Dynamic quantitative early warning method for risk of sepsis coagulopathy based on combined detection of multiple biomarkers

Abstract

The present application relates to the technical field of sepsis coagulopathy early warning, and discloses a sepsis coagulopathy risk dynamic quantification early warning method based on combined detection of multiple biomarkers, which takes urinary thrombomodulin, urinary procalcitonin and urinary C-reactive protein as detection objects, and constructs a double-reference deviation quantity by using individual and group statistics; after unifying the direction of the deviation quantity, it is mapped into a single-index risk value, a dynamic weight is constructed based on the area under the ROC curve and the dispersion trend item, and a static fusion risk value and an early screening evidence accumulation item in a sliding time window are respectively established; an interactive feature and a joint super-threshold value gating rule are constructed to form a multi-dimensional dynamic feature vector input into a cost-sensitive ensemble learning model, the risk probability is inferred and mapped into a basic score, and then the final score is output by combining the gating signal and the upward trend rule, so that hierarchical early warning and closed-loop retesting are performed. The method can capture weak abnormal signals in the prodromal stage and improve the sensitivity of early screening of sepsis coagulopathy.

Description

Technical Field

[0001] This invention relates to the field of risk warning technology for septic coagulopathy, specifically a dynamic quantitative early warning method for the risk of septic coagulopathy based on the combined detection of multiple biomarkers. Background Technology

[0002] Sepsis is a systemic inflammatory response disorder syndrome triggered by infection. During its progression, it is often accompanied by endothelial dysfunction, abnormal release of inflammatory factors, and coagulation system disturbances. In the early stages of sepsis, it can further develop into septic coagulopathy. Septic coagulopathy typically manifests as microthrombus formation, coagulation factor depletion, and impaired organ perfusion. Its early manifestations are often subtle, and a single clinical indicator or test result is insufficient to reflect the disease progression in a timely manner. With the development of laboratory testing techniques, biomarkers such as urinary thrombomodulin (U-TM), urinary procalcitonin (PCT), and urinary C-reactive protein (CRP) are widely used to assess infection status, inflammation severity, and the risk of endothelial / coagulation damage. Among them, U-TM reflects endothelial damage and abnormal coagulation regulation, urinary PCT reflects bacterial infection and systemic inflammatory burden, and urinary CRP reflects the intensity of the body's inflammatory response. Combined screening based on multiple non-invasive biomarkers has clear clinical significance for the early identification and dynamic assessment of the risk of septic coagulopathy.

[0003] In existing technologies, risk identification for septic coagulopathy typically employs various approaches, including single-indicator thresholding, multi-biomarker joint scoring, ROC curve-based statistical discriminant analysis, and machine learning classification methods. Single-indicator thresholding primarily judges risk based on whether a biomarker exceeds a preset reference range, suitable for scenarios with clear rules and large abnormal amplitudes. Joint scoring methods weight and fuse multiple indicators according to empirical or statistical weights to improve overall discriminant capability. Statistical discriminant analysis usually determines the optimal cutoff value using indicators such as ROC curves, AUC, and Youden index, and constructs a classification model accordingly. Machine learning methods further incorporate algorithms such as logistic regression, support vector machines, random forests, and gradient boosting trees to model multi-indicator inputs and output risk probabilities or classification results. Some solutions also combine time window sliding, continuous monitoring, trend analysis, or multimodal clinical data fusion to enhance responsiveness to changes in disease progression.

[0004] However, existing technologies for screening septic coagulopathy still have room for improvement in terms of adaptability: First, early-stage septic coagulopathy often presents only as mild, scattered, and asynchronous abnormalities. Single-point threshold judgments are easily affected by individual baseline differences and test fluctuations, making it difficult to reliably capture prodromal risks. Second, most existing combined models use static features as input and rarely explicitly consider the time lag relationship between different biomarkers in the course of the disease. In actual clinical practice, inflammatory responses, endothelial damage, and coagulation abnormalities often evolve asynchronously. Directly fusing them at the same time may weaken the ability to reflect the risk transformation process. Third, although some models can output risk probabilities, for early screening tasks where the cost of missed diagnoses is higher, if the training objective still leans towards overall classification accuracy, the sensitivity to high-risk prodromal states may not be ideal. Therefore, how to combine the diagnostic value, temporal evolution characteristics, and missed detection sensitivity requirements of multiple biomarkers in early screening for septic coagulopathy, and construct a technical solution that can dynamically quantify weak abnormalities and achieve graded early warning to solve the above problems. Summary of the Invention

[0005] To address the problems in related technologies, this invention provides a dynamic quantitative early warning method for the risk of septic coagulopathy based on the joint detection of multiple biomarkers, thereby overcoming the aforementioned technical problems in existing related technologies.

[0006] To solve the aforementioned technical problem, the present invention is achieved through the following technical solution: In a first aspect, embodiments of the present invention provide a method for dynamic quantitative early warning of septic coagulopathy risk based on the joint detection of multiple biomarkers, specifically including: collecting biomarker detection data of target patients at multiple time points within a preset monitoring period, and performing time alignment and confidence assessment to obtain preprocessed data with confidence coefficients; based on the preprocessed data, constructing a dual-benchmark deviation using individual historical window statistics, population statistics, and adaptive fusion coefficients, and obtaining the effective deviation of each biomarker through noise floor constraints; determining time lag compensation parameters according to the temporal correlation of each biomarker in the course of the disease, performing time lag compensation on the effective deviation to obtain a time lag aligned deviation; unifying the risk direction of the time lag aligned deviation and mapping it to a single-indicator risk value; and based on the area under the receiver operating characteristic curve and the discrete trend term... Early screening sensitive weights are constructed, and the dynamic weights of each biomarker are determined. The risk values ​​of the single indicators are weighted using the dynamic weights to construct a static fusion risk value and an early screening evidence accumulation term based on a sliding time window. An interaction feature representing the synergistic enhancement relationship of multiple indicators is constructed, and an independent joint over-threshold gating rule is set. The single indicator risk value, static fusion risk value, early screening evidence accumulation term, interaction feature, and discrete trend term are combined into a multi-dimensional dynamic feature vector, which is input into an ensemble learning model trained with a cost-sensitive loss function to infer the risk probability of sepsis coagulopathy. A basic risk score is obtained based on the risk probability mapping. The risk is then adjusted upwards by combining the gating rule with the monotonically increasing trend of the score within a continuous time window. The final risk score is output, and a graded early warning is executed. Retesting and closed-loop updates are triggered based on the early warning level.

[0007] As a preferred embodiment of the dynamic quantitative early warning method for septic coagulopathy based on the joint detection of multiple biomarkers described in this invention, the time alignment and confidence assessment include: Biomarker detection data from different sources are uniformly mapped to the same time reference frame, missing data are repaired by linear interpolation, and outlier values ​​are filtered by robust statistical rules. A confidence coefficient is assigned to each indicator at each time point, where the confidence coefficient for measured data is 1, and the confidence coefficient for interpolated or carried-over data is determined exponentially based on the compensation span. The biomarker detection data include urinary thrombomodulin, urinary procalcitonin, and urinary C-reactive protein.

[0008] As a preferred embodiment of the dynamic quantitative early warning method for septic coagulopathy risk based on the joint detection of multiple biomarkers described in this invention, the construction of the dual-benchmark deviation includes: Extract the mean and standard deviation of each indicator within the individual patient's historical window, as well as the mean and standard deviation of the corresponding indicators in the group training samples; The individual baseline deviation component and the group baseline deviation component are weighted and combined using adaptive fusion coefficients; A lower bound for noise floor is constructed using instrument measurement error and the lower bound for population physiological fluctuations. The individual standard deviation is truncated and the larger value is taken to obtain the effective individual standard deviation, which is then used to replace the original individual standard deviation in the deviation calculation. The adaptive fusion coefficient is dynamically determined based on the individual variation coefficient, the number of valid historical sample points, and the preset upper limit of indicator specificity, so that when the data is abundant and the individual is in a steady state, the individual baseline is emphasized, and when the data is scarce or the individual fluctuates drastically, the group baseline is emphasized.

