Acute kidney injury early-stage intelligent prediction method fusing multi-modal data

Abstract

The invention belongs to the technical field of medical artificial intelligence, and relates to an acute kidney injury early intelligent prediction method fusing multi-modal data, which comprises the following steps: generating a two-dimensional clinical feature flow containing slow variable trend features and instantaneous impact event vectors; generating a parallel risk path trajectory set containing the independent risk evolution trajectory of each path; extracting the current state value of each independent risk evolution trajectory, and generating a pathway state snapshot representing the current risk level of each pathogenic pathway; performing dynamic correction on each independent risk evolution trajectory in the parallel risk path trajectory set, and generating a collaborative enhancement risk trajectory set fusing cross-path influence; identifying and locking the trajectory with the most significant deterioration trend, and determining the trajectory as the current dominant risk pathway; and matching and generating a structured early warning instruction containing a risk type qualitative conclusion and a targeted intervention suggestion. The problem that risk fluctuation caused by treatment intervention or disease change cannot be reflected in real time is solved.

Description

Technical Field

[0001] This invention belongs to the technical field of medical artificial intelligence and relates to an early intelligent prediction method for acute kidney injury that integrates multimodal data. Background Technology

[0002] Acute kidney injury (AKI) is a common complication in critically ill patients, characterized by its insidious onset and rapid progression. Early identification and intervention are crucial for improving patient prognosis. The core challenge in current clinical practice lies in the fact that AKI is a complex result of the dynamic interaction of multiple pathogenic factors, rather than simply a manifestation of a single abnormal indicator. Massive amounts of heterogeneous clinical data, including continuous vital signs, intermittent laboratory results, unstructured text records, and discrete treatment events, collectively constitute a complete informational picture for assessing patient risk. However, effectively integrating this multimodal data and gaining insights into the underlying mechanisms and future trends of disease progression remains a technological bottleneck for achieving accurate early warning.

[0003] The commonly used solutions in the industry rely on threshold monitoring of key physiological indicators and traditional clinical scoring systems. Physicians and nurses assess renal function by observing changes in lagging biomarkers such as serum creatinine and urine output, and many electronic medical record systems also have built-in simple rule-based alarm functions based on such single or a few indicators. Static scoring models are used to stratify the risk of admitted patients. These models typically use baseline data at admission for a one-time assessment, providing initial risk references for clinicians.

[0004] Based on the above problems, traditional methods have obvious drawbacks. Static risk scoring models cannot adapt to the dynamic evolution of a patient's condition during hospitalization, and cannot reflect the risk fluctuations caused by treatment interventions or changes in the condition in real time. Their clinical guidance value diminishes rapidly over time. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides an intelligent early prediction method for acute kidney injury that integrates multimodal data.

[0006] An intelligent prediction method for early acute kidney injury that integrates multimodal data includes the following steps: S1. Acquire patients' structured physiological data, unstructured text records, and clinical intervention event data to generate a two-dimensional clinical feature stream containing slow variable trend features and instantaneous impact event vectors; S2. Receive the two-dimensional clinical feature stream, and independently simulate it through multiple preset parallel risk pathway simulators to generate a set of parallel risk pathway trajectories containing the independent risk evolution trajectories of each pathway. S3. Analyze the set of parallel risk pathway trajectories, extract the current state value of each independent risk evolution trajectory, and generate a snapshot of the pathway state representing the current risk level of each pathogenic pathway. S4. Call the path state snapshot and the parallel risk path trajectory set to establish a synergistic enhancement relationship between different risk paths, dynamically correct each independent risk evolution trajectory in the parallel risk path trajectory set, and generate a synergistic enhancement risk trajectory set that integrates cross-path influence. S5. Based on the collaborative enhancement risk trajectory set, perform forward-looking slope analysis on each risk trajectory to identify and lock in the trajectory with the most significant deterioration trend, and determine it as the current dominant risk path. S6. Based on the identified dominant risk pathway, match and generate structured early warning instructions that include qualitative conclusions about risk types and targeted intervention suggestions.

[0007] A further aspect of this invention generates a two-dimensional clinical feature stream containing slow variable trend features and instantaneous shock event vectors, comprising the following steps: Based on structured physiological data and unstructured text records, we calculate the slow variable trend characteristics that reflect the long-term evolution trend of patient status. Identify discrete high-risk procedures from clinical intervention event data and quantify them into instantaneous impact event vectors that include event type, intensity, and occurrence time; Based on timestamps, slow variable trend features and instantaneous shock event vectors are integrated to form a two-dimensional clinical feature stream.

[0008] A further aspect of this invention involves generating a set of parallel risk pathway trajectories containing independent risk evolution trajectories for each pathway, comprising the following steps: The slow variable trend features and transient impact event vectors related to hemodynamics in the two-dimensional clinical feature flow are input into the perfusion-driven pathway simulator to simulate an independent perfusion risk evolution trajectory. The slow variable trend features and transient shock event vectors related to baseline renal function and exposure to nephrotoxic substances in the two-dimensional clinical feature stream are input into the toxicity-driven pathway simulator to simulate independent toxicity risk evolution trajectories. The evolution trajectories of perfusion risk and toxicity risk are combined to form a set of parallel risk pathway trajectories.

