Blood transfusion adverse reaction risk prediction system based on multi-modal data fusion

Abstract

The invention relates to the technical field of medical data mining, in particular to a blood transfusion adverse reaction risk prediction system based on multi-modal data fusion. Firstly, the blood transfusion rate and monitoring data are obtained; the independent transaction degree of each monitoring feature in the corresponding time period is further obtained, and the abnormal time period is screened out; further combining the abnormal time periods of the two monitoring features, analyzing the change relevance of the independent transaction degrees of the two monitoring features, and obtaining the transaction relevance; further combining the correlation analysis ranges of time domain intersection, analyzing the independent transaction degree and transaction correlation of the monitoring features in the combined range, and obtaining a data support weight in combination with the blood transfusion rate; and finally, the prediction influence weight of the monitoring features is adjusted based on the data support weight, and the blood transfusion risk is predicted. According to the method, through independent transaction analysis, abnormal time period screening, correlation fusion and weight compensation, multi-feature dynamic correlation modeling is enhanced, and the accuracy and individual adaptability of blood transfusion adverse reaction risk prediction are improved.

Description

Technical Field

[0001] This invention relates to the field of medical data mining technology, specifically to a system for predicting adverse transfusion reaction risks based on multimodal data fusion. Background Technology

[0002] Blood transfusion therapy is an important measure for clinical emergency and disease support, but the associated risks of adverse reactions (such as hemolytic reactions, allergic reactions, circulatory overload, etc.) can seriously threaten patient safety. Dynamic changes in the patient's vital signs (such as body temperature, heart rate, blood pressure, respiratory rate, etc.) are often early warning signals of adverse reactions.

[0003] Real-time vital sign monitoring is the core of early warning for adverse transfusion reactions. In existing technologies, the values ​​of actual monitoring characteristic data are directly compared with the standard value range. However, due to the individual differences in the physical condition of different patients, there is a certain deviation in the prediction of adverse reactions during the transfusion process. Furthermore, during the real-time monitoring of vital signs during transfusion, the introduction of foreign blood may cause the patient to have a normal immune response, affecting the risk prediction of adverse transfusion reactions. Summary of the Invention

[0004] To address the technical problem of unsatisfactory risk prediction results during blood transfusion due to patient individuality and the influence of normal immune responses, the present invention aims to provide a blood transfusion adverse reaction risk prediction system based on multimodal data fusion. The specific technical solution adopted is as follows: Monitoring module: Acquires transfusion rate during blood transfusion and monitoring data of various patient characteristics, and performs time-domain segmentation; Anomaly Analysis Module: Based on the poor performance of monitoring data in a certain period, obtain the degree of independent anomaly of each monitoring feature in the corresponding period; based on the temporal distribution of the degree of independent anomaly of each monitoring feature, obtain the abnormal period; combine the abnormal period sets of two monitoring features to form the correlation analysis range; within the correlation analysis range, analyze the correlation of changes in the degree of independent anomaly of the two monitoring features to obtain the anomaly correlation. Compensation analysis module: merges the correlation analysis ranges that intersect in the time domain, and obtains data support weights based on the degree of independent anomaly and the anomaly correlation of the monitoring features corresponding to each merged time domain range, combined with the blood transfusion rate; Prediction module: Based on the data, the prediction influence weight of the monitoring features is adjusted to predict the risk of blood transfusion.

[0005] Furthermore, the monitoring data of the monitoring features are divided into numerical data and textual symptom data.

[0006] Furthermore, the method for obtaining the degree of independent anomaly includes: When the monitoring data is numerical, the degree of independent variation of the monitoring features in the corresponding time period is obtained based on the standard deviation of the data and the mean of the data at the maximum point. When the monitoring data is text-based, the symptoms are segmented to obtain a segmented sequence; based on the TF-IDF value of each segmented sequence in the monitoring data, and combined with the semantic similarity between the segmented sequence and the preset set of segmented sequences of adverse symptoms, the degree of independent change of the monitoring features in the corresponding time period is obtained.

[0007] Furthermore, the method for obtaining the abnormal time period includes: For each of the monitoring features, the degree of independent anomaly is arranged chronologically. The time periods corresponding to the minimum and maximum values ​​in the arrangement, as well as the time periods corresponding to the degree of independent anomaly that are higher than the mean of the degree of independent anomaly, are marked as abnormal segments. Continuous abnormal segments in the time domain are connected to form abnormal time periods.