[0009] As a preferred embodiment of the dynamic quantitative early warning method for septic coagulopathy risk based on the joint detection of multiple biomarkers described in this invention, the time lag compensation includes: A set of candidate time delays is pre-defined. On the training samples, with the goal of maximizing the discriminative ability of the deviation amount on the outcome of septic coagulopathy, the optimal time delay compensation parameters are selected for each indicator. The effective deviations of each indicator are shifted according to the time lag compensation parameter, so that the abnormal signals of inflammatory response, endothelial injury and coagulation abnormality markers are aligned in the order of biological evolution on the disease time axis.

[0010] As a preferred embodiment of the dynamic quantitative early warning method for septic coagulopathy based on the joint detection of multiple biomarkers described in this invention, the time lag alignment deviation is unified in risk direction and mapped to a single-indicator risk value, including: Based on the direction of indicator risk, the deviation of low-risk indicators is reversed, while the original direction of high-risk indicators is retained, thus obtaining a unified direction quantity. The direction unification quantity is mapped to the interval between 0 and 1 using a Sigmoid function, and the mapping steepness is controlled by a sensitivity adjustment parameter to obtain the initial risk value; Multiply the initial risk value at each time point by the confidence coefficient to obtain the confidence-corrected single-index risk value.

[0011] As a preferred embodiment of the dynamic quantitative early warning method for sepsis coagulopathy risk based on the joint detection of multiple biomarkers described in this invention, the method involves constructing early screening sensitivity weights based on the area under the receiver operating characteristic curve and the dispersion trend term, and determining the dynamic weights of each biomarker, including: The basic weights are determined based on the gain of the area under the receiver operating characteristic curve (AUC) of each indicator relative to the random discriminant reference value. Construct a discrete trend term using the rate of change of single-indicator risk values ​​at adjacent time points; The trend enhancement coefficient is used to correct the trend of the basic weights and normalize them to obtain the dynamic weights; The trend enhancement coefficient is determined offline by maximizing the early warning time gain function, based on the optimal feasible region determined by the biological half-life of each indicator.

[0012] As a preferred embodiment of the dynamic quantitative early warning method for the risk of septic coagulopathy based on the joint detection of multiple biomarkers described in this invention, the static fusion risk value is the weighted sum of the dynamic weight and the single indicator risk value after confidence correction. The early screening evidence accumulation term is the time-decayed weighted cumulative sum of the weighted sums at each moment within the sliding time window; The sliding time window length is determined based on the upper limit of the typical physiological latency period of septic coagulopathy and the sampling interval. The time decay weight is obtained by normalization based on the exponential decay of the physical time span and the forgetting decay rate. The forgetting decay rate is determined offline by maximizing the trajectory resolution gain function with physiological prior regularization penalty term, based on the physiological reference decay rate determined by combining the biological half-life of each biomarker.

[0013] As a preferred embodiment of the dynamic quantitative early warning method for sepsis coagulopathy risk based on the joint detection of multiple biomarkers described in this invention, the interactive feature is the sum of the products of the risk values ​​of different indicators; the joint threshold gate control rule is: when the risk value of each indicator exceeds the corresponding preset risk threshold, the gate control signal is triggered, otherwise it is not triggered.

[0014] As a preferred embodiment of the dynamic quantitative early warning method for sepsis coagulopathy based on the joint detection of multiple biomarkers described in this invention, the ensemble learning model is a gradient boosting tree model. During training, a cost-sensitive loss function is used to assign a higher penalty coefficient to positive samples progressing to sepsis coagulopathy than to negative samples, and the sample confidence weight is the product of the confidence coefficients of each indicator. The step of outputting the final risk score and executing tiered early warning includes: mapping the risk probability to a preset scoring interval to obtain a base score; when the gating signal is met and the base score shows a strictly monotonically increasing trend over multiple consecutive time windows, the final score is forcibly set to the upper limit of the score; otherwise, the base score is used as the final score; based on the low, medium, and high risk threshold intervals into which the final score falls, routine monitoring, continuous observation early warning, and the highest-level early warning are executed respectively, triggering corresponding levels of retesting and closed-loop updates.

[0015] Secondly, embodiments of the present invention provide a dynamic quantitative early warning system for the risk of septic coagulopathy based on the joint detection of multiple biomarkers, comprising: a data preprocessing module for collecting biomarker detection data of target patients at multiple time points within a preset monitoring period, and performing time alignment and confidence assessment to obtain preprocessed data with confidence coefficients; a dual-benchmark deviation construction module for constructing dual-benchmark deviations and obtaining the effective deviation of each biomarker through noise floor constraints; a time lag compensation module for determining time lag compensation parameters based on the temporal correlation of each biomarker in the course of the disease, performing time lag compensation on the effective deviation to obtain time lag aligned deviation; a single-indicator risk mapping module for unifying the risk direction of the time lag aligned deviation and mapping it to a single-indicator risk value; and an early screening sensitivity weight construction module for constructing weights based on subject artificial intelligence. The system employs several modules: a feature curve area under the curve and a discrete trend term to construct sensitive weights for early screening, determining the dynamic weights of each biomarker; a static and dynamic feature construction module to weight single-indicator risk values ​​using dynamic weights, constructing a static fusion risk value and an early screening evidence accumulation term based on a sliding time window; an interaction feature and gating module to construct interaction features representing the synergistic enhancement relationship of multiple indicators and set independent joint over-threshold gating rules; a risk inference and scoring module to form a multi-dimensional dynamic feature vector, input it into an ensemble learning model, infer the risk probability of sepsis coagulopathy, obtain a basic risk score based on the risk probability mapping, perform risk up-adjustment based on gating rules and the monotonically increasing trend of the score within a continuous time window, output the final risk score, and execute graded early warning; and a closed-loop control module to trigger retesting and closed-loop updates based on the early warning level.

[0016] The present invention has the following beneficial effects: 1. This invention constructs a dual-benchmark deviation for early screening scenarios, introduces individual variation coefficients to dynamically and adaptively adjust the fusion weight of individual baselines and population baselines, and uses a background noise constraint mechanism to truncate excessively small variances. This enables the system to identify abnormal signals deviating from individual steady state during early screening for sepsis coagulopathy, even when the indicators do not trigger the population absolute threshold. At the same time, it avoids noise amplification caused by excessive fluctuations in the patient's individual baseline or inherent measurement errors of the instrument. This enables the quantification of mild and scattered abnormal changes in the prodromal period, solving the problem in existing technologies where single-time-point threshold judgment is easily affected by individual baseline differences and detection fluctuations, making it difficult to stably capture weak risks in the prodromal period.

[0017] 2. This invention compensates for the time lag in the pathological manifestations of different biomarkers to align the time axis of asynchronous evolution. It also extracts discrete trend terms and constructs dynamic evidence accumulation terms based on a sliding time window mechanism, enabling different types of risk signals to be fused in series according to their actual biological evolution order. This reflects the gradual cumulative effect of multidimensional indicators in the time dimension, overcomes the defect of weakened ability to reflect the risk transformation process caused by directly fusing static features at the same time, and solves the problem that existing joint models mostly use static features as input and do not explicitly consider the asynchronous evolution relationship of inflammatory response, endothelial injury and coagulation abnormalities in the course of the disease.

[0018] 3. This invention introduces an early screening sensitive weight mechanism and a cost-sensitive loss function during the model building stage, assigning a higher penalty coefficient to positive samples that progress to sepsis coagulopathy. It also sets a safety gating rule based on joint over-threshold independently of the model output, making the system's judgment boundary tend to capture early high-risk states. Furthermore, it provides a fallback logic for mandatory identification of extreme high-risk indicator combinations, meeting the requirement for low false negative rates in early screening scenarios for sepsis coagulopathy. This solves the problem that conventional classification algorithms, which use overall classification accuracy as the training objective, are not sensitive enough to high-risk precursor states in early screening tasks where the cost of missed diagnoses is higher.

