Device and method for assessing crush risk

Abstract

This invention provides a crush risk assessment device and method, relating to the field of data processing technology. The device includes: an acquisition module for acquiring N sets of data to be detected for at least one object to be assessed in a target scenario; an extraction module for extracting features from the N sets of data to be detected using feature extractors corresponding to each of the N modal types, to obtain N sets of risk features; a generation module for inputting the N sets of risk features into risk assessment models corresponding to each of the N modal types, to generate N sets of risk assessment information; and a first feature fusion module for fusing features from the N sets of risk assessment information according to the correlation between the N modal types, to obtain first target risk assessment information, and to generate a processing scheme applicable to at least one object to be assessed based on the first target risk assessment information.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically to a crush risk assessment device and method. Background Technology

[0002] Crush syndrome refers to a series of clinical symptoms, primarily characterized by hyperkalemia, myoglobinuria, and acute renal failure, that occur when or after the compression of muscular areas (such as the limbs and trunk) is subjected to prolonged compression by a heavy object. This occurs due to ischemic necrosis of muscle tissue and the release of intracellular substances into the bloodstream. The occurrence of this syndrome is closely related to the intensity and duration of the compression, as well as the muscle content of the compressed area, and its severity directly affects the patient's prognosis. Therefore, accurate identification and assessment of crush syndrome, and timely intervention, are crucial for reducing morbidity and mortality.

[0003] In realizing the concept of this invention, at least the following problems exist: Existing single-data-based assessment methods can only reflect local tissue morphological changes, making it difficult to comprehensively assess the extent and degree of muscle damage as well as the state of systemic inflammatory response. This does not conform to the pathophysiological characteristics of multi-system involvement in crush syndrome, which can lead to biased assessment results and thus affect the timeliness of intervention. Summary of the Invention

[0004] In view of the above problems, the present invention provides a device and method for assessing crush risk.

[0005] This invention provides a crush risk assessment device, comprising: an acquisition module for acquiring N sets of test data for at least one object to be assessed in a target scenario, wherein the N sets of test data correspond to target biomarkers of N modalities, and N is a positive integer greater than 2; the acquisition module includes a biochemical detection data acquisition unit, a proteomics data acquisition unit, a metabolomics data acquisition unit, a blood pressure data acquisition unit, an electrocardiogram data acquisition unit, and an electrical impedance imaging data acquisition unit; and at least one of the following relationships exists among the N modalities: a causal relationship characterizing the abnormality of a biomarker in the first modality as a cause of the abnormality of a biomarker in the second modality; and a relationship characterizing the occurrence time of the abnormality of a biomarker in the third modality earlier than that in the fourth modality. The system identifies the temporal sequence of abnormal occurrences of markers and the synergistic relationship between markers representing the fifth and sixth modal types, showing a predetermined trend of synchronous change. An extraction module is used to extract features from N sets of data to be detected using feature extractors corresponding to each of the N modal types, resulting in N sets of risk features. A generation module inputs these N sets of risk features into risk assessment models corresponding to each of the N modal types, generating N sets of risk assessment information. A first feature fusion module fuses the N sets of risk assessment information based on the correlation between the N modal types to obtain first target risk assessment information, which is then used to generate a processing scheme suitable for the object to be assessed.

[0006] According to an embodiment of the present invention, the N sets of data to be tested include at least data to be tested for proteomics modalities and data to be tested for metabolomics modalities. A proteomics data acquisition unit is used to acquire the data to be tested for proteomics modalities of the object to be evaluated. The target biomarkers for the proteomics modalities include at least one of: phosphoglycerate mutase 2, lactate dehydrogenase A, malate dehydrogenase 1, fructose-1,6-bisphosphate aldolase, pyruvate kinase M, glyceraldehyde-3-phosphate dehydrogenase, phosphoglycerate kinase, phosphoglucose mutase, and carbonic anhydrase 3. A metabolomics data acquisition unit is used to acquire the data to be tested for metabolomics modalities of the object to be evaluated. The target biomarkers for the metabolomics modalities include at least one of: galactose metabolism data, phenylalanine metabolism data, fructose and mannose metabolism data, amino sugar and nucleotide sugar metabolism data, phenylalanine, tyrosine, and tryptophan biosynthesis data, and ascorbic acid and malonic acid metabolism data. A biochemical detection data acquisition unit is used to acquire data for the object to be evaluated. The target biochemical markers for the biochemical testing modality include at least the following: white blood cell count, red blood cell count, hemoglobin concentration, hematocrit, mean corpuscular volume (MCV), mean corpuscular hemoglobin (MCH) content, mean corpuscular hemoglobin concentration, red blood cell distribution width (RDW), platelet count, mean platelet volume (MCV), platelet distribution width (RDW), plateletcrit, arterial blood pH, partial pressure of carbon dioxide (SPC), partial pressure of oxygen (SPO), blood oxygen saturation, lactate, sodium ions, potassium ions, chloride ions, creatinine, urea, uric acid, prothrombin time, and activated partial thromboplastin time (APT). Thrombin time, fibrinogen, creatine kinase, and creatine kinase isoenzymes; a blood pressure data acquisition unit for collecting target data of the blood pressure modality of the subject to be evaluated, with target biomarkers including at least: systolic blood pressure, diastolic blood pressure, mean arterial pressure, pulse rate, pulse pressure, and blood pressure variability; and an electrocardiogram (ECG) data acquisition unit for collecting target data of the ECG modality of the subject to be evaluated, with target biomarkers including at least: heart rate, time interval between two adjacent ventricular peaks, duration of ventricular depolarization waves, and time from the onset of ventricular depolarization to the end of repolarization. The time interval, corrected total ventricular repolarization time, initial ventricular repolarization potential changes, and waveform characteristics at the end of ventricular repolarization; the electrical impedance imaging data acquisition unit is used to acquire the detection data of the electrical impedance imaging mode of the object to be evaluated. The target markers of the electrical impedance imaging mode include at least: global conductivity change, conductivity change range, conductivity change standard deviation and maximum conductivity change, regional weighted average conductivity, conductivity change time slope, multi-frequency electrical impedance response characteristics, local maximum area ratio, conductivity distribution center offset, and symmetry index.

[0007] According to an embodiment of the present invention, the first feature fusion module includes: a calculation submodule, configured to calculate the initial weight value of any one of the N sets of risk assessment information to obtain the initial weight value of each of the N sets of risk assessment information; an adjustment submodule, configured to adjust the N initial weight values ​​according to the correlation between the N modal types to obtain N target weight values, wherein, for the first modal type and the second modal type that have a causal relationship, the first target weight value of the risk assessment information corresponding to the first modal type is greater than the second target weight value of the risk assessment information corresponding to the second modal type; for the third modal type and the fourth modal type that have a temporal relationship, the third target weight value of the risk assessment information corresponding to the third modal type is greater than the fourth target weight value of the risk assessment information corresponding to the fourth modal type; for the fifth modal type and the sixth modal type that have a cooperative change relationship, the fifth target weight value of the risk assessment information corresponding to the fifth modal type is greater than the sixth target weight value of the risk assessment information corresponding to the sixth modal type; and a first obtaining submodule, configured to perform a weighted summation of the N sets of risk assessment information according to the N target weight values ​​to obtain the first target risk assessment information.

[0008] According to an embodiment of the present invention, the calculation submodule includes: a first calculation unit, configured to calculate a first evaluation value and a second evaluation value of the risk assessment model for each risk assessment information, wherein the first evaluation value is used to characterize the ability of the risk assessment model to distinguish between positive and negative samples, and the second evaluation value is used to characterize the comprehensive balance between precision and recall of the risk assessment model; a second calculation unit, configured to calculate a target evaluation value of the risk assessment model based on the first evaluation value and the second evaluation value of the risk assessment model; and a determination unit, configured to determine the weight value of the risk assessment model based on the proportion of the target evaluation value of the risk assessment model to the cumulative evaluation value, wherein the cumulative evaluation value is determined based on the target evaluation values ​​of each of the N risk assessment models.

[0009] According to an embodiment of the present invention, the extrusion risk assessment device further includes: a second feature fusion module, used to perform feature fusion on X sets of risk assessment information when any set of risk assessment information in N sets of risk assessment information does not meet predetermined conditions, to obtain updated second target risk assessment information, wherein X < N and X is a positive integer greater than or equal to 0.

[0010] According to an embodiment of the present invention, the second feature fusion module includes: an update submodule, configured to update the respective weight values ​​of X groups of risk assessment information when any group of risk assessment information does not meet predetermined conditions, to obtain updated X weight values; and a second obtaining submodule, configured to perform weighted summation on the X groups of risk assessment information based on the updated X weight values, to obtain second target risk assessment information.