[0008] Furthermore, the method for obtaining the anomaly correlation includes: Based on the Pearson correlation coefficient of the degree of independent anomaly of the two monitored features, and combined with the difference in the magnitude of the change of the degree of independent anomaly, the anomaly correlation of the two monitored features is obtained.

[0009] Furthermore, the data supports methods for obtaining weights, including: Select one of the merged time domain ranges as the target time domain range, and the set of monitoring features corresponding to the target time domain range is taken as the target set; Within the target set in the target time domain, based on the degree of independent anomaly and the correlation of anomaly for each monitored feature, and in conjunction with the transfusion rate, a transfusion influencing factor is obtained. Based on the temporal fluctuations of the degree of independent aberration of each of the monitored features, the probability of normal immunity is obtained; By integrating the degree of independent anomaly, the correlation of anomaly, the transfusion influencing factors, and the immunization probability of all the monitoring features, the data support weight of each monitoring feature in the target set within the target time domain is obtained.

[0010] Furthermore, the method for obtaining the transfusion influencing factors includes: By integrating the sum of the independent anomalies of all monitored features, the sum of the anomaly correlations between the monitored features, and the average value of the transfusion rate, a transfusion influencing factor for the target time domain is obtained.

[0011] Furthermore, the method for obtaining the probability of normal immunity includes: Within the target set in the target time domain, the probability of normal immunity is obtained based on the difference between the first and last independent anomalies of each monitoring feature and the largest independent anomaly, combined with the difference between the first and last independent anomalies.

[0012] Furthermore, the method for predicting transfusion risk by adjusting the predictive influence weights of the monitoring features based on the data support weights includes: For each monitoring feature, the product of the degree of independent change in the merged time domain range and the data support weight is used as the prediction influence weight, which is then input into the neural network prediction model along with the monitoring data to obtain the prediction result of the risk of adverse transfusion reactions.

[0013] Furthermore, the length of the time-domain segment is 2 minutes.

[0014] The present invention has the following beneficial effects: This invention first acquires transfusion rate and monitoring data and segments them in the time domain to provide an analytical basis and facilitate the analysis of subtle fluctuations in the data. It then acquires the degree of independent variation of each monitoring feature in the corresponding time period, dynamically assessing the individual's actual performance during that period and reducing misjudgments due to patient individuality. Next, it filters out abnormal time periods and merges abnormal time periods from two monitoring features. By filtering and merging their respective abnormal time periods, it obtains the correlation of variation and extracts potential joint abnormal segments, improving the effectiveness of subsequent correlation analysis and the sensitivity of risk identification. Furthermore, it merges the correlation analysis ranges where the time domains intersect, analyzing the degree of independent variation and the correlation of variation of monitoring features within the merged range. Combined with the transfusion rate, it obtains data support weights, reducing the predictive weight of data related to the patient's normal immune response, providing a basis for the final adjustment of predictive weights. Finally, based on the data support weights, it adjusts the predictive influence weights of monitoring features to predict transfusion risk. This invention enhances multi-feature dynamic association modeling through independent variation analysis, abnormal time period screening, correlation fusion, and weight compensation, improving the accuracy and individual adaptability of predicting adverse transfusion reaction risks. Attached Figure Description

[0015] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A system block diagram of a transfusion adverse reaction risk prediction system based on multimodal data fusion provided in one embodiment of the present invention; Figure 2An example diagram of respiratory rate monitoring data provided in one embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the degree of independent aberration in body temperature and blood pressure, provided as an embodiment of the present invention. Figure 4 A flowchart illustrating a method for obtaining data support weights according to an embodiment of the present invention; Figure 5 This is a comparison diagram of adverse transfusion reactions and normal immune responses provided in one embodiment of the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a transfusion adverse reaction risk prediction system based on multimodal data fusion proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0019] The following description, in conjunction with the accompanying drawings, details a specific scheme for a transfusion adverse reaction risk prediction system based on multimodal data fusion provided by the present invention.

[0020] Please see Figure 1 The diagram shows a system block diagram of a transfusion adverse reaction risk prediction system based on multimodal data fusion provided by an embodiment of the present invention. The system includes: a monitoring module 101, an anomaly analysis module 102, a compensation analysis module 103, and a prediction module 104.

[0021] Monitoring module 101: Acquires the transfusion rate during blood transfusion and monitoring data of various patient monitoring characteristics and performs time-domain segmentation.

[0022] Since the transfusion time for patients is not fixed and usually ranges from 1 to 4 hours, directly analyzing and processing all the data that have occurred may overlook a short-term fluctuation. Therefore, it is necessary to appropriately segment the patient's transfusion process over a timescale and divide the physiological indicators of the patient's changes in all continuous segments into fixed-length segments to provide an analytical basis and facilitate the analysis of subtle fluctuations in the data.