[0019] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, the drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 The present invention provides a flowchart of a method for dynamic quantitative early warning of septic coagulopathy risk based on the joint detection of multiple biomarkers.

[0022] Figure 2 This is a schematic diagram of the process S9 provided by the present invention.

[0023] Figure 3 This is a schematic diagram of a module for a dynamic quantitative early warning system for the risk of septic coagulopathy based on the joint detection of multiple biomarkers, provided by the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Example 1 In the early screening scenario of sepsis coagulopathy, patients in the prodromal stage usually only show weak signal changes such as mild inflammatory response, endothelial damage and coagulation dysfunction. The static threshold determination of a single biomarker is difficult to reflect the trend of disease progression in a timely manner, and conventional general classification algorithms are usually more suitable for identifying cases that have formed significant abnormalities, and are difficult to apply to risk advance warning in the early screening stage.

[0026] To solve the above technical problems, such as Figure 1 As shown, Embodiment 1 of this invention provides a dynamic quantitative early warning method for the risk of septic coagulopathy based on the joint detection of multiple biomarkers. This method uses urinary thrombomodulin (U-TM), urinary procalcitonin (PCT), and urinary C-reactive protein (CRP) as joint detection targets. Through dual-benchmark deviation modeling, time lag compensation, AUC-driven weighting, weak evidence accumulation, and cost-sensitive inference, it constructs a dynamic early warning architecture that crosses multiple dimensions including biochemical testing, disease progression time series, and machine learning, forming a closed-loop tracking and early warning framework adapted to early screening scenarios, thus achieving dynamic quantitative early warning for septic coagulopathy. Specifically, it includes the following steps: S1. Collect biomarker detection data of target patients at multiple time points within a preset monitoring period, and perform time alignment and confidence assessment to obtain preprocessed data with confidence coefficients; S2. Based on preprocessed data, dual-benchmark deviations are constructed using individual historical window statistics, population statistics, and adaptive fusion coefficients, and the effective deviations of each biomarker are obtained through noise floor constraints. S3. Determine the time delay compensation parameters based on the temporal correlation of each biomarker in the course of the disease, perform time delay compensation on the effective deviation, and obtain the time delay alignment deviation. S4. Unify the risk direction of the time lag alignment deviation and map it into a single indicator risk value; S5. Construct early screening sensitivity weights based on the area under the receiver operating characteristic curve and the dispersion trend term, and determine the dynamic weights of each biomarker. S6. Use dynamic weights to weight the risk values ​​of single indicators and construct a static fusion risk value and an early screening evidence accumulation term based on a sliding time window; S7. Construct interactive features that represent the synergistic enhancement relationship of multiple indicators, and set independent joint over-threshold gating rules; S8. The single-indicator risk value, static fusion risk value, early screening evidence accumulation term, interaction feature and discrete trend term are combined into a multi-dimensional dynamic feature vector, which is then input into the ensemble learning model trained with a cost-sensitive loss function to infer the risk probability of sepsis coagulopathy. S9. Obtain the basic risk score based on the risk probability mapping, and adjust the risk upward by combining the gating rules and the monotonically increasing trend of the score within the continuous time window. Output the final risk score and execute the graded early warning. S10. Perform closed-loop updates and retesting iterations based on the warning level.

[0027] In this invention, the time alignment completed in step S1 is the data foundation for all subsequent dynamic analyses; the dual benchmarks and time lag deviations constructed in steps S2 to S4 are physical benchmarks to overcome individual differences and asynchronous disease progression; the dynamic weights and evidence accumulation terms calculated in steps S5 and S6 are core steps for capturing weak signals in the prodromal period and maintaining the model's sensitivity to mild abnormalities; the gating rules in step S7 provide a safety net for extremely high-risk combinations; and steps S8 and S9 are the execution steps for achieving risk quantification and graded output. These steps work synergistically, breaking the limitations of traditional static single-indicator cutoff value judgments and achieving joint early warning of the highly concealed prodromal state of sepsis-related coagulopathy.

[0028] It is important to note that the evolution of septic coagulopathy (SIC) begins with occult damage to the microcirculatory system. In the prodromal phase, while systemic macroscopic coagulation parameters have not yet deviated significantly, substantial structural damage and loss of anticoagulant function of microvascular endothelial cells have already occurred. Thrombomotor proteins (TMs) are transmembrane glycoproteins widely expressed on the surface of vascular endothelial cells and are key receptors for maintaining local anticoagulation homeostasis in the vascular bed. When systemic inflammatory responses lead to endothelial barrier damage, TMs on the cell surface are abnormally cleaved by inflammatory proteases and shed into the bloodstream, forming soluble thrombomodulin (sTMs). The resulting free fragments are ultimately excreted via the renal filtration system, forming urinary thrombomodulin (U-TMs). Abnormal shedding of TMs is clearly defined as direct molecular evidence of endothelial structural damage, constituting an upstream leading event in the initiation of the microthrombus formation and SIC cascade.

[0029] In characterizing the depth of vascular endothelial barrier damage, U-TM exhibits clear biological equivalence to circulating sTM. Clinical validation has shown that the concentration of detached TM fragments is significantly positively correlated with the severity of sepsis, the Sequential Organ Failure Assessment (SOFA) score, and the incidence of secondary disseminated intravascular coagulation (DIC). When constructing a dynamic quantitative early warning system, compared to serum TM or other conventional endothelial molecules, selecting U-TM as the observation indicator offers the following pathophysiological and engineering extraction advantages: Sampling frequency and perturbation resistance: U-TM's retention is non-invasive, supporting high-frequency and even continuous sampling by the early warning system during the monitoring period, meeting the rigid requirements of sliding time windows for high-density time series data. In addition, the urine matrix naturally eliminates background interference from a large number of complex protein substrates and immune complexes in the blood, resulting in a high signal-to-noise ratio in the detection link, which is more conducive to the model capturing weak abnormal deviations at the edge of the background noise.

[0030] Target microenvironment sensitivity shifts earlier: In the early stages of sepsis-induced systemic microcirculatory disturbances, the renal microvascular bed endothelium is prone to early damage. Detached TM molecular fragments accumulate locally during glomerular filtration and tubular excretion. Therefore, the initial jump in urinary U-TM concentration, compared to the slow increase in serum concentration after dilution by total body fluid volume, can further advance the warning time window to the very early stage of microcirculatory hemodynamic disturbances.

[0031] Furthermore, to better illustrate the technical solution of Embodiment 1 of the present invention, a detailed description is provided of the method for dynamic quantitative early warning of septic coagulopathy risk based on the joint detection of multiple biomarkers, specifically including the following: First, S1 completes the time alignment of multi-timepoint biomarker detection data by reconstructing the time axis and assessing confidence, including the following sub-steps: S11. Within the preset monitoring period, perform continuous sampling at multiple time points on the target patient to obtain the raw test values ​​of urinary thrombomodulin U-TM, urinary PCT, and urinary CRP. Let the first... A biomarker at time The original detection value is ,in These correspond to U-TM, urinary PCT, and urinary CRP, respectively.

[0032] S12. To address issues such as inconsistent sampling times, delayed test results, and missing samples in actual clinical scenarios, the original data undergoes timeline reconstruction to uniformly map test results from different sources to the same time reference frame. For missing data, when the interval between adjacent time points is within the allowable range, the last observation carryover method is used. Alternatively, linear interpolation can be used for repair; for outliers, robust statistical rules can be used for filtering.