[0011] According to an embodiment of the present invention, the crush risk assessment device further includes: a first determining module, used to identify missing values ​​in N sets of data to be tested and determine the modal type of the missing data; and a filling module, used to fill in the missing data according to the modal type of the missing data.

[0012] According to an embodiment of the present invention, the data filling module includes: a first data filling submodule, used to fill in missing data based on historical detection data of a historical object when the modality type of the missing data is at least one of biochemical detection modality, blood pressure modality, electrocardiogram modality, and electrical impedance imaging modality, wherein the historical object is similar to the object to be evaluated; a second data filling submodule, used to fill in missing data using mean imputation when the missing data is a proteomics modality; and a third data filling submodule, used to fill in missing data using multiple imputation when the missing data is a metabolomics modality.

[0013] According to an embodiment of the present invention, the target scenario is an emergency scenario, and the crush risk assessment device further includes: a second determining module, used to determine a variety of initial markers related to the target scenario; and a screening module, used to screen a variety of target markers from the variety of initial markers based on the data of the various initial markers of historical objects and the historical risk assessment information of historical objects.

[0014] This invention provides a method for assessing crush risk, comprising: acquiring N sets of test data for at least one object to be assessed in a target scenario, wherein the N sets of test data correspond to target biomarkers of N modalities, and N is a positive integer greater than 2; the modalities include biochemical detection modality, proteomics modality, metabolomics modality, blood pressure modality, electrocardiogram modality, and electrical impedance imaging modality; and at least one of the following relationships exists among the N modalities: a causal relationship that characterizes the abnormality of a biomarker in the first modality as a cause of the abnormality of a biomarker in the second modality; and a relationship that characterizes the occurrence time of the abnormality of a biomarker in the third modality as earlier than that in the fourth modality. The study investigates the temporal sequence of abnormal marker occurrences and the synergistic relationship between markers representing the fifth and sixth modal types, showing a predetermined trend of synchronous change. Using feature extractors corresponding to each modal type, features are extracted from N sets of data to be detected, resulting in N sets of risk features. These N sets of risk features are then input into risk assessment models corresponding to each modal type, generating N sets of risk assessment information. Based on the correlation between the N modal types, feature fusion is performed on the N sets of risk assessment information to obtain the first target risk assessment information. A processing plan suitable for the assessed object is then generated based on this first target risk assessment information.

[0015] According to embodiments of the present invention, an acquisition module collects N sets of different modal data to capture risk information related to target markers from multiple dimensions. For example, in disaster relief, multimodal data of crush injuries can be acquired simultaneously to avoid omissions or misjudgments due to incomplete information. Furthermore, since the characteristics of different modal data vary greatly, the extraction module matches a corresponding feature extractor and risk assessment model for each modality, which can more accurately retain the core risk characteristics of each modality. The first feature fusion module integrates the N sets of risk assessment information into a unified first target risk assessment information, avoiding the bias of single-modality assessment. The multimodal fusion of correlations not only achieves comprehensive information coverage, but also improves the temporal sensitivity, causal interpretability, and robustness of anomaly identification in risk assessment by mining the intrinsic connections between modalities, more accurately capturing risk evolution trends and reducing the risk of delays caused by assessment lags or biases. This improves the comprehensiveness and accuracy of risk assessment, making it particularly suitable for scenarios with limited resources and tight deadlines, such as disaster relief. It can quickly determine the processing priority based on the risk assessment information of the primary objective, reducing the subjectivity and delay of human judgment, so as to provide a more practical solution for the object to be assessed. Attached Figure Description

[0016] Figure 1 A structural block diagram of a crush risk assessment device according to an embodiment of the present invention is shown.

[0017] Figure 2 A structural block diagram of the first feature fusion module according to an embodiment of the present invention is shown.

[0018] Figure 3 A structural block diagram of a computing submodule according to an embodiment of the present invention is shown.

[0019] Figure 4 A data flow diagram illustrating the determination of weight values ​​for risk assessment information according to an embodiment of the present invention is shown.

[0020] Figure 5 A flowchart of a crush risk assessment method according to an embodiment of the present invention is shown. Detailed Implementation

[0021] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the invention. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the invention for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0022] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. Terms such as include, comprise, etc., as used herein indicate the presence of features, steps, operations, and / or components, but do not preclude the presence or addition of one or more other features, steps, operations, or components.

[0023] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0024] In the technical solution of this invention, the user information (including but not limited to user personal information, user image information, user device information, such as location information) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of related data all comply with relevant laws, regulations, and standards, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entry points for users to choose to authorize or refuse.

[0025] In the context of disaster medicine, crush syndrome (CS) is a condition caused by prolonged pressure on a part of the body, leading to ischemia and necrosis of skeletal muscle. This results in the release of muscle cell contents into the bloodstream, potentially inducing acute kidney injury (AKI) and even death. Early identification and intervention of CS at disaster sites are crucial for improving treatment success rates. However, disaster sites face challenges such as difficulty in data acquisition, limited medical resources, and data gaps. Currently, there is a lack of systematic and intelligent risk assessment technologies to support rapid clinical decision-making.

[0026] Currently, there are two main types of technologies used to assess the risk of crush syndrome in victims at disaster sites: The first is based on a single physiological or biochemical indicator (such as creatine kinase, serum potassium, or serum creatinine). While this method is simple to operate, it has the following drawbacks from a technical perspective: limited sensitivity and specificity (a single indicator cannot comprehensively reflect the progression of CS, leading to a high risk of misdiagnosis); and an inability to provide a systematic assessment (lacking the ability to jointly model and recognize patterns among multiple indicators). In essence, this type of approach is a static rule-driven method, lacking the data modeling, feature fusion, and evaluation algorithms found in information and communication technologies, resulting in an reliance on experience and insufficient adaptability and scalability. The second type relies on assessment tools based on static scoring systems, which have stringent prerequisites and require continuous and stable data, making them difficult to implement at disaster sites.

[0027] Based on this, embodiments of the present invention provide a crush risk assessment device.

[0028] Figure 1 A structural block diagram of a crush risk assessment device according to an embodiment of the present invention is shown.

[0029] like Figure 1 As shown, the crush risk assessment device 100 includes an acquisition module 110, an extraction module 120, a generation module 130, and a first feature fusion module 140.

[0030] The acquisition module 110 is used to acquire N sets of test data for at least one object to be evaluated in a target scenario. The N sets of test data correspond to target biomarkers of N modalities, and N is a positive integer greater than 2. The modalities include biochemical detection modality, proteomics modality, metabolomics modality, blood pressure modality, electrocardiogram modality, and electrical impedance imaging modality. There are at least one of the following relationships among the N modalities: a causal relationship that characterizes the abnormality of the biomarker of the first modality as the cause of the abnormality of the biomarker of the second modality; a temporal relationship that characterizes the occurrence time of the abnormality of the biomarker of the third modality as earlier than the occurrence time of the abnormality of the biomarker of the fourth modality; and a synergistic relationship that characterizes the biomarker of the fifth modality and the biomarker of the sixth modality as changing synchronously with a preset trend.

[0031] According to embodiments of the present invention, the target scenario is a disaster relief scenario, such as the scene of an accident like an earthquake, mudslide, or building collapse. These scenarios often involve a large number of trapped or injured people, with complex rescue environments and limited resources, requiring rapid assessment of the injured's condition to prioritize the treatment of critically ill patients. The subjects to be assessed are injured persons at the disaster scene, specifically those at risk of crush syndrome. N sets of data to be detected come from different modalities to cover multiple dimensions of risk, such as images, text, physiological signals, and omics data.

[0032] According to embodiments of the present invention, N sets of test data are collected from the subjects to be assessed at the disaster site using portable on-site rapid testing devices. Portable on-site rapid testing devices include, but are not limited to, portable devices for rapid detection of blood potassium and creatinine levels, portable blood gas and electrolyte analyzers, dry handheld blood gas analyzers, blood analysis equipment, blood gas and electrolyte analyzers, test strips, wearable vital sign monitoring devices, portable electrical impedance imaging devices, etc. Furthermore, basic information about the subjects to be assessed is obtained, including but not limited to gender, age, height, weight parameters, and past history of being crushed.

[0033] According to embodiments of the present invention, proteomics and metabolomics modalities are causally related. When the body undergoes pathological changes, such as crush syndrome leading to muscle damage, muscle cell-specific necrosis markers, such as titin fragments and troponin I isoforms, will show abnormalities in proteomics detection. This muscle cell necrosis triggers a series of metabolic changes, leading to the accumulation of muscle catabolism products in metabolomics, such as branched-chain amino acids. Therefore, the abnormal changes in muscle damage-related proteins in proteomics are the cause, driving the changes in metabolites such as branched-chain amino acids in metabolomics; a clear causal relationship exists between the two.