[0023] In one embodiment of the present invention, the length of the time domain segment is 2 minutes, and the time domain is segmented with a fixed length; wearable devices or monitors are used to acquire the monitoring characteristic data of the patient during blood transfusion. The monitoring data of each monitoring characteristic includes numerical data, such as respiratory rate, body temperature, blood pressure, heart rate, and SpO2 (blood oxygen saturation); it also includes textual symptom data, such as chills, rash, dyspnea, and other clinical record data (including the recording time to match the corresponding time segment).

[0024] It should be noted that the analysis process is the same for each patient, and only one case is described here; for the earliest time periods in the time domain, it can be set as the data accumulation stage for risk prediction, where only data is collected and not analyzed, such as the first 5 time periods (first 10 minutes) of blood transfusion.

[0025] Please see Figure 2 The diagram illustrates an example of respiratory rate monitoring data provided by an embodiment of the present invention. Figure 2 The horizontal axis represents time, and the vertical axis represents respiratory rate data, with the unit being breaths per minute.

[0026] Anomaly Analysis Module 102: Based on the poor performance of monitoring data in a certain period, obtain the degree of independent anomaly of each monitoring feature in the corresponding period; based on the temporal distribution of the degree of independent anomaly of each monitoring feature, obtain the abnormal period; merge the abnormal periods of two monitoring features as the scope of correlation analysis; within the scope of correlation analysis, analyze the correlation of changes in the degree of independent anomaly of the two monitoring features to obtain the anomaly correlation.

[0027] During blood transfusion, fluctuations or abnormal records in the values ​​of each monitoring characteristic within a certain time period may represent potential physiological stress responses. Therefore, based on the adverse performance of monitoring data within a time period, the degree of independent abnormality of each monitoring characteristic in the corresponding time period is obtained. By quantifying the adverse performance of each characteristic within that time period and emphasizing the independence of characteristic dimensions, the abnormal performance of that characteristic itself can be more accurately reflected. Instead of relying on fixed standard ranges, the actual performance of individuals in that time period is dynamically assessed, reducing misjudgments caused by individual patient differences.

[0028] Preferably, in one embodiment of the present invention, when the monitoring data is numerical, considering that the larger the standard deviation of the data, the more unstable the numerical performance; the larger the value of the data at the maximum point, the more significant the data deviation of the current segment, and the more likely it is to indicate the occurrence of adverse reactions in patients, and the higher the degree of independent variation; therefore, the degree of independent variation of the monitoring feature in the corresponding time period is obtained according to the standard deviation of the data and the mean of the data at the maximum point. As an example, the product of the standard deviation of the data and the mean of the data at the maximum point is used as the degree of independent variation of the corresponding monitoring feature in the corresponding time period. The poor performance of the monitoring data is shown through the standard deviation of the numerical data and the mean of the data at the maximum point.

[0029] When the monitoring data is text-based, it is analyzed using natural language processing technology. First, the symptoms are segmented into word sequences, such as using the existing JieBa word segmentation tool. Considering that the larger the TF-IDF value, the stronger the representativeness and the higher the reference value of the word sequence, the higher the semantic similarity between the word sequence and the preset set of word sequences of adverse symptoms, the more it indicates the occurrence of adverse reactions and the higher the degree of independent change. Based on this, by combining the TF-IDF value of each segmented sequence in the monitoring data with the semantic similarity between the segmented sequence and the set of segmented sequences of preset adverse symptoms, the degree of independent change of the corresponding monitoring feature in the corresponding time period is obtained.

[0030] As an example, the semantic similarity between the segmented word sequence and the preset set of segmented word sequences representing adverse symptoms is obtained using the Jaccard algorithm. The sum of the products of the TF-IDF values ​​and semantic similarities of all segmented word sequences in the monitoring data is used as the degree of independent variation of the corresponding monitoring feature in the corresponding time period. By using the semantic similarity between the segmented word sequences of textual data and the preset set of segmented word sequences representing adverse symptoms, and by applying weighted correction with TF-IDF values, the adverse performance of the monitoring data is represented.

[0031] In another embodiment of the present invention, considering that the greater the difference between the numerical monitoring data and the patient's calm state before transfusion, the greater the patient's reaction, the mean of the data at the maximum point of a time period before transfusion can also be obtained as a reference benchmark value for the numerical monitoring data; the product of the absolute value of the difference between the mean of the data at the maximum point of a time period during transfusion and the reference benchmark value of the corresponding dimension and the standard deviation of the data is used as the degree of independent change of the corresponding monitoring feature in the corresponding time period.

