameter array for subsequent spatial transformations.

[0091] Table 3: Calculation of Dimensional Sequence Correlation and Rotation Angle

[0092]

[0093] As shown in Table 3, the data in the table is calculated based on the time series of 100 sampling points. The rotation parameter directly corresponds to the geometric angle that needs to be adjusted in the feature space and is used to guide the rotation operation of the vector.

[0094] S302: Call the rotation parameter array and optimize the medical term embedding set, perform angle correction based on each vector, adjust the vector value sequence according to the radian value in the rotation parameter array, and aggregate the rotated vectors to obtain the rotation vector matrix;

[0095] Calling the rotation parameter array And the pre-generated optimized medical terminology embedding set, for each time step in the embedding set. Four-dimensional feature vectors A hierarchical rotation correction is performed, a process that utilizes the Givens rotation matrix principle to sequentially eliminate redundant correlations between adjacent dimensions, constructing a two-dimensional plane rotation formula. Taking the rotation of a plane as an example, let's define the rotation matrix:

[0096] ;

[0097] Among them, the correction angle Set as rotational parameter The negative value is used to perform inverse decorrelation, i.e. Calculate the values ​​of trigonometric functions , The advantage of the formula lies in redistributing the feature weights through linear transformation, making the transformed dimension closer to the independent principal components. This is achieved by selecting a vector at a specific time point. The first two components Substitute into the calculation, and the updated components Updated components The results show that after rotation, the main energy is concentrated in the first dimension, and the value in the second dimension is significantly reduced. After completing the first level of correction, the second rotation parameter is called. Perform the same matrix operations on the second and third dimensions of the updated vector, and repeat this logic recursively until the correction of adjacent dimensions is completed, thus transforming the vector after multiple levels of rotation. Re-aggregate according to the original time index to form a rotated vector matrix with orthogonal characteristics.

[0098] S303: Call the rotation vector matrix to perform dimension reconstruction on the vector sequence, integrate adjacent sequences according to the position alignment rules, establish the aligned vector set structure framework, and perform convergence processing on the vectors to obtain the aligned embedded vector data structure.

[0099] Call the rotation vector matrix The algorithm performs dimension reconstruction and convergence evaluation on the vector sequence within the matrix. The alignment rule for the vector set structure is set as the "maximum energy compression ratio" principle, requiring the projected variance of the transformed vector set in the low-dimensional subspace to exceed a preset standard. The variance distribution of the reconstructed vector sequence is calculated, and a convergence determination mechanism is introduced. This mechanism quantifies the degree of convergence by calculating the average of the squared Euclidean distances between the current vector set and the preset target sparse vector. A convergence threshold of 0.01 is set. The algorithm iterates through the vector data within the entire time window, subtracting the values ​​of the rotated vector and the target vector at corresponding positions at each time step, and then squaring the result. The sum of these squared differences is divided by the total number of time points to obtain the average error value. Substituting this into a practical example, it assumes... The sum of squared distances is Then the average error value ,determination Once the convergence condition is met, the current vector numerical state is locked, and position alignment and integration are performed. The numerical sequence of the rotated dimensions is strictly aligned according to the timestamp, removing minor floating-point errors generated during the calculation process (truncated to 4 decimal places). For example, ... Revised to Establish a standardized aligned embedded vector data structure.

[0100] Please see Figure 5 The specific steps of S4 are as follows:

[0101] S401: Based on the aligned embedded vector data structure and the timestamp sequence in the patient's vital signs, perform a matching judgment operation, obtain the time interval difference, compare the difference with a preset synchronization threshold, and if the difference is lower than the threshold, mark the synchronization time point and generate a time synchronization index matrix.

[0102] Based on the aligned embedding vector data structure of the output, retrieve the corrected time index sequence contained therein. Simultaneously, the original timestamp sequence of the patient's vital signs collection device is read from the original database. Set the time pointers for the two sequences respectively. and Perform time-aligned matching and judgment operations for each embedding vector's time point. Search for the closest time point in the vital signs sequence. Call the difference calculation formula The absolute difference in time interval between the two is calculated. This calculation aims to quantify the degree of offset between the processed high-dimensional feature data and the original physiological signal on the time axis, and a preset synchronization threshold is set. The determination of this threshold is based on the sampling cycle of the vital signs monitoring equipment. ms (i.e., 1Hz frequency), hardware synchronization accuracy of the acquisition channel milliseconds and calibration error of data transmission link Based on the principle of distinguishable intervals between adjacent timestamps, a threshold calculation formula is constructed. Substituting the numerical values, we get ms, the calculated time interval difference With this threshold Perform a numerical comparison, if determined If the time interval is ms, it confirms that both belong to the same valid monitoring time, marks this time point as a synchronization state, and sets the corresponding index pair. and offset Write the time synchronization index matrix. Taking a specific match as an example, if... s, s, then s (i.e., 150ms) to determine If the synchronization requirements are met, the entire timeline is traversed according to this logic, and isolated nodes that cannot be matched are removed, and finally a time synchronization index matrix with a defined mapping relationship is generated.