[0034] According to embodiments of the present invention, there is a causal relationship between the biochemical detection modality and the blood pressure modality. When the renin-angiotensin system is activated in the body, this can be reflected in biochemical detection through an increase in substances such as angiotensinogen. Activation of the renin-angiotensin system promotes vasoconstriction and water and sodium retention, ultimately leading to increased blood pressure and affecting blood pressure modality data. In other words, changes in renin-angiotensin system-related indicators in the biochemical detection modality are an important cause of increased blood pressure, and the two are causally linked.

[0035] According to embodiments of the present invention, the proteomics modality and the biochemical detection modality have a temporal sequence. In the early stages of kidney injury, neutrophil gelatinase-associated lipocarboxylase (NGAL), a marker of early tubular injury in proteomics, will show an abnormal increase first, generally detectable about 2 hours after kidney injury. Creatinine, as a final indicator of declining kidney function in biochemical detection, typically only shows a significant increase about 24 hours after the kidney injury has progressed to a more severe stage. Therefore, in the development of kidney injury, changes in NGAL in proteomics occur significantly earlier than changes in creatinine in biochemical detection.

[0036] According to embodiments of the present invention, the metabolomics modality and the electrocardiogram (ECG) modality have a temporal sequence. In cases of electrolyte disturbances caused by crush syndrome or other reasons, metabolomics can detect abnormal changes in electrolyte metabolites such as serum potassium. When serum potassium concentration begins to rise, after a certain period, it affects the electrophysiological activity of the heart, which is reflected in the ECG modality as abnormalities such as prolonged corrected total ventricular repolarization time. Changes in electrolyte metabolites such as serum potassium occur first in the metabolomics process, and only after a certain period of physiological and pathological process do corresponding changes appear in the ECG modality.

[0037] According to embodiments of the present invention, the metabolomics modality and the biochemical detection modality exhibit a synergistic relationship. When crush syndrome induces muscle necrosis, metabolomics can detect abnormal fluctuations in metabolites such as creatine metabolites and inflammation-related lipids; simultaneously, biochemical indicators reflecting muscle damage, such as creatine kinase and lactate dehydrogenase, and indicators reflecting renal function impairment, such as serum potassium and serum creatinine, show a synchronous upward trend with the aforementioned metabolites. For example, an increase in creatine kinase (biochemical detection modality) is often accompanied by the accumulation of inosine (metabolomics modality), and the two synergistically enhance each other, jointly quantifying the degree of muscle necrosis and secondary organ damage.

[0038] According to embodiments of the present invention, the electrical impedance tomography (EIT) modality and the blood pressure modality exhibit a synergistic relationship. When crush syndrome leads to ischemic necrosis of muscle tissue, local tissue edema and increased vascular permeability occur. The EIT modality can reflect the extent and degree of muscle damage by detecting a decrease in the electrical impedance value of the damaged area. Simultaneously, inflammatory factors released from muscle necrosis trigger systemic vasodilation and a decrease in effective circulating blood volume, resulting in a decreasing trend in the blood pressure modality data. The lower the electrical impedance value in the muscle injury area, the greater the decrease in blood pressure, collectively reflecting the progressive state of "local tissue damage - systemic circulatory disturbance" caused by crush syndrome.

[0039] The extraction module 120 is used to extract features from N sets of data to be detected by using feature extractors corresponding to each modality type, so as to obtain N sets of risk features.

[0040] According to embodiments of the present invention, a corresponding feature extractor is designed in advance for each modality of data. For example, a convolutional neural network can be used to extract lesion features from image data. Natural language processing can be used to rapidly extract keywords or semantic vectors from text data. Time-domain or frequency-domain analysis can be used to extract fluctuation features, such as heart rate variability, from time-series data. Through feature extraction, N sets of risk features are obtained, each corresponding to a data modality. For example, image features include tumor size and density. Text features include keyword frequencies in symptom descriptions. Physiological features include heart rate variability indices.

[0041] The generation module 130 is used to input N sets of risk characteristics into the risk assessment model corresponding to each modality type to generate N sets of risk assessment information.

[0042] According to an embodiment of the present invention, historical detection data of different modalities of historical objects are used as sample data, and historical risk assessment information of historical objects are used as labels to optimize the initial assessment model corresponding to each modality, thereby obtaining the risk assessment model corresponding to each modality type.

[0043] According to an embodiment of the present invention, a trained risk assessment model is used to independently assess the risk characteristics of each modality. The input to each risk assessment model is a set of data to be tested for that modality, and the output is the risk assessment information obtained from evaluating the data to be tested for that modality. The risk assessment information is used to indicate the risk level of the assessed object for crush syndrome, for example: high risk or low risk.

[0044] For example, when N=3, there are 3 sets of risk characteristics (Group 1, Group 2, and Group 3) and 3 risk assessment models (Group 1, Group 2, and Group 3). Inputting Group 1 risk characteristics into Group 1 risk assessment model outputs Group 1 risk assessment information as high risk; inputting Group 2 risk characteristics into Group 2 risk assessment model outputs Group 2 risk assessment information as high risk; inputting Group 3 risk characteristics into Group 3 risk assessment model outputs Group 3 risk assessment information as low risk.

[0045] The first feature fusion module 140 is used to perform feature fusion on N sets of risk assessment information according to the correlation between N modal types to obtain the first target risk assessment information, so as to generate a processing plan suitable for the object to be assessed based on the first target risk assessment information.

[0046] According to embodiments of the present invention, based on different modal data (such as physiological signals, biochemical indicators, imaging data, etc.), and taking into account the intrinsic correlation between modalities, a more accurate comprehensive evaluation result is formed by strengthening the weight of key modalities and compensating for the limitations of single modalities. This results in directly matching suitable processing solutions for the object to be evaluated, thereby achieving multi-dimensional risk assessment to obtain the first target risk assessment information.

[0047] Specifically, when using a voting method for fusion, the voting weights are adjusted based on the causal, temporal, or synergistic relationships between modalities. For example, for proteomics modalities (causes) and metabolomics modalities (outcomes) with causal relationships, the former has a higher voting weight than the latter. If, out of five risk information sets, three key cause modalities are judged as "high-risk" and two outcome modalities are judged as "medium-risk," then the overall assessment after weight calculation is high-risk. When using a weighted summation fusion method, the weight allocation directly reflects the association relationship. For example, for proteomics modalities with temporal relationships (early... In the early stages of assessment, the risk assessment modality (early warning) and biochemical detection modality (progression indicators) have different weights. The former (e.g., 0.4) has a higher weight than the latter (e.g., 0.3). If the risk value of proteomics is 0.8 and the risk value of biochemical detection is 0.5, and then combined with the electrical impedance imaging modality (weight 0.3, risk value 0.3) which has a synergistic change relationship, the first target risk assessment information is calculated as 0.4×0.8+0.3×0.5+0.3×0.3=0.58. If the threshold of 0.5 is high risk, the risk level corresponding to the first target risk assessment information is determined to be high risk.

[0048] According to embodiments of the present invention, the first target risk assessment information is used to characterize the probability that the subject under assessment has crush syndrome in the target scenario, or it can also be the probability of having acute kidney injury risk and short-term mortality risk. For example, a high-risk level is obtained through blood biochemical indicators (such as creatine kinase and serum potassium); a high-risk level is obtained through physiological signals (such as heart rate and blood pressure); and a low-risk level is obtained through imaging data (such as ultrasound showing the extent of muscle damage). Combining these three risk levels, a final risk level of high risk may be obtained, which more comprehensively and accurately reflects the overall risk status of the injured person suffering from crush syndrome.

[0049] According to embodiments of the present invention, intervention recommendations are automatically generated based on the final risk level obtained through comprehensive analysis. For example, based on a preset rule base or a pre-trained decision model, corresponding targeted and operable treatment measures can be output for different risk levels to guide on-site rescue personnel to quickly carry out rescue work. Alternatively, priority can be given to treating injured persons whose first target risk assessment information indicates a high risk.

[0050] For example, if the initial risk assessment indicates high risk, the treatment plan could be to establish intravenous access, rapidly administer fluids to promote the excretion of harmful substances such as myoglobin, monitor blood potassium levels, and if hyperkalemia occurs, promptly lower potassium levels (e.g., through intravenous infusion of calcium gluconate or insulin); and arrange for transport as soon as possible. If the initial risk assessment indicates low risk, routine trauma treatment should be administered, with observation for 24 hours; urine output and basic vital signs should be monitored; if no abnormalities are observed during this period, the patient can be transferred to a general treatment area for further observation.