[0032] It should be noted that the natural language processing techniques used are all existing technologies. The pre-defined set of word segments for adverse symptoms includes {"fever", "elevated body temperature", "abnormal body temperature", "high fever"; "chills", "shivering", "trembling", "chills"; "rashes", "itchy skin", "rash", "allergic rash"; "difficulty breathing", "rapid breathing", "shortness of breath", "chest tightness"; "rapid heart rate", "palpitations", "arrhythmia", "tachycardia"; "low blood pressure", "unstable blood pressure", "hypotension", "hypertensive reaction"; "nausea", "vomiting", "stomach discomfort"; "laryngeal edema", "wheezing", "throat tightness"}. The set uses multiple synonymous phrases to describe the same adverse symptom to accommodate different expression habits, which the implementer can adjust as needed. The analysis process is consistent for each time period and will not be repeated.

[0033] Thus, the degree of independent aberration of different monitoring features in different segments on the same time axis has been quantified. However, during the blood transfusion process of patients, if corresponding adverse reactions occur, multiple monitoring features will be associated, and the monitoring data of the features will have certain patterns.

[0034] During blood transfusion, when viewed longitudinally (across multiple monitoring feature dimensions), if the trends of independent changes in the degree of change among the monitoring features are consistent, it is possible to conduct a joint analysis of the patient's physiological changes during the transfusion process. This is beneficial for predicting adverse reactions and has a positive correlation with the risk prediction results of adverse reactions. Therefore, it is advisable to combine different monitoring features for analysis.

[0035] Considering the potential temporal overlap of abnormal responses from different monitoring features during blood transfusion, filtering and merging their respective abnormal time periods can extract potential joint abnormal segments, improving the effectiveness of subsequent correlation analysis and the sensitivity of risk identification. Merging abnormal time periods as the scope of correlation analysis helps focus on key periods where physiological linkages may exist, enhancing the ability of multidimensional monitoring features to collaboratively identify adverse reaction risks.

[0036] Furthermore, considering that the temporal distribution of the degree of independent anomaly reflects the temporal changes in the degree of independent anomaly and some abnormal distribution states, abnormal time periods are obtained based on the temporal distribution of the degree of independent anomaly of each monitoring feature. The abnormal time periods of two monitoring features are then merged and used as the scope of correlation analysis. Within the scope of correlation analysis, the correlation between the changes in the degree of independent anomaly of the two monitoring features is analyzed to obtain anomaly correlation, which facilitates subsequent multimodal data fusion for risk prediction and improves prediction accuracy.

[0037] Preferably, in one embodiment of the present invention, for each monitoring feature, the degree of independent change is arranged in time sequence. Considering that the time period corresponding to the minimum point to the maximum point in the arrangement corresponds to the stage where the degree of independent change continues to rise, reflecting the worsening trend of the adverse reaction, it is marked as an abnormal segment; the time period corresponding to the degree of independent change higher than the mean of the degree of independent change corresponds to the period of significant abnormality, so it is also marked as an abnormal segment. Based on this, abnormal time periods are identified by marking periods with a continuous upward trend and significant anomalies in the time-domain distribution as abnormal segments. The time periods corresponding to the minimum and maximum values ​​in the distribution, as well as the time periods corresponding to the degree of independent anomalies exceeding the mean, are also marked as abnormal segments. Connecting consecutive abnormal segments in the time domain constitutes abnormal time periods.

[0038] It should be noted that when adjacent anomalous segments have intervals (discontinuous) in the time domain, they are divided into different anomalous time periods. Implementers can also limit the anomalous time periods to only those anomalous segments that correspond to the markers from the minimum to the maximum point, so that they are not isolated in the time domain. That is, the length of such anomalous time periods is at least two time periods, in order to eliminate instantaneous noise interference.

[0039] Preferably, in one embodiment of the present invention, please refer to Figure 3 It illustrates a schematic diagram of the degree of independent variability in body temperature and blood pressure provided in an embodiment of the present invention. Figure 3 In the graph, the horizontal axis represents time, and the vertical axis represents the degree of independent variation of the monitored features. The solid line corresponds to body temperature, and the dashed line corresponds to blood pressure. Considering that the Pearson correlation coefficient can quantify the correlation strength between data, the larger the absolute value of the Pearson correlation coefficient, the stronger the correlation between changes. Furthermore, the smaller the difference in the magnitude of the change in the degree of independent variation, the stronger the correlation between changes from the perspective of the magnitude of change. Based on this, the correlation of anomalies between the two monitoring features is obtained by combining the Pearson correlation coefficient of the degree of independent anomaly between the two monitoring features with the difference in the magnitude of the change in the degree of independent anomaly.