[0033] S13. To avoid the compensated data and the measured data having the same weight, a confidence coefficient is defined for each time point and each indicator. If the data at that moment is directly measured, then take... If the data at that moment is interpolated or carried over, then take... In the formula, For this indicator at time The compensation span; In order to target the The specific decay coefficient of each biomarker, which is based on the biological half-life of each biomarker. Confirmed, the expression is: When the compensation span of the interpolated data When the inherent half-life of this indicator is reached, its confidence coefficient automatically decays to 0.5. In real-world complex conditions (such as patients with concomitant renal or hepatic clearance dysfunction), a decay hysteresis factor greater than 1 is introduced based on the patient's actual glomerular filtration rate or liver function rating. Multiplicative corrections are performed to match the slower pathological metabolic rate and data expiration rate in organ failure states. In subsequent steps, all features generated from the compensated data are multiplied by their corresponding confidence coefficients.

[0034] Furthermore, S2 constructs a dual-benchmark deviation for early screening scenarios by integrating individual and group differences, including the following sub-steps: S21. Extract the first [item] from the patient's individual history window. The mean and standard deviation of each biomarker are denoted as follows: and Simultaneously, the mean and standard deviation of the corresponding indicators for the group training samples are extracted and denoted as follows: and .

[0035] S22. Convert the raw concentration values ​​into a dual-benchmark deviation that reflects the degree of deviation from the individual baseline and the population baseline. The specific expression is: ; In the formula, The fusion coefficient between the individual baseline and the group baseline; To prevent extremely small positive numbers with a denominator of zero.

[0036] Specifically, for example: Although the absolute value of a patient's urine CRP has not reached the generally accepted strong positive threshold (population out of bounds), it has increased sharply compared to the extremely low baseline level on the first day of admission (individual baseline). In this case, a higher individual baseline fusion coefficient can be assigned. The system can keenly detect weak deterioration signals that have not crossed the boundary but have significantly deviated from the individual's steady state, thereby enabling early warning.

[0037] Among them, the fusion coefficient in step S22 A dynamic adaptive determination mechanism can be used, which specifically includes the following steps: The inherent fluctuations in patients' physiological indicators vary greatly. To quantify the reliability and stability of such historical data, this embodiment introduces the coefficient of variation. At each evaluation time point, the adjusted effective individual standard deviation is used. Compared with individual baseline mean Perform the calculation: ; In the formula, To prevent extremely small positive numbers with a denominator of zero, the larger the coefficient of variation, the more drastic the fluctuation of this indicator within the historical window, and the weaker the reference value of the individual baseline.

[0038] Furthermore, this embodiment comprehensively considers the statistical confidence brought by the amount of effective historical data and the physiological stability brought by the coefficient of variation, and constructs a nonlinear adaptive weight mapping equation to update the fusion coefficient in real time: ; In the formula: This represents the maximum permissible individual fusion weight for this type of biomarker, due to the endothelial injury index. Individual specificity is usually stronger than systemic inflammatory markers. Different upper limits can be set for different indicators; This represents the number of valid historical sample points acquired at the current moment. Set the preset historical window to full capacity; This is the data richness gain coefficient, used to control the rate at which sample accumulation positively increases the weights; This is the variability decay coefficient, used to control the exponential penalty of high-frequency fluctuations on the weights.

[0039] Based on the above mapping equation, the fusion coefficient It has a two-way adjustment capability: when data is scarce, the weights are forcibly reduced to prevent overfitting; when data is sufficient but the patient's baseline condition is extremely unstable, the dependence on the chaotic individual baseline is reduced through the decay term; only when data is sufficient and the individual is in a relatively stable state will the model accept the individual's deviation from the two baselines to the greatest extent.

[0040] For example: Both critically ill patient A and patient B received treatment in the ICU for 5 consecutive days. Monitoring to ensure full-load windows That is, the data richness is the same.

[0041] For patient A (stable type): whose historical urinary PCT level has remained consistently low, the coefficient of variation was calculated. Due to the abundance and high stability of the data, the system calculates the fusion coefficients through the mapping equation. The value can reach as high as 0.85. If a small but significant increase in urinary PCT occurs at this point, the system will highly trust the individual's abnormal signal and promptly trigger an early warning.

[0042] For patient B (fluctuating pattern): This patient has chronic systemic inflammation, and their historical urinary PCT values ​​fluctuate repeatedly within a high range. The coefficient of variation was calculated. Due to the extremely chaotic historical data, the system uses an exponential decay term to adjust the fusion coefficient. The automatic compression is reduced to 0.30. At this point, the model will automatically tilt the evaluation caliber towards the population baseline, requiring that the deviation must reach a more stringent population limit before issuing an alarm, thereby avoiding frequent false alarms caused by excessive baseline fluctuations.

[0043] In this embodiment, by way of example: History window full capacity The calculation formula is: In the formula, The clinical observation period required to establish physiological homeostasis characteristics in critically ill patients. This is the standard sampling time interval in real-world medical scenarios. This is a rounding function. Under continuous monitoring conditions in the ICU for septic coagulopathy, The value is typically between 72 and 120 hours. It typically takes 12 to 24 hours. Based on this calculation, the result is... The recommended value range is 3 to 10.

[0044] Data fullness gain coefficient This is used to control the non-linear growth rate of the system's confidence in an individual's baseline as effective samples accumulate. It ensures that when the sample size reaches full capacity (i.e., When the gain term smoothly approaches the desired maximum saturation threshold, the gain term can smoothly approach the maximum saturation threshold. (Usually set between 0.90 and 0.95). The expression is: In actual deployment, The standard configuration is typically between 2.5 and 3.5.

[0045] Variation attenuation coefficient Used for analyzing historical data with significant fluctuations (i.e., coefficient of variation) Individuals with excessively high confidence levels are penalized. The value is determined based on the maximum tolerable physiological coefficient of variation for different biomarkers. Compared to the preset maximum penalty retention rate Confirmed, the expression is: In the formula, To extract the first from clinical retrospective statistical prior data The extreme values ​​are determined by the disordered fluctuations of each indicator. Because the individual baseline fluctuations of endothelial damage markers (such as U-TM) are relatively small, The standard value is generally set at 0.3 to 0.4; however, systemic inflammatory markers (such as urinary PCT or urinary CRP) are prone to significant fluctuations due to nonspecific stress responses. and The value is typically set between 0.6 and 0.8. This is used to determine the individual confidence penalty retention rate at which the maximum volatility tolerance value is reached. (That is, only 10% of the individual weights are retained), and the endothelial injury markers can be calculated. The value range is approximately 5.7 to 7.6, indicating an inflammatory marker. The value range is approximately 2.8 to 3.8.

[0046] As an optional embodiment, in step S21, for patients who are newly diagnosed or lack historical window data, this embodiment supplements individual parameters by introducing a population prior-guided individual baseline evolution mechanism, specifically including the following sub-steps: S211. For cases where the number of historical monitoring points is less than the preset window length For newly diagnosed patients, this embodiment uses either the population mean mapping method or the first-test value anchoring method for initialization: Strategy A (Group Prior Method): Set the initial individual baseline. Initial individual standard deviation In the absence of data, it is assumed that the patient is at the median level for that age / gender / disease group.

[0047] Strategy B (First Clinical Test Method): Using the high signal-to-noise ratio detection value of the first sample taken after the patient's admission. As the initial baseline And set a large initial deviation tolerance value. This is to avoid false triggering of a single high value during the initial diagnosis.

[0048] S212. To address the jump problem during the transition from group baseline to individual baseline, a method based on the number of observed samples is introduced. Dynamic confidence factor Redefine its fusion baseline: ; In the formula, This represents the current number of valid sampling points. The preset historical window has a full load length. This mechanism ensures that as the monitoring time extends, the system evaluation indicators automatically and smoothly evolve from general group standards to individual sensitive standards.

[0049] S213. In complex situations such as emergency rooms, if a patient's initial test value upon admission is at an extremely high-risk level, such as a significantly abnormal urinary PCT, to prevent pathologically high values ​​from being mistaken for normal individual baselines and thus leading to a decrease in subsequent warning sensitivity, this embodiment sets a baseline saturation cutoff logic: like Then forcibly lock This ensures that for patients admitted with severe symptoms, the early warning logic reverts to the criteria for identifying highly sensitive groups; among which... The threshold for significant anomalies defined for the population.

