Squeezing risk assessment device and method

Abstract

The invention provides an extrusion risk assessment device and method, and relates to the technical field of data processing, and the device comprises an obtaining module which is used for obtaining N groups of to-be-detected data of at least one to-be-assessed object in a target scene; the extraction module is used for performing feature extraction on the N groups of to-be-detected data by using feature extractors corresponding to the N modal types to obtain N groups of risk features; the generation module is used for inputting the N groups of risk features into risk assessment models corresponding to the N modal types respectively to generate N groups of risk assessment information; and the first feature fusion module is used for performing feature fusion on the N groups of risk assessment information according to the association relationship among the N modal types to obtain first target risk assessment information, so as to generate a processing scheme suitable for at least one to-be-assessed object according to the first target risk assessment information.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a crush risk assessment device and method. BACKGROUND

[0002] Crush syndrome refers to a series of clinical syndromes characterized by hyperkalemia, myoglobinuria, and acute renal failure, which are caused by ischemic necrosis of muscle tissue and release of intracellular substances into the blood circulation after long-term compression of body parts rich in muscle (such as limbs and torso) by heavy objects. The occurrence of this syndrome is closely related to the intensity and duration of the crush and the muscle content of the compressed part, and its severity directly affects the prognosis of the patient. Therefore, accurate identification and assessment of crush syndrome and timely intervention are crucial to reduce the incidence and mortality rate.

[0003] In the process of implementing the present application concept, at least the following problems exist: The existing single data-based assessment method can only reflect local histological changes, making it difficult to comprehensively assess the extent of muscle damage and the state of systemic inflammatory response, which does not conform to the pathophysiological characteristics of crush syndrome with multiple system involvement, and may lead to one-sided assessment results, thereby affecting the timeliness of intervention. SUMMARY

[0004] In view of the above problems, the present application provides a crush risk assessment device and method.

[0005] The application provides an extrusion risk assessment device, comprising: an acquisition module configured to acquire N groups of to-be-detected data of at least one to-be-evaluated object in a target scene, wherein the N groups of to-be-detected data correspond to target markers of N modal types, N is a positive integer greater than 2, the acquisition module comprises 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 there is at least one of the following association relationships between the N modal types: a cause-and-effect relationship for representing that marker abnormality of a first modal type is a cause of marker abnormality of a second modal type, a time sequence relationship for representing that marker abnormality of a third modal type appears earlier than marker abnormality of a fourth modal type, and a synergistic change relationship for representing that a marker of a fifth modal type and a marker of a sixth modal type change synchronously in a preset trend; an extraction module configured to use a feature extractor corresponding to each of the N modal types to perform feature extraction on the N groups of to-be-detected data respectively, and obtain N groups of risk features; a generation module configured to input the N groups of risk features into a risk assessment model corresponding to each of the N modal types respectively, and generate N groups of risk assessment information; and a first feature fusion module configured to perform feature fusion on the N groups of risk assessment information according to the association relationship between the N modal types, and obtain first target risk assessment information, so as to generate a processing scheme suitable for the to-be-evaluated object according to the first target risk assessment information.

[0006] According to an embodiment of the present application, the N groups of to-be-detected data at least include to-be-detected data of a proteomics modality and to-be-detected data of a metabolomics modality, a proteomics data acquisition unit is configured to acquire to-be-detected data of a proteomics modality of the to-be-evaluated subject, target markers of the proteomics modality at least include one of phosphoglycerate mutase 2, lactate dehydrogenase A, malate dehydrogenase 1, fructose-1, 6-bisphosphate aldolase, pyruvate kinase M type, glyceraldehyde-3-phosphate dehydrogenase, phosphoglycerate kinase, phosphoglucomutase and carbonic anhydrase 3; a metabolomics data acquisition unit is configured to acquire to-be-detected data of a metabolomics modality of the to-be-evaluated subject, target markers of the metabolomics modality at least include 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, ascorbic acid and malonic acid metabolism data; a biochemical detection data acquisition unit is configured to acquire to-be-detected data of a biochemical detection modality of the to-be-evaluated subject, target markers of the biochemical detection modality at least include 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 volume distribution width, platelet count, mean platelet volume, platelet distribution width, platelet hematocrit, arterial blood acid-base balance, carbon dioxide partial pressure, oxygen partial pressure, blood oxygen saturation, lactic acid, sodium ion, potassium ion, chloride ion, creatinine, urea, uric acid, prothrombin time, activated partial thromboplastin time, thrombin time, fibrinogen, creatine kinase and creatine kinase isozyme; a blood pressure data acquisition unit is configured to acquire to-be-detected data of a blood pressure modality of the to-be-evaluated subject, target markers of the blood pressure modality at least include systolic pressure, diastolic pressure, mean arterial pressure, pulse rate, pulsatility pressure and blood pressure variability; an electrocardiogram data acquisition unit is configured to acquire to-be-detected data of an electrocardiogram modality of the to-be-evaluated subject, target markers of the electrocardiogram modality at least include heart rate, time interval between adjacent two ventricular wave peaks, ventricular depolarization wave group duration, time interval from ventricular depolarization start to repolarization end, corrected total ventricular repolarization time, ventricular repolarization early potential change, waveform characteristics of ventricular repolarization terminal stage; an electrical impedance imaging data acquisition unit is configured to acquire to-be-detected data of an electrical impedance imaging modality of the to-be-evaluated subject, target markers of the electrical impedance imaging modality at least include global conductivity variation amount, conductivity variation range, conductivity variation standard deviation and maximum conductivity variation amount, regional weighted average conductivity, time slope of conductivity variation, multi-frequency electrical impedance response characteristics, local maximum value area ratio, conductivity distribution center offset amount, symmetry index.