[0009] A further aspect of this invention involves simulating independent perfusion risk evolution trajectories and independent toxicity risk evolution trajectories, comprising the following steps:

[0010] in, This represents the risk value of a specific risk pathway p at the current time t; It is the risk value of a specific risk pathway p at the previous time point; It is the normalized value of the slow variable trend feature related to pathway p extracted from the two-dimensional clinical feature stream at time point t; It is the quantized intensity value of the transient impact event vector related to pathway p at time point t, extracted from the two-dimensional clinical feature stream; These represent the risk forgetting factor, the weight of slow variable influence, and the weight of instantaneous shock influence, respectively.

[0011] A further aspect of this invention involves generating a pathway state snapshot characterizing the current risk level of each pathogenic pathway, comprising the following steps: Read the perfusion risk evolution trajectory in the parallel risk path trajectory set, obtain its risk activation intensity at the current time point, and use it as the perfusion risk state value; Read the toxicity risk evolution trajectory from the parallel risk pathway trajectory set and obtain its risk activation intensity at the current time point as the toxicity risk state value. Combine perfusion risk status values ​​and toxicity risk status values ​​to generate a pathway status snapshot.

[0012] A further aspect of this invention involves generating a set of synergistically enhanced risk trajectories that integrate cross-pathway influences, comprising the following steps: Based on the perfusion risk status value in the pathway status snapshot, a susceptibility gain coefficient is generated; The susceptibility gain coefficient is applied to nonlinearly amplify and pre-correct the toxicity risk evolution trajectory in the parallel risk pathway trajectory set, resulting in the corrected toxicity risk trajectory. By integrating the revised toxicity risk trajectory with the perfusion risk evolution trajectory, a set of synergistically enhanced risk trajectories that incorporate cross-pathway effects is generated.

[0013] A further aspect of the present invention, obtaining a modified toxicity risk trajectory, includes the following steps: The risk value at future time points in the toxicity risk evolution trajectory is multiplied by the susceptibility gain coefficient to achieve a nonlinear amplification of the risk peak. Dividing the time difference between a future point in time and the current point in the toxicity risk evolution trajectory by the susceptibility gain coefficient allows for the advance correction of the risk peak occurrence time.

[0014] A further aspect of this invention, identifying the current dominant risk pathway, includes the following steps: For each risk trajectory in the collaborative enhancement risk trajectory set, calculate its deterioration slope within a preset future time window; By comparing all the calculated deterioration slopes, the pathogenesis corresponding to the risk trajectory with the largest deterioration slope value that exceeds the preset trigger threshold is determined as the dominant risk pathway.

[0015] A further aspect of the present invention involves calculating the deterioration slope within a future preset time window, including the following steps: Obtain the predicted risk value of the risk trajectory at the end of a future preset time window; Calculate the difference between the predicted risk value and the current risk value, and divide the difference by the length of the future preset time window to obtain the deterioration slope.

[0016] A further aspect of this invention involves matching and generating a structured early warning instruction that includes a qualitative conclusion on the risk type and targeted intervention recommendations, comprising the following steps: Extract the type identifier of the dominant risk pathway and use it as a qualitative conclusion of the risk type; Using type identifiers as search keywords, the system matches and retrieves corresponding clinical intervention recommendations from a pre-defined knowledge base. Combine qualitative conclusions on risk types with clinical intervention recommendations to generate structured early warning instructions.

[0017] In summary, the present invention has the following beneficial technical effects: 1. By employing a dual analysis of multimodal data, this approach integrates the patient's long-term, slow-moving trends with the impact of sudden high-risk events, constructing a more comprehensive and dynamic stream of clinical characteristics. This method transcends the reliance on isolated data points in traditional approaches, simultaneously capturing the gradual accumulation of risk, such as the continuous deterioration of vital signs, and the sudden risk impact, such as the use of nephrotoxic drugs, thereby providing a deeper understanding of the complete evolution of the patient's condition.

[0018] 2. An innovative parallel risk pathway simulation mechanism is introduced, decomposing the complex pathogenesis of acute kidney injury into multiple parallel risk evolution pathways consistent with clinical medical mechanisms. This design not only enhances the interpretability of the predictive model, enabling clinicians to understand the sources of risk, but also allows for the independent simulation of the dynamic effects of different pathogenic factors. By simulating the risk trajectories of different pathways such as perfusion and toxicity, the independent contribution of different risk factors to patient status is clearly revealed, achieving a shift from "black box" prediction to "white box" analysis.

[0019] 3. A dynamic correction and dominant pathway identification mechanism is established. By analyzing the synergistic enhancement effects among different risk pathways, independent risk trajectories are nonlinearly corrected. This process simulates the real-world scenario of multiple shocks leading to exacerbated damage in clinical practice, significantly improving the accuracy of predictions in complex cases. Based on this, prospective slope analysis is used to identify the dominant risk pathway with the most significant deterioration trend. This allows the system not only to predict the occurrence of risks but also to pinpoint the most significant current risk drivers, providing a clear focus for clinical decision-making. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. The drawings are used to provide a further understanding of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A flowchart illustrating an embodiment of this application is disclosed.

[0022] Figure 2 Structural schematic diagrams of embodiments of this application are disclosed. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] The following is in conjunction with the appendix Figures 1-2 A preferred description of the present invention is provided below.