[0103] S402: Call the time synchronization index matrix, traverse the timestamp positions, obtain the corresponding vital sign data frames, calculate the difference between adjacent parameters, and compare it with the preset smoothing benchmark value. When the difference is lower than the benchmark value, perform a weighted operation, aggregate multiple parameter data, and obtain a smoothing parameter vector set.

[0104] Call the valid index pairs recorded in the time synchronization index matrix According to the index Directly locate and extract the corresponding vital sign data frames to obtain the parameter values ​​at the current moment. (e.g., systolic blood pressure value) and the parameter value at the previous moment. Perform the calculation of the difference between adjacent parameters, using the following formula: This step evaluates the instantaneous fluctuations of the signal by calculating the first-order difference and sets a preset smoothing reference value. The baseline value is set based on the noise range of this type of vital sign parameter acquisition. mmHg, equipment measurement accuracy mmHg and the normal physiological fluctuation range of the human body mmHg, set through statistical fitting Substituting the numerical values, we get mmHg, the difference calculated in real time Compared with the benchmark value Compare, if determined This indicates that the current fluctuations are mainly caused by sensor noise or minor interference, requiring weighted aggregation to smooth the data; otherwise, the original values ​​are retained to reflect sudden pathological changes, as shown in Table 4, in the time index. At this point, the original systolic blood pressure abruptly changes from 120 to 122, a difference of 2, which is less than the baseline value of 3.5, triggering the smoothing mechanism and setting the weight of the current value. Original value weight Perform aggregate calculations The above logic is executed one by one on the parameters such as heart rate, blood pressure, and body temperature contained in the data frame, and the processed values ​​are repackaged to obtain a smooth parameter vector set after denoising.

[0105] Table 4: Example Table of Smoothing Processing for Vital Signs Parameters

[0106]

[0107] As shown in Table 4, this table demonstrates how, during continuous monitoring, the data can be dynamically weighted and smoothed or kept at its original mutation value based on the comparison between the calculated difference and the preset baseline value, so as to retain key pathological features while filtering out noise.

[0108] S403: Based on the smooth parameter vector set, perform cross-source aggregation operation on the parameter vector, execute alignment and embedding vector data structure mapping calculation, concatenate and merge the mapped data with the original data, and splice the data frames to obtain cross-platform joint monitoring data stream;

[0109] Based on the smoothing parameter vector set With the generated aligned embedded vector data structure To perform cross-source aggregation, the first step is to parse the dimensional information of the aligned embedded vector data structure to obtain its feature dimensions. At the same time, confirm the feature dimensions of the smoothing parameter vector set. Based on the generated time synchronization index matrix, the same time... Smoothed physiological parameter vectors (representing systolic blood pressure, diastolic blood pressure, body temperature, and blood oxygen saturation, respectively) and medical terminology embedding vectors Perform alignment mapping, execute data frame concatenation, and construct a joint vector using a cascade union approach. This involves concatenating two vectors end-to-end to generate a vector with dimension 1. New data frame:

[0110] ;

[0111] During this process, the concatenated data stream undergoes format validation to ensure that the data type of the numerical fields is uniformly double-precision floating-point, and that the missing term vector portion (i.e., the time point for which there is no corresponding medical record) is represented by an all-zero vector. The process involves filling in and completing data frames sequentially according to time, ultimately outputting a cross-platform joint monitoring data stream that includes objective physiological indicators and subjective semantic features.

[0112] Please see Figure 6 The specific steps of S5 are as follows:

[0113] S501: Based on cross-platform joint monitoring data stream, the intelligent risk discrimination model compares the heart rate, blood pressure and body temperature values ​​in vital signs with the corresponding vital sign thresholds at each time step, performs judgment calculations based on the difference magnitude, and generates deviation matrix values.