[0007] According to an embodiment of the present application, the first feature fusion module comprises: a calculation submodule, configured to calculate an initial weight value of any one of the N groups of risk assessment information, to obtain the initial weight value of each of the N groups 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 a first modal type and a second modal type having a causal relationship, a first target weight value of the risk assessment information corresponding to the first modal type is greater than a second target weight value of the risk assessment information corresponding to the second modal type; for a third modal type and a fourth modal type having a time sequence relationship, a third target weight value of the risk assessment information corresponding to the third modal type is greater than a fourth target weight value of the risk assessment information corresponding to the fourth modal type; for a fifth modal type and a sixth modal type having a synergistic change relationship, a fifth target weight value of the risk assessment information corresponding to the fifth modal type is greater than a sixth target weight value of the risk assessment information corresponding to the sixth modal type; and a first obtaining submodule, configured to perform weighted summation on the N groups 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 application, the calculation submodule comprises: a first calculation unit, configured to calculate a first evaluation value and a second evaluation value of a risk assessment model for each risk assessment information, wherein the first evaluation value is used to represent the distinguishing ability of the risk assessment model for positive and negative samples, and the second evaluation value is used to represent the comprehensive balance ability of the risk assessment model between precision and recall; a second calculation unit, configured to calculate a target evaluation value of the risk assessment model according to the first evaluation value and the second evaluation value of the risk assessment model; and a determination unit, configured to determine a weight value of the risk assessment model according to the proportion of the target evaluation value of the risk assessment model in a cumulative evaluation value, wherein the cumulative evaluation value is determined according to the target evaluation value of each of the N risk assessment models.

[0009] According to an embodiment of the present application, the extrusion risk assessment device further comprises: a second feature fusion module, configured to perform feature fusion on X groups of risk assessment information to obtain updated second target risk assessment information in a case where any one of the N groups of risk assessment information does not meet a predetermined condition, wherein X < N and X is a positive integer greater than or equal to 0.

[0010] According to an embodiment of the present application, the second feature fusion module comprises: an updating submodule, configured to update the weight value of each of the X groups of risk assessment information to obtain updated X weight values in a case where any one of the risk assessment information does not meet the predetermined condition; and a second obtaining submodule, configured 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.

[0011] According to an embodiment of the present application, the extrusion risk assessment device further comprises: a first determination module configured to perform missing value identification on the N groups of to-be-detected data, and determine the modal type of the missing data; and a filling module configured to fill the missing data according to the modal type of the missing data.

[0012] According to an embodiment of the present application, the filling module comprises: a first filling submodule configured to, in a case where the modal type of the missing data is at least one of a biochemical detection modal, a blood pressure modal, an electrocardiogram modal, and an electrical impedance imaging modal, fill the missing data according to historical detection data of a historical subject similar to the to-be-evaluated subject; a second filling submodule configured to, in a case where the missing data is a proteomics modal, fill the missing data by using a mean imputation method; and a third filling submodule configured to, in a case where the missing data is a metabolomics modal, fill the missing data by using a multiple imputation method.

[0013] According to an embodiment of the present application, the target scene is an emergency scene, and the extrusion risk assessment device further comprises: a second determination module configured to determine a plurality of initial markers related to the target scene; and a screening module configured to screen a plurality of target markers from the plurality of initial markers according to the plurality of initial marker data of the historical subject and the historical risk assessment information of the historical subject.

[0014] The present application provides an extrusion risk assessment method, comprising: obtaining N groups of to-be-detected data of at least one to-be-evaluated subject in a target scene, wherein the N groups of to-be-detected data correspond to target markers of N modal types, and N is a positive integer greater than 2, the modal types include a biochemical detection modal, a proteomics modal, a metabolomics modal, a blood pressure modal, an electrocardiogram modal, and an electrical impedance imaging modal, and there is at least one of the following association relationships between the N modal types: a causal relationship for representing that marker abnormality of a first modal type is a cause of marker abnormality of a second modal type, a time sequence relationship for representing that marker abnormality of a third modal type appears earlier than marker abnormality of a fourth modal type, and a cooperative change relationship for representing that a marker of a fifth modal type and a marker of a sixth modal type change synchronously in a preset trend; performing feature extraction on the N groups of to-be-detected data by using a feature extractor corresponding to each modal type, respectively, to obtain N groups of risk features; inputting the N groups of risk features into a risk assessment model corresponding to each modal type, respectively, to generate N groups of risk assessment information; performing feature fusion on the N groups of risk assessment information according to the association relationship between the N modal types, to obtain first target risk assessment information, so as to generate a processing scheme suitable for the to-be-evaluated subject according to the first target risk assessment information.