[0051] According to embodiments of the present invention, an acquisition module collects N sets of different modal data to capture risk information related to target markers from multiple dimensions. For example, in disaster relief, multimodal data of crush injuries can be acquired simultaneously to avoid omissions or misjudgments due to incomplete information. Furthermore, since the characteristics of different modal data vary greatly, the extraction module matches a corresponding feature extractor and risk assessment model for each modality, which can more accurately retain the core risk characteristics of each modality. The first feature fusion module integrates the N sets of risk assessment information into a unified first target risk assessment information, avoiding the bias of single-modality assessment. The multimodal fusion of correlations not only achieves comprehensive information coverage, but also improves the temporal sensitivity, causal interpretability, and robustness of anomaly identification in risk assessment by mining the intrinsic connections between modalities, more accurately capturing risk evolution trends and reducing the risk of delays caused by assessment lags or biases. This improves the comprehensiveness and accuracy of risk assessment, making it particularly suitable for scenarios with limited resources and tight deadlines, such as disaster relief. It can quickly determine the processing priority based on the risk assessment information of the primary objective, reducing the subjectivity and delay of human judgment, so as to provide a more practical solution for the object to be assessed.

[0052] According to embodiments of the present invention, the risk assessment model corresponding to each modality type can be any of the following: Support Vector Machine, Logistic Regression, Random Forest, Extreme Gradient Boosting, Lightweight Gradient Boosting, K-Nearest Neighbors Classifier, Naive Bayes, Decision Tree, Adaptive Boosting Algorithm (Ada Boost), Bagging, and Ensemble Voting, etc. Before training the initial evaluation model using historical detection data, the data for each modality type is standardized to eliminate the influence of different data volumes, and missing values ​​are imputed to avoid data incompleteness. Then, the dataset is divided into training and test sets. During training, 5-fold cross-validation is used to ensure model stability. Simultaneously, grid search and Bayesian optimization are used to systematically adjust the model's hyperparameters to find the optimal parameter combination for model performance. Finally, the model most suitable for this type of modality data is selected from multiple algorithms.

[0053] According to embodiments of the present invention, the risk assessment model required for each modality can be determined by calculating a first evaluation value and a second evaluation value. The first evaluation value characterizes the risk assessment model's ability to distinguish between positive and negative samples, i.e., the area under the curve (AUC). AUC is calculated by placing the area under the Receiver Operating Characteristic Curve (ROC) and reflects the model's ability to identify positive and negative classes at different thresholds. The AUC value ranges from 0 to 1; the closer the value is to 1, the more accurately the model can distinguish between positive and negative samples (i.e., the probability of evaluating the positive class is higher than the probability of evaluating the negative class). The second evaluation value characterizes the risk assessment model's overall balance between precision and recall, i.e., the F1 score, representing the model's overall balance between precision and recall. Precision measures the proportion of samples that the model evaluates as positive but are actually positive (avoiding false positives), while recall measures the proportion of samples that are actually positive but were correctly evaluated by the model (avoiding false negatives). The F1-score is the harmonic mean of the two, ranging from 0 to 1. The closer the value is to 1, the better the model performs in balancing precision and recall, effectively avoiding the limitations of a single indicator.

[0054] According to an embodiment of the present invention, the crush risk assessment device further includes: a second determining module for determining multiple initial markers related to the target scenario; and a screening module for selecting multiple target markers from the multiple initial markers based on the multiple initial marker data of historical objects and the historical risk assessment information of historical objects.

[0055] According to an embodiment of the present invention, the N sets of data to be detected include at least the data to be detected in the proteomics modality and the data to be detected in the metabolomics modality. The modality type includes one of the following: biochemical detection modality, proteomics modality, metabolomics modality, blood pressure modality, electrocardiogram modality and electrical impedance imaging modality.

[0056] Biochemical testing modalities include: white blood cell count (WBC), red blood cell count (RBC), hemoglobin concentration (HGB), hematocrit (HCT), mean corpuscular volume (MCV), mean corpuscular hemoglobin content (MCH), mean corpuscular hemoglobin concentration (MCHC), red blood cell distribution width (RDW), platelet count (PLT), mean platelet volume (MPV), platelet distribution width (PDW), and plateletcrit (PCT) – all routine blood parameters. Simultaneously, blood gas and electrolyte parameters are collected, including: arterial blood pH, partial pressure of carbon dioxide (SPC), and other parameters. ), oxygen partial pressure ( ), blood oxygen saturation ( ), lactic acid (Lactate), sodium ions ( ), potassium ions ( ), chloride ions ( In terms of renal function and metabolic parameters, creatinine (CREA), urea (UREA), and uric acid (UA) were collected. Additionally, coagulation function indicators included prothrombin time (PT), activated partial thromboplastin time (APTT), thrombin time (TT), and fibrinogen (FIB); and tissue damage and myocardial markers included creatine kinase (CK) and creatine kinase isoenzyme (CK-MB). Nineteen target biomarkers were identified from these initial biomarkers using extreme gradient boosting (XG Boost) feature importance ranking. Principal component analysis (PCA) and linear discriminant analysis (LDA) can also be used. For scenarios with high modality dimensionality but limited sample size, recursive feature elimination (RFE), L1 regularized linear regression (Lasso), and maximum correlation minimum redundancy (mRMR) methods can be used to improve the interpretability and stability of feature selection.

[0057] Potential proteomics biomarkers were identified through data-dependent analysis (DDA) of mass spectrometry data acquisition and library searches, followed by analysis and screening. Nine proteins were selected as target biomarkers. These proteomics modalities included: phosphoglycerate mutase 2 (PGAM2), lactate dehydrogenase A (LDHA), malate dehydrogenase 1 (MDH1), fructose-1,6-bisphosphate aldolase (ALDOA), pyruvate kinase M (PKM), glyceraldehyde-3-phosphate dehydrogenase (GAPDH), phosphoglycerate kinase (PGK), phosphoglucose mutase (PGM), and carbonic anhydrase 3 (CA3).

[0058] Traditional CS risk identification relies on creatine kinase (CK) and myoglobin, but these indicators only significantly increase 6-12 hours after muscle injury and are greatly affected by non-CS factors (such as exercise and minor trauma). However, within hours after crush injury, proteomics biomarkers can detect changes in the expression of structural and functional proteins (such as PGAM2 and LDHA) after muscle injury. These changes usually occur earlier than the time window for the elevation of biochemical indicators such as creatine kinase (CK), allowing for earlier detection of tissue damage signals. High-throughput screening of these biomarkers using proteomics can advance the CS warning window, avoiding delays in intervention due to traditional indicators not meeting the required levels. On the other hand, the core harm of CS is acute kidney injury (AKI), but traditional indicators such as creatinine and blood urea nitrogen only significantly increase after a certain level of kidney function decline. Proteomics can detect proteins released from early damage to renal tubular epithelial cells; these proteins can increase as early as 2 hours after kidney injury and can distinguish between "reversible" and "irreversible" damage.

[0059] Potential metabolomics biomarkers were screened and identified through differential metabolite analysis and metabolic pathway analysis, resulting in the identification of six target biomarkers. The metabolomics modalities included: galactose metabolism, phenylalanine metabolism, fructose and mannose metabolism, amino sugar and nucleotide sugar metabolism, phenylalanine, tyrosine and tryptophan biosynthesis, and ascorbate and malonate metabolism.

[0060] Metabolomics biomarkers can differentiate between pseudorhabdomyolysis and true CS. Traditional indicators (such as elevated creatinine kinase) may be caused by non-compression factors (such as drug-induced myopathy or infection), while metabolomics can differentiate through a "characteristic metabolite combination": true CS patients will simultaneously exhibit a triple metabolic abnormality of "enhanced creatine-to-creatinine metabolic pathway" (a characteristic of muscle breakdown), "lactate accumulation" (tissue hypoxia), and "elevated uric acid" (impaired renal excretion), while other myopathy patients usually only have a single pathway abnormality, thus improving the accuracy of risk assessment. Metabolomics modalities reflect cellular metabolic network disorders, such as lactate accumulation, amino acid metabolism disorders, and electrolyte fluctuations. Abnormalities in these metabolic pathways can precede traditional electrolyte tests, providing molecular-level early warning capabilities for predicting high-risk events such as hyperkalemia and acute kidney injury in CS.

[0061] Without proteomics and metabolomics data, the model will rely on physiological signal modalities in the early stages of CS pathogenesis, potentially leading to delayed detection of crush syndrome and missed risk identification windows. Therefore, protein and metabolic modalities provide the earliest and most discriminative information sources in the temporal dimension, improving the overall model's detection timeliness.

[0062] Blood pressure modalities include: systolic blood pressure (SBP), diastolic blood pressure (DBP), mean arterial pressure (MAP), pulse rate (Pulse Rate), pulsatility (PP), and blood pressure variability (BPV).

