Prewarning and judging method for organ function damage of sepsis patient based on machine learning

Abstract

The invention discloses a sepsis patient organ function damage early warning judgment method based on machine learning, and the method comprises the steps: obtaining the multi-modal clinical data of a sepsis patient, including vital sign time sequence data, inspection data, treatment intervention data and static patient basic information; after data standardization processing, organ function associated features are extracted through a multi-scale feature fusion strategy, and an organ-level feature set is constructed; the method comprises the following steps of: obtaining a multi-organ collaborative early warning model, inputting the multi-organ collaborative early warning model into a pre-trained multi-organ collaborative early warning model, respectively constructing an inter-organ compensatory relation map by double branches of the model, quantifying organ injury risk contribution degree, and outputting a multi-organ functional injury risk matrix by combining with attention mechanism weighted fusion; and generating an early warning result based on the risk matrix and the organ specificity early warning threshold, including the injury risk level of each organ and the dominant risk key feature identifier, monitoring data update in real time, and dynamically adjusting the interval to update the early warning result. According to the invention, the limitation of traditional single-source data and single-organ early warning is broken through, and the early warning accuracy and real-time performance are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical early warning, in particular to a method for early warning and judgment of organ function injury of sepsis patients based on machine learning. BACKGROUND

[0002] Sepsis is a silent killer in the field of clinical critical illness, and its core harm is that it is often accompanied by multiple organ function injury in the progression of the disease, and the injury occurs with the characteristics of concealment and suddenness. Once obvious clinical symptoms appear, it has entered the irreversible stage, resulting in a significant increase in patient mortality. According to epidemiological data, about 40%-50% of sepsis patients worldwide will develop at least one organ dysfunction, and the mortality rate of patients with multiple organ failure is more than 60%. In clinical practice, medical staff traditionally rely on static laboratory indicators (such as serum creatinine and bilirubin) and subjective experience to judge the organ function status, but such indicators have the defects of long detection interval and strong lag, which makes it difficult to capture the dynamic change trend of organ function in real time, especially to early warn potential damage in the compensatory phase, resulting in delayed intervention opportunity and missed best treatment window.

[0003] The existing organ function injury early warning technology has obvious limitations in data processing, which is difficult to adapt to the multi-modal characteristics of sepsis patient clinical data. The clinical data of sepsis patients include continuous high-frequency vital signs (such as heart rate and blood pressure), intermittent laboratory test results (such as blood routine and biochemical indicators), dynamically adjusted treatment interventions (such as fluid resuscitation and vasoactive drug use), and static basic information (such as age and underlying diseases). The collection frequency and dimension characteristics of different types of data differ greatly. Current technology mostly uses single data source or simple splicing to construct features, failing to fully explore the correlation between different modal data, such as ignoring the dynamic influence of treatment interventions on organ function indicators, or not considering the synergistic effect of short-term fluctuations and long-term trends in time series data on injury warning, resulting in insufficient feature expression ability and limited accuracy and generalization of the early warning model.

[0004] In the model construction layer, traditional early warning methods generally lack consideration of the core pathological mechanism of multi-organ synergistic effect, making it difficult to accurately quantify the correlation between organ compensation and injury. Under sepsis state, human organs do not exist in isolation, but form a complex compensatory network through neural-humoral regulation. For example, kidney function injury causes compensatory blood pressure rise in the circulatory system, which in turn increases the cardiac load, forming a chain reaction of "injury-compensation-secondary injury". Existing models mostly use single-organ independent early warning mode, which only assesses the risk based on the changes in single organ indicators, and cannot identify the mutual influence between organs, resulting in misjudgment due to overestimation of organ risk in compensatory state, and missed judgment due to neglect of the transmission effect of other organ injury, which is difficult to meet the demand of clinical overall risk assessment of multiple organs.

[0005] With the deepening of the application of machine learning technology in the medical field, although some studies attempt to use it for sepsis-related early warning, there are still shortcomings in clinical practicability and dynamic adaptability. On the one hand, most models rely on fixed early warning thresholds and update cycles, without considering individual differences of patients (such as the influence of underlying diseases and age on organ tolerance) and dynamic changes in the condition, such as using the same update frequency for stable patients and high-risk patients, which not only increases the cost of invalid calculation, but also causes early warning lag due to untimely updates in the high-risk stage. On the other hand, the output results of existing models are mostly single risk probability values, lacking identification of dominant risk characteristics, making it difficult for medical staff to quickly locate the key indicators that cause the risk to rise (such as whether the risk of kidney function in a patient is due to creatinine fluctuation or reduced urine output), and unable to provide clear guidance for precise intervention, limiting the transformation and landing of technology to clinical practice.

[0006] Chinese patent (publication number: CN120823994A) discloses an AI-based early warning method and system for severe patients with sepsis. Although it analyzes the multi-organ interaction mechanism through a dynamic organ-level causal network, integrates high-frequency physiological waveforms, asynchronous laboratory indicators, and treatment intervention data, and adjusts the early warning threshold based on treatment coverage and generates a hierarchical clinical action chain, it has made some progress in the field of early warning of sepsis, but still has limitations. Its multi-modal data fusion focuses on local and long-range feature extraction of high-frequency signals and causal network topology transformation, without designing a multi-scale feature fusion strategy for the organ function correlation characteristics of multi-modal data of sepsis patients. Moreover, when quantifying organ damage risk, it does not combine attention mechanisms to accurately weight and fuse organ compensatory relationships and risk contributions, making it difficult to output a multi-organ function damage risk matrix and dominant risk key feature identification that better meets clinical needs, and unable to fully meet the actual needs of precise early warning of organ function damage in sepsis patients. Therefore, there is an urgent need for a more optimized machine learning-based early warning method for organ function damage in sepsis patients. SUMMARY

[0007] Based on the above technical problems, the present application discloses a machine learning-based early warning method for organ function damage in sepsis patients, specifically including:

[0008] Obtaining multi-modal clinical data of sepsis patients, including continuous monitoring of vital sign time series data, intermittent collection of laboratory test data, dynamic recording of treatment intervention data, and static storage of patient basic information;

[0009] The multi-modal clinical data is standardized, organ function association features are extracted through a multi-scale feature fusion strategy, and an organ-level feature set is constructed; the multi-scale feature fusion strategy includes extracting short-term fluctuation features and long-term trend features from vital sign time series data, calculating index coefficient of variation and dynamic growth rate from laboratory test data, and constructing intervention effect association features from treatment intervention data;

[0010] The organ-level feature set is input into a pre-trained multi-organ collaborative early warning model, the multi-organ collaborative early warning model constructs an inter-organ compensation relationship graph through a first branch, quantifies organ damage risk contribution degree through a second branch, combines attention mechanism to weightedly fuse two types of outputs, and outputs a multi-organ function damage risk matrix within a preset time length in the future of the patient;

[0011] Based on the multi-organ function damage risk matrix and organ-specific early warning threshold, an organ function damage early warning judgment result is generated, the early warning judgment result includes damage risk levels of each organ and key feature identifiers of dominant risks;

[0012] Real-time monitoring of the update of the multi-modal clinical data, re-executing feature extraction and risk matrix calculation every interval dynamic adjustment time, and updating the organ function damage early warning judgment result.