[0025] See attached document Figure 1 This invention proposes an intelligent early prediction method for acute kidney injury that integrates multimodal data, comprising the following steps: S1. Acquire patients' structured physiological data, unstructured text records, and clinical intervention event data to generate a two-dimensional clinical feature stream containing slow variable trend features and instantaneous impact event vectors; S2. Receive the two-dimensional clinical feature stream, and independently simulate it through multiple preset parallel risk pathway simulators to generate a set of parallel risk pathway trajectories containing the independent risk evolution trajectories of each pathway. S3. Analyze the set of parallel risk pathway trajectories, extract the current state value of each independent risk evolution trajectory, and generate a snapshot of the pathway state representing the current risk level of each pathogenic pathway. S4. Call the path state snapshot and the parallel risk path trajectory set to establish a synergistic enhancement relationship between different risk paths, dynamically correct each independent risk evolution trajectory in the parallel risk path trajectory set, and generate a synergistic enhancement risk trajectory set that integrates cross-path influence. S5. Based on the collaborative enhancement risk trajectory set, perform forward-looking slope analysis on each risk trajectory to identify and lock in the trajectory with the most significant deterioration trend, and determine it as the current dominant risk path. S6. Based on the identified dominant risk pathway, match and generate structured early warning instructions that include qualitative conclusions about risk types and targeted intervention suggestions.

[0026] In one embodiment of the present invention, step S1 includes the following steps: Structured physiological data and unstructured text records of patients are extracted, and slow variable trend features reflecting patient status are calculated. The slow variable trend features include the time series slope of vital signs and the frequency growth rate of status words in the text description. High-risk operations that occur discretely in clinical intervention event data are identified and quantified into instantaneous impact event vectors containing event type, intensity and occurrence time. The slow variable trend features and instantaneous impact event vectors are integrated and aligned according to timestamps to form a two-dimensional clinical feature stream.

[0027] Specifically, the data interface program retrieves all of a patient's medical records within a certain period after admission from the hospital's electronic medical record system. These records are categorized into three data sources: structured physiological data, unstructured text records, and clinical intervention event data. Structured physiological data refers to numerical data collected periodically from monitors or laboratory systems. The data structure consists of key-value pairs with timestamps, such as heart rate, blood pressure, blood oxygen saturation, urine output, and serum creatinine levels. Unstructured text records refer to descriptive text information handwritten or entered by medical staff. The data structure consists of free text paragraphs with timestamps, such as nursing records, ward round records, and progress summaries. Clinical intervention event data refers to data recording specific medical procedures. The data structure consists of discrete entries containing the event name, execution time, and related parameters, such as contrast agent injection, initiation of ventilator support, or administration of high-dose diuretics.

[0028] After acquiring three types of data sources, a dual-analysis process is performed: parallel processing of gradual and discrete event information. On one hand, structured physiological data and unstructured text records are analyzed to extract slow-variable trend features reflecting the long-term evolution of the patient's physiological state. These slow-variable trend features are a quantitative description of long-term changes in the patient's state, including the time-series slope of vital signs and the growth rate of the frequency of state-related terms. The time-series slope of vital signs is calculated through linear regression analysis of blood pressure and heart rate data points within the observation time window, satisfying the following formula:

[0029] in, The slope of a time series representing vital signs; Indicates a point in time t Observed vital sign values; Indicates the start time of the observation time window; Indicates the end time of the observation time window.

[0030] The rate of change in the frequency of state words is calculated by identifying and statistically analyzing unstructured text records from past consecutive time periods containing words related to kidney injury, such as "edema," "oliguria," and "decreased filtration rate," and the result satisfies the following formula:

[0031] in, The growth rate of the frequency of state words; Indicates the time from the start of hospitalization to the point in time. t Up to that point, specific keywords The cumulative number of times it appears; This indicates the starting point for calculating the growth rate of the frequency of words in the state. This indicates the point in time at which the growth rate of the frequency of words in a given state is calculated to end.

[0032] On the other hand, high-risk procedures that may have an immediate impact on kidney function are identified in clinical intervention event data and transformed into transient impact event vectors in a standardized format. The transient impact event vector is a standardized representation of high-risk medical procedures, containing a data structure with three elements: event type, intensity, and occurrence time. The event type is an integer obtained by classifying and encoding according to a pre-defined list of kidney injury risk procedures, such as "use of nephrotoxic drugs" or "blood loss during major surgery." The intensity is a dimensionless value between 0 and 1 assigned based on the inherent risk level of the procedure or the drug dosage, based on a retrospective analysis of relevant events in 500 historical cases of acute kidney injury. The occurrence time is the precise timestamp at which the procedure was recorded.

[0033] The two types of processed information—slow variable trend features and transient impact event vectors—are merged and sorted according to their respective timestamps to form a two-dimensional clinical feature stream with time as the axis and a unified format. The two-dimensional clinical feature stream is a data sequence arranged in chronological order, representing that each time point contains the current slow variable trend features and a vector of possible transient impact events. If no high-risk operation occurs at the current time point, the transient impact event vector is null.

[0034] For example, suppose we obtain patient A's data from 24 hours to 72 hours after admission. In the structured physiological data, the mean arterial pressure linearly decreased from 95 mmHg to 75 mmHg. In the unstructured text records, the frequency of the phrase "decreased urine output" increased from 2 times to 6 times. In the clinical intervention event data, a "contrast-enhanced CT scan" was recorded at 60 hours. Calculating the slow variable trend characteristics, the time series slope of vital signs is (75-95) / 48 = -0.42, and the frequency growth rate of state words is (6-2) / 2 = 2.