[0015] According to an embodiment of the present application, N sets of different modal data are collected by the acquisition module to capture risk information related to the target marker from multiple dimensions. For example, in disaster rescue, multi-modal data of a crush victim is acquired simultaneously to avoid missed or misjudged information due to one-sidedness. Further, due to the large differences in characteristics of different modal data, the extraction module matches a corresponding feature extractor and risk assessment model for each modal, which can more accurately retain the core risk features of each modal data. The first feature fusion module integrates N sets of risk assessment information into unified first target risk assessment information to avoid the deviation of single modal assessment. The multi-modal fusion of the correlation relationship not only achieves comprehensive coverage of information, but also improves the time sensitivity, causal interpretability and abnormal recognition robustness of risk assessment by mining the internal relationship between modalities, more accurately captures the risk evolution trend, and reduces the delay risk caused by assessment lag or deviation. Thus, the comprehensiveness and accuracy of risk assessment are improved, which is especially suitable for disaster rescue and other resource-limited and time-critical scenarios, can quickly obtain the processing priority according to the first target risk assessment information, reduces the subjectivity and delay of human judgment, and provides a more practical processing scheme for the object to be evaluated. BRIEF DESCRIPTION OF DRAWINGS

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

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

[0018] Figure 3 A structural block diagram of a calculation sub-module according to an embodiment of the present application is shown.

[0019] Figure 4 A data flow diagram for determining a weight value of risk assessment information according to an embodiment of the present application is shown.

[0020] Figure 5 A flowchart of a crush risk assessment method according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0021] Hereinafter, embodiments of the present application will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present application. In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the embodiments of the present application. However, it will be apparent to one skilled in the art that one or more embodiments can be practiced without these specific details. In addition, in the following description, descriptions of well-known structures and techniques have been omitted to avoid unnecessarily obscuring the concept of the present application.

[0022] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. As used herein, the terms "comprises", "comprising", "includes", "including" and the like are inclusive of the stated features, steps, operations and / or components, but are not limited to those features, steps, operations and / or components.

[0023] All terms used herein including technical and scientific terms have the meanings commonly understood by one of ordinary skill in the art unless otherwise defined. It should be noted that the terms used herein are defined in accordance with the meanings consistent with the context of the present specification, and should not be interpreted in an idealized or overly formal manner.

[0024] In the technical solutions of the present application, the user information (including but not limited to user personal information, user image information, user equipment information, such as location information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved are all information and data authorized by the user or authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of related data comply with relevant laws, regulations and standards, necessary security measures are taken, public order and good customs are not violated, and corresponding operation entrances are provided for users to choose authorization or refusal.

[0025] In the context of disaster medicine, crush syndrome (CS) is a condition in which a part of the body is under pressure for a long time, leading to skeletal muscle ischemic necrosis, and muscle cell contents enter the blood circulation, thereby inducing acute kidney injury (AKI) and even death. Early identification and intervention of CS at the disaster site is crucial to improve the success rate of treatment, but there are problems such as difficulty in data acquisition, limited diagnosis and treatment resources, and data missing at the accident site, and there is currently a lack of systematic and intelligent risk assessment technology to support rapid clinical decision-making.

[0026] Currently, there are two main types of technology that can be used to assess whether a crushed object at the disaster site has a crush syndrome risk: one is a judgment method based on a single physiological or biochemical indicator (such as creatine kinase, blood potassium, serum creatinine). This type of method has the advantage of being simple to operate, but from a technical point of view, it has the following limitations: limited sensitivity and specificity: a single indicator cannot fully reflect the development of CS, and the risk of misjudgment is high; unable to provide systematic evaluation: lack of joint modeling and pattern recognition capabilities between multiple indicators. From a theoretical point of view, this type of solution is essentially a static rule-driven method, without introducing data modeling, feature fusion and evaluation algorithms in information and communication technology, resulting in its evaluation effect relying on experience, and insufficient adaptability and scalability. The application premise of the second type of evaluation tool based on a static scoring system is harsh, requiring continuous and stable data, which is difficult to achieve at the disaster site.

[0027] Based on this, the embodiment of the present application provides a crush risk assessment device.

[0028] Figure 1 The structural block diagram of the crush risk assessment device according to the embodiment of the present application is shown.

[0029] As Figure 1 shown, the crush risk assessment device 100 comprises 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 configured to acquire N groups of to-be-detected data of at least one to-be-evaluated object in a target scene, wherein the N groups of to-be-detected data correspond to target markers of N modal types, and N is a positive integer greater than 2, the modal types include a biochemical detection modal, a proteomics modal, a metabolomics modal, a blood pressure modal, an electrocardiogram modal and an electrical impedance imaging modal, and there are at least one of the following association relationships between the N modal types: a causal relationship for representing that a marker abnormality of a first modal type is a cause of a marker abnormality of a second modal type, a time sequence relationship for representing that a marker abnormality of a third modal type appears earlier than a marker abnormality of a fourth modal type, and a cooperative change relationship for representing that a marker of a fifth modal type and a marker of a sixth modal type change synchronously in a preset trend.