[0013] Preferably, the standardized processing of the multi-modal clinical data includes: correcting vital signs and laboratory test data by using a normalization method based on patient individual baseline values, mapping data of different collection frequencies to a unified time axis by using a time series alignment algorithm, filling short-time data missing by using an interpolation method, and processing data exceeding a physiologically reasonable range by using an abnormal value correction rule based on clinical logic;

[0014] The abnormal value correction rule based on clinical logic for processing data exceeding a physiologically reasonable range is specifically: calculating a corrected data for vital sign data, the formula is: , wherein is an abnormal value exceeding a physiological range, is normal data at a previous collection time of the abnormal value, is normal data at a next collection time of the abnormal value, , are respective weight coefficients.

[0015] Preferably, the intervention effect association features constructed from the treatment intervention data are specifically: obtaining treatment intervention data to calculate intervention effect features, the formula is: , wherein is a value of an organ function association index before intervention, is a mean value of the index within a preset time length after intervention, is the intervention duration; when there are multiple interventions, the total intervention effect feature is calculated, and the formula is: wherein is the total number of interventions, is the weight of the th intervention, is the effect feature of the th intervention.

[0016] Preferably, the first branch of the multi-organ collaborative early warning model constructs an inter-organ compensation relationship graph, including: calculating the Euclidean distance between the th and the th organ function feature sequence by the formula , wherein , are the th function feature of the th organ, is the number of function features; constructing an adjacency matrix, and the formula is: wherein is the distance decay coefficient, is the adjacency matrix element; inputting the adjacency matrix into a graph convolution network to output an inter-organ compensation relationship strength matrix, and the matrix element represents the compensation influence strength of the th organ function change on the th organ.

[0017] Preferably, the second branch of the multi-organ collaborative early warning model quantifies the organ damage risk contribution degree, including: calculating an initial risk value, and the formula is: wherein is the normal reference value of the th organ function feature, is the weight of the th feature; combining the inter-organ compensation relationship to calculate a compensation risk value, and the formula is: wherein is the total number of organs, is the compensation strength of the th organ on the th organ, is the quantified damage risk contribution degree of the th organ.

[0018] Preferably, the combination attention mechanism weights and fuses two types of outputs, including: calculating a local correlation weight, and the formula is: wherein is the organ compensation relationship strength, is the th organ compensation risk value; calculating a global attention weight, and the formula is: ,in For the first The mean of local correlation weights of each organ; through the formula The fusion features are obtained, among which For the first The compensatory relationship feature vector of an individual organ Features are fused to target specific features.

[0019] Preferably, determining the organ-specific early warning threshold includes: calculating a primary threshold, using the formula: The formula for calculating the secondary threshold is: ,in The first in historical data The mean risk value of each organ under normal conditions For the corresponding standard deviation, , These are the corresponding threshold coefficients; the thresholds are updated based on newly included clinical data, using the following formula: ,in For historical data volume, For new data volume, The old threshold, , The mean and standard deviation of the new data. The threshold is used to update the coefficients.

[0020] Preferably, the key feature identifier of the dominant risk includes: calculating feature importance, using the formula: ,in For the first The minute change in risk value after organ compensation For the first The first organ The minute changes in each characteristic For this feature value; select Features with the highest absolute values ​​are designated as key features, and the direction of feature change is indicated.

[0021] Preferably, the re-execution of feature extraction and risk matrix calculation at each dynamically adjusted time interval includes: calculating the average risk of multiple organs, using the formula: The adjustment interval is determined by the following formula: ,in , These are the minimum and maximum update intervals, respectively. , These are the minimum and maximum values ​​of the average risk in historical data, respectively; when the risk value of any organ exceeds the secondary threshold, the update interval is forcibly set to [value missing]. .

[0022] Preferably, the output patient future preset length of time multi-organ function damage risk matrix comprises: The input multi-output full connection layer is calculated by formula The damage risk probability of the first Organ, wherein Is the weight vector corresponding to the first Organ, Is a bias term; the risk probabilities of all organs are arranged in rows or columns to form a multi-organ function damage risk matrix, and each element in the matrix corresponds to the future damage risk probability of a single organ.

[0023] Compared with the prior art, the technical scheme of the present application has the following technical effects:

[0024] The present application effectively breaks through the limitations of traditional single-source data through multi-modal clinical data integration and multi-scale feature fusion strategy, realizes deep extraction of organ function correlation characteristics of sepsis patients, and specifically processes different types of data characteristics, captures short-term fluctuations and long-term trends from vital sign time series data, combines the variation coefficient of laboratory test data and the effect characteristics of treatment intervention data, and constructs a comprehensive organ-level feature set.

[0025] The present application realizes accurate quantification of compensatory relationship and damage risk among organs relying on the double-branch design and attention fusion mechanism of the multi-organ collaborative early warning model; the first branch constructs a compensatory relationship graph to clearly determine the mutual influence strength of organs, the second branch combines the compensatory relationship to correct the risk value, and then dynamically adjusts the weights of the two types of outputs through attention weights, so as to generate a multi-organ function damage risk matrix.

[0026] The present application realizes clinical accurate interpretation and targeted guidance of early warning results based on organ-specific early warning threshold and key feature identification technology, dynamically updates the threshold through historical data and new data, adapts to individual differences and changes in patient condition, and ensures the scientificity of the threshold; at the same time, the feature importance is calculated to locate the dominant risk features, and medical staff can quickly determine the source of risk.

[0027] The application balances the real-time of early warning determination and the high efficiency of resource utilization by means of a dynamically adjusted update interval mechanism, calculates the update interval according to the average value of multi-organ risk, automatically shortens the update interval to the minimum interval to ensure real-time early warning in a high-risk state, appropriately prolongs the update interval to reduce invalid calculation in a low-risk state, avoids early warning lag in a high-risk stage, reduces the consumption of computing resources compared with fixed cycle update, ensures that the model continuously and stably plays a role in a clinical scene, and improves the practicality of the technology.