[0063] The electrocardiogram modality includes: heart rate, the time interval between two adjacent ventricular peaks (RR interval), the duration of ventricular depolarization complexes (QRS duration), the time interval from the onset of ventricular depolarization to the end of repolarization (QT interval), the corrected total ventricular repolarization time (QTc interval), the initial ventricular repolarization potential changes (ST segment changes), the waveform characteristics at the end of ventricular repolarization (T wave morphology), and two heart rate variability indicators, standard deviation (SDNN) and root mean square deviation (RMSSD).

[0064] The electrical impedance tomography modality includes fundamental statistical parameters such as global conductivity change (ΔGCV), conductivity range (ΔCRV), standard deviation of conductivity change (ΔSD.CV), and maximum conductivity change (ΔMCV); it also includes dynamic and spatial structural features such as region-weighted average conductivity (RWMC), conductivity change temporal slope (CCS), multi-frequency impedance response characteristics (FRS), local maximum area ratio (MCAR), conductivity distribution center shift (Centroid Shift), symmetry index, and nonlinear fit residuals. The data features of this modality can be used to identify the degree of tissue compression, the spread trend of damaged areas, and the potential risk of necrosis.

[0065] According to embodiments of the present invention, by integrating data from six modalities—biochemical detection, proteomics, metabolomics, blood pressure, electrocardiography, and electrical impedance tomography—key risk factors are comprehensively covered, overcoming the problem of insufficient information coverage in traditional single-modality models and improving the breadth of pathological mechanism identification. Furthermore, to address the challenges of significant differences between modal data and the difficulty of unified modeling, each modality is trained independently, and the optimal model is selected through cross-validation and multi-index evaluation, significantly improving the modeling accuracy and generalization ability of each modality.

[0066] According to embodiments of the present invention, in terms of feature selection, highly adaptable modeling feature selection strategies can be adopted for the above six modalities, such as feature importance ranking based on XG Boost, cluster analysis, pathway enrichment, etc., to extract a set of core variables highly correlated with crush syndrome from different physiological systems. Based on this, to further improve the model's adaptability and scalability, alternative strategies including RFE, Lasso, and mRMR can be used to optimize feature subset selection. In modalities with high dimensionality or limited samples, Lasso can eliminate redundant features through a penalty mechanism, enhancing the sparsity and generalization ability of the model; RFE is suitable for progressively optimizing feature sets in linear or tree model frameworks, improving stability; mRMR can perform variable selection in unsupervised scenarios based on mutual information, especially suitable for metabolomics, proteomics, and other omics modalities. Furthermore, for scenarios requiring improved model interpretability and deployment performance, feature masking methods based on perturbation effects (such as Permutation Importance) can be used, or a joint feature scoring mechanism can be introduced in multimodal ensemble modeling to select the globally optimal feature subset across modalities through weighted scoring across different modalities.

[0067] According to an embodiment of the present invention, the crush risk assessment device 100 further includes a first determining module and a filling module. The first determining module is used to identify missing values ​​in N sets of data to be tested and determine the modal type of the missing data; the filling module is used to fill in the missing data according to the modal type of the missing data.

[0068] According to an embodiment of the present invention, all N sets of data to be detected are traversed, and the missing rate (number of missing values ​​ / total data volume) of each modality is calculated. The specific modality type in which the missing values ​​are located is marked. Based on the characteristics of different modality data (such as continuity, temporal sequence, and distribution characteristics), the most suitable imputation method is selected to avoid data distortion caused by general methods. For example, for missing QTc intervals in electrocardiogram signals, linear interpolation can be performed using data from the first 5 time points and the last 5 time points.

[0069] According to embodiments of the present invention, the data filling module includes a first filling submodule, a second filling submodule, and a third filling submodule. The first filling submodule is used to fill in missing data based on historical detection data of a historical object when the missing data modality type is at least one of biochemical detection modality, blood pressure modality, electrocardiogram modality, and electrical impedance imaging modality, wherein the historical object is similar to the object to be evaluated. For example, K-nearest neighbor interpolation is used. The second filling submodule is used to fill in missing data using mean interpolation when the missing data is a proteomics modality. The third filling submodule is used to fill in missing data using multiple interpolation when the missing data is a metabolomics modality.

[0070] According to embodiments of the present invention, in handling missing data values, differentiated imputation is performed for different data types, temporal structures, and acquisition scenarios to improve modeling effectiveness and data consistency. Considering the complexities of data loss and incomplete modalities at disaster sites, various missing value handling strategies can be adopted to enhance model robustness. For example, collaborative filtering methods can be introduced to infer missing values ​​based on the feature similarity among historical patients, which is suitable for structured data modalities (such as biochemistry, blood pressure, etc.). For omics modalities, nonlinear reconstruction techniques such as autoencoders can be combined to restore and reconstruct high-dimensional features, preserving the potential expression relationships within the omics data. For continuously acquired time-series signals, such as ECG and blood pressure modalities, predictive models based on sliding window modeling can be used for fragment completion, such as time-series-based K-nearest neighbor regression models. Furthermore, for extreme cases of missing key modalities, multimodal conditional generation models can be constructed from acquired modalities to assist in predicting the feature expression of missing modalities, thereby maintaining the system's end-to-end inference capability.

[0071] According to embodiments of the present invention, by identifying missing modalities and matching imputation strategies, the reduction in sample size caused by directly deleting samples containing missing values ​​is avoided. This reduces data loss and preserves the integrity of multimodal data. Different modalities exhibit significant differences in characteristics, and general imputation methods may disrupt the inherent patterns of the data. Modal-specific processing, however, can maintain the dynamic trend of time-series modalities using interpolation, avoiding the loss of time-series features caused by static value imputation. For high-dimensional modalities, matrix factorization preserves the synergistic relationships between variables. This reduces the risk of data distortion and improves the reliability of subsequent modeling.

[0072] Figure 2 A structural block diagram of the first feature fusion module according to an embodiment of the present invention is shown.

[0073] According to an embodiment of the present invention, the first feature fusion module 140 includes a calculation submodule 210, an adjustment submodule 220, and a first acquisition submodule 230.

[0074] The calculation submodule 210 is used to calculate the initial weight value of any one of the N sets of risk assessment information, thus obtaining the initial weight value of each of the N sets of risk assessment information. The sum of the initial weight values ​​of the N sets of risk assessment information is equal to 1. For example, if N equals 3, the initial weight values ​​of the three sets of risk assessment information are 0.4, 0.3, and 0.3, respectively.

[0075] The adjustment submodule 220 is used to adjust N initial weight values ​​according to the correlation between N modal types to obtain N target weight values. Specifically, for the first and second modal types that have a causal relationship, the first target weight value of the risk assessment information corresponding to the first modal type is greater than the second target weight value of the risk assessment information corresponding to the second modal type; for the third and fourth modal types that have a temporal relationship, the third target weight value of the risk assessment information corresponding to the third modal type is greater than the fourth target weight value of the risk assessment information corresponding to the fourth modal type; and for the fifth and sixth modal types that have a cooperative change relationship, the fifth target weight value of the risk assessment information corresponding to the fifth modal type is greater than the sixth target weight value of the risk assessment information corresponding to the sixth modal type.

[0076] Based on the example above, the three risk assessment information modalities correspond to proteomics, metabolomics, and biochemical detection modalities, respectively. The initial weight values ​​for the risk assessment information corresponding to the proteomics modal are adjusted to the highest, followed by the metabolomics modal, and then the biochemical detection modal, with the lowest initial weight values, for example, 0.45, 0.29, and 0.26.

[0077] The first submodule 230 is used to perform weighted summation of N sets of risk assessment information based on N target weight values ​​to obtain the first target risk assessment information, referring to formula (1).

[0078] Formula (1)

[0079] in, Information for the primary objective risk assessment; Let i be the target weight value for the i-th risk assessment information; This is the i-th risk assessment information.

[0080] According to the above example, if the weights of the three risk assessment information are 0.45, 0.29, and 0.26 respectively, and the corresponding three risk assessment values ​​are 0.7, 0.6, and 0.3 respectively, the first target risk assessment information is calculated according to formula (1), that is, 0.45×0.7+0.29×0.6+0.26×0.3=0.567.

[0081] According to an embodiment of the present invention, the first target risk assessment information can be used to determine the risk level using a fixed threshold: for example, in a binary classification result, when... A value ≥0.5 is defined as "high risk," otherwise as "low risk." Multiple threshold grading can also be used; for example, when... A value ≥0.8 is considered the first risk level. As the second risk level, when It is classified as the third risk level, when <0.2 indicates the fourth risk level.