[0040] As an example, within the scope of correlation analysis, the absolute value of the Pearson correlation coefficient corresponding to the degree of independent change of two monitoring features is used as the numerator, the range of the degree of independent change is used as the amplitude of change, and the sum of the absolute value of the difference between the amplitudes of change of the two monitoring features and the preset positive parameter divided by zero is used as the denominator. The preset positive parameter divided by zero is 0.01, and the fractional ratio is used as the correlation of change.

[0041] Among them, the data differences are represented by the absolute value of the difference; the correlation of the degree of independent anomaly is represented by the Pearson correlation coefficient and the difference in the magnitude of change.

[0042] It should be noted that the correlation analysis range merging and anomaly correlation analysis calculation are performed for any two monitoring features. This is only described with one example and will not be repeated here. Considering that the abnormal time periods merged by the binary pairs of some monitoring features have multiple segments, such as [20,25], [42,46], [69,74], where the numbers represent the time period numbers, the anomaly correlation of each abnormal time period is calculated. Then, the average of the anomaly correlations of multiple time ranges is taken as the final anomaly correlation, or the anomaly correlations are weighted by the time proportion to obtain the final anomaly correlation. In another embodiment of the present invention, the degree of independent anomaly of each monitoring feature can be linearly normalized within its respective data dimension before analyzing the correlation of anomalies; multiple abnormal time periods can be merged for analysis over a longer time range, such as [20,25], [42,46], and [69,74] being merged into [20,74].

[0043] Compensation Analysis Module 103: Merges the correlation analysis ranges of the intersecting time domains, and obtains the data support weights based on the degree of independent anomaly and the anomaly correlation of the monitoring features corresponding to each merged time domain range, combined with the blood transfusion rate.

[0044] When using neural network-based prediction models to predict the risk of adverse reactions in patients during blood transfusions, the monitoring feature data includes normal immune responses caused by the patient's body receiving foreign blood components. In this case, the weight of such data in the prediction model should be reduced to obtain accurate risk prediction results.

[0045] Considering that a patient's normal immune response usually causes changes in multiple monitoring features simultaneously, the correlation analysis ranges that intersect in the time domain are merged. At this time, the monitoring features corresponding to the correlation analysis ranges are also merged into a single set.

[0046] In one embodiment of the present invention, when merging correlation analysis ranges that intersect in the time domain, the intersection depth of the correlation analysis ranges is also limited. That is, the intersection of two correlation analysis ranges on the time axis is limited to be at least greater than half the time length of one of the correlation analysis ranges. This helps to eliminate false intersections caused only by short-term occasional disturbances, ensuring that the merged ranges have significant temporal overlap, and enhancing the stability and reliability of subsequent multi-feature collaborative analysis. Range merging is transitive. For example, if A and B are merged, and B and C are merged, then the three correlation analysis ranges A, B, and C are ultimately merged. In other embodiments of the present invention, the implementer can adjust the merging restrictions as needed.

[0047] Considering that normal immune responses and adverse reactions differ in the degree of independent dysregulation and the resulting dysregulation correlations, and that normal immune responses are affected by transfusion rate, we obtain data support weights based on the degree of independent dysregulation and dysregulation correlation of the monitoring features corresponding to each combined time domain, combined with transfusion rate, to provide a basis for finally adjusting the prediction weights and improve the accuracy of the prediction results.

[0048] Preferably, in one embodiment of the present invention, please refer to Figure 4 The flowchart illustrates a method for obtaining data support weights according to an embodiment of the present invention, specifically including: Step S301: Select one merged time domain range after another as the target time domain range, and the set of monitoring features corresponding to the target time domain range as the target set.

[0049] First, select the target time domain and target set to facilitate analysis on a case-by-case basis.

[0050] Please see Figure 5 It shows a comparison diagram of adverse transfusion reactions and normal immune responses provided in an embodiment of the present invention. Figure 5 The horizontal axis represents time, and the vertical axis represents the degree of independent variability of a monitored feature. Solid lines correspond to adverse transfusion reactions, and dashed lines correspond to normal immune responses. Figure 5 It can be seen that when the degree of independent abnormality of the monitored characteristics shows a small fluctuation and a corresponding decrease, it may correspond to a normal immune response during the transfusion process. This corresponds to the patient experiencing brief and mild discomfort during the process, but maintaining a stable state of vital signs. Essentially, it is the recipient's immune system's physiological recognition and low-intensity response to foreign components, which falls within the normal range of transfusion-related immune responses.