[0031] According to the embodiment of the present application, the target scene is a disaster rescue scene, such as an earthquake, a mudslide, a building collapse and the like, in which there are often a large number of trapped or injured personnel, the rescue environment is complex and the resources are limited, and it is necessary to quickly evaluate the condition of the injured personnel to preferentially treat the critically ill. The to-be-evaluated object is an injured personnel at a disaster scene, i.e., an injured personnel who may have a crush syndrome risk. The N groups of to-be-detected data come from different modalities to cover multiple dimensions of the risk. For example, image, text, physiological signal, omics data and the like.

[0032] According to the embodiment of the present application, the N groups of to-be-detected data of the to-be-evaluated object are collected by a portable on-site rapid detection device at the disaster scene. The portable on-site rapid detection device includes but is not limited to a portable device for rapidly detecting potassium and creatinine content, a portable blood gas and electrolyte analyzer, a dry handheld blood gas analyzer, a blood analysis device, a blood gas electrolyte analyzer, a test strip, a wearable vital sign monitoring device, a portable electrical impedance imaging device and the like. In addition, basic information of the to-be-evaluated object is acquired, including but not limited to gender, age, height, weight parameters, past crush history and the like.

[0033] According to an embodiment of the present application, the proteomics modality has a causal relationship with the metabolomics modality. When the body is diseased, such as muscle damage caused by crush syndrome, muscle cell-specific necrosis markers, such as fragments of muscle-associated protein, troponin I subtypes, etc., will appear abnormal in proteomics detection. The necrosis of these muscle cells will trigger a series of metabolic changes, causing muscle catabolism products such as branched-chain amino acids to accumulate in metabolomics. Therefore, the abnormal changes of muscle injury-related proteins in proteomics are the cause, driving the changes of metabolites such as branched-chain amino acids in metabolomics, and there is a clear causal relationship between the two.

[0034] According to an embodiment of the present application, the biochemical detection modality has a causal relationship with the blood pressure modality. When the renin-angiotensin system in the body is activated, it can be reflected in biochemical detection by the increase of angiotensinogen and other substances. After the renin-angiotensin system is activated, it will cause vasoconstriction, water and sodium retention, etc., ultimately leading to an increase in blood pressure, affecting the blood pressure modality data. That is, the changes in the renin-angiotensin system-related indicators in the biochemical detection modality are an important cause of the increase in blood pressure, and there is a causal relationship between the two.

[0035] According to an embodiment of the present application, the proteomics modality has a time sequence relationship with the biochemical detection modality. In the early stage of kidney damage, the early damage marker of the renal tubule in proteomics, neutrophil gelatinase-associated lipocalin (NGAL), will first appear abnormal and increase, and changes can be detected about 2 hours after kidney injury. Creatinine in biochemical detection, as a terminal indicator of decreased kidney function, will not significantly increase until about 24 hours after kidney injury progresses to a certain stage. Therefore, during the development of kidney damage, the change in NGAL in proteomics is significantly earlier than the change in creatinine in biochemical detection.

[0036] According to an embodiment of the present application, the metabolomics modality has a time sequence relationship with the electrocardiogram modality. When electrolyte disorders are caused by crush syndrome or other reasons, metabolomics can detect abnormal changes in electrolyte metabolites such as potassium. When the concentration of potassium begins to rise, after a certain period of time, it will affect the electrical activity of the heart, which is reflected in the electrocardiogram modality, showing abnormalities such as prolonged corrected total ventricular repolarization time. Changes in electrolyte metabolites such as potassium in metabolomics occur first, and after a certain period of physiological and pathological process, the electrocardiogram modality changes accordingly.

[0037] According to an embodiment of the present application, the metabolomics modality has a synergistic change relationship with the biochemical detection modality. When crush syndrome causes muscle necrosis, the metabolomics can detect abnormal fluctuations in metabolites such as creatine metabolites and inflammation-related lipids; at the same time, the creatine kinase, lactate dehydrogenase reflecting muscle damage, and blood potassium, serum creatinine reflecting kidney function damage in biochemical detection will show a synchronous increasing trend with the above metabolites. For example, the increase of creatine kinase (biochemical detection modality) is often accompanied by the accumulation of inosine (metabolomics modality), and the two are synergistically enhanced, quantifying the degree of muscle necrosis and secondary organ damage.

[0038] According to an embodiment of the present application, the electrical impedance imaging modality has a synergistic change relationship with the blood pressure modality. When crush syndrome causes muscle tissue ischemic necrosis, local tissue edema and increased vascular permeability, the electrical impedance imaging modality can detect the decrease of electrical impedance value in the damaged area, reflecting the range and degree of muscle damage; at the same time, the inflammatory factors released by muscle necrosis will cause systemic vasodilation and decrease of effective circulating blood volume, resulting in a decreasing trend of blood pressure modality data. The lower the electrical impedance value of the muscle damage area, the greater the decrease in blood pressure, which together reflects the progressive state of "local tissue damage-systemic circulation disorder" caused by crush syndrome.