[0028] The above description is only a summary of the technical solutions of the present application. In order to more clearly understand the technical means of the present application, the contents of the specification can be implemented, and in order to make the above and other purposes, characteristics and advantages of the present application more obvious and easy to understand, the following will be described in detail with the preferred embodiments of the present application and with the help of the accompanying drawings.

[0029] According to the detailed description of the specific embodiments of the present application in the following text combined with the accompanying drawings, those skilled in the art will more clearly understand the above and other purposes, advantages and characteristics of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0030] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiment or the prior art description will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creating any creative labor. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, each element or part is not necessarily drawn according to the actual proportion.

[0031] According to the description of the drawings in the file and the corresponding technical content, the titles of the drawings are as follows:

[0032] Figure 1 Flow chart of organ function injury early warning determination method based on machine learning for sepsis patients;

[0033] Figure 2 Schematic diagram for generating double-branch structure and risk matrix of multi-organ collaborative early warning model;

[0034] Figure 3 Overall architecture and data interaction diagram of sepsis organ function injury early warning determination system;

[0035] Figure 4 Comparison line graph of multi-organ injury early warning accuracy rate with early warning time point change;

[0036] Figure 5 Scatter plot of early warning lead time and 28-day survival rate;

[0037] Figure 6A heat map is generated for key signatures of multi-organ functional impairment. DETAILED DESCRIPTION

[0038] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments of the present application. In the following description, specific details such as specific configurations and components are provided only to help a comprehensive understanding of the embodiments of the present application. Therefore, those skilled in the art should clearly understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present application. In addition, in order to be clear and concise, the description of known functions and structures is omitted in the embodiments.

[0039] It should be understood that the "one embodiment" or "the embodiment" mentioned throughout the specification means that the specific features, structures or characteristics related to the embodiment are included in at least one embodiment of the present application. Therefore, "one embodiment" or "the embodiment" appearing throughout the specification does not necessarily mean the same embodiment. In addition, these specific features, structures or characteristics can be combined in one or more embodiments in any suitable manner.

[0040] In addition, reference numerals and / or letters can be repeated in different examples in the present application. Such repetition is for the purpose of simplification and clarity, and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0041] The term "and / or" herein is only a description of the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B can mean that A exists alone, B exists alone, and A and B exist together. The term "and" herein is a description of another association relationship of the associated objects, which means that there can be two relationships, for example, A and B can mean that A exists alone and A and B exist together. In addition, the character " / " herein generally means that the associated objects before and after are in an "or" relationship.

[0042] The term "at least one" herein is only a description of the association relationship of the associated objects, which means that there can be three relationships, for example, at least one of A and B can mean that A exists alone, A and B exist together, and B exists alone.

[0043] It should also be noted that the relationship terms such as first and second in the present document are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion.

[0044] Embodiment 1

[0045] This embodiment mainly describes a machine learning-based organ function injury early warning method for sepsis patients, as shown in the following formula: Figure 1 Specifically, the method comprises the following steps:

[0046] Obtaining multi-modal clinical data of a sepsis patient, the multi-modal clinical data comprising continuous monitoring vital sign time series data, intermittent acquisition laboratory test data, dynamic recording treatment intervention data and static storage patient basic information;

[0047] Standardizing the multi-modal clinical data, extracting organ function related features through a multi-scale feature fusion strategy, and constructing an organ-level feature set; the multi-scale feature fusion strategy comprising extracting short-term fluctuation features and long-term trend features from the vital sign time series data, calculating index coefficient of variation and dynamic growth rate from the laboratory test data, and constructing intervention effect related features from the treatment intervention data;

[0048] Inputting the organ-level feature set into a pre-trained multi-organ collaborative early warning model, the multi-organ collaborative early warning model comprising a first branch for constructing an inter-organ compensation relationship graph, a second branch for quantifying organ injury risk contribution degree, and a combination of attention mechanism for weighted fusion of the two types of outputs, and outputting a multi-organ function injury risk matrix of the patient within a preset time length in the future;

[0049] Based on the multi-organ function injury risk matrix and organ-specific early warning threshold, generating an organ function injury early warning judgment result, the early warning judgment result comprising injury risk level of each organ and key feature identifier of dominant risk;

[0050] Real-time monitoring the update of the multi-modal clinical data, re-executing feature extraction and risk matrix calculation every interval dynamic adjustment time, and updating the organ function injury early warning judgment result.

[0051] Further, the multi-modal clinical data standardization processing and multi-scale feature fusion are as follows:

[0052] The multi-modal clinical data comes from different sources, has different collection frequencies and formats, and needs to be standardized to eliminate data heterogeneity. The specific processing process includes three aspects: first, a normalization method based on individual patient baseline values is used, with stable vital signs (such as heart rate, blood pressure) and laboratory test data (such as serum creatinine, white blood cell count) at the time of patient admission as the baseline, to convert subsequent data collection into a deviation ratio relative to the baseline value, avoiding reduced data comparability due to individual baseline differences; second, a time alignment algorithm is used to map continuous monitoring vital signs time series data (collection frequency usually 1-5 minutes / time) and intermittent laboratory test data (collection frequency usually 4-24 hours / time) to a unified time axis (such as 10 minutes as a time unit), ensuring that different types of data can be matched in the time dimension; third, for data missing and outliers, interpolation is used to fill short-term data missing (such as linear interpolation for vital signs data missing for no more than 30 minutes), and clinical logic-based outlier correction rules are used to handle data outside the physiological reasonable range.

[0053] Among them, the outlier correction based on clinical logic needs to be realized by a specific formula to calibrate the data rationality, and the formula is: , in the formula, represents the outlier value outside the physiological range (such as systolic pressure suddenly dropping to 50 mmHg, which is outside the reasonable range of adult systolic pressure 90-139 mmHg), is the normal data at the previous collection time of the outlier value (such as systolic pressure 110 mmHg 10 minutes before the outlier value appears), is the normal data at the next collection time of the outlier value (such as systolic pressure 105 mmHg 10 minutes after the outlier value appears), , is the weight coefficient (usually set according to the time interval of the normal data before and after the outlier value, the shorter the interval, the greater the weight), and through the formula, the corrected outlier value can be fitted to the clinical data trend;

[0054] Multi-scale feature fusion extracts multi-dimensional correlation features that can reflect the functional state of organs by analyzing the characteristics of different types of data. For time series data of vital signs (such as heart rate and blood oxygen saturation), a sliding window method is used to extract short-term fluctuation features and long-term trend features. Short-term fluctuation features are calculated by a 30-minute sliding window to reflect the short-term stability of organ function. Long-term trend features are calculated by a 4-hour sliding window to reflect the long-term change trend of organ function. For laboratory test data (such as liver function indicators such as bilirubin and kidney function indicators such as urea nitrogen), the coefficient of variation (the ratio of the standard deviation to the mean of multiple test results for the same indicator) and the dynamic growth rate (the difference between the current test value and the previous test value divided by the previous test value, then divided by the test interval) are calculated. The coefficient of variation of bilirubin can reflect the fluctuation degree of liver function indicators, and the dynamic growth rate can reflect the progress speed of liver function damage.