[0035] "Contrast-enhanced CT scan" was identified as a high-risk procedure, and a transient impact event vector was generated. The event type was coded as 3 (representing "exposure to nephrotoxic substances"), the intensity was set to 0.8 according to a predetermined rule, and the occurrence time was 60 hours. This information was integrated to form a two-dimensional clinical feature stream. This data stream, at the 60-hour time point, not only recorded the slow variable trend features with values ​​of -0.42 and 2, but also included the transient impact event vector with values ​​{3, 0.8, 60h}.

[0036] In one embodiment of the present invention, step S2 includes the following steps: The slow variable trend features and instantaneous impact event vectors related to hemodynamics in the two-dimensional clinical feature stream are input into the "perfusion-driven" pathway simulator to simulate independent perfusion risk evolution trajectories; the slow variable trend features and instantaneous impact event vectors related to renal function baseline and nephrotoxic substance exposure in the two-dimensional clinical feature stream are input into the "toxicity-driven" pathway simulator to simulate independent toxicity risk evolution trajectories; the independent risk evolution trajectories generated by all pathway simulators are collected to form a parallel risk pathway trajectory set.

[0037] Specifically, multiple parallel risk pathway simulators targeting different pathogenic mechanisms are pre-defined. These simulators are independent computational models, each designed to simulate a specific pathogenesis of acute kidney injury. The risk pathway simulator is a computational program built upon a specific medical mechanism. Its function is to receive relevant clinical feature data and simulate the evolution of the corresponding risk over time according to built-in algorithmic logic. Upon receiving a two-dimensional clinical feature stream, the information in the stream is distributed to the corresponding risk pathway simulator. For example, slow-moving variable trend features and instantaneous impact event vectors related to hemodynamic changes are fed into the "perfusion-driven" pathway simulator. This simulator generates a perfusion risk evolution trajectory over time based on these inputs. The perfusion risk evolution trajectory is a time-series data output by the "perfusion-driven" pathway simulator, recording the changes in the risk of kidney injury caused by hemodynamic instability over time.

[0038] Features and vectors related to basic kidney function and exposure to nephrotoxic substances are fed into a "toxicity-driven" pathway simulator to simulate independent toxicity risk evolution trajectories. These toxicity risk evolution trajectories are time-series data output by the "toxicity-driven" pathway simulator, recording the changes in the risk of kidney damage caused by exposure to nephrotoxic substances over time.

[0039] The formula for calculating the risk evolution trajectory satisfies the following formula: , This represents the risk value of a specific risk pathway p at the current time t; It is the risk value of a specific risk pathway p at the previous time point, reflecting the cumulative effect of risk; It is the normalized value of the slow variable trend feature related to pathway p extracted from the two-dimensional clinical feature stream at time point t; It is the quantized intensity value of the transient impact event vector related to pathway p at time point t, extracted from the two-dimensional clinical feature stream.

[0040] The weights representing risk amnesia factors, slow variable influence weights, and instantaneous impact influence weights are derived from model fitting and parameter optimization based on historical data from 1000 patients diagnosed with acute kidney injury. For example, for the "perfusion-driven" pathway, the β value is relatively high to emphasize the effect of sustained blood pressure decline; for the "toxicity-driven" pathway, the γ value is set higher to highlight the immediate impact of nephrotoxic drug events.

[0041] The process of distributing information from the two-dimensional clinical feature stream to the corresponding risk pathway simulators is parallel, meaning that multiple simulators can process their respective related data simultaneously. After all pathway simulators have completed their calculations, the independent risk evolution trajectories they generate are aggregated to form a parallel risk pathway trajectory set. The parallel risk pathway trajectory set contains multiple independent risk evolution trajectories, each consisting of a series of time-ordered risk values, accompanied by an identifier indicating the type of risk pathway it belongs to, such as "perfusion" or "toxicity".

[0042] For example, a two-dimensional clinical feature stream containing the event of patient A undergoing a contrast-enhanced CT scan at time 60 hours is received. The slow-varying trend feature (-0.42 mmHg / h) reflecting a sustained decrease in mean arterial pressure is input to the perfusion-driven pathway simulator. This simulator calculates and updates the risk value from time 24 to time 72 based on the corresponding perfusion risk evolution trajectory data, showing a slowly rising curve. Simultaneously, the instantaneous impact event vector {3, 0.8, 60h} at time 60 hours is input to the toxicity-driven pathway simulator. This simulator calculates and generates a sharply rising risk peak at time 60 hours based on the corresponding toxicity risk evolution trajectory data. The slowly rising perfusion risk evolution trajectory and the toxicity risk evolution trajectory with a sudden change at time 60 hours are packaged together to form a parallel risk pathway trajectory set.

[0043] In one embodiment of the present invention, step S3 includes the following steps: Read the evolution trajectory of the "infusion-driven" pathway in the parallel risk pathway trajectory set, obtain its risk activation intensity at the current time point, and use it as the infusion risk state value; read the evolution trajectory of the "toxicity-driven" pathway in the parallel risk pathway trajectory set, obtain its risk activation intensity at the current time point, and use it as the toxicity risk state value; combine the infusion risk state value and the toxicity risk state value to generate a pathway state snapshot.