[0114] Based on the output cross-platform joint monitoring data stream, the data stream is first decoded and separated to extract real-time numerical sequences containing four key dimensions: heart rate, systolic blood pressure, diastolic blood pressure, and body temperature. Simultaneously, a raw data retrieval request is sent to the hospital's data center to obtain information about the target monitoring object in the past. The original health monitoring records within the day are used to establish an individualized health benchmark database. Statistical feature extraction is performed for each dimension, and the mean calculation formula is called:

[0115] ;

[0116] Formula for calculating standard deviation:

[0117] ;

[0118] Calculate the baseline mean of each physiological parameter. with discrete standard deviation As shown in Table 5, taking systolic blood pressure as an example, the original mean value was calculated. mmHg, standard deviation mmHg, setting dynamic threshold fluctuation coefficient This coefficient is set based on the physiological lethality sensitivity of the parameter, with higher sensitivity for systolic blood pressure. Construct a formula for calculating the threshold range of vital signs. Substituting the values ​​into the numerical calculation, the safe range for systolic blood pressure is... Then, the systolic blood pressure value was read from the real-time data stream. For mmHg, perform an interval comparison to confirm that the value is within the safe range, and then construct a deviation determination formula. The advantage of this formula lies in eliminating evaluation barriers between physiological parameters of different dimensions by mapping the absolute difference to a standardized ratio relative to the individual's range of fluctuation, thus achieving risk quantification in a unified dimension. Substituting the data into the calculation:

[0119] ;

[0120] Similarly, calculate heart rate parameters, such as real-time heart rate. bpm, raw mean , ,coefficient ,but This result indicates that although the heart rate did not exceed the general threshold for the population, it deviated significantly from the patient's own baseline. Based on this logic, the body temperature was calculated. , , , Calculated The calculated deviation values ​​are arranged by parameter index according to dimensions such as (e.g., ...), generating a result containing... The high-dimensional deviation matrix values.

[0121] Table 5: Statistics and Deviation Calculation of Individualized Physiological Parameters

[0122]

[0123] As shown in Table 5, this table lists the personalized statistical baseline calculated based on 30 days of patient data, as well as the standardized deviation calculated from the real-time data stream. A deviation of more than 1.0 in the center rate suggests a potential abnormal trend.

[0124] S502: Call the deviation matrix value, calculate the similarity between the deviation sequence and the corresponding vital sign threshold sequence based on the distance method, and perform normalization processing based on the similarity results to obtain the associated risk parameter set;

[0125] Call the deviation matrix value A weighted similarity calculation mechanism is introduced to assess the association strength between deviation components and potential risk patterns. First, a risk sign threshold sequence is defined. The sequence is composed of typical deviation features from the medical expert database for acute failure symptoms, such as... (Corresponding to early shock models of rapid heart rate increase, stable blood pressure, and slight increase in body temperature, respectively), targeting real-time deviation vectors With typical risk vector Perform a dimension-wise weighted similarity estimation and construct a normalized matching degree calculation formula:

[0126] ;

[0127] This formula uses a normalized exponential decay model to characterize the matching degree. This is the real-time deviation value. The advantage of the formula, which is the risk threshold, is that it amplifies the weight of the similarity of features through an exponential function, while using a scaling factor to retain the influence of the original amplitude. The values ​​are substituted into the formula for calculation, specifically for the heart rate dimension (index 1). absolute value of difference exponent term proportional term Calculate the normalized similarity score Regarding the blood pressure dimension (index 2). Difference exponent term proportional term ,calculate Regarding the body temperature dimension (index 3). Difference exponent term proportional term ,calculate The calculated dimensional similarity score By integrating these parameters, a set of associated risk parameters is obtained that characterizes the degree of similarity between the current state and the preset risk pattern.

[0128] S503: Based on the associated risk parameter set, determine the size relationship between the risk parameters and the risk judgment benchmark value, filter out risk parameters whose magnitude is greater than the risk judgment benchmark value, aggregate the parameter index and encode it, and generate real-time risk warning status information;

[0129] According to the associated risk parameter set The quantile analysis module is activated to determine the dynamic risk discrimination benchmark value. It retrieves the associated risk parameter records of confirmed cases from the original database and extracts the sample size. The risk distribution data was used to perform numerical sorting and percentile location, and the high-risk attention interval was set as the percentile. to Extract the sample values ​​within this interval and calculate their arithmetic mean as the risk assessment benchmark. Construct the benchmark value calculation formula:

[0130] ;