[0055] For treatment intervention data (such as fluid resuscitation volume and vasoactive drug dosage), intervention effect correlation features are constructed to quantify the impact of treatment measures on organ function. The effect feature of a single intervention measure is calculated using the formula: , where is the numerical value of the organ function correlation indicator before intervention (such as mean arterial pressure before fluid resuscitation, which is 65 mmHg), is the mean value of the indicator within a predetermined time period after intervention (such as mean arterial pressure rising to 75 mmHg 2 hours after fluid resuscitation), is the duration of intervention (such as fluid resuscitation lasting 1.5 hours), and the formula result can reflect the improvement efficiency of intervention measures on organ function indicators per unit time. When there are multiple intervention measures (such as fluid resuscitation and vasoactive drug therapy), the total intervention effect feature is calculated using the formula: , where is the total number of intervention measures, is the weight of the th intervention measure (set according to the priority of intervention measures in clinical guidelines, is the effect feature of the th intervention measure. The total intervention effect feature can be used to comprehensively evaluate the overall impact of multiple treatment measures on organ function. All the above features are classified by organ (such as heart rate and mean arterial pressure related features for the heart, and bilirubin and albumin related features for the liver) to form an organ-level feature set.

[0056] Further, the multi-organ collaborative early warning model and risk matrix output are constructed, as shown in Figure 2 , specifically as follows:

[0057] The multi-organ collaborative early warning model comprises two core branches, which respectively construct an inter-organ compensation relationship map and quantify the contribution of organ damage risk. When constructing the inter-organ compensation relationship map in the first branch, the similarity of organ functional feature sequences is calculated using Euclidean distance, with the following formula: In the formula , The first , The first organ One functional feature For the number of functional features, Euclidean distance The smaller the value, the more similar the trends in the functional characteristics of the two organs are, and the stronger the potential compensatory relationship.

[0058] The adjacency matrix is ​​constructed based on Euclidean distance, and the formula is as follows: In the formula This is the distance attenuation coefficient (calibrated based on historical clinical data to ensure that the values ​​of adjacent matrix elements are between 0 and 1). For elements in the adjacency matrix, the closer the value is to 1, the higher the value. The organ and the first The higher the degree of correlation between organs, the better. The adjacency matrix is ​​input into a graph convolutional network (GCN), which performs feature aggregation and optimization of organ-to-organ correlations through 2-3 layers of graph convolution operations, outputting a matrix of compensatory relationship strength between organs. Indicates the first Changes in organ function affect the first The intensity of compensatory effects on individual organs;

[0059] When quantifying the contribution of organ injury risk in the second branch, the initial risk value of a single organ is calculated using the following formula: In the formula For the first Normal reference values ​​for the functional characteristics of each organ (e.g., kidney function). The normal reference range for serum creatinine is 44-133. The average value is 88.5. ), For the first The weights of each feature (set according to the degree of influence of the feature on organ function impairment, initial risk value) The larger the value, the greater the degree to which a single organ deviates from its normal state;

[0060] Because there is inter-organ compensation in sepsis, the initial risk value needs to be corrected by combining the compensation relationship strength matrix output from the first branch to obtain the compensated risk value (i.e., the contribution of organ damage risk), as shown in the formula: In the formula The total number of organs, For the first The organ affects the first The compensatory strength of individual organs, For the first Initial risk value for each organ;

[0061] By combining the attention mechanism with weighted fusion of the dual-branch outputs, and through the calculation of local correlation weights and global attention weights, dynamic focusing on important features is achieved. The formula for calculating the local correlation weights is as follows: In the formula For the strength of organ compensation relationship, For the first Risk value after organ compensation, For the first Risk value after organ compensation, local correlation weight It can reflect the first The organ is related to the first Local impact weighting in individual organ risk assessment;

[0062] The global attention weight is calculated based on the local association weights, using the following formula: In the formula For the first The mean local correlation weight of each organ As an exponential function, this formula allows organs with high risk values ​​and significant local correlation effects to receive higher global attention weights.

[0063] After completing the weight calculation, use the formula The fusion features are obtained, where For the first The compensatory relationship feature vector of the _th organ (derived from the _th ... (Line element composition). To integrate the target characteristics of organ compensation relationship and damage risk; to integrate the characteristics The input-multiple-output fully connected layer calculates the probability of damage to each organ using the sigmoid activation function, with the following formula: In the formula For the first The weight vectors corresponding to each organ (obtained through backpropagation optimization using training set data). This is a bias term (calibrated based on organ damage label data). The value ranges from 0 to 1; the closer the value is to 1, the more significant it is. The higher the probability of functional damage to an organ within a predetermined timeframe, the more likely it is to cause functional damage. The risk probabilities of all organs are arranged by rows (each row corresponds to one organ) or columns (each column corresponds to one time point) to form a multi-organ functional damage risk matrix.

[0064] Further, the organ-specific early warning threshold determines the generation of early warning determination results, specifically:

[0065] The determination of the organ-specific early warning threshold is based on historical clinical data to construct the organ-specific early warning threshold, and dynamically updated combined with new data to adapt to the physiological characteristics of different organs and individual differences of patients. First, calculate two-level early warning thresholds: the first threshold (low-risk early warning threshold) formula is , and the second threshold (high-risk early warning threshold) formula is ; in the formula is the mean of the risk value of the th organ in the normal state in the historical data (for example, the mean of the compensatory risk value of the heart in the normal state in the historical data is 0.2), is the corresponding standard deviation, , are the corresponding threshold coefficients (set according to the clinical misjudgment rate requirement, for example, to control the false negative rate, take 1.2, take 2.0, for example, the heart, the first threshold =0.2+1.2×0.08=0.296, and the second threshold =0.2+2.0×0.08=0.36, when the compensatory risk value of the heart is between and , it is determined to be low risk; when it exceeds , it is determined to be high risk.