[0050] Specifically, for example: Let's assume a patient in the emergency department is a first-time visitor with zero historical data points. After collecting the first urine CRP test value, the system detects that the historical window does not meet the calculation conditions. At this point, the system automatically activates the cold start strategy: During the first evaluation period: the group mean will be temporarily used. (like () serves as its individual baseline.

[0051] With continuous monitoring after hospitalization: Number of sampling points Increase. When (Window length) When individual data accounts for only a small percentage of the total weight, the weight of individual data is only a small percentage of the total weight. The system still largely references group standards.

[0052] When the data points are full: the system switches to a pure individual baseline assessment based on the patient's past 24 test values.

[0053] The above design can prevent the early warning blind spot in newly diagnosed patients during the historical gap period, and at the same time solve the problem of baseline elevation that may be caused by excessively high initial test values ​​in critically ill patients upon admission, thereby improving the universality of the program in dynamic screening in emergency, ICU and outpatient departments.

[0054] As an optional embodiment, in step S22, the individual standard deviation In the early stages, the deviation may be very small, causing the individual deviation to be amplified dramatically, generating noise. Noise-resistant stability optimization can be performed, which includes the following steps: S221, Regarding individual standard deviation To address the instability caused by minimal values, this embodiment introduces an effective individual standard deviation with practical clinical physical significance. Define the lower bound of regularization: In the formula, The physiological fluctuation tolerance coefficient is used to characterize the minimum physiological fluctuation limit inherent in the indicators of this population. This refers to the inherent measurement error of the biomarker detection instrument, such as the fluctuation of the reagent kit's limit of detection or the absolute error determined by the coefficient of variation (CV).

[0055] S222, Original individual standard deviation Compare with the calculated lower bound of the noise floor, and forcibly truncate excessively small variances: ; Through the above operations, even if a patient's individual historical data presents a straight line, that is... The denominator will also be supported by a reasonable physiological / instrumental background noise, thus cutting off the path of mathematical amplification of minute noise.

[0056] S223. Using the corrected effective individual standard deviation Replace the original individual standard deviation in the original formula Reconstruct the formula for calculating the deviation between the two benchmarks: ; Through the above regularization process, both the extremely high sensitivity to real abnormal mutations in individuals is preserved, and mathematical divergence caused by abnormally small denominators during the stationary period is suppressed. Specifically, for example: a certain ICU patient's urinary PCT was extremely stable for the three days prior to admission, with consecutive tests showing zero values. The calculated individual baseline mean Individual standard deviation .group The standard deviation is At this point, due to non-infectious stress such as turning over and suctioning, the patient experienced [a condition] on the fourth day. Slight fluctuations to .

[0057] If no noise floor constraint is applied: the deviation calculation includes... ,like This tiny fluctuation will be amplified to Extremely abnormal signals can lead to the highest level of false alarms in the system.

[0058] After introducing dual noise floor constraints: the system extracts instrument detection errors. lower limit of group physiology The system uses the maximum value as the noise floor, i.e. At this point, the individual deviation component becomes This value reasonably reflects slight fluctuations in the patient's condition, but is far from reaching a level of deviation that warrants an alarm, successfully filtering out artifact noise.

[0059] Furthermore, S3, by mining temporal correlation patterns, performs time-lag compensation on different indicators to adapt to the asynchronous evolution of the disease course, including the following sub-steps: In the pathological cascade of sepsis, pathogen invasion and local infection first induce a systemic inflammatory response, subsequently leading to widespread endothelial cell damage, and ultimately triggering an imbalance in the coagulation and anticoagulation systems (septic coagulopathy). This physiological evolution determines the asynchronous dynamic characteristics of the expression of different biomarkers in blood and urine: Urinary procalcitonin (PCT), a sensitive indicator of early systemic infection and inflammatory burden, begins rapid transcription and release within 3 to 6 hours after bacterial toxin stimulation, typically reaching its peak at 12 to 24 hours. Its biological half-life... It takes approximately 20 to 24 hours. Urinary PCT is characterized by rapid onset and early peak, serving as a leading warning signal for the initiation of a pathological cascade.

[0060] Urinary C-reactive protein (CRP), an acute-phase protein synthesized by the liver, is regulated in transcription by downstream inflammatory factors such as interleukin-6 (IL-6), exhibiting a significant transcriptional delay. Urinary CRP typically rises 6 to 8 hours after infection, peaks at 36 to 50 hours, and has a half-life of [missing information]. It takes approximately 19 hours. Urinary CRP exhibits characteristics of delayed response and delayed peak.

[0061] The concentration level of thrombomodulin U-TM directly reflects the degree of damage and shedding of vascular endothelial cells and coagulation abnormalities in the microvascular bed. Endothelial injury is usually a direct consequence of severe systemic inflammatory response; therefore, the abnormal rise of U-TM often lags behind the initial surge in urinary PCT over time, and exhibits persistently high expression levels as coagulation dysfunction worsens.

[0062] In traditional static multi-index joint models, the same sampling time is directly used. Feature fusion of the original detection values ​​can cause a phase misalignment between the decay period of the early signal (urinary PCT) and the rise period of the late signal (U-TM), weakening the model's ability to discriminate prodromal risks.

[0063] Therefore, this embodiment sets time delay compensation parameters for each biomarker. And define the time-delay alignment deviation as: ;in, Based on the temporal correlation between changes in various indicators and the occurrence of outcomes in the training samples, the discriminative ability of the outcome can be estimated by maximizing the corresponding features, i.e.: ;in, For the candidate time delay set, These are the labels for the training samples.

[0064] This step allows risk signals from different indicators to be chained into the model in the order of their biological evolution, such as the first increase in inflammatory markers followed by changes in endothelial damage markers.

[0065] Furthermore, S4 uses risk semantic space transformation to unify the direction of the deviation after time lag compensation and map it to a single-index risk value, including the following sub-steps: S41. Since different biomarkers have different risk orientations, the deviation is first based on the time lag compensation. Constructing a unified quantity in the direction: ; in, The corresponding U-TM is a low-risk indicator, therefore, when its detected value decreases, The value is negative; after negation, A positive value indicates a higher risk; urinary PCT and urinary CRP are high-risk indicators, so their original deviations are retained.

[0066] S42. Further, define the first The initial risk value for a single biomarker is: ; in, ; The sensitivity adjustment parameter can be set through cross-validation or empirically. For the early weak signal indicator urinary thrombomodulin (U-TM), the deviation in the prodromal phase is small, so a higher sensitivity is needed to amplify the weak abnormality. The value range can be set to Regarding the intermediate-stage inflammatory evolution marker urinary procalcitonin (PCT), its response speed is moderate. The value range can be set to For urinary C-reactive protein (CRP), an indicator that is late-stage and susceptible to non-specific interference, its basal physiological fluctuations are significant. Therefore, a gentler mapping slope is needed to suppress false-positive noise. The value range can be set to .

[0067] Meanwhile, to mitigate the misleading effects of compensated data and low-confidence samples, a confidence-corrected risk value is introduced: ;in, The confidence coefficient is defined in step S1.

[0068] The above mapping transforms U-TM, urinary PCT, and urinary CRP from their original detection values ​​into standardized features within the same risk semantic space, facilitating subsequent joint modeling.

[0069] Furthermore, S5 constructs early screening sensitivity weights based on AUC by introducing overall discriminative power and early change trends, including the following sub-steps: S51. Based on the area under the ROC curve for each indicator and random discrimination reference value Define effective discrimination gain and construct initial weights .

[0070] S52. To adapt to the risk evolution characteristics of mild abnormalities but continuous increases in early screening scenarios, a discrete trend term is defined: In the formula, The interval between adjacent evaluation times.

[0071] S53, Define the intermediate weights after trend correction The early screening sensitivity weights were obtained by normalization. In the formula, This is the trend enhancement coefficient. This design allows the model to be more biased towards capturing dynamic weak signals in the early stages of risk shift.