[0082] Figure 3 A structural block diagram of a computing submodule according to an embodiment of the present invention is shown. Figure 4 A data flow diagram illustrating the determination of weight values ​​for risk assessment information according to an embodiment of the present invention is shown. The following is in conjunction with... Figure 3 and Figure 4 Please provide a detailed explanation.

[0083] According to an embodiment of the present invention, the calculation submodule 210 includes a first calculation unit 310, a second calculation unit 320, and a determination unit 330.

[0084] The first calculation unit 310 is used to calculate a first evaluation value and a second evaluation value of the risk assessment model for each risk assessment information. The first evaluation value is used to characterize the risk assessment model's ability to distinguish between positive and negative samples, and the second evaluation value is used to characterize the risk assessment model's comprehensive balance between precision and recall.

[0085] According to embodiments of the present invention, such as Figure 4As shown, two independent performance metrics are calculated for each modality of the risk assessment model 401-1, 401-2...401-N to measure the model's discriminative ability and precision-recall balance, ensuring a comprehensive evaluation of the model's reliability during subsequent weighted fusion. For example, risk assessment model 401-1 can output a first evaluation value 402-1 and a second evaluation value 402-2; risk assessment model 401-2 can output a first evaluation value 402-3 and a second evaluation value 402-4; risk assessment model 401-N can output a first evaluation value 402-(2N-1) and a second evaluation value 402-2N. For instance, if the first evaluation value 402-1 of risk assessment model 401-1 is AUC=0.85, it means that when randomly selecting a positive sample and a negative sample, risk assessment model 401-1 has an 85% probability of ranking the positive sample before the negative sample. This evaluates the model's ability to balance recall and precision, which is particularly suitable for imbalanced scenarios, such as in disaster relief where severely ill patients are in the minority. If the second evaluation value 402-2 of the risk assessment model 401-1 is F1-score=0.747, it indicates that the risk assessment model 401-1 achieves a good balance between "the proportion of patients predicted to be severely ill who are actually severely ill" (precision) and "the proportion of patients who are actually severely ill who are correctly predicted to be severely ill" (recall). It neither overly favors "fewer missed diagnoses" leading to a large number of misdiagnoses (high recall but low precision), nor overly pursues "fewer misdiagnoses" missing truly severely ill patients (high precision but low recall). In scenarios such as disaster relief, which require both accuracy and comprehensiveness in identification, it can better balance the needs of "no wrong diagnosis" and "no missed diagnosis".

[0086] The second calculation unit 320 is used to calculate the target evaluation value of the risk assessment model based on the first evaluation value and the second evaluation value of the risk assessment model.

[0087] According to an embodiment of the present invention, the target evaluation value can be a weighted combination of a first evaluation value (AUC) and a second evaluation value (F1-score), with each weight set according to business needs, for example, each accounting for 50%. For example: Target evaluation value = 0.5 × AUC + 0.5 × F1-score. By integrating the two indicators, the limitations of a single indicator are avoided. For example, a model with a high AUC but a low F1-score may perform poorly in practical applications, thus reflecting the accuracy of the model more comprehensively. If the AUC of any risk assessment model is 0.80 and the F1-score is 0.78, then the target evaluation value = 0.5 × 0.80 + 0.5 × 0.78 = 0.79.

[0088] The determining unit 330 is used to determine the weight value of a risk assessment model based on the proportion of the target assessment value of any one of the risk assessment models, such as 403-1, 403-2, or 403-N, to the cumulative assessment value 404. The cumulative assessment value is determined based on the target assessment values ​​of each of the N risk assessment models. For example, the weight value 405-1 of risk assessment model 401-1 is determined based on the proportion of the target assessment value 403-1 to the cumulative assessment value 404; the weight value 405-2 of risk assessment model 401-2 is determined based on the proportion of the target assessment value 403-2 to the cumulative assessment value 404; and the weight value 405-N of risk assessment model 401-N is determined based on the proportion of the target assessment value 403-N to the cumulative assessment value 404.

[0089] According to an embodiment of the present invention, the cumulative evaluation value refers to the sum of the target evaluation values ​​of all N risk assessment models. For example, if the target evaluation values ​​of three models are 0.85, 0.79, and 0.82 respectively, then the cumulative evaluation value = 0.85 + 0.79 + 0.82 = 2.46. The weight value of each model = the target evaluation value of that model ÷ the cumulative evaluation value. The higher the weight value, the better the overall performance of the model, and the greater the influence of its risk assessment information during fusion.

[0090] For example, the biochemical detection modality uses XG Boost to balance model performance and feature interpretability; the proteomics modality chooses a random forest model to suit its high-dimensionality and small-sample features; the metabolomics modality uses a Light Gradient Boosting Machine (Light GBM) for efficient modeling and fast inference; the blood pressure modality uses a logistic regression model due to its small feature dimensionality and high deployment requirements; the electrocardiogram modality uses categorical Boosting (Cat Boost) after time-domain feature extraction to improve the ability to distinguish between class boundaries; and the electrical impedance tomography modality uses XG Boost for modeling after statistical feature expression to balance accuracy and robustness. As shown in Table 1, each modality model exhibits good predictive performance.

[0091] The models selected for each of the above modalities are merely illustrative examples and are not limited to these in practice. For example, graph neural network structures can be used for proteomics modalities, and sequence models such as Long Short-Term Memory (LSTM) networks can be used for electrocardiography modalities if the acquisition accuracy is high and complete electrocardiogram waveform sequences are included. If the image features are obvious, models other than gradient boosting methods based on image feature vectors, such as support vector machines, can also be used for electrical impedance imaging modalities.

[0092] Table 1. Correspondence between evaluation values ​​and weight values ​​for different modal types

[0093]

[0094] Among them, the biochemical detection modality had an AUC of 0.91 and an F1-score of 0.88, demonstrating excellent overall performance; the proteomics modality had an AUC of 0.89 and an F1-score of 0.85, maintaining robust performance even in high-dimensional feature environments; the metabolomics modality had an AUC of 0.86 and an F1-score of 0.83, balancing model accuracy and computational efficiency; the blood pressure modality had an AUC of 0.84 and an F1-score of 0.80; the electrocardiogram modality had an AUC of 0.81 and an F1-score of 0.77, effectively identifying physiological fluctuation risks such as heart rate variability; and the electrical impedance tomography modality had an AUC of 0.78 and an F1-score of 0.75. The target assessment value of the risk assessment model was obtained by multiplying the first and second assessment values. The weight value for each modality represents the proportion of the target assessment value to the cumulative assessment value. For example, in Table 1, the cumulative evaluation value of the biochemical detection modality, proteomics modality, metabolomics modality, blood pressure modality, electrocardiogram modality, and electrical impedance imaging modality is the sum of all target evaluation values, i.e., 0.8008 + 0.7565 + 0.7138 + 0.6720 + 0.6237 + 0.5850 = 4.1518. The weight of the biochemical detection modality is 0.8008 / 4.1518 = 0.19. The proportions of other modalities are calculated using the same method, which will not be elaborated here.

[0095] According to an embodiment of the present invention, the extrusion risk assessment device further includes a second feature fusion module, which is used to perform feature fusion on X sets of risk assessment information to obtain updated second target risk assessment information when any set of risk assessment information in N sets of risk assessment information does not meet predetermined conditions.

[0096] According to embodiments of the present invention, in multimodal risk assessment scenarios, considering the complex disaster site environment, insufficient equipment stability, and the possibility that some modal data may not be available in a timely manner, when a set of risk assessment information fails to meet predetermined standards, such as data loss, acquisition failure, or model malfunction, the system automatically excludes that set of information and performs feature fusion only on the remaining X sets of valid information. The predetermined standards are pre-set to determine the validity of a set of risk assessment information, ensuring the reliability of the data input to the fusion model or the model output results. The predetermined standards may include the following: First, data integrity standards: specifying that the missing rate of a set of data must not exceed a threshold, such as a single modal data missing rate ≤ 5%, or that key indicators must be complete, such as creatine kinase and blood potassium levels in crush syndrome assessment. Second, data quality standards: including the accuracy and reasonableness range of the data. For example, the normal reference range for serum potassium concentration in biochemical testing modalities is 3.5-5.5 mmol / L. If the collected data exceeds the extreme threshold (e.g., >8 mmol / L or <2 mmol / L) without a reasonable clinical explanation, it may be judged as a collection error and does not meet the standard. Similarly, if the imaging data has severe noise or artifacts that prevent the extraction of effective features, it will also be excluded. Third, collection process standards: If the data collection process does not conform to the preset specifications, such as sample collection time exceeding the window period or incorrect operation procedures, even if the data itself has no obvious abnormalities, it may be judged as invalid.