[0051] Step S302: Within the target set in the target time domain, based on the degree of independent anomaly and the correlation of anomaly of each monitored feature, and combined with the transfusion rate, obtain the transfusion influencing factors.

[0052] Considering that the higher the transfusion rate, the greater the impact on the patient; the larger the sum of the degree of independent dysregulation of the monitoring characteristics, and the smaller the sum of the correlation of dysregulation, the more likely it is that a small number of abnormal change patterns of the monitoring characteristics have occurred, which is more likely to be a normal reaction caused by transfusion, and the greater the transfusion influencing factor.

[0053] Based on this, in one embodiment of the present invention, the sum of the independent anomalies of all monitored features, the sum of the anomaly correlations between monitored features, and the average value of the transfusion rate are fused to obtain the transfusion influencing factor for the target time domain.

[0054] As an example, within the target set of the target time domain, the product of the sum of the independent anomalies of all monitored features and the average transfusion rate is used as the numerator, the sum of the anomaly correlations between all monitored features is used as the denominator, and the fractional ratio is used as the transfusion influencing factor.

[0055] Step S303: Obtain the probability of normal immunity based on the temporal fluctuations of the degree of independent aberration of each monitored feature.

[0056] In any sequence of independent anomalies of a monitoring feature, the smaller the difference between the first (first in time series) data and the last (last in time series) data, the more it reflects that after the fluctuation within the abnormal range, the monitoring feature has rebounded and returned to the initial level. At the same time, the smaller the difference between the first and last independent anomalies of the time series and the largest independent anomaly, the smaller the fluctuation amplitude of the monitoring feature, which is more likely to be a small fluctuation of the monitoring feature caused by a normal immune response. Based on this, in one embodiment of the present invention, within the target set in the target time domain, the probability of normal immunity is obtained by combining the difference between the first and last independent anomalies of each monitoring feature and the largest independent anomaly, and the difference between the first and last independent anomalies.

[0057] As an example, within the target set in the target time domain, the average of the absolute values ​​of the differences between the first and last independent anomalies of each monitoring feature and the largest independent anomaly is taken as the first factor, the absolute value of the difference between the first and last independent anomalies of the time series is taken as the second factor, the reciprocal of the product of the first and second factors is taken as the normal factor for each monitoring feature, and the sum of all normal factors is taken as the normal immune probability.

[0058] Among them, the data difference is represented by the absolute value of the difference, and the time series fluctuation of the degree of independent anomaly is represented by the difference between the first and last data of the time series and the maximum value of the data, combined with the difference between the degree of independent anomaly of the first and last data of the time series.

[0059] Step S304: Integrate the degree of independent anomaly, anomaly correlation, transfusion influencing factors and immunization probability of all monitoring features to obtain the data support weight of each monitoring feature in the target set within the target time domain.

[0060] When a patient consistently maintains a high independent variability in a certain monitoring characteristic during blood transfusion, and the stronger the correlation with changes in other monitoring characteristics, then for the current patient, this persistently high abnormal indicator may indicate the occurrence of adverse transfusion reactions, and the higher the weight of the data support.

[0061] The greater the transfusion-related influencing factors and the greater the likelihood of an immune response, the more likely it is to be a normal immune response triggered by a transfusion, and the lower the weight of the data supporting it. Based on this, within the target set in the target time domain, the product of transfusion influencing factors and immunization probability is used as the denominator, the average value of the independent variability of all monitored features is used as the first supporting factor, the sum of the moving correlations of the binary pairs of all monitored features in the target set within their correlation analysis range is used as the second supporting factor, the product of the first supporting factor and the second supporting factor is used as the numerator, and the fractional ratio is used as the data support weight of the monitored features in the target set within the target time domain.

[0062] It should be noted that the analysis process for each merged time domain range is the same and will not be repeated. The update interval for abnormal periods is one time domain segment. Whenever a new abnormal period appears, the correlation analysis range, abnormal correlation, merged time domain range, and data support weight are updated.

[0063] Prediction Module 104: Based on data-supported weight adjustment of the predictive impact weight of monitoring features, it predicts transfusion risk.

[0064] By using data-supported weights to characterize the monitoring data of the monitoring features within the corresponding time domain, and after assessing their ability to support risk prediction, the predictive influence weights of the monitoring features are adjusted based on the data-supported weights to predict transfusion risks. This allows the model to adaptively focus on the most representative and clinically significant feature data according to the actual monitoring situation, thereby improving the accuracy and robustness of risk prediction.