[0039] The extraction module 120 is configured to extract features from the N groups of to-be-detected data respectively by using feature extractors corresponding to the respective modalities, to obtain N groups of risk features.

[0040] According to an embodiment of the present application, the feature extractors corresponding to each modality data are designed in advance. For example, a convolutional neural network can be used to extract lesion features in image data. Natural language processing can be used to quickly extract keywords or semantic vectors in text data. Time domain or frequency domain analysis can be used to extract fluctuation features in time series data, such as heart rate variability. Through feature extraction, N groups of risk features are obtained, each corresponding to a data modality. For example, image features include tumor size, density, etc. Text features include keyword frequency in symptom description. Physiological features include heart rate variability indicators, etc.

[0041] The generation module 130 is configured to input the N groups of risk features into risk assessment models corresponding to the respective modalities respectively, to generate N groups of risk assessment information.

[0042] According to an embodiment of the present application, the historical detection data of different modalities of the historical object are used as sample data, and the historical risk assessment information of the historical object is used as a label, and the initial assessment models of each modality are optimized to obtain the risk assessment models corresponding to each modality type.

[0043] According to an embodiment of the present application, each modality risk feature is independently risk evaluated by a trained risk evaluation model. The input of each risk evaluation model is a set of to-be-detected data of the modality, and the output is risk evaluation information evaluated by the to-be-detected data of the modality. The risk evaluation information is used to represent the risk level of the to-be-evaluated object having crush syndrome, for example, high risk or low risk.

[0044] Exemplarily, when N = 3, there are 3 groups of risk features (the first group, the second group, and the third group) and 3 risk evaluation models (the first risk evaluation model, the second risk evaluation model, and the third risk evaluation model). The first group of risk features is input into the first risk evaluation model, and the output of the first group of risk evaluation information is high risk. The second group of risk features is input into the second risk evaluation model, and the output of the second group of risk evaluation information is high risk. The third group of risk features is input into the third risk evaluation model, and the output of the third group of risk evaluation information is low risk.

[0045] The first feature fusion module 140 is configured to perform feature fusion on the N groups of risk evaluation information according to the correlation between the N modality types, to obtain first target risk evaluation information, and to generate a processing scheme suitable for the to-be-evaluated object according to the first target risk evaluation information.

[0046] According to an embodiment of the present application, according to different modality data (such as physiological signals, biochemical indicators, and imaging data), and based on the internal correlation between the modalities, a more accurate comprehensive evaluation result is formed by strengthening the weight of key modalities and making up for the limitations of a single modality, and then a processing scheme suitable for the to-be-evaluated object is directly matched, and a multi-dimensional risk evaluation is realized to obtain the first target risk evaluation information.

[0047] Specifically, when the voting method is used for fusion, the voting weights will be adjusted according to the causal, temporal or coordinated change relationship between the modalities. For example, for proteomics modalities (inducers) and metabolomics modalities (results) with causal relationships, the voting weight of the former is higher than that of the latter. If, among the five groups of risk information, three groups of key inducer modalities are judged as "high risk" and two groups of result modalities are judged as "medium risk", then the comprehensive judgment after combining the weight calculation is high risk; when weighted summation fusion is used, the weight distribution directly reflects the correlation relationship, such as for proteomics modalities with temporal relationships (early Early warning) and biochemical detection modality (progression indicator), in the early stage of assessment, the weight of the former (such as 0.4) is higher than the latter (such as 0.3). If the proteomics risk value is 0.8 and the biochemical detection risk value is 0.5, combined with the electrical impedance imaging modality with a synergistic change relationship (weight 0.3, risk value 0.3), the fusion calculation obtains the first target risk assessment information = 0.4×0.8+0.3×0.5+0.3×0.3=0.58. If the threshold value of 0.5 is high risk, the comprehensive judgment of the first target risk assessment information corresponding to the risk level is high risk.

[0048] According to an embodiment of the present invention, the first target risk assessment information is used to represent the probability of the assessed subject experiencing crush syndrome in the target scenario, or alternatively, the probability of acute kidney injury and short-term mortality. For example, blood biochemical indicators (such as creatine kinase and blood potassium) are used to assess the risk level as high; physiological signals (such as heart rate and blood pressure) are used to assess the risk level as high; and imaging data (such as ultrasound showing the extent of muscle injury) are used to assess the risk level as low. Combining these three risk levels may yield a final high risk level, which more comprehensively and accurately reflects the patient's overall risk profile for CS.

[0049] According to an embodiment of the present invention, intervention recommendations are automatically generated based on the final risk level. For example, based on a preset rule library or pre-trained decision-making model, corresponding targeted and actionable treatment measures can be output for different risk levels to guide on-site rescue personnel to quickly carry out treatment. Alternatively, priority can be given to those with high-risk risk assessment information.

[0050] For example, if the first-target risk assessment indicates high risk, the treatment plan may include establishing intravenous access and rapidly rehydrating to promote the excretion of harmful substances such as myoglobin; monitoring blood potassium levels; and promptly lowering potassium levels (e.g., intravenous calcium gluconate, insulin, etc.) if hyperkalemia occurs; and arranging transfer as soon as possible. If the first-target risk assessment indicates low risk, standard trauma management and 24-hour observation are performed; urine output and basic vital signs are monitored; if there are no abnormalities during this period, the patient can be transferred to a general emergency room for further observation.