[0072] Among them, the trend enhancement coefficient The methods for determining this include: S531. The metabolic kinetics of different biomarkers vary greatly. For example, the half-life of urinary PCT is approximately 20-24 hours, while the half-life of urinary CRP is approximately 19 hours with a significant peak delay. To avoid overfitting caused by purely data-driven approaches, we first combine prior clinical knowledge to extract the first... The average biological half-life of a biomarker in the human body Based on half-life construction Offline optimization of feasible domains (candidate space) : In the formula, This is an empirical scaling constant.

[0073] S532. During the offline training phase before model deployment, extract electronic medical records and laboratory information system (LIS) data from patients undergoing continuous historical monitoring in the Intensive Care Unit (ICU) or Emergency Intensive Care Unit (EICU). Positive sample set. Defined as a score according to the International Society for Thrombosis and Haemostasis (ISTH) criteria within 72 hours of admission. Score, or Sequential Organ Failure Assessment (SOFA) coagulation system score. Patients diagnosed with septic coagulopathy (SIC) were identified through a negative sample set. Defined as infected patients admitted at the same time but who have not developed SIC.

[0074] In order to evaluate different The combined effect allows us to construct an early warning time gain function that relies solely on the weighting of single-indicator features. Its expression is as follows: ; In the formula, Let be the coefficient vector to be optimized; The number of real positive samples of sepsis coagulopathy in the training set; For the first The clinical moment when a positive sample is diagnosed as septic coagulopathy; For a given Below, relying solely on weighted single-indicator risk and The moment when the basic statistical threshold is first exceeded; This represents the number of false positives (false alarms) caused by this set of parameters in negative samples; The weighting for false alarm penalties.

[0075] S533, in the feasible region Within this framework, an orthogonal grid search or Bayesian optimization algorithm is used to solve for the objective function that maximizes the time gain, thereby obtaining the optimal trend enhancement coefficient vector. ; After the solution is completed, (Right now It can be burned or solidified as a static configuration parameter into the configuration file of the online early warning system.

[0076] Furthermore, S6 constructs a static fusion risk value and a dynamic evidence accumulation term based on a sliding time window mechanism, including the following sub-steps: S61. After obtaining the risk values ​​and weights of each indicator, construct the static fusion risk value: ; S62. Considering that early screening often involves the gradual accumulation of minor abnormalities over multiple time windows, rather than a single significant breach, a dynamic evidence accumulation term is further constructed. Let the sliding time window length be... The time decay weight is Then the early screening evidence integral term is defined as: ;in, This is used to emphasize recent evidence signals while preserving the cumulative effect of previous anomalies.

[0077] Among them, the sliding time window length and time decay weight The methods for determining this include: S621, Time Window The physical meaning lies in how long the system can remember an abnormal state from before. This embodiment links it to prior knowledge of the pathological evolution of septic coagulopathy (SIC). It extracts the upper limit of the typical physiological latency period for the evolution of the target disease from a mildly abnormal prodromal phase to a clinically diagnosed state. For septic coagulopathy, this cycle is typically 24 to 48 hours. This should be combined with the actual sampling frequency of the medical institution, i.e., the interval between adjacent assessment times. Calculate and solidify the time window length: In the formula, This is the floor function.

[0078] S622. Define the physical time span from the current evaluation time. Construct unnormalized weights: The time decay weights are obtained after normalization: In the formula, This represents the time step from the current time. This represents the forgetting decay rate.

[0079] S623. Utilize the dynamic weights of each biomarker determined in step S5. and the average biological half-life of each pre-extracted biomarker in the human body. Calculate the physiological reference decay rate of the current multi-indicator combined system. : ; During the offline training phase, a trajectory resolution gain function incorporating a physiological prior regularization penalty term is constructed. : ; In the formula, To calculate the integral term of early screening evidence against the true outcome label for all training samples at each time step. The area under the receiver operating characteristic curve; This is the regularization penalty coefficient. The optimal forgetting decay rate is determined by maximizing the objective function. ; Finally Substitute into the formula to calculate the static time decay weighted reassembly And it is embedded into the online system.

[0080] Furthermore, S7 characterizes the collaborative relationships of multi-dimensional indicators, constructs interactive features, and sets independent security gating rules, including the following sub-steps: S71. To further characterize the synergistic enhancement relationship among inflammatory response, endothelial injury, and coagulation abnormalities, the following interactive features are constructed: This feature is used to characterize the enhanced coupling effect when inflammation, endothelial damage, and coagulation abnormalities occur simultaneously.

[0081] S72. To avoid logical conflicts between hard rules and model output, the joint over-threshold judgment is no longer used as a model input feature, but is set as an independent security gating rule. The joint over-threshold gating variable is defined as follows: ; in, The risk threshold is the corresponding indicator.

[0082] Will As a fallback gating signal after the model output, it is used to enforce the identification of extremely high-risk combinations.

[0083] Furthermore, S8 integrates multi-dimensional dynamic feature vectors and introduces a cost-sensitive loss function for model inference, including the following sub-steps: S81. Combine the above-mentioned single-indicator risk value, static fusion risk value, dynamic evidence accumulation term, interaction feature, and trend term to form a dynamic feature vector. : ; S82, from historical case samples As input, whether a patient progresses to septic coagulopathy within a preset prediction window is used as a label to construct a supervised training set, which is then trained using an ensemble learning model.

[0084] Preferably, the ensemble learning model is a gradient boosting tree model. The model outputs the probability of septic coagulopathy risk. for: ;in, For the trained ensemble model, For the Sigmoid function, These are the model parameters.

[0085] S83. Considering that the cost of false negatives is significantly higher than the cost of false positives in early screening scenarios, a cost-sensitive loss function is introduced during model training to assign a higher penalty coefficient to positive samples: ; in, For sample labels, To predict probabilities, The penalty coefficient for positive samples. Let the sample confidence weights be: ; Through this training strategy, the model will focus more on identifying high-risk samples in the early stages during the optimization process, thereby improving its ability to predict the prodromal state of sepsis coagulopathy.

[0086] Furthermore, S9 combines model output with gating rules to output a risk score and execute tiered early warnings, such as... Figure 2 As shown, it includes the following sub-steps: S91. Mapping risk probability to a unified risk score : ;in, This is the maximum score.

[0087] S92. Set high-risk thresholds and medium risk threshold And satisfy The corresponding warning result will be output when the following conditions are met: when When the situation is deemed high-risk, the highest level of warning is triggered. when When the situation is determined to be medium-risk, a continuous monitoring and early warning system is triggered. when At that time, it was determined to be a low-risk state, and routine monitoring was maintained.

[0088] At the same time, establish risk escalation rules for independent security gating. Define the most recent consecutive... The increasing risk trend within a time window satisfies: That is, the basic risk score is in continuous Strictly monotonically increasing within a time window. When simultaneously satisfying... And under the aforementioned strictly monotonically increasing condition, the output gating upgrade score Otherwise Final output score: .

[0089] As an optional embodiment, in step S92, the high-risk threshold... and medium risk threshold The determination steps include: S921. During the offline calibration phase before model deployment, extract a clinically independent validation set that is strictly isolated from the training set. Through steps S8 and S91, a uniform risk score is calculated for each sample in the validation set. For any candidate threshold Define a statistical evaluation function on the validation set: sensitivity ( ): Characterized by the ability to capture truly high-risk patients; False positive rate (FPR): Characterized by the degree of false alarm perturbation in normal patients; Positive predictive value (PPV): This indicates the probability that a patient will actually develop the disease once the system issues an alarm.