[0065] Preferably, in one embodiment of the present invention, for each monitoring feature, the product of the degree of independent change in the merged time domain range and the data support weight is used as the prediction influence weight, and is input into the neural network prediction model along with the monitoring data to obtain the prediction result of the risk of adverse transfusion reactions.

[0066] In one embodiment of the present invention, for a monitoring feature that does not exist within a certain merged time domain range, the corresponding prediction influence weight is set to 0; All monitoring feature data and their predictive impact weights are input into a neural network prediction model. Deep learning algorithms are used to automatically mine the complex relationships between the data and deeply integrate multidimensional information to finally obtain the prediction result of adverse transfusion risk at this time. The results of adverse transfusion reaction risk prediction are then imported into the prediction system, which classifies the risk level based on pre-set risk thresholds (e.g., low risk: prediction probability < 10%; medium risk: 10% < prediction probability < 30%; high risk: prediction probability > 30%).

[0067] In one embodiment of the present invention, the neural network prediction model is a multi-layer perceptron (MLP) structure, comprising: An input layer is used to receive the raw monitoring data and the corresponding predicted influence weight for each monitoring feature; There are at least three hidden layers, each containing 64, 128, and 64 neurons, respectively, and the ReLU (Rectified Linear Unit) activation function is used to enhance the nonlinear expressive power of the model. An output layer outputs a probability value between 0 and 1, representing the risk probability of the patient experiencing an adverse transfusion reaction in the current time period. This output layer uses the Sigmoid activation function. During model training, the Binary Cross Entropy Loss function is used as the objective function, the Adam optimizer is used for optimization, the learning rate is set to 0.001, and an early stopping mechanism is used to prevent overfitting. The input features are constructed in vector form. The data value of each monitoring feature and its predicted influence weight together form a set of dual-channel inputs, which are spliced ​​and fused in the model to ensure their joint expression in the data space and weight space. Optionally, to enhance the stability and generalization ability of the model, a BatchNormalization normalization operation can be added before the input layer, and a Dropout layer (with a dropout rate of 0.3) can be added between the hidden layers to reduce the risk of overfitting.

[0068] Neural network prediction models are existing technology. In other embodiments of the present invention, implementers may choose other models for prediction, which will not be elaborated further.

[0069] For patients with different risk levels, the system implements differentiated early warning and intervention strategies: low-risk patients trigger pop-up prompts, reminding medical staff to increase monitoring frequency; medium-risk patients activate the automatic alarm function, simultaneously pushing a detailed early warning report containing basic patient information, abnormal indicators, and risk probability; high-risk patients immediately generate a recommendation to stop transfusion and link with the medical workstation to trigger the emergency plan activation process, prompting medical staff to review the patient's symptoms and prepare for resuscitation measures. Through this hierarchical intelligent early warning mechanism, accurate identification and efficient handling of transfusion risks are achieved, maximizing patient safety.

[0070] It should be noted that this solution is only used to generate warnings and reminders to assist relevant personnel in monitoring the patient's transfusion safety, and does not involve treatment or diagnosis.

[0071] In summary, to address the technical problem of unsatisfactory risk prediction results during blood transfusions due to patient individuality and the influence of normal immune responses, this invention proposes a blood transfusion adverse reaction risk prediction system based on multimodal data fusion. This invention first acquires transfusion rate and monitoring data; then, it acquires the degree of independent variation of each monitoring feature in the corresponding time period, filtering out abnormal time periods; further, it merges the abnormal time periods of two monitoring features, analyzes the correlation of changes in the degree of independent variation of the two monitoring features, and obtains the variation correlation; further, it merges the correlation analysis ranges of intersecting time domains, analyzes the degree of independent variation and variation correlation of monitoring features within the merged range, and obtains data support weights in conjunction with the transfusion rate; finally, it adjusts the predictive influence weights of monitoring features based on the data support weights to predict transfusion risk. This invention enhances multi-feature dynamic correlation modeling through independent variation analysis, abnormal time period screening, correlation fusion, and weight compensation, improving the accuracy and individual adaptability of blood transfusion adverse reaction risk prediction.