[0066] Due to the timeliness and accumulation of clinical data, the threshold needs to be updated regularly based on new clinical data to ensure the accuracy and adaptability of the threshold. The update formula is: ; in the formula is the amount of historical data, is the amount of new data, is the old threshold (for example, the old first threshold of the heart is 0.296), , are the mean and standard deviation of the risk value of the th organ in the normal state in the new data, is the threshold update coefficient (set according to the consistency of new data and historical data);

[0067] Based on the multi-organ functional injury risk matrix and the organ-specific early warning threshold, the early warning determination results containing the organ injury risk level and the dominant risk key feature identifier are generated. First, the risk level is determined: compare the risk probability (or compensatory risk value) of each organ in the risk matrix with the corresponding first and second thresholds, if the risk value is less than , it is determined to be "no risk"; between and between 0 and 1, determined as “low risk”; greater than 1, determined as “high risk”, and the risk level is recorded according to the organ (e.g. heart: low risk, kidney: high risk).

[0068] Meanwhile, the key features that dominate the risk are identified by calculating the feature importance, and the formula is: ; wherein is the small change amount of the risk value of the i-th organ after compensation (e.g. the risk value of the kidney after compensation changes from 0.32 to 0.33, = 0.01), is the small change amount of the j-th feature of the i-th organ (e.g. the j-th feature of the kidney, serum creatinine, changes from 150 μmol / L to 155 μmol / L, = 5), is the feature value (150 μmol / L), is the risk value after compensation (0.32), The greater the absolute value of , the higher the degree of influence of the feature on the risk of organ damage, and the top 1-5 features with the largest absolute value are selected as the key features (e.g. the key features of the kidney are serum creatinine, urine volume, and urea nitrogen), and the direction of feature change is marked according to the positive and negative relationship between and (e.g. the increase of serum creatinine leads to the increase of the risk value, and is marked as “serum creatinine ↑”), and the risk level and the key feature identification are integrated to form a complete early warning determination result. Further, the real-time monitoring of the clinical data and the dynamic updating of the early warning result are specifically:

[0069]

[0070] ​​​​Real-time monitoring of the update of multi-modal clinical data requires the establishment of a data real-time acquisition and transmission link to cover all sources of clinical data. For continuous monitoring of time series data of vital signs (such as heart rate collected by an electrocardiogram monitor), the hospital Internet of Things (IoT) system is used to transmit real-time data to the data processing module, and the transmission frequency is consistent with the acquisition frequency (1-5 minutes / time), ensuring that the data is obtained without delay. For intermittent laboratory test data, after the test equipment completes the detection and generates a report (such as the output of a biochemical analyzer in the clinical laboratory to detect creatinine), the hospital laboratory information system (LIS) automatically pushes the data to the data processing module, with a delay of less than 5 minutes. For dynamically recorded treatment intervention data (such as nurses performing fluid resuscitation and doctors adjusting the dosage of vasoactive drugs), the real-time input function of the electronic medical record system (EMR) is used to synchronize the data to the data processing module immediately after the medical staff completes the operation record. For static storage of patient basic information (such as age, underlying diseases), although real-time updating is not required, it needs to be imported completely and once for all when the patient is included in the early warning system, and the update mechanism is triggered when the basic information changes (such as the addition of an allergy history).

[0071] The data processing module needs to verify the validity of the real-time received data, including whether the data format meets the preset specifications (such as heart rate data should be an integer, ranging from 30 to 200 times per minute), and whether there are packet loss or errors in data transmission (such as if heart rate data is not received for three consecutive times, it is determined that there is a transmission error). For data with format errors or transmission errors, a retry mechanism is triggered immediately (such as requesting the collection device to send data again), and if valid data cannot be obtained after three retries, it is marked as “data anomaly” and the medical staff is notified to troubleshoot equipment or link problems, ensuring that the data input into the subsequent feature extraction link is valid.

[0072] The feature extraction and risk matrix calculation are performed again at each interval, and the core is to adjust the update cycle adaptively according to the current multi-organ risk level of the patient, balancing the real-time of early warning and resource consumption. First, calculate the average value of multi-organ risk, the formula is: , where is the total number of organs, is the risk value of each organ after compensation;

[0073] The adjustment interval is determined based on the average value of multi-organ risk, the formula is: ; where , are the minimum and maximum update intervals (set according to clinical needs), , are the minimum and maximum values of the average risk in the historical data;

[0074] When any organ risk value exceeds the secondary threshold value (such as the post-compensation risk value of the kidney 0.38 exceeding the secondary threshold value 0.36), the update interval is forcibly set to ${{{T}}_{{min}}}$ (15 minutes), ensuring that the organ function changes in a high-risk state can be captured in real time. When each update interval arrives, the data processing module automatically calls the feature extraction algorithm (such as the multi-scale feature fusion strategy) to extract features from the latest multi-modal clinical data, generate a new organ-level feature set, and input the multi-organ collaborative early warning model to recalculate the risk matrix. The new early warning judgment result is compared with the updated organ-specific early warning threshold value, and a new early warning judgment result is generated. The result is pushed to the medical staff terminal in real time through the hospital information system (such as EMR, nurse station display screen), and the historical early warning results are automatically covered to ensure that the medical staff always obtains the latest organ function damage early warning information.

[0075] The embodiment details the construction of an organ-level feature set by integrating multi-modal clinical data, standardizing processing, and multi-scale feature fusion. Then, with the help of a multi-organ collaborative early warning model, the organ compensation relationship and damage risk are quantified. The risk matrix is output by combining the attention mechanism weighted fusion. The dynamic threshold and key feature identification are used to realize precise early warning of multi-organ damage, which can capture organ function abnormalities in advance, improve the early warning accuracy and time, and dynamically update the early warning results to provide scientific basis for clinical intervention.

[0076] Embodiment 2

[0077] The embodiment details the organ function damage early warning system for sepsis patients based on machine learning, which includes a data acquisition unit, a data processing unit, a model operation unit, an early warning judgment unit, and a dynamic updating unit. The units are connected in sequence through a data transmission link and work cooperatively to execute the organ function damage early warning judgment method for sepsis patients based on machine learning.

[0078] The data acquisition unit is communicatively connected with the hospital clinical data system (including the electrocardio monitoring system, the laboratory information system, and the electronic medical record system) to obtain multi-modal clinical data of sepsis patients. The multi-modal clinical data includes continuous monitoring vital sign time series data, intermittent acquisition laboratory test data, dynamic recording treatment intervention data, and static storage patient basic information.