[0097] This dynamic adjustment mechanism generates updated secondary target risk assessment information by recalculating weights and performing weighted summation, ensuring that the final result is not affected by low-quality data and improving the reliability and robustness of the assessment. For example, in disaster relief, if the blood potassium data from a rapid testing device is unreliable due to operational errors, the system will prioritize relying on data from other modalities (such as electrocardiogram and imaging) to complete a comprehensive judgment, avoiding the influence of erroneous information on the assessment results.

[0098] According to an embodiment of the present invention, the second feature fusion module includes an update submodule and a second acquisition submodule.

[0099] The update submodule is used to update the weight values ​​of each of the X groups of risk assessment information when any group of risk assessment information does not meet the predetermined conditions, so as to obtain the updated X weight values; and the second obtaining submodule is used to perform weighted summation on the X groups of risk assessment information according to the updated X weight values ​​to obtain the second target risk assessment information.

[0100] According to an embodiment of the present invention, if a certain mode cannot output a predicted probability due to missing data, acquisition failure, or model malfunction, The system will automatically identify this mode as invalid and remove it from the fusion process. At this point, the remaining X valid modes will form a set. To maintain a total fusion weight of 1, the system proportionally normalizes the weights of the remaining effective modes. The updated weight values ​​are then calculated. Refer to formula (2):

[0101] Formula (2)

[0102] in, Let be the target weight value for the j-th risk assessment information.

[0103] Furthermore, the fusion model only calculates the weighted probability of the effective modes, and the calculation method for the second objective risk assessment information is based on formula (3):

[0104] Formula (3)

[0105] When the fusion probability meets the following condition, the system outputs the risk assessment information for the second objective: A value ≥0.5 is considered "high risk"; otherwise, it is considered "low risk".

[0106] One approach is to use a "soft voting weighting" method to fuse the evaluation results of different modalities. Specifically, fixed weights are pre-assigned based on the overall performance of each modal model (e.g., accuracy, reliability). Simultaneously, valid modal information is automatically identified, and the weights are adjusted to maintain a total weight of 1, ensuring stable prediction results even if data for some modalities is missing. To make this fusion method more adaptable to different situations, techniques such as "stacked fusion" and "hybrid fusion" from ensemble learning can be employed. These techniques automatically learn the nonlinear relationships between the outputs of different modalities in higher-level models, enhancing the expressive power of the fused model. Alternatively, an "attention-based fusion strategy" can be used, dynamically adjusting the importance weights of each modality according to the specific circumstances of different injured individuals, adapting to the differences in the roles of different modalities in different samples.

[0107] According to embodiments of the present invention, this fault-tolerance mechanism supports scenarios with an arbitrary number of missing modalities. The system can automatically identify valid modalities and dynamically adjust the fusion strategy to ensure uninterrupted prediction functionality. This mechanism does not rely on a fixed upper limit on the number of modalities and theoretically supports prediction calculations as long as at least one valid modality is present. In verification experiments simulating missing modalities, the maximum decrease in AUC is no more than 2%, and the maximum decrease in F1-score is no more than 1.5%, with overall performance still superior to any single-modal model, demonstrating good robustness and adaptability to practical deployment.

[0108] According to embodiments of the present invention, in adapting to scenarios where disaster site data collection is incomplete, the present invention designs a performance-weighted soft-voting fusion strategy, introducing a modality validity identification mechanism and weight normalization logic. Even when single-modality or multi-modality data is missing, the system can automatically adjust the fusion strategy to ensure uninterrupted prediction. Experiments show that the AUC and F1-score of the fusion model decrease by no more than 2%, improving robustness. Furthermore, it not only outputs the overall risk probability of crush syndrome but also supports individualized assessments of complications such as acute kidney injury and mortality risk, providing support for medical personnel.

[0109] The crush risk assessment method of this invention can also be used to simultaneously assess the risk probability of crush syndrome, the risk probability of acute kidney injury, and the risk probability of short-term death. Specifically, during the model training phase, historical test data of historical subjects are used as sample data, and the crush syndrome assessment information of historical subjects is used as labels to train the crush syndrome risk assessment model; similar historical test data of the same subjects are used as sample data, and the acute kidney injury risk assessment information of historical subjects is used as labels to train the acute kidney injury risk assessment model; similar historical test data of the same subjects are used as sample data, and the short-term death risk assessment information of historical subjects is used as labels to train the short-term death risk assessment model. When it is actually necessary to conduct risk assessments on the subjects to be assessed, the test data of the subjects to be assessed are input into the crush syndrome risk assessment model, the acute kidney injury risk assessment model, and the short-term death risk assessment model, respectively, to obtain the risk probability of the subject to be assessed having crush syndrome, the risk probability of acute kidney injury, and the risk probability of short-term death, respectively.

[0110] The output of the primary objective risk assessment information is all structured data, which can be integrated into electronic medical record systems, mobile terminals at disaster sites, or remote command platforms through standard formats. It also demonstrates which modalities and models were used to complete the assessment.

[0111] It should be noted that the primary or secondary risk assessment information output by the aforementioned model is merely a reference for medical personnel during on-site rescue of the injured. It serves as an intermediate result of processing parameters to prioritize the treatment of high-risk individuals and is not intended as a direct diagnostic result. Furthermore, the entire process is an information processing method implemented by computers or similar devices, requiring no medical personnel involvement. Its direct purpose is not to obtain a diagnosis or health status, and the information obtained cannot directly lead to a diagnosis or health status.

[0112] Figure 5 A flowchart of a crush risk assessment method according to an embodiment of the present invention is shown.

[0113] like Figure 5As shown, the crush risk assessment method of this embodiment includes steps S510 to S540.

[0114] Step S510: Obtain N sets of detection data for at least one object to be evaluated in the target scene. The N sets of detection data are related to the target marker, and N is a positive integer greater than 2. The N sets of detection data are data of different modal types.

[0115] Step S520: Using the feature extractors corresponding to each modality type, feature extraction is performed on the N sets of data to be detected to obtain N sets of risk features.

[0116] Step S530: Input the N sets of risk characteristics into the risk assessment models corresponding to the modality types respectively to generate N sets of risk assessment information.

[0117] Step S540: Perform feature fusion on N sets of risk assessment information to obtain first target risk assessment information, and generate a processing plan suitable for the object to be assessed based on the first target risk assessment information.

[0118] According to embodiments of the present invention, risk information related to target markers is captured from multiple dimensions by collecting N sets of different modal data. For example, in disaster relief, multimodal data of crush injuries can be acquired simultaneously to avoid omissions or misjudgments due to incomplete information. Furthermore, since the characteristics of different modal data vary greatly, matching a corresponding feature extractor and risk assessment model for each modality can more accurately retain the core risk characteristics of each modality. By integrating N sets of risk assessment information into unified first-target risk assessment information through feature fusion technology, the bias of single-modality assessment is avoided. Multimodal fusion of correlations not only achieves comprehensive information coverage, but also improves the temporal sensitivity, causal interpretability, and robustness of anomaly identification in risk assessment by mining the intrinsic connections between modalities, more accurately capturing risk evolution trends, and reducing the risk of delays caused by assessment lags or biases. Thus, the comprehensiveness and accuracy of risk assessment are improved, especially suitable for scenarios with limited resources and tight time, such as disaster relief. It can quickly determine the processing priority based on the first-target risk assessment information, reducing the subjectivity and delay of human judgment, so as to provide a processing solution that is more in line with actual needs for the assessed object.

[0119] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A crush risk assessment device, characterized in that, include: The acquisition module is used to acquire N sets of test data for at least one object to be evaluated in an emergency scenario. The N sets of test data correspond to N modalities of target biomarkers, and N is a positive integer greater than 2. The acquisition module includes a biochemical detection data acquisition unit, a proteomics data acquisition unit, a metabolomics data acquisition unit, a blood pressure data acquisition unit, an electrocardiogram data acquisition unit, and an electrical impedance imaging data acquisition unit. At least one of the following relationships exists among the N modalities: a causal relationship characterizing the abnormality of a biomarker in the first modality as a cause of an abnormality in a biomarker in the second modality; a biomarker characterizing the abnormality of a biomarker in the third modality. The time sequence of abnormal occurrence of biomarkers in the fourth modality type is considered, as is the synergistic change relationship between biomarkers in the fifth and sixth modality types, which exhibit a predetermined trend of synchronous change. The N sets of data to be tested include at least data from proteomics and metabolomics modalities. The proteomics data acquisition unit is used to collect the data from the proteomics modalities of the object to be evaluated. The target biomarkers for the proteomics modalities include at least: phosphoglycerate mutase 2, lactate dehydrogenase A, malate dehydrogenase 1, fructose-1,6-bisphosphate aldolase, and pyruvate kinase M. The metabolomics data acquisition unit is used to collect the target data of the metabolomics modality of the subject to be evaluated. The target biomarkers of the metabolomics modality include at least one of the following: galactose metabolism data, phenylalanine metabolism data, fructose and mannose metabolism data, amino sugar and nucleotide sugar metabolism data, phenylalanine, tyrosine and tryptophan biosynthesis data, ascorbic acid and malonic acid metabolism data. The extraction module is used to extract features from the N sets of data to be detected by using feature extractors corresponding to the N modal types respectively, so as to obtain N sets of risk features; The generation module is used to input the N sets of risk characteristics into the risk assessment models corresponding to the N modal types respectively, and generate N sets of risk assessment information. The first feature fusion module is used to perform feature fusion on the N sets of risk assessment information according to the correlation between the N modal types to obtain the first target risk assessment information, so as to generate a processing scheme applicable to the at least one object to be assessed based on the first target risk assessment information.