[0131] in For the original high-risk samples falling within the top 25% quantile interval, assume the mean of the higher quantiles of the original data is... This setting uses statistical long-tail distribution characteristics to lock in the lower limit of risk requiring immediate response, performs a magnitude relationship determination operation, traverses every value in the current set of associated risk parameters, and for parameter 1 (heart rate association), makes a determination. If the result is true, it is marked as a risk item. For parameter 2 (related to blood pressure), a judgment is made. The result is false and is ignored. For parameter 3 (body temperature correlation), a judgment is made. If the result is true, it is marked as a risk item, and the index of parameters marked as true is aggregated. The status code is generated according to the preset binary encoding rules. The encoding length is set to 3 bits, corresponding to 3 dimensions. The risk bits are mainly set to 1 and minor bits are set to 0. And convert it into a hexadecimal risk signature code. The results showed that the patient's current heart rate and body temperature exhibited a high degree of risk correlation, and the final output included real-time risk warning status information containing the feature code and the corresponding risk level description.

[0132] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for real-time monitoring and intelligent risk warning of multi-source cross-platform medical data, characterized in that, Includes the following steps: S1: Obtain patient vital signs and medical terminology texts from multiple medical institutions, convert the texts into corresponding medical terminology embedding vectors, and combine them with patient vital signs to generate a preliminary fused medical dataset. The specific steps of S1 are as follows: S101: Obtain patient vital sign data frames from multiple medical institutions, perform interval discretization calculations on heart rate, blood pressure, and body temperature values, and perform serialization rearrangement based on the vital sign collection timestamps. Then, aggregate and normalize the rearranged values ​​according to the time index to generate a vital sign sequence. S102: Call the vital signs sequence, obtain medical terminology text through a medical institution, extract medical terminology units, perform mapping calculations between the extracted character sequence and the medical terminology embedding vector, and perform alignment calculations between the mapping result and the vital signs sequence to obtain the terminology embedding alignment sequence. S103: Based on the term embedding alignment sequence, call the normalized index in the vital signs sequence, perform weighted aggregation calculation on the vectors of the two sequences at the same time index position, and integrate the aggregated vectors in time order to establish a preliminary fused medical dataset; S2: Call the aforementioned preliminary fused medical dataset, use the mutual information calculation method to analyze the semantic information contribution of the medical term embedding vectors, filter the dimensions whose contribution meets the contribution threshold, and label the heterogeneous dimensions between different platforms to generate an optimized medical term embedding set. The specific steps of S2 are as follows: S201: Based on the medical term embedding vectors in the preliminary fused medical dataset, for each embedding vector's dimension and term semantic annotation value, the mutual information calculation method is used to quantize the probability distribution of the dimension and annotation value, perform joint distribution quantization and difference measurement, and generate a semantic contribution matrix. S202: Call the semantic contribution matrix, compare the contribution value of each dimension with the contribution threshold, filter the dimensions whose contribution value is greater than the contribution threshold, and aggregate the corresponding indices to obtain the semantic contribution dimension set; S203: Based on the semantic contribution dimension set, call the medical term embedding vectors in the preliminary fused medical dataset, perform consistency marking, mark the differences in the values ​​of embedding vectors from different platforms on the same dimension, bind the platform identifier and dimension index, and finally obtain the optimized medical term embedding set. S3: Based on the optimized medical terminology embedding set, rotate and adjust the coordinate axes of the embedding vectors for different platforms, align the cross-platform embedding space by calculating the coordinate rotation angle, and output the aligned embedding vector data structure. The specific steps for S3 are as follows: S301: Based on the optimized medical term embedding set, select cross-platform semantically consistent medical term embedding vectors, obtain the vector representation in the embedding space, calculate the angle difference between vectors, extract rotation parameters, and generate a rotation parameter array. S302: Call the rotation parameter array and the optimized medical term embedding set, perform angle correction based on each vector, adjust the vector value sequence according to the radian value in the rotation parameter array, and aggregate the rotated vectors to obtain the rotation vector matrix; S303: Call the rotation vector matrix to perform dimension reconstruction on the vector sequence, integrate adjacent sequences according to the position alignment rules, establish the aligned vector set structure framework, and perform convergence processing on the vectors to obtain the aligned embedded vector data structure; S4: Based on the aligned embedded vector data structure and the patient's vital signs, perform time synchronization processing, smooth the vital sign parameter data, and obtain a cross-platform joint monitoring data stream; S5: Invoke the cross-platform joint monitoring data stream input to the intelligent risk discrimination model, compare vital signs with preset vital sign thresholds at each time step, calculate the associated risk score based on the statistical deviation, and generate real-time risk warning status information.