[0079] The data processing unit is communicatively connected with the data acquisition unit to receive the multi-modal clinical data transmitted by the data acquisition unit, standardize the multi-modal clinical data, and extract organ function correlation features through a multi-scale feature fusion strategy to construct an organ-level feature set. The multi-scale feature fusion strategy includes extracting short-term fluctuation features and long-term trend features from vital sign time series data, calculating index coefficient of variation and dynamic growth rate from laboratory test data, and constructing intervention effect correlation features from treatment intervention data.

[0080] The model operation unit is built-in with a pre-trained multi-organ collaborative early warning model and is in communication connection with the data processing unit, for receiving the organ-level feature set transmitted by the data processing unit, constructing an inter-organ compensation relationship graph through a first branch of the multi-organ collaborative early warning model, quantifying the organ damage risk contribution degree through a second branch, combining the attention mechanism to weight and fuse the two types of outputs, and outputting a multi-organ function damage risk matrix of the patient within a preset time length in the future;

[0081] The early warning determination unit is in communication connection with the model operation unit, for receiving the multi-organ function damage risk matrix transmitted by the model operation unit, combining a preset organ-specific early warning threshold, generating an organ function damage early warning determination result, and the early warning determination result including the damage risk level of each organ and the key feature identifier of the dominant risk;

[0082] The dynamic updating unit is in communication connection with the data acquisition unit, the data processing unit, the model operation unit and the early warning determination unit respectively, for monitoring the multi-modal clinical data updating transmitted by the data acquisition unit in real time, triggering the data processing unit to re-execute feature extraction, the model operation unit to recalculate the risk matrix, and the early warning determination unit to regenerate the early warning determination result according to the dynamically adjusted time interval, and realizing the dynamic updating of the early warning determination result.

[0083] Further, the data acquisition unit includes a time series data acquisition module, a test data acquisition module, an intervention data acquisition module and a basic information acquisition module, each module corresponding to acquiring different types of multi-modal clinical data:

[0084] The time series data acquisition module is connected with the hospital electrocardiogram monitoring system and the bedside monitor through wired or wireless communication protocol, and real-time acquisition of heart rate, blood pressure, blood oxygen saturation and other continuous monitoring vital sign time series data is realized, the acquisition frequency is consistent with the monitoring frequency of the monitoring equipment, and the data breakpoint transmission is supported to avoid data loss;

[0085] The test data acquisition module is interfaced with the hospital laboratory information system, receives the laboratory test data such as blood routine, biochemical indicators and coagulation function collected intermittently through the standardized data interface, completes the data reception and format conversion within 10 minutes after the test result is generated, and ensures the timeliness of the data;

[0086] The intervention data acquisition module is in communication connection with the treatment record module of the electronic medical record system, and real-time grabbing of the dynamically recorded treatment intervention data such as fluid resuscitation amount, blood active drug usage dose and mechanical ventilation parameters is realized, manual supplementary recording of missing intervention operation records by medical staff is supported, and the supplementary data is time-stamped calibrated;

[0087] The basic information acquisition module acquires patient age, gender, basic disease history, allergy history and other static stored patient basic information through a hospital patient information system, and completes the complete acquisition of the basic information when the patient is first included in the early warning system. If the patient's basic information changes, the system can be manually triggered to update or automatically synchronized to update.

[0088] Further, the system further comprises a threshold management unit and a result display unit, the threshold management unit is in communication connection with the early warning determination unit, and the result display unit is in communication connection with the early warning determination unit and the dynamic updating unit respectively:

[0089] The threshold management unit is used for storing and updating organ-specific early warning thresholds, supporting the import of historical clinical data to calculate the initial first-level threshold and second-level threshold, and regularly updating the threshold according to newly included clinical data. The historical threshold version is retained during the updating process, which facilitates the tracing of the influence of threshold adjustment on the early warning results. At the same time, medical staff can manually fine-tune the threshold coefficient within a preset range according to the needs of clinical practice, in order to adapt to the diagnosis and treatment characteristics of different departments.

[0090] The result display unit includes a terminal display module and an early warning pushing module. The terminal display module is deployed on a doctor workstation and a nurse station display screen to visually display the damage risk level of each organ and the key feature identifier of the dominant risk, and supports medical staff to click to view the feature change details. The early warning pushing module pushes early warning information to responsible medical staff through a hospital message system and a mobile nursing APP. High-risk early warning information is triggered in the form of a pop-up window and voice prompt, and low-risk early warning information is triggered in the form of message notification. All early warning information is attached with risk determination basis, which facilitates medical staff to quickly understand and make decisions.

[0091] The embodiment describes in detail that the multi-modal clinical data is acquired by the data acquisition unit, the organ-level feature set is constructed by the data processing unit standardization and multi-scale feature fusion, the model operation unit outputs the risk matrix relying on the multi-organ collaborative early warning model, the early warning determination unit generates the early warning result in combination with the threshold, the dynamic updating unit triggers the update in real time, and the threshold management and result display unit are also provided. The system realizes efficient integration of multi-modal data, accurate determination of multi-organ damage risk and dynamic updating of early warning results, can push clear early warning information to medical staff in real time, provides stable and efficient system support for organ function damage early warning of sepsis patients, and helps clinical timely intervention.

[0092] Based on Example 1 or 2, this embodiment describes the specific implementation effect of the machine learning-based early warning method for organ function damage in sepsis patients in detail. Among them, 960 sepsis patients admitted to the ICU of 3 domestic first-class hospitals were selected, including 528 males and 432 females, aged 24-87 years, with an average age of (56.8±13.2) years; the infection types covered 386 cases of pulmonary infection, 264 cases of abdominal infection, 192 cases of urinary system infection, and 118 cases of infection in other parts, all patients met the diagnostic criteria of “Guidelines for Diagnosis and Treatment of Sepsis (2021 Edition)”, and cases with chronic organ failure and immune deficiency before admission were excluded; the study used a randomized controlled design, and the patients were divided into a test group (using the early warning method of the present application) and a control group (using the traditional SOFA score combined with single organ index monitoring), each group had 480 cases, and the test period was from admission to discharge or observation for 28 days. Clinical data were collected in real time through the hospital intensive care information system (CIS), vital signs (heart rate, blood oxygen saturation, mean arterial pressure) were recorded every 5 minutes, laboratory indicators (creatinine, bilirubin, lactic acid, and prothrombin time) were detected every 6-24 hours, treatment intervention (fluid resuscitation, use of vasoactive drugs) data were entered within 5 minutes after execution, and static basic information (age, underlying disease) was collected completely at admission to ensure data continuity and authenticity.