[0051] According to an embodiment of the present application, N sets of different modal data are collected by the acquisition module to capture risk information related to the target marker from multiple dimensions. For example, in disaster rescue, multi-modal data of a crush victim is acquired simultaneously to avoid missed or incorrect judgments due to one-sided information. Further, since the characteristics of different modal data differ greatly, the extraction module matches a corresponding feature extractor and risk assessment model for each modality, which can more accurately retain the core risk features of each modal data. The first feature fusion module integrates the N sets of risk assessment information into unified first target risk assessment information, avoiding the deviation of single modal assessment. The multi-modal fusion of the correlation relationship not only achieves comprehensive coverage of information, but also improves the time sensitivity, causal interpretability and abnormal recognition robustness of risk assessment by mining the internal relationship between modalities, more accurately captures the risk evolution trend, and reduces the delay risk caused by assessment lag or deviation. Thus, the comprehensiveness and accuracy of risk assessment are improved, which is particularly suitable for scenes such as disaster rescue where resources are limited and time is urgent, can quickly obtain a processing priority according to the first target risk assessment information, reduces the subjectivity and delay of human judgment, and provides a more practical processing scheme for the object to be evaluated.

[0052] According to an embodiment of the present application, the risk assessment model corresponding to each modality type can be any one of the following: support vector machine, logistic regression, random forest, extreme gradient boosting, light gradient boosting, K nearest neighbor classifier, naive Bayes, decision tree, adaptive boosting algorithm (Ada Boost), bagging method (Bagging) and integrated voting method, etc. Before training the initial assessment model using historical detection data, each type of modal data is uniformly standardized to eliminate the influence of different data magnitudes, and missing value filling is performed to avoid incomplete data problems, and then the data is divided into a training set and a test set. During the training process, 5-fold cross-validation is used to ensure the stability of the model, and at the same time, the hyperparameters of the model are systematically adjusted through grid search and Bayesian optimization methods to find the optimal parameter combination for the performance of the model, and finally the most suitable model for this type of modal data is selected from a variety of algorithms.

[0053] According to an embodiment of the present application, the risk assessment model required to be used by each modality can be determined by calculating a first evaluation value and a second evaluation value of the risk assessment model. The first evaluation value is used to represent the distinguishing ability of the risk assessment model for positive and negative samples, that is, the Area Under the Curve (AUC). The AUC is calculated by calculating the area under the Receiver Operating Characteristic Curve (ROC), which reflects the identification ability of the model for positive and negative classes at different thresholds. The AUC value ranges from 0 to 1, and the closer the value is to 1, the more accurately the model can distinguish positive and negative samples (i.e., the evaluation probability of the positive class is higher than that of the negative class). The second evaluation value is used to represent the comprehensive balance ability of the risk assessment model between precision and recall, that is, the F1-score, which represents the comprehensive balance ability of the model between precision and recall. The precision measures the proportion of samples actually positive in the samples evaluated as positive by the model (to avoid false positives), and the recall measures the proportion of samples actually positive that are correctly evaluated by the model (to avoid false negatives). The F1-score is the harmonic mean of the two, and the value ranges from 0 to 1. The closer the value is to 1, the better the model performs in terms of precision and recall, effectively avoiding the limitations of a single indicator.

[0054] According to an embodiment of the present application, the extrusion risk assessment device further comprises a second determination module for determining a plurality of initial markers related to the target scene, and a screening module for screening a plurality of target markers from the plurality of initial markers according to the plurality of initial marker data of the historical object and the historical risk assessment information of the historical object.

[0055] According to an embodiment of the present application, the N groups of to-be-detected data at least include to-be-detected data of a proteomics modality and to-be-detected data of a metabolomics modality, and the modality type includes one of a biochemical detection modality, a proteomics modality, a metabolomics modality, a blood pressure modality, an electrocardiogram modality, and an electrical impedance imaging modality.

[0056] The biochemical detection modality includes 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 volume distribution width (RDW), platelet count (PLT), mean platelet volume (MPV), platelet distribution width (PDW), and platelet hematocrit (PCT) blood routine parameters. Blood gas and electrolyte indicators are also collected, including arterial blood pH, carbon dioxide partial pressure (PCO2), oxygen partial pressure (PO2), and blood oxygen saturation (SaO2). ​​), Lactate, Sodium (Na+) ), Potassium (K+) ), Chloride (Cl-) ); in terms of renal function and metabolic parameters, CREA, UREA, UA are collected; in addition, PT, APTT, TT, FIB are also included as coagulation function indicators; and CK and CK-MB are included as tissue damage and myocardial markers. Among them, 19 target markers are determined from the above initial markers based on XG Boost feature importance ranking. PCA, LDA, etc. can also be used. For scenarios with high modal dimension but limited sample size, RFE, Lasso, mRMR methods, etc. can be used to improve the interpretability and stability of feature selection.