[0090] S922. The clinical significance of the intermediate-risk status lies in screening and observation, triggering subsequent retesting mechanisms. Therefore, its core requirement is high sensitivity (i.e., extremely high negative predictive value NPV) to ensure the true safety of patients classified as low-risk. A minimum clinically acceptable sensitivity threshold should be set. (like Medium risk threshold The determination logic is as follows: while ensuring that the sensitivity meets the standard, the threshold should be increased as much as possible to filter out safe samples. Its expression is: If the validation set data is limited, making it impossible to satisfy the requirements... Then, the highest score corresponding to the maximum sensitivity is taken as... .

[0091] S923. The clinical significance of a high-risk status lies in the immediate initiation of clinical intervention, which consumes core medical resources and brings potential drug side effects. Therefore, the core requirement of a high-risk threshold is extremely high specificity and positive predictive value. A maximum upper limit should be set for the false positive rate that clinical resources can tolerate. (usually taken) to and minimum confidence requirements High-risk threshold The determination logic is as follows: .

[0092] At the same time, to ensure logical consistency, additional constraints are added. Among them, the minimum hierarchical buffer Based on the rating cap The average standard error of the model's output probabilities on the validation set Jointly determined: This constraint prevents unnatural jumps in patient risk levels due to biomarker detection noise or inherent model errors.

[0093] S924. In actual continuous monitoring, if a patient's risk score fluctuates slightly around a threshold, for example... exist Repeated jumps between up and down can lead to frequent pop-ups and cancellations on healthcare terminals. To address this, a hysteresis interval, similar to that of a Schmitt trigger in electronics, is introduced. : Upgrade logic: When At that time, a high-risk alert is triggered.

[0094] Degradation logic: Only when Furthermore, a risk level can only be downgraded from high to medium if the monitoring continues for at least one monitoring window.

[0095] And so on, for Set the lower boundary of hysteresis .

[0096] In this embodiment, the clinical significance of a medium-risk status is as follows: the captured abnormal signals have substantially broken through the dual noise floor constraints of individuals and the population, making it impossible to safely rule out the risk of septic coagulopathy, but the confidence level of the evidence for disease progression has not yet reached the threshold for initiating high-load medical intervention. When the system experiences threshold conflicts during offline calibration, such as a low area under the overall receiver operating characteristic curve (AUC) of the model, the initially determined high-risk threshold calculated based on constraints may be lower than... Implement a constraint relaxation and degradation solution strategy: lock the medium-risk threshold determined in step S922. To ensure that no bottom line is missed in the early stages; the objective function is gradually relaxed with a fixed step size. and Limit until the buffer constraint is met. Thus, a viable option is re-established. In actual early warning implementation, patients falling into the medium-risk range are handled through the closed-loop update mechanism in step S10: the alarm is temporarily suspended and the retesting cycle is automatically shortened. The system relies on the accumulated early screening evidence continuously calculated in step S6 within subsequent time windows. The time integral accumulation, or the independent safety gating signal triggered by the extreme combination in step S7. This will drive its final risk score. directional evolution.

[0097] Finally, S10 performs closed-loop update and retest iteration through a closed-loop feedback mechanism, including the following sub-steps: After the system outputs the warning result, it pushes the current risk level and corresponding clinical recommendations to the medical staff terminal. If the patient is in a medium or high risk state, the retesting mechanism is triggered. U-TM, urine PCT and urine CRP data are collected again in the next preset time window, and steps S1 to S9 are repeated. This allows the risk assessment results to be continuously updated as the patient's disease progresses, forming a closed-loop warning mechanism.

[0098] Through the above embodiments, the present invention is not only for conventional joint classification, but also for the specific scenario of early screening of sepsis coagulopathy. It has made targeted optimizations to conventional algorithms, such as dual benchmark deviation, time delay compensation, direction unification, weak evidence accumulation, trend weighting, and cost-sensitive training, so that it can be adapted to the clinical application environment of weak signals in the prodromal period, asynchronous evolution, and high underreporting cost.

[0099] Example 2 As a second embodiment of the present invention, such as Figure 3 As shown in Example 1, this example also discloses a dynamic quantitative early warning system for the risk of septic coagulopathy based on the joint detection of multiple biomarkers, specifically including: The data preprocessing module is used to collect biomarker detection data of target patients at multiple time points within a preset monitoring period. And perform time alignment and confidence assessment to obtain a confidence coefficient. Preprocessed data; The dual-benchmark deviation construction module is used to construct deviations from preprocessed data using individual historical window statistics, population statistics, and adaptive fusion coefficients. We constructed a dual-benchmark deviation and obtained the effective deviation of each biomarker by using noise floor constraints. ; The time delay compensation module is used to determine the time delay compensation parameters based on the temporal correlation of each biomarker during the course of the disease. For effective deviation Perform time delay compensation to obtain the time delay alignment deviation; The single-indicator risk mapping module is used to unify the risk direction of time lag alignment deviations and obtain the direction unification quantity. The risk value after confidence correction is output after mapping. ; The early screening sensitivity weighting construction module is used to construct weights based on the area under the receiver operating characteristic curve and the dispersion term. Construct early screening sensitivity weights and determine the early screening sensitivity weights of each biomarker. ; Static and dynamic feature building modules are used to utilize early screening sensitivity weights. Risk value after confidence correction Weighting is performed to construct a static fusion risk value. Compared with the accumulation of early screening evidence based on sliding time windows ; The interaction feature and gating module is used to construct interaction features that represent the synergistic enhancement relationships of multiple indicators. And set independent joint over-threshold gating rules to generate gating signals. ; The risk reasoning and scoring module is used to calculate the confidence-adjusted risk value. Static fusion risk value Early screening evidence accumulation item Interaction features and discrete trend term Composition of multidimensional dynamic feature vectors Inputting the data into an ensemble learning model, the model infers the probability of septic coagulopathy risk. And based on risk probability Mapping yields the basic risk score Combined with gating signals The risk score is adjusted upwards based on the monotonically increasing trend of the score within a continuous time window, and the final risk score is output. And implement tiered early warning systems; The closed-loop control module is used to trigger retesting and closed-loop updates based on the warning level.

[0100] In the specific implementation of the above embodiment 2, the data preprocessing module will process the multi-time point biomarker detection data. Convert to standardized input features. The dual-benchmark deviation construction module combines individual coefficients of variation. Generate adaptive fusion coefficients Together with the time lag compensation module, it eliminates errors caused by individual differences in baseline conditions and asynchronous temporal evolution during the disease course. The single-indicator risk mapping module and the early screening sensitive weight construction module transform the underlying signal into a signal with a trend enhancement coefficient. The risk mapping value. The static and dynamic feature construction module, along with the interactive feature and gating module, establishes security upgrade rules for complex and extreme situations by combining multi-dimensional indicators and sliding time window features. The risk reasoning and scoring module analyzes multi-dimensional dynamic feature vectors. The system performs logical calculations and outputs the final quantified risk, which is then used for monitoring iterations via a closed-loop control module. These modules work together to extract and classify the risk characteristics of the disease.

[0101] The preferred embodiments of the invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A method for dynamic quantitative early warning of septic coagulopathy risk based on the combined detection of multiple biomarkers, characterized in that, include: Collect biomarker detection data of target patients at multiple time points within a preset monitoring period, and perform time alignment and confidence assessment to obtain preprocessed data with confidence coefficients; Based on the preprocessed data, a dual-benchmark deviation is constructed using individual historical window statistics, population statistics, and adaptive fusion coefficients, and the effective deviation of each biomarker is obtained through noise floor constraints. Based on the temporal correlation of each biomarker in the course of the disease, the time delay compensation parameter is determined, and the effective deviation is compensated for by time delay to obtain the time delay alignment deviation. The time lag alignment deviation is unified in risk direction and mapped to a single indicator risk value; Early screening sensitivity weights are constructed based on the area under the receiver operating characteristic curve and the dispersion trend term, and the dynamic weights of each biomarker are determined. The risk values ​​of the single indicators are weighted using the dynamic weights to construct a static fusion risk value and an early screening evidence accumulation term based on a sliding time window; Construct interactive features that characterize the synergistic enhancement relationship of multiple indicators, and set independent joint over-threshold gating rules; The single-indicator risk value, static fusion risk value, early screening evidence accumulation term, interaction feature and discrete trend term are combined into a multi-dimensional dynamic feature vector, which is then input into an ensemble learning model trained with a cost-sensitive loss function to infer the risk probability of sepsis coagulopathy. The basic risk score is obtained based on the risk probability mapping. The risk is then adjusted upwards by combining the gating rules with the monotonically increasing trend of the score within a continuous time window. The final risk score is output and a graded early warning is executed. Retesting and closed-loop updates are triggered based on the early warning level.