[0044] Specifically, the evolutionary trajectory data of the "perfusion-driven" pathway in the parallel risk pathway trajectory set is read. The value corresponding to the last time point in this time series data is read. This value represents the current risk activation intensity and is marked as the perfusion risk state value. Risk activation intensity is a synonym for the current state value, used to describe the degree to which the risk pathway is triggered. The perfusion risk state value is the current state value extracted from the perfusion risk evolution trajectory, directly reflecting the risk level caused by insufficient blood perfusion. The current state value refers to the complete risk evolution trajectory time series, corresponding to the specific dimensionless risk value at the current analysis time point. The evolutionary trajectory of the "toxicity-driven" pathway is processed in the same way, obtaining its risk activation intensity at the current time point and marking it as the toxicity risk state value. The toxicity risk state value is the current state value extracted from the toxicity risk evolution trajectory, quantifying the immediate risk caused by nephrotoxic factors.

[0045] After the current status values ​​of all risk pathways are extracted, these independent values ​​are combined into a pathway status snapshot, such as perfusion risk status value and toxicity risk status value. The pathway status snapshot is a collection of multiple key-value pairs that centrally present the risk level of all monitored pathogenic pathways at the same time, where the key is the pathway name and the value is the corresponding risk status value.

[0046] For example, assuming the current time point is the 72nd hour after the patient's admission, a set of parallel risk pathway trajectories containing the patient A's risk evolution data is received. The evolution trajectory of the "perfusion-driven" pathway is parsed, and the risk value at the 72nd hour is read. Since the blood pressure continues to decrease slowly, this value is 0.45, which is extracted as the perfusion risk status value.

[0047] The evolutionary trajectory of the "toxicity-driven" pathway was analyzed. Although a peak was observed at hour 60, the risk value dropped back to 0.30 by hour 72 due to drug metabolism, and this value was extracted as the toxicity risk state value. These two values ​​were combined to generate a pathway state snapshot, containing {perfusion risk state value: 0.45, toxicity risk state value: 0.30}.

[0048] In one embodiment of the present invention, step S4 includes the following steps: Based on the perfusion risk status value in the pathway status snapshot, a susceptibility gain coefficient is generated. The susceptibility gain coefficient is applied to the toxicity risk evolution trajectory in the parallel risk pathway trajectory set to nonlinearly amplify and pre-correct its future risk peak and occurrence time, resulting in a corrected toxicity risk trajectory. The corrected toxicity risk trajectory is integrated with other uncorrected risk trajectories to form a synergistically enhanced risk trajectory set.

[0049] Specifically, based on the perfusion risk status value in the pathway status snapshot, a susceptibility gain coefficient is calculated using a preset function. The susceptibility gain coefficient is used to quantify the increased susceptibility of the kidneys caused by current insufficient blood perfusion, and satisfies the following formula: , It is the susceptibility gain coefficient; This represents the synergistic enhancement adjustment coefficient, used to adjust the intensity of the enhancement effect. It is set based on statistical regression analysis of 500 clinical case data with simultaneous underperfusion and nephrotoxicity exposure, for example, set to 0.5. This represents the perfusion risk state value in the pathway state snapshot. The susceptibility gain coefficient function quantifies the degree to which one risk state increases the sensitivity of another risk pathway. The calculated susceptibility gain coefficient is applied to the future part of the toxicity risk evolution trajectory in the parallel risk pathway trajectory set, including two aspects: first, nonlinearly amplifying the possible future risk peak to make its value higher; second, correcting the expected occurrence time of this peak in advance to make it occur earlier.

[0050] The revised formula for calculating future risk value is as follows: The time correction relationship is as follows: , It is the revised future time point; This indicates the revised toxicity risk trajectory at future time points. The risk value; It is the risk value at a future time point t, representing the original trajectory of toxicity risk evolution. It is the current time point; It is any point in the future; ensuring that when When the value is greater than 1, the risk value is amplified, and the timeline of future events is compressed, thereby amplifying and bringing forward the risk.

[0051] After this correction, a revised toxicity risk trajectory reflecting the composite risk is obtained. This revised trajectory is a new time series data generated by nonlinearly amplifying and pre-correcting the original toxicity risk evolution trajectory using a susceptibility gain coefficient. The revised toxicity risk trajectory is then recombined with other uncorrected risk trajectories in the parallel risk pathway trajectory set, such as the perfusion risk evolution trajectory, to form a synergistic enhancement risk trajectory set that integrates cross-pathway effects. The content of the synergistic enhancement risk trajectory set is similar to that of the parallel risk pathway trajectory set, containing at least the risk evolution trajectory corrected for synergistic enhancement. Synergistic enhancement refers to the phenomenon in medicine where, when two or more pathogenic factors coexist, the overall harm is greater than the sum of the individual harms of each factor.

[0052] For example, a pathway state snapshot with the content {perfusion risk state value: 0.45, toxicity risk state value: 0.30} is retrieved, and a susceptibility gain coefficient is calculated based on the perfusion risk state value of 0.45, assuming a synergistic enhancement modulation coefficient. If it is 0.5, then the susceptibility gain coefficient =1 + 0.5 + 0.45 = 1.225. Assuming the original toxicity risk evolution trajectory predicts a risk peak of 0.6 after 12 hours (i.e., at hour 84), a susceptibility gain coefficient of 1.225 is applied for dynamic correction. The corrected risk peak is non-linearly amplified to 0.6 + 1.225 = 0.735.