[0057] After the mass spectrum data is collected and identified by data dependency (DDA), the potential proteomic markers are analyzed and screened, and 9 proteins are determined as target markers. The proteomic modalities include: PGAM2, LDHA, MDH1, ALDOA, PKM, GAPDH, PGK, PGM, and CA3.

[0058] Traditional CS risk identification relies on creatine kinase (CK), myoglobin, etc., but these indicators do not significantly increase until 6-12 hours after muscle injury, and are greatly affected by non-CS factors (such as exercise, mild trauma). However, within hours after crush injury, proteomic markers can detect changes in the expression of structural and functional proteins (such as PGAM2, LDHA, etc.) after muscle injury, such as actin and glycolytic enzymes. These changes are usually earlier than the time window for the increase of biochemical indicators such as creatine kinase (CK), and can capture tissue damage signals earlier. Through high-throughput screening of proteomics for such markers, the CS warning window can be advanced, avoiding delays in intervention due to "non-compliance" of traditional indicators. On the other hand, the core danger of CS is acute kidney injury (AKI), but traditional indicators such as creatinine and urea nitrogen do not significantly increase until kidney function declines to a certain extent. Proteomics can detect proteins released by early damage to renal tubular epithelial cells, which can increase within 2 hours after kidney injury and can distinguish between "reversible" and "irreversible" damage.

[0059] Six target markers were identified by screening and identifying differential metabolites and metabolic pathway analysis of potential metabolomics markers. Metabolomics modalities include: galactose metabolism, phenylalanine metabolism, fructose and mannose metabolism, amino sugar and nucleotide sugar metabolism, phenylalanine, tyrosine and tryptophan biosynthesis, ascorbic acid and malonate metabolism.

[0060] Metabolomics markers can distinguish between "false myolysis" and true CS. Traditional indicators (such as CK elevation) can be caused by non-crush factors (such as drug-induced myopathy, infection), while metabolomics can identify "characteristic metabolite combinations": true CS patients will have three metabolic abnormalities of "muscle breakdown" (muscle breakdown), "lactic acid accumulation" (tissue hypoxia), and "uric acid elevation" (kidney excretion disorder), while other myopathy patients usually only have single pathway abnormalities, to improve the accuracy of risk assessment. Metabolomics modalities reflect cellular metabolic network disorders, such as lactic acid accumulation, amino acid metabolism disorders, and electrolyte fluctuations. These metabolic pathway abnormalities can precede traditional electrolyte tests, providing molecular-level warning capabilities for predicting high-risk events such as CS-associated hyperkalemia and acute kidney injury.

[0061] If proteomic data and metabolomic data are missing, the model will rely on physiological signal modalities in the early stage of CS onset, which may not be timely in detecting crush syndrome, resulting in a delayed risk identification window. Therefore, protein and metabolic modalities provide the earliest high-discrimination information source in the time dimension, improving the timeliness of the overall model detection.

[0062] The blood pressure modality includes: systolic blood pressure (SBP), diastolic blood pressure (DBP), mean arterial pressure (MAP), pulse rate (PulseRate), pulse pressure (PP), and blood pressure variability (BPV).

[0063] The electrocardiogram modality includes: heart rate (Heart Rate), time interval between adjacent two ventricular wave peaks (RR interval), ventricular depolarization wave group duration (QRS wave group duration), time interval from ventricular depolarization start to repolarization end (QT interval), corrected total ventricular repolarization time (QTc interval), ventricular repolarization initial potential change (ST segment change), ventricular repolarization terminal wave form (T wave form), and two heart rate variability indicators, standard deviation (SDNN) and root mean square difference (RMSSD).

[0064] The electrical impedance imaging modality includes: global conductivity change amount (AGCV), conductivity change range (ACRV), conductivity change standard deviation (ASD.CV), and maximum conductivity change amount (AMCV) and other basic statistical parameters; and includes regional weighted mean conductivity (RWMC), conductivity change time slope (CCS), multi-frequency electrical impedance response characteristics (FRS), local maximum area ratio (MCAR), conductivity distribution center shift (Centroid Shift), symmetry index (SymmetryIndex), and nonlinear fit residuals (Nonlinear Fit Residuals) and other dynamic and spatial structure characteristics. The data characteristics of this modality can be used to identify the compression degree of the tissue, the damage area diffusion trend, and the potential necrosis risk.

[0065] According to the embodiments of the present application, by integrating the six types of modality data of biochemical detection modality, proteomic modality, metabolomic modality, blood pressure modality, electrocardiogram modality, and electrical impedance imaging modality, the key risk factors are comprehensively covered, the problem of insufficient coverage of traditional single modality information is overcome, and the breadth of pathological mechanism identification is improved. At the same time, in order to cope with the large difference between different modalities and the difficulty of unified modeling, each modality is independently trained, the optimal model is selected through cross-validation and multi-index evaluation, and the modeling accuracy and generalization ability of each modality are significantly improved.