2. The method for dynamic quantitative early warning of septic coagulopathy risk based on the combined detection of multiple biomarkers according to claim 1, characterized in that, The time alignment and confidence assessment includes: Biomarker detection data from different sources are uniformly mapped to the same time reference frame, missing data are repaired by linear interpolation, and outlier values ​​are filtered by robust statistical rules. A confidence coefficient is assigned to each indicator at each time point, where the confidence coefficient for measured data is 1, and the confidence coefficient for interpolated or carried-over data is determined exponentially based on the compensation span. The biomarker detection data include urinary thrombomodulin, urinary procalcitonin, and urinary C-reactive protein.

3. The method for dynamic quantitative early warning of septic coagulopathy risk based on the combined detection of multiple biomarkers according to claim 1, characterized in that, The constructed dual-benchmark deviation includes: Extract the mean and standard deviation of each indicator within the individual patient's historical window, as well as the mean and standard deviation of the corresponding indicators in the group training samples; The individual baseline deviation component and the group baseline deviation component are weighted and combined using adaptive fusion coefficients; A lower bound for noise floor is constructed using instrument measurement error and the lower bound for population physiological fluctuations. The individual standard deviation is truncated and the larger value is taken to obtain the effective individual standard deviation, which is then used to replace the original individual standard deviation in the deviation calculation. The adaptive fusion coefficient is dynamically determined based on the individual variation coefficient, the number of valid historical sample points, and the preset upper limit of indicator specificity, so that when the data is abundant and the individual is in a steady state, the individual baseline is emphasized, and when the data is scarce or the individual fluctuates drastically, the group baseline is emphasized.

4. The method for dynamic quantitative early warning of septic coagulopathy risk based on the combined detection of multiple biomarkers according to claim 1, characterized in that, The time delay compensation includes: A set of candidate time delays is pre-defined. On the training samples, with the goal of maximizing the discriminative ability of the deviation amount on the outcome of septic coagulopathy, the optimal time delay compensation parameters are selected for each indicator. The effective deviations of each indicator are shifted according to the time lag compensation parameter, so that the abnormal signals of inflammatory response, endothelial injury and coagulation abnormality markers are aligned in the order of biological evolution on the disease time axis.

5. The method for dynamic quantitative early warning of septic coagulopathy risk based on the combined detection of multiple biomarkers according to claim 1, characterized in that, The time lag alignment deviation is unified in risk direction and mapped to a single indicator risk value, including: Based on the direction of indicator risk, the deviation of low-risk indicators is reversed, while the original direction of high-risk indicators is retained, thus obtaining a unified direction quantity. The direction unification quantity is mapped to the interval between 0 and 1 using a Sigmoid function, and the mapping steepness is controlled by a sensitivity adjustment parameter to obtain the initial risk value; Multiply the initial risk value at each time point by the confidence coefficient to obtain the confidence-corrected single-index risk value.

6. The method for dynamic quantitative early warning of septic coagulopathy risk based on the combined detection of multiple biomarkers according to claim 1, characterized in that, Early screening sensitivity weights are constructed based on the area under the receiver operating characteristic (ROC) curve and the dispersion term, and the dynamic weights of each biomarker are determined, including: The basic weights are determined based on the gain of the area under the receiver operating characteristic curve (AUC) of each indicator relative to the random discriminant reference value. Construct a discrete trend term using the rate of change of single-indicator risk values ​​at adjacent time points; The trend enhancement coefficient is used to correct the trend of the basic weights and normalize them to obtain the dynamic weights; The trend enhancement coefficient is determined offline by maximizing the early warning time gain function, based on the optimal feasible region determined by the biological half-life of each indicator.

7. The method for dynamic quantitative early warning of septic coagulopathy risk based on the combined detection of multiple biomarkers according to claim 1, characterized in that: The static fusion risk value is the weighted sum of the dynamic weights and the single-indicator risk values ​​after confidence correction. The early screening evidence accumulation term is the time-decayed weighted cumulative sum of the weighted sums at each moment within the sliding time window; The sliding time window length is determined based on the upper limit of the typical physiological latency period of septic coagulopathy and the sampling interval. The time decay weight is obtained by normalization based on the exponential decay of the physical time span and the forgetting decay rate. The forgetting decay rate is determined offline by maximizing the trajectory resolution gain function with physiological prior regularization penalty term, based on the physiological reference decay rate determined by combining the biological half-life of each biomarker.

8. The method for dynamic quantitative early warning of septic coagulopathy risk based on the combined detection of multiple biomarkers according to claim 1, characterized in that: The interaction feature is the sum of the products of different indicator risk values; The joint threshold gating rule is as follows: when the risk value of each indicator exceeds the corresponding preset risk threshold, a gating signal is triggered; otherwise, it is not triggered.

9. The method for dynamic quantitative early warning of septic coagulopathy risk based on the combined detection of multiple biomarkers according to claim 1, characterized in that: The ensemble learning model is a gradient boosting tree model. During training, the cost-sensitive loss function assigns a higher penalty coefficient to positive samples that progress to sepsis coagulopathy than to negative samples, and the sample confidence weight is the product of the confidence coefficients of each indicator. The step of outputting the final risk score and executing tiered early warning includes: mapping the risk probability to a preset scoring range to obtain a basic score; When the gating signal is met and the base score shows a strict monotonically increasing trend over multiple consecutive time windows, the final score is forcibly set to the upper limit of the score; otherwise, the base score is used as the final score. Based on the low, medium, and high risk threshold ranges into which the final score falls, regular monitoring, continuous observation and early warning, and the highest level early warning are executed respectively, and corresponding level retesting and closed-loop updates are triggered.

10. A dynamic quantitative early warning system for the risk of septic coagulopathy based on the combined detection of multiple biomarkers, employing the method described in any one of claims 1 to 9, characterized in that, include: The data preprocessing module is used to collect biomarker detection data of target patients at multiple time points within a preset monitoring period, and to perform time alignment and confidence assessment to obtain preprocessed data with confidence coefficients. The dual-benchmark deviation construction module is used to construct dual-benchmark deviations and obtain the effective deviation of each biomarker through noise floor constraints. The time delay compensation module is used to determine the time delay compensation parameters based on the temporal correlation of each biomarker in the course of the disease, to perform time delay compensation on the effective deviation, and to obtain the time delay alignment deviation. The single-indicator risk mapping module is used to unify the risk direction of time lag alignment deviation and map it to a single-indicator risk value. The early screening sensitivity weight construction module is used to construct early screening sensitivity weights based on the area under the receiver operating characteristic curve and the dispersion trend term, and to determine the dynamic weights of each biomarker. The static and dynamic feature construction module is used to weight the risk value of a single indicator using dynamic weights, and to construct a static fusion risk value and an early screening evidence accumulation item based on a sliding time window. The interaction feature and gating module is used to construct interaction features that characterize the synergistic enhancement relationship of multiple indicators and set independent joint over-threshold gating rules; The risk reasoning and scoring module is used to form a multidimensional dynamic feature vector, input it into the ensemble learning model, reason to obtain the risk probability of sepsis coagulopathy, obtain a basic risk score based on the risk probability mapping, combine the gating rules and the monotonically increasing trend of the score within the continuous time window to adjust the risk upward, output the final risk score and execute the graded warning. The closed-loop control module is used to trigger retesting and closed-loop updates based on the warning level.

Images