[0072] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0073] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A transfusion adverse reaction risk prediction system based on multimodal data fusion, characterized in that, The system includes: Monitoring module: Acquires transfusion rate during blood transfusion and monitoring data of various patient characteristics, and performs time-domain segmentation; Anomaly Analysis Module: Based on the poor performance of monitoring data in a certain period, obtain the degree of independent anomaly of each monitoring feature in the corresponding period; based on the temporal distribution of the degree of independent anomaly of each monitoring feature, obtain the abnormal period; combine the abnormal period sets of two monitoring features to form the correlation analysis range; within the correlation analysis range, analyze the correlation of changes in the degree of independent anomaly of the two monitoring features to obtain the anomaly correlation. Compensation analysis module: merges the correlation analysis ranges that intersect in the time domain, and obtains data support weights based on the degree of independent anomaly and the anomaly correlation of the monitoring features corresponding to each merged time domain range, combined with the blood transfusion rate; Prediction module: Based on the data, the prediction influence weight of the monitoring features is adjusted to predict the risk of blood transfusion.

2. The transfusion adverse reaction risk prediction system based on multimodal data fusion according to claim 1, characterized in that, The monitoring data for the aforementioned monitoring features are divided into numerical data and text-based symptom data.

3. The transfusion adverse reaction risk prediction system based on multimodal data fusion according to claim 2, characterized in that, The method for obtaining the degree of independent anomaly includes: When the monitoring data is numerical, the degree of independent variation of the monitoring features in the corresponding time period is obtained based on the standard deviation of the data and the mean of the data at the maximum point. When the monitoring data is text-based, the symptoms are segmented to obtain a segmented sequence; based on the TF-IDF value of each segmented sequence in the monitoring data, and combined with the semantic similarity between the segmented sequence and the preset set of segmented sequences of adverse symptoms, the degree of independent change of the monitoring features in the corresponding time period is obtained.

4. The transfusion adverse reaction risk prediction system based on multimodal data fusion according to claim 1, characterized in that, The method for obtaining the abnormal time period includes: For each of the monitoring features, the degree of independent anomaly is arranged chronologically. The time periods corresponding to the minimum and maximum values ​​in the arrangement, as well as the time periods corresponding to the degree of independent anomaly that are higher than the mean of the degree of independent anomaly, are marked as abnormal segments. Continuous abnormal segments in the time domain are connected to form abnormal time periods.

5. The transfusion adverse reaction risk prediction system based on multimodal data fusion according to claim 1, characterized in that, The method for obtaining the anomaly correlation includes: Based on the Pearson correlation coefficient of the degree of independent anomaly of the two monitored features, and combined with the difference in the magnitude of the change of the degree of independent anomaly, the anomaly correlation of the two monitored features is obtained.

6. The transfusion adverse reaction risk prediction system based on multimodal data fusion according to claim 1, characterized in that, The data supports methods for obtaining weights, including: Select one of the merged time domain ranges as the target time domain range, and the set of monitoring features corresponding to the target time domain range is taken as the target set; Within the target set in the target time domain, based on the degree of independent anomaly and the correlation of anomaly for each monitored feature, and in conjunction with the transfusion rate, a transfusion influencing factor is obtained. Based on the temporal fluctuations of the degree of independent aberration of each of the monitored features, the probability of normal immunity is obtained; By integrating the degree of independent anomaly, the correlation of anomaly, the transfusion influencing factors, and the immunization probability of all the monitoring features, the data support weight of each monitoring feature in the target set within the target time domain is obtained.

7. The transfusion adverse reaction risk prediction system based on multimodal data fusion according to claim 6, characterized in that, The methods for obtaining the transfusion-affecting factors include: By integrating the sum of the independent anomalies of all monitored features, the sum of the anomaly correlations between the monitored features, and the average value of the transfusion rate, a transfusion influencing factor for the target time domain is obtained.

8. The transfusion adverse reaction risk prediction system based on multimodal data fusion according to claim 6, characterized in that, The method for obtaining the probability of normal immunity includes: Within the target set in the target time domain, the probability of normal immunity is obtained based on the difference between the first and last independent anomalies of each monitoring feature and the largest independent anomaly, combined with the difference between the first and last independent anomalies.

9. A transfusion adverse reaction risk prediction system based on multimodal data fusion according to claim 1, characterized in that, The method for predicting transfusion risk by adjusting the predictive influence weights of the monitoring features based on the data support weights includes: For each monitoring feature, the product of the degree of independent change in the merged time domain range and the data support weight is used as the prediction influence weight, which is then input into the neural network prediction model along with the monitoring data to obtain the prediction result of the risk of adverse transfusion reactions.

10. A transfusion adverse reaction risk prediction system based on multimodal data fusion according to claim 1, characterized in that, The length of the time-domain segment is 2 minutes.

Images