2. The apparatus according to claim 1, characterized in that, The N sets of data to be detected include at least the data to be detected in the biochemical detection modality, the blood pressure modality, the electrocardiogram modality, and the electrical impedance tomography modality. The biochemical detection data acquisition unit is used to acquire the detection data of the biochemical detection modality of the object to be evaluated. The target biomarkers of the biochemical detection modality include at least: white blood cell count, red blood cell count, hemoglobin concentration, hematocrit, mean corpuscular volume, mean corpuscular hemoglobin content, mean corpuscular hemoglobin concentration, red blood cell distribution width, platelet count, mean platelet volume, platelet distribution width, plateletcrit, arterial blood pH, partial pressure of carbon dioxide, partial pressure of oxygen, blood oxygen saturation, lactate, sodium ion, potassium ion, chloride ion, creatinine, urea, uric acid, prothrombin time, activated partial thromboplastin time, thrombin time, fibrinogen, creatine kinase and creatine kinase isoenzyme; The blood pressure data acquisition unit is used to acquire the target data of the blood pressure modality of the subject to be evaluated. The target markers of the blood pressure modality include at least: systolic blood pressure, diastolic blood pressure, mean arterial pressure, pulse rate, pulsatility, and blood pressure variability. The electrocardiogram data acquisition unit is used to acquire the detection data of the electrocardiogram modality of the subject to be evaluated. The target markers of the electrocardiogram modality include at least: heart rate, time interval between two adjacent ventricular peaks, duration of ventricular depolarization wave group, time interval from the start of ventricular depolarization to the end of repolarization, corrected total ventricular repolarization time, ventricular repolarization potential changes in the early stage of ventricular repolarization, and waveform characteristics in the late stage of ventricular repolarization. The electrical impedance imaging data acquisition unit is used to acquire the detection data of the electrical impedance imaging mode of the object to be evaluated. The target markers of the electrical impedance imaging mode include at least: global conductivity change, conductivity change range, conductivity change standard deviation and conductivity change, regional weighted average conductivity, conductivity change time slope, multi-frequency electrical impedance response characteristics, local area ratio, conductivity distribution center offset, and symmetry index.

3. The apparatus according to claim 1, characterized in that, The first feature fusion module includes: The calculation submodule is used to calculate the initial weight value of any one of the N sets of risk assessment information, and obtain the initial weight value of each of the N sets of risk assessment information. An adjustment submodule is used to adjust the N initial weight values ​​according to the correlation between the N modal types to obtain N target weight values. Specifically, for the first and second modal types with a causal relationship, the first target weight value of the risk assessment information corresponding to the first modal type is greater than the second target weight value of the risk assessment information corresponding to the second modal type; for the third and fourth modal types with a temporal relationship, the third target weight value of the risk assessment information corresponding to the third modal type is greater than the fourth target weight value of the risk assessment information corresponding to the fourth modal type; and for the fifth and sixth modal types with a cooperative change relationship, the fifth target weight value of the risk assessment information corresponding to the fifth modal type is greater than the sixth target weight value of the risk assessment information corresponding to the sixth modal type. The first obtaining submodule is used to perform a weighted summation of the N sets of risk assessment information based on the N target weight values ​​to obtain the first target risk assessment information.

4. The apparatus according to claim 3, characterized in that, The computational submodule includes: The first calculation unit is used to calculate a first evaluation value and a second evaluation value of the risk assessment model for each of the risk assessment information, wherein the first evaluation value is used to characterize the ability of the risk assessment model to distinguish between positive and negative samples, and the second evaluation value is used to characterize the ability of the risk assessment model to balance between precision and recall. The second calculation unit is used to calculate the target evaluation value of the risk assessment model based on the first evaluation value and the second evaluation value of the risk assessment model. A determining unit is used to determine the weight value of the risk assessment model based on the proportion of the target assessment value of the risk assessment model to the cumulative assessment value, wherein the cumulative assessment value is determined based on the target assessment values ​​of each of the N risk assessment models.

5. The apparatus according to claim 1, characterized in that, The device further includes: The second feature fusion module is used to perform feature fusion on X sets of risk assessment information when any one set of risk assessment information in the N sets of risk assessment information does not meet the predetermined conditions, so as to obtain updated second target risk assessment information, wherein X < N and X is a positive integer greater than or equal to 0.

6. The apparatus according to claim 5, characterized in that, The second feature fusion module includes: The update submodule is used to update the weight values ​​of the X groups of risk assessment information when any one of the groups of risk assessment information does not meet the predetermined conditions, so as to obtain the updated X weight values. The second obtaining submodule is used to perform a weighted summation of the X groups of risk assessment information based on the updated X weight values ​​to obtain the second target risk assessment information.

7. The apparatus according to claim 1, characterized in that, The device further includes: The first determining module is used to identify missing values ​​in the N sets of data to be detected and determine the modality type of the missing data. The filling module is used to fill in the missing data according to the modality type of the missing data.

8. The apparatus according to claim 7, characterized in that, The filling module includes: The first filling submodule is used to fill in the missing data based on the historical detection data of the historical object when the modal type of the missing data is at least one of biochemical detection modality, blood pressure modality, electrocardiogram modality, and electrical impedance imaging modality, wherein the historical object is similar to the object to be evaluated; The second imputation submodule is used to imput the missing data using mean imputation when the missing data is a proteomics modality. The third imputation submodule is used to impute the missing data using multiple imputation methods when the missing data is a metabolomics modality.

9. The apparatus according to claim 1, characterized in that, The crush risk assessment device also includes: The second determination module is used to determine multiple initial markers related to the target scene; The filtering module is used to filter out multiple target markers from the multiple initial markers based on the historical object's data and the historical risk assessment information of the historical object.

10. A method for assessing crush risk, characterized in that, The method includes: At least N sets of test data for at least one object to be evaluated in an emergency scenario are acquired. These N sets of test data correspond to target biomarkers of N modalities, where N is a positive integer greater than 2. The N sets of test data are acquired by a biochemical detection data acquisition unit, a proteomics data acquisition unit, a metabolomics data acquisition unit, a blood pressure data acquisition unit, an electrocardiogram data acquisition unit, and an electrical impedance imaging data acquisition unit. At least one of the following relationships exists among the N modalities: a causal relationship characterizing abnormal biomarkers of the first modality as a cause of abnormal biomarkers of the second modality; a biomarker characterizing a third modality... The time sequence of abnormal occurrence of biomarkers in the fourth modality type is considered, as is the synergistic change relationship between biomarkers in the fifth and sixth modality types, which exhibit a predetermined trend of synchronous change. The N sets of data to be tested include at least data from proteomics and metabolomics modalities. The proteomics data acquisition unit is used to collect the data from the proteomics modalities of the object to be evaluated. The target biomarkers for the proteomics modalities include at least: phosphoglycerate mutase 2, lactate dehydrogenase A, malate dehydrogenase 1, fructose-1,6-bisphosphate aldolase, and pyruvate kinase M. The metabolomics data acquisition unit is used to collect the target data of the metabolomics modalities of the subject to be evaluated. The target biomarkers of the metabolomics modalities include at least one of the following: galactose metabolism data, phenylalanine metabolism data, fructose and mannose metabolism data, amino sugar and nucleotide sugar metabolism data, phenylalanine, tyrosine and tryptophan biosynthesis data, ascorbic acid and malonic acid metabolism data; using the feature extractors corresponding to each of the N modal types, features are extracted from the N sets of target data to obtain N sets of risk features; the N sets of risk features are input into the risk assessment models corresponding to each of the N modal types to generate N sets of risk assessment information; Based on the correlation between the N modal types, feature fusion is performed on the N sets of risk assessment information to obtain first target risk assessment information, so as to generate a processing scheme applicable to the at least one object to be assessed based on the first target risk assessment information.

Images