2. The method for real-time monitoring and intelligent risk warning of multi-source cross-platform medical data according to claim 1, characterized in that, The preliminary fused medical dataset includes a unified data identifier, a feature integration domain, and a parameter association box. The optimized medical terminology embedding set includes a semantic filtering domain, a weight distribution set, and a structural annotation layer. The aligned embedding vector data structure includes a coordinate correspondence table, a spatial reference box, and an alignment transformation domain. The cross-platform joint monitoring data stream includes a temporal continuous band, a parameter fusion channel, and a joint observation trajectory. The real-time risk warning status information includes a risk indication level, a threshold comparison domain, and a warning trigger signal.

3. The method for real-time monitoring and intelligent risk warning of multi-source cross-platform medical data according to claim 1, characterized in that, The contribution threshold is determined based on the statistical characteristics, data distribution patterns, and the requirements of the tasks involved in the preliminary fusion of medical term embedding vectors in the medical dataset. The selection of dimensions with contribution values ​​greater than the contribution threshold involves comparing the contribution value of each dimension with the contribution threshold one by one. If the contribution value is greater than the threshold, the dimension index is included in the semantic contribution dimension set.

4. The method for real-time monitoring and intelligent risk warning of multi-source cross-platform medical data according to claim 1, characterized in that, The specific steps of S4 are as follows: S401: Based on the aligned embedded vector data structure and the timestamp sequence in the patient's vital signs, perform a matching judgment operation to obtain the time interval difference. Compare the difference with a preset synchronization threshold. If the difference is lower than the threshold, mark the synchronization time point and generate a time synchronization index matrix. S402: Call the time synchronization index matrix, traverse the timestamp positions, obtain the corresponding vital sign data frames, calculate the difference between adjacent parameters, and compare it with the preset smoothing benchmark value. When the difference is lower than the benchmark value, perform a weighted operation, aggregate multiple parameter data, and obtain a smoothing parameter vector set. S403: Based on the smoothing parameter vector set, perform cross-source aggregation operation on the parameter vector, execute alignment embedding vector data structure mapping calculation, perform concatenation and merging of the mapped data with the original data, perform data frame splicing, and obtain cross-platform joint monitoring data stream.

5. The method for real-time monitoring and intelligent risk warning of multi-source cross-platform medical data according to claim 4, characterized in that, The preset synchronization threshold is set based on the sampling period of the timestamp sequence, the accuracy of the acquisition channel, the time step span, and the synchronization link calibration error, and the synchronization limit is set according to the distinguishable interval between adjacent timestamp markers. The preset smoothing baseline value is obtained by statistical fitting based on the acquisition noise range, equipment accuracy, parameter fluctuation range and change gradient of vital sign parameters, through the amplitude distribution of the difference between adjacent parameters.

6. The method for real-time monitoring and intelligent risk warning of multi-source cross-platform medical data according to claim 4, characterized in that, The specific steps of S5 are as follows: S501: Based on the cross-platform joint monitoring data stream, the intelligent risk discrimination model compares the heart rate, blood pressure and body temperature values ​​in vital signs with the corresponding vital sign thresholds at each time step, performs judgment calculations based on the difference magnitude, and generates deviation matrix values. S502: Call the deviation matrix value, calculate the similarity between the deviation sequence and the corresponding vital sign threshold sequence based on the distance method, and perform normalization processing based on the similarity result to obtain the associated risk parameter set; S503: Based on the associated risk parameter set, determine the magnitude relationship between the risk parameters and the risk judgment benchmark value, filter out risk parameters whose magnitude is greater than the risk judgment benchmark value, aggregate the parameter index and encode it, and generate real-time risk warning status information.

7. The method for real-time monitoring and intelligent risk warning of multi-source cross-platform medical data according to claim 6, characterized in that, The vital sign thresholds are obtained by collecting the original health data of the target monitoring object and performing statistical analysis. The mean and standard deviation of the collected heart rate data, blood pressure data and body temperature data are calculated respectively, and the interval formed by adding and subtracting the standard deviation multiple of the mean is used as the vital sign threshold range of the corresponding vital signs. The risk discrimination benchmark value is determined by performing quantile analysis on the deviation matrix values, extracting samples whose values ​​in the set of associated risk parameters are located in a set percentile interval, and calculating the sample mean as the risk discrimination benchmark value.

Images