[0093] To verify the effectiveness of the multi-modal data standardization processing and feature extraction of the present application, the statistical results of the two groups of data were first processed, and the core indicators are shown in the following table. This table focuses on the original indicators and extracted features directly related to organ function, such as the long-term trend slope of heart rate and the short-term fluctuation standard deviation of mean arterial pressure, which reflect heart function, the dynamic growth rate of creatinine and the coefficient of variation of urine volume, which reflect kidney function, the coefficient of variation of bilirubin and the dynamic change of γ-glutamyl transpeptidase (γ-GT), which reflect liver function, and the effect of treatment intervention features (mean arterial pressure change rate after fluid resuscitation). The data show that through abnormal value correction based on clinical logic (such as weighted correction combined with normal data before and after the time) and multi-scale feature fusion, the test group's feature matching rate with organ damage clinical diagnosis is all above 90%, among which the long-term trend slope of heart rate is 94.5%, and the dynamic growth rate of creatinine is 93.8%, and it can effectively capture the influence of treatment intervention on organ function (fluid resuscitation effect feature matching rate 92.3%); while the control group only uses static indicator mean, the feature matching rate is generally lower than 75%, and it cannot extract intervention effect related features. This difference shows that the data processing and feature extraction strategy of the present application can accurately mine organ function related information, providing high-quality input for subsequent early warning models.

[0094] Table 1 Comparison of feature extraction results

[0095] Data processing and feature types Specific indicators / features Test group Control group Raw data processing Deviation rate after correction of outliers and clinical practice 0.8 2.5 Completeness rate after data filling 99.6 97.2 Cardiac function features Match rate of long-term trend slope of heart rate (4h window) 94.5 72.6 Match rate of standard deviation of short-term fluctuation of mean arterial pressure (30min window) 93.9 71.8 Renal function features Match rate of dynamic growth rate of blood creatinine (24h interval) 93.8 73.5 Match rate of coefficient of variation of urine volume (8h window) 92.7 72.3 Liver function features Match rate of coefficient of variation of bilirubin (24h interval) 93.2 74.1 Match rate of dynamic change of γ-glutamyltransferase (γ-GT) 91.9 70.5 Therapeutic intervention effect features Match rate of change rate of mean arterial pressure after fluid resuscitation 92.3 (no feature) Match rate of blood pressure change feature after intervention of vasoactive drugs 91.5 (no feature)

[0096] As Figure 4 shown, Figure 4 The broken line chart is the comparison of the accuracy of multi-organ injury warning. As can be seen from the chart, 6 hours before the injury, the heart warning accuracy of the test group reached 81.8%, the kidney 80.2%, and the liver 81.1%, while the control group was only 50.5%, 48.9%, and 49.6% respectively. As the time approached the injury time, the accuracy of the test group continued to improve. 2 hours before the injury, the accuracy of the heart, kidney, and liver was 92.9%, 91.5%, and 92.2% respectively, and 0 hours before the injury, it reached 96.3%, 95.5%, and 95.9% respectively. The accuracy of the control group during the same period was only 73.2%, 71.0%, and 72.5% respectively. According to the above table data, it is known that it is the heart rate trend and the dynamic change of blood creatinine and other characteristics accurately extracted by the test group that support earlier and more accurate multi-organ warning, solving the defects of the control group that static indicators are lagging and cannot cover multiple organs.

[0097] As Figure 5 shown, Figure 5 The scatter plot is the correlation between the warning advance time and the 28-day survival rate. The red scatter points are the test group, and the blue ones are the control group. The warning advance time of the test group is concentrated in 3.5-6.0 hours, corresponding to a survival rate of 78.2%-92.5%, and there is a strong positive correlation between the advance time and the survival rate (r=0.84). The advance time of the control group is concentrated in 0.6-2.0 hours, corresponding to a survival rate of 52.5%-70.8%, and the correlation is weak (r=0.53). To quantify the improvement effect of prognosis, the core prognostic indicators of the two groups are as shown in the following table. The table includes the 28-day cumulative survival rate, the organ injury reversal rate, and the specific organ function recovery time (such as the time for the mean arterial pressure of heart injury patients to recover to 80-100 mmHg, and the time for the blood creatinine of kidney injury patients to decrease to 44-133 hours). The data shows that the 28-day survival rate of the test group is 76.9%, which is 21.6 percentage points higher than that of the control group (55.3%); the heart, kidney, and liver injury reversal rates are 82.1%, 79.4%, and 80.2% respectively, which are significantly higher than those of the control group (58.2%, 55.0%, and 56.5%); the organ function recovery time is shortened by more than 10 hours, and the difference in survival rate between the two groups is statistically significant (P<0.001) by Log-rank test, fully proving the improvement value of the warning method of the present application on clinical prognosis.

[0098] Table 2 Comparison of core indicators of organ injury prognosis

[0099] Prognosis indicators Test group Control group Difference P value 28-day cumulative survival rate (%) 76.9 55.3 +21.6 <0.001 Cardiac injury reversal rate (%) 82.1 58.2 +23.9 <0.001 Renal injury reversal rate (%) 79.4 55.0 +24.4 <0.001 Liver injury reversal rate (%) 80.2 56.5 +23.7 <0.001 Cardiac injury: mean arterial pressure recovery time (h) 18.2 27.5 -9.3 <0.001 Renal injury: blood creatinine recovery time (h) 42.1 57.8 -15.7 <0.001 Liver injury: bilirubin recovery time (h) 38.4 50.8 -12.4 <0.001 Multiple organ injury incidence rate (%) 18.5 32.2 -13.7 <0.001

[0100] As Figure 6 shown, Figure 6The heat map is identified for the key features, the horizontal axis is the organ type (heart, kidney, liver), the vertical axis is the organ function feature, and the color depth represents the feature importance (G value). As can be seen from the figure, when the heart is damaged, the long-term trend slope of the heart rate and the short-term fluctuation standard deviation of the mean arterial pressure have the highest G value (the deepest color), which is completely matched with the clinically diagnosed dominant factor of heart function abnormality; when the kidney is damaged, the dynamic growth rate of blood creatinine and the coefficient of variation of urine volume are the core features; when the liver is damaged, the coefficient of variation of bilirubin and the dynamic change of gamma-GT are the key features, and the positioning accuracy of the key features of the test group reaches 92.4%.

[0101] The above are only preferred embodiments of the present application, and are not intended to limit the protection scope of the present application. The present application can have various changes and modifications for those skilled in the art; any changes, modifications, replacements, integrations and parameter changes to these embodiments within the spirit and principles of the present application, which can realize the same functions without departing from the principles and spirit of the present application, all fall within the protection scope of the present application.