[0053] The peak occurrence time was revised to approximately 9.8 hours after the current time (12 / 1.225 ≈ 9.8 hours), generating a revised toxicity risk trajectory. This new trajectory is integrated with the original perfusion risk evolution trajectory to form a synergistically enhanced risk trajectory set.

[0054] In one embodiment of the present invention, step S5 includes the following steps: Calculate the deterioration slope of each trajectory in the synergistic enhancement risk trajectory set within a preset time window; compare all calculated deterioration slopes, and determine the pathogenic mechanism corresponding to the risk trajectory with the largest slope value that exceeds the preset trigger threshold as the dominant risk pathway.

[0055] Specifically, a forward-looking analysis is performed on each risk evolution trajectory included in the collaboratively enhanced risk trajectory set. The purpose of forward-looking slope analysis is to predict the severity of the development trend by calculating the rate of change of time series data over a future period. For each trajectory in the collaboratively enhanced risk trajectory set, the deterioration slope within a preset future time window satisfies the formula:

[0056] in, It is the deterioration slope of the risk pathway p, used to describe the average growth rate of the risk trajectory within a preset time window in the future, with the unit being risk value / hour; It is a collaborative enhancement of the risk trajectory concentration path p at the current time The risk value; This indicates that the risk pathway p will be in the future within a predetermined time window. The predicted risk value at the end; It is the length of the future preset time window, which is set based on the key observation period for the evolution of acute kidney injury in clinical practice, for example, 24 hours.

[0057] The calculation process is performed on all trajectories to obtain the deterioration slope value corresponding to different pathogenic mechanisms. All calculated deterioration slopes are compared to find the maximum value. The maximum slope value is compared with a preset trigger threshold; a final determination is made only when the maximum slope value exceeds the set trigger threshold. The determination criteria for the dominant risk pathway are as follows: and , Indicates the deterioration slope of the identified dominant risk pathway; This represents the preset trigger threshold, an empirical value derived from statistical analysis of historical case data. It is used to distinguish between normal risk fluctuations and rapidly deteriorating trends that require vigilance. For example, it is set to 0.05 risk value / hour. The setting is based on ensuring a balance between the sensitivity and specificity of the warning, and is derived from the analysis of receiver operating characteristic curves of backtesting data from 2000 historical cases.

[0058] The result of the determination is that the pathogenic mechanism corresponding to the risk trajectory with the largest and over-trigger threshold slope is identified as the dominant risk factor, i.e., the dominant risk pathway, which is the pathogenic mechanism that contributes the most to the deterioration of the patient's condition and has the most obvious trend at a specific point in time.

[0059] For example, receiving a set of co-enhanced risk trajectories containing the modified toxicity risk trajectory and the original perfusion risk evolution trajectory, assuming the current time point... At hour 72, with a preset time window Δt of 12 hours, the evolution trajectory of perfusion risk was analyzed. The current risk value was 0.45, and the predicted risk value after 12 hours (hour 84) was 0.51. The deterioration slope was calculated as (0.51-0.45) / 12=0.005. Analyzing the corrected toxicity risk trajectory, the current risk value was 0.30, and the predicted risk value after 12 hours (hour 84) was 0.70 (because the peak occurred earlier and was amplified). The deterioration slope was calculated as (0.70-0.30) / 12≈0.033.

[0060] Comparing these two deterioration slopes, we can see that 0.033 is greater than 0.005. Assuming the preset trigger threshold is 0.02, since 0.033 is not only the maximum value but also exceeds the trigger threshold of 0.02, the current dominant risk pathway is ultimately determined to be the "toxicity-driven" pathway.

[0061] In one embodiment of the present invention, step S6 includes the following steps: Based on the type identifier of the dominant risk pathway, the system retrieves and matches corresponding clinical intervention recommendations from a pre-set knowledge base. For example, for "perfusion-driven", the system matches "recommendation to focus on fluid management and circulatory stability". The system combines qualitative conclusions and intervention recommendations to generate structured early warning instructions and outputs them to terminal devices.

[0062] Specifically, upon receiving the determination result of the dominant risk pathway, such as "toxicity-driven," the type identifier of this result, namely "toxicity-driven type," is directly extracted, and "toxicity-driven type" is used as the qualitative conclusion of the risk type. The qualitative conclusion of the risk type indicates the most important source of pathogenic risk at present. Based on the type identifier of the dominant risk pathway, including "perfusion-driven type" or "toxicity-driven type," the type identifier is a standardized label extracted from the name of the dominant risk pathway and used as a key value for pre-set knowledge base retrieval.

[0063] Referring to Appendix 1, the "toxicity-driven" identifier is used as the keyword. The preset knowledge base is a database that stores multiple risk types and corresponding intervention measures. The knowledge base is a pre-configured database or lookup table that stores the mapping relationship between "type identifier" and "intervention recommendation" and is established based on clinical guidelines and expert experience.

[0064] Appendix 1:

[0065] The system retrieves and matches corresponding clinical intervention recommendations from a pre-defined knowledge base. It identifies entries in the knowledge base that perfectly match "toxicity-driven" and reads the pre-defined targeted intervention recommendations under those entries. The qualitative risk type conclusion obtained in the first step and the targeted intervention recommendations retrieved in the second step provide specific and actionable reference opinions for clinical decision-making. These are combined into a data package with a standard format, which is the structured warning instruction. A structured warning instruction contains standardized digital information with at least two fixed fields: one for storing the qualitative risk type conclusion and the other for storing the targeted intervention recommendation, facilitating parsing and clear display by terminal devices. After generating the structured warning instruction, it is sent through an internal communication interface and ultimately displayed on the terminal devices used by medical staff to complete the warning process. Terminal devices refer to hardware devices capable of receiving and presenting information to users, such as computer monitors at medical workstations, central monitoring screens at nurse stations, or mobile smart devices held by doctors.

[0066] For example, if the received dominant risk pathway assessment result is "toxicity-driven," its type identifier "toxicity-driven" is first extracted and used as the qualitative conclusion of the risk type. A search is then performed in a pre-defined knowledge base using "toxicity-driven" as the keyword, matching a pre-defined targeted intervention suggestion: "It is recommended to review the records of nephrotoxic drug use and consider adjusting the dosage or changing the medication." These two parts are combined to generate a structured warning instruction, containing the following content: {Qualitative conclusion of risk type: "Increased risk of toxicity-driven acute kidney injury," and targeted intervention suggestion: "It is recommended to review the records of nephrotoxic drug use and consider adjusting the dosage or changing the medication"}. This structured warning instruction is then pushed to the central monitoring system screen in the intensive care unit.

[0067] See appendix Figure 2 This invention also proposes an early intelligent prediction system for acute kidney injury that integrates multimodal data, comprising the following modules: The two-dimensional clinical feature stream generation module is used to acquire patients' structured physiological data, unstructured text records, and clinical intervention event data, and generate a two-dimensional clinical feature stream containing slow variable trend features and instantaneous impact event vectors. The parallel risk pathway trajectory set generation module is used to receive a two-dimensional clinical feature stream, perform independent simulations through multiple preset parallel risk pathway simulators, and generate a parallel risk pathway trajectory set containing the independent risk evolution trajectory of each pathway. The pathway state snapshot generation module is used to parse the set of parallel risk pathway trajectories, extract the current state value of each independent risk evolution trajectory, and generate a pathway state snapshot that represents the current risk level of each pathogenic pathway. The collaborative enhancement risk trajectory set generation module is used to call the path state snapshot and the parallel risk path trajectory set to establish a collaborative enhancement relationship between different risk paths, dynamically correct each independent risk evolution trajectory in the parallel risk path trajectory set, and generate a collaborative enhancement risk trajectory set that integrates cross-path effects. The dominant risk pathway determination module, based on the collaborative enhancement risk trajectory set, performs forward slope analysis on each risk trajectory to identify and lock in the trajectory with the most significant deterioration trend, and determines it as the current dominant risk pathway. The structured early warning instruction generation module matches and generates structured early warning instructions that include qualitative conclusions about risk types and targeted intervention suggestions, based on the identified dominant risk pathways.

[0068] It should be noted that the formulas described above, through the principle of dimensional consistency and mathematical standardization methods (such as normalization, dimensionless parameter conversion, or unit system unification), can translate physical quantities with different properties into unitless standard values ​​or superimposed parameters of the same dimension. This eliminates the interference of different dimensions on the computational logic, allowing the formulas to retain the original data distribution characteristics while possessing mathematical rationality and adaptability to objective laws. The descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the invention.

[0069] Each of the modules can be implemented in whole or in part through software, hardware, or a combination thereof. It supports hardware embedded in or independent of the processor in the computer device, and also supports software stored in the memory of the computer device, so that the processor can call and execute the operations corresponding to each of the above modules.

[0070] It should be noted that the human information (including but not limited to human device information and personal information) and data (including but not limited to data used for analysis, data stored and data displayed) involved in this invention are all information and data authorized by the human body or fully authorized by all parties. The collection, use and processing of related data require relevant legal standards.

[0071] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for early intelligent prediction of acute kidney injury by fusing multimodal data, characterized in that, Includes the following steps: S1. Acquire patients' structured physiological data, unstructured text records, and clinical intervention event data to generate a two-dimensional clinical feature stream containing slow variable trend features and instantaneous impact event vectors; S2. Receive the two-dimensional clinical feature stream, and independently simulate it through multiple preset parallel risk pathway simulators to generate a set of parallel risk pathway trajectories containing the independent risk evolution trajectories of each pathway. S3. Analyze the set of parallel risk pathway trajectories, extract the current state value of each independent risk evolution trajectory, and generate a snapshot of the pathway state representing the current risk level of each pathogenic pathway. S4. Call the path state snapshot and the parallel risk path trajectory set to establish a synergistic enhancement relationship between different risk paths, dynamically correct each independent risk evolution trajectory in the parallel risk path trajectory set, and generate a synergistic enhancement risk trajectory set that integrates cross-path influence. S5. Based on the collaborative enhancement risk trajectory set, perform forward-looking slope analysis on each risk trajectory to identify and lock in the trajectory with the most significant deterioration trend, and determine it as the current dominant risk path. S6. Based on the identified dominant risk pathway, match and generate structured early warning instructions that include qualitative conclusions about risk types and targeted intervention suggestions.