Claims

1. A method for early warning determination of organ dysfunction in sepsis patients based on machine learning, characterized in that, The application relates to a method for predicting organ function injury of a sepsis patient, comprising the following steps: acquiring multi-modal clinical data of a sepsis patient, wherein the multi-modal clinical data comprises continuously monitored vital sign time series data, intermittently collected laboratory test data, dynamically recorded treatment intervention data and statically stored patient basic information; standardizing the multi-modal clinical data, extracting organ function correlation features through a multi-scale feature fusion strategy, and constructing an organ-level feature set; the multi-scale feature fusion strategy comprises extracting short-time fluctuation features and long-time trend features from the vital sign time series data, calculating index coefficient of variation and dynamic growth rate from the laboratory test data, and constructing intervention effect correlation features from the treatment intervention data; inputting the organ-level feature set into a pre-trained multi-organ collaborative early warning model, wherein the multi-organ collaborative early warning model constructs an inter-organ compensation relationship graph through a first branch, quantifies organ injury risk contribution degrees through a second branch, combines attention mechanisms to weightedly fuse two types of outputs, and outputs a multi-organ function injury risk matrix of the patient within a preset time length in the future; generating an organ function injury early warning judgment result based on the multi-organ function injury risk matrix and organ-specific early warning threshold, wherein the early warning judgment result comprises injury risk grades of each organ and key feature identifiers of dominant risks; real-time monitoring of update conditions of the multi-modal clinical data, re-executing feature extraction and risk matrix calculation every interval of a dynamically adjusted time, and updating the organ function injury early warning judgment result. 2.The machine learning-based early warning determination method for organ function impairment of sepsis patients according to claim 1, characterized in that, The standardization processing of the multi-modal clinical data comprises the following steps: correcting the vital signs and the laboratory test data by adopting a normalization method based on patient individual basic values, mapping data of different collection frequencies to a unified time axis through a time series alignment algorithm, filling short-time data missing through an interpolation method, and processing data beyond a physiological reasonable range through an abnormal value correction rule based on clinical logic. The abnormal value correction rule based on clinical logic is used to process data beyond the physiological reasonable range, specifically, the vital sign data is calculated to obtain corrected data, and the formula is: wherein is an abnormal value beyond the physiological range, is normal data at a previous collection time of the abnormal value, is normal data at a next collection time of the abnormal value, , are corresponding weight coefficients, respectively. 3.The machine learning based early warning decision method for organ function impairment of sepsis patients according to claim 1, characterized in that, The treatment intervention data is used to construct an intervention effect correlation feature, specifically: intervention effect features are calculated by obtaining the treatment intervention data, and the formula is: wherein is the value of the organ function correlation index before intervention, is the mean value of the index within the preset time period after intervention, is the intervention duration; When there are multiple interventions, the total intervention effect profile is calculated as: where is the total number of interventions, is the weight of the th intervention, is the effect profile of the th intervention. 4.The method of claim 1, wherein, The first branch of the multi-organ collaborative early warning model constructs an inter-organ compensation relationship map, including: using the formula Calculate the first The and the first The Euclidean distance of the functional feature sequences of each organ, where , The first The first organ One functional feature To determine the number of functional features, construct an adjacency matrix using the following formula: ,in This is the distance attenuation coefficient. The adjacency matrix elements are used as inputs to a graph convolutional network, which outputs a matrix of compensatory relationship strength between organs, with the matrix elements as follows: Indicates the first Changes in organ function affect the first The intensity of compensatory effects on individual organs.

5. The machine learning-based early warning decision method for organ function impairment of sepsis patients according to claim 4, characterized in that, The second branch of the multi-organ collaborative early warning model quantifies the contribution of organ damage risk, including: calculating the initial risk value, using the formula: ,in For the first Normal reference values ​​for the functional characteristics of individual organs For the first The weights of each feature are used to calculate the post-compensation risk value, taking into account the compensatory relationship between organs. The formula is as follows: ,in The total number of organs, For the first The organ affects the first The compensatory strength of individual organs, For the quantified first Contribution of individual organ damage risk.

6. The machine learning-based early warning decision method for organ function impairment of sepsis patients according to claim 5, characterized in that, The weighted fusion of the two types of outputs using the attention mechanism includes: calculating the local correlation weight, as shown in the formula: ,in For the strength of organ compensation relationship, For the first Risk value after individual organ compensation; calculate global attention weight using the following formula: ,in For the first The mean of local correlation weights of each organ; through the formula The fusion features are obtained, among which For the first The compensatory relationship feature vector of an individual organ Features are fused to target specific features. 7.The machine learning based early warning decision method of organ function impairment of sepsis patient according to claim 1, characterized in that, The determination of the organ-specific early warning threshold value comprises: calculating a first threshold value, with a formula of: ; calculating a second threshold value, with a formula of: , wherein is a mean value of risk values of the i-th organ in a normal state in historical data, is a corresponding standard deviation, , , are respective threshold value coefficients; updating the threshold value based on newly included clinical data, with a formula of: , wherein is a historical data amount, is a new data amount, is an old threshold value, , are a mean value and a standard deviation of the new data, is a threshold value updating coefficient. 8.The machine learning based early warning decision method of organ function impairment of sepsis patient according to claim 1, characterized in that, The key characteristics of the dominant risk include: calculating the importance of the characteristics, using the following formula: ,in For the first The minute change in risk value after organ compensation For the first The first organ The minute changes in each characteristic For this feature value; select Features with the highest absolute values ​​are designated as key features, and the direction of feature change is indicated. 9.The machine learning based early warning decision method of organ function impairment of sepsis patient according to claim 1, characterized in that, The time of the dynamic adjustment per interval re-executes the feature extraction and risk matrix calculation, including: calculating the multi-organ risk average value, the formula is: , determining the adjustment interval, the formula is: , wherein , The minimum and maximum update intervals are respectively , The minimum and maximum values of the risk average value in the historical data are respectively; when there is any organ risk value exceeding the secondary threshold value, the update interval is forced to be . 10.The machine learning based early warning decision method for organ function impairment of sepsis patient according to claim 1, characterized in that, The output patient future preset time length multi-organ function injury risk matrix comprises the following steps: fusing features Input multi-output full connection layer, through formula Calculate the injury risk probability of the first Organ, wherein is the weight vector corresponding to the first Organ, is a bias term; arrange the risk probabilities of all organs in rows or columns to form a multi-organ function injury risk matrix, and each element in the matrix corresponds to the future injury risk probability of a single organ.

Images