2. The method for early intelligent prediction of acute kidney injury by fusing multimodal data according to claim 1, characterized in that, Generating a two-dimensional clinical feature stream that includes slow variable trend features and transient shock event vectors includes the following steps: Based on structured physiological data and unstructured text records, we calculate the slow variable trend characteristics that reflect the long-term evolution trend of patient status. Identify discrete high-risk procedures from clinical intervention event data and quantify them into instantaneous impact event vectors that include event type, intensity, and occurrence time; Based on timestamps, slow variable trend features and instantaneous shock event vectors are integrated to form a two-dimensional clinical feature stream.

3. The method for early intelligent prediction of acute kidney injury by fusing multimodal data according to claim 1, characterized in that, Generating a set of parallel risk pathway trajectories containing independent risk evolution trajectories for each pathway includes the following steps: The slow variable trend features and transient impact event vectors related to hemodynamics in the two-dimensional clinical feature flow are input into the perfusion-driven pathway simulator to simulate an independent perfusion risk evolution trajectory. The slow variable trend features and transient shock event vectors related to baseline renal function and exposure to nephrotoxic substances in the two-dimensional clinical feature stream are input into the toxicity-driven pathway simulator to simulate independent toxicity risk evolution trajectories. The evolution trajectories of perfusion risk and toxicity risk are combined to form a set of parallel risk pathway trajectories.

4. The method for early intelligent prediction of acute kidney injury by fusing multimodal data according to claim 3, characterized in that, Simulating independent perfusion risk evolution trajectories and independent toxicity risk evolution trajectories includes the following steps: in, This represents the risk value of a specific risk pathway p at the current time t; It is the risk value of a specific risk pathway p at the previous time point; It is the normalized value of the slow variable trend feature related to pathway p extracted from the two-dimensional clinical feature stream at time point t; It is the quantized intensity value of the transient impact event vector related to pathway p at time point t, extracted from the two-dimensional clinical feature stream; These represent the risk forgetting factor, the weight of slow variable influence, and the weight of instantaneous shock influence, respectively.

5. The method for early intelligent prediction of acute kidney injury by fusing multimodal data according to claim 1, characterized in that, Generate a pathway status snapshot representing the current risk level of each pathogenic pathway, including the following steps: Read the perfusion risk evolution trajectory in the parallel risk path trajectory set, obtain its risk activation intensity at the current time point, and use it as the perfusion risk state value; Read the toxicity risk evolution trajectory from the parallel risk pathway trajectory set and obtain its risk activation intensity at the current time point as the toxicity risk state value. Combine perfusion risk status values ​​and toxicity risk status values ​​to generate a pathway status snapshot.

6. The method for early intelligent prediction of acute kidney injury by fusing multimodal data according to claim 1, characterized in that, Generating a set of synergistically enhanced risk trajectories that integrate cross-pathway effects includes the following steps: Based on the perfusion risk status value in the pathway status snapshot, a susceptibility gain coefficient is generated; The susceptibility gain coefficient is applied to nonlinearly amplify and pre-correct the toxicity risk evolution trajectory in the parallel risk pathway trajectory set, resulting in the corrected toxicity risk trajectory. By integrating the revised toxicity risk trajectory with the perfusion risk evolution trajectory, a set of synergistically enhanced risk trajectories that incorporate cross-pathway effects is generated.

7. The method for early intelligent prediction of acute kidney injury by fusing multimodal data according to claim 6, characterized in that, The revised toxicity risk trajectory is obtained by the following steps: The risk value at future time points in the toxicity risk evolution trajectory is multiplied by the susceptibility gain coefficient to achieve a nonlinear amplification of the risk peak. Dividing the time difference between a future point in time and the current point in the toxicity risk evolution trajectory by the susceptibility gain coefficient allows for the advance correction of the risk peak occurrence time.

8. The method for early intelligent prediction of acute kidney injury by fusing multimodal data according to claim 1, characterized in that, The identification of the current dominant risk pathway includes the following steps: For each risk trajectory in the collaborative enhancement risk trajectory set, calculate its deterioration slope within a preset future time window; By comparing all the calculated deterioration slopes, the pathogenesis corresponding to the risk trajectory with the largest deterioration slope value that exceeds the preset trigger threshold is determined as the dominant risk pathway.

9. The method for early intelligent prediction of acute kidney injury by fusing multimodal data according to claim 8, characterized in that, Calculating its deterioration slope over a future preset time window includes the following steps: Obtain the predicted risk value of the risk trajectory at the end of a future preset time window; Calculate the difference between the predicted risk value and the current risk value, and divide the difference by the length of the future preset time window to obtain the deterioration slope.

10. The method for early intelligent prediction of acute kidney injury by fusing multimodal data according to claim 1, characterized in that, Matching and generating structured early warning instructions that include qualitative conclusions about risk types and targeted intervention recommendations includes the following steps: Extract the type identifier of the dominant risk pathway and use it as a qualitative conclusion of the risk type; Using type identifiers as search keywords, the system matches and retrieves corresponding clinical intervention recommendations from a pre-defined knowledge base. Combine qualitative conclusions on risk types with clinical intervention recommendations to generate structured early warning instructions.

Images