[0066] According to an embodiment of the present application, in terms of feature selection, the above-mentioned six types of modalities can respectively adopt a modeling feature screening strategy with strong adaptability, such as feature importance sorting based on XG Boost, cluster analysis, pathway enrichment, etc. to extract a core variable set highly related to crush syndrome from different physiological systems. On this basis, to further improve the adaptability and scalability of the model, strategies such as RFE, Lasso, mRMR, etc. can be used to optimize feature subset selection. In modalities with high dimension or limited samples, Lasso can remove redundant features through a penalty mechanism to enhance the sparsity and generalization ability of the model; RFE is suitable for step-by-step optimization of feature sets in linear or tree model frameworks to improve stability; mRMR can select variables in an unsupervised scenario based on mutual information, especially suitable for omics modalities such as metabolism and protein. In addition, for scenarios that require improved model interpretability and deployment performance, feature masking methods based on perturbation impact (such as Permutation Importance) can be used, or a joint feature scoring mechanism can be introduced in multi-modal integrated modeling to filter out the globally optimal feature subset across modalities through scoring weighting between different modalities.

[0067] According to an embodiment of the present application, the crush risk assessment device 100 further comprises a first determination module and a filling module. The first determination module is configured to identify missing values in the N groups of to-be-detected data and determine the modal type of the missing data. The filling module is configured to fill the missing data according to the modal type of the missing data.

[0068] According to an embodiment of the present application, all N groups of to-be-detected data are traversed to calculate the missing rate (number of missing values / total data amount) of each modality. The specific modal type of the missing value is marked. According to the characteristics (such as continuity, time sequence, distribution characteristics) of different modal data, the most suitable filling method is selected to avoid data distortion caused by a universal method. For example, the QTc interval of an electrocardiogram signal is missing, which can be linearly interpolated by the data of the previous 5 and the next 5 time points.

[0069] According to an embodiment of the present application, the filling module comprises a first filling submodule, a second filling submodule and a third filling submodule. The first filling submodule is configured to fill the missing data according to the historical detection data of a historical object similar to the to-be-evaluated object when the modal type of the missing data is at least one of a biochemical detection modality, a blood pressure modality, an electrocardiogram modality and an electrical impedance imaging modality. For example, a K-neighbor interpolation method is used. The second filling submodule is configured to fill the missing data using a mean interpolation method when the missing data is a proteomics modality. The third filling submodule is configured to fill the missing data using a multiple interpolation method when the missing data is a metabolomics modality.

[0070] According to an embodiment of the present application, in data missing value processing, different filling strategies are used for different data types, time sequence structures and collection scenarios to improve modeling effectiveness and data consistency. Considering the complex situations such as data loss and incomplete modalities in disaster scenes, various missing value processing strategies can be used to enhance the robustness of the model. For example, a collaborative filtering method can be introduced to infer missing values based on the feature similarity between historical patients, which is suitable for structured data modalities (such as biochemical data, blood pressure, etc.); for omics modalities, a self-encoder or other nonlinear reconstruction technology can be used to recover and reconstruct high-dimensional features, preserving the internal potential expression relationship of omics data; for continuously collected time sequence signals such as electrocardiogram and blood pressure, a sliding window modeling prediction model can be used for segment completion, such as a K-nearest neighbor algorithm regression model based on time sequence. In addition, for the extreme case of missing key modalities, a multi-modal conditional generation model can be constructed based on the collected modalities to assist in predicting the feature expression of the missing modalities, thereby maintaining the end-to-end inference ability of the system.

[0071] According to an embodiment of the present application, by identifying missing modalities and matching filling strategies, the reduction of sample size caused by directly deleting samples with missing values is avoided. Thus, data information loss is reduced and the integrity of multi-modal data is preserved. Different modalities have significant differences in characteristics, and a general filling method may destroy the inherent rules of the data. However, the sub-modality processing can maintain the dynamic trend of the time sequence modality using interpolation method, avoiding the loss of time sequence characteristics caused by static value filling. For high-dimensional modalities, the matrix decomposition method is used to preserve the correlation between variables. This can reduce the risk of data distortion and improve 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 application is shown.

[0073] According to an embodiment of the present application, the first feature fusion module 140 includes a calculation sub-module 210, an adjustment sub-module 220 and a first obtaining sub-module 230.

[0074] The calculation sub-module 210 is configured to calculate an initial weight value of any one of the N groups of risk assessment information, to obtain the initial weight value of each of the N groups of risk assessment information. The sum of the initial weight values of the N groups of risk assessment information is equal to 1. For example, if N is equal to 3, the initial weight values of the three groups of risk assessment information are 0.4, 0.3 and 0.3 respectively.

[0075] The adjusting sub-module 220 is configured to adjust the N initial weight values according to the correlation among the N modal types to obtain N target weight values, wherein for the first modal type and the second modal type having 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 having a time sequence 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 modal type and the sixth modal type having a synergistic 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] According to the above example, if the modal corresponding to the risk assessment information of the three groups is proteomics, metabolomics, and biochemical detection, respectively, the initial weight value of the risk assessment information corresponding to the proteomics is adjusted to be the maximum, the initial weight value of the risk assessment information corresponding to the metabolomics is the second, and the initial weight value of the risk assessment information corresponding to the biochemical detection is the minimum, for example, 0.45, 0.29, and 0.26.

[0077] The first obtaining sub-module 230 is configured to weight and sum the N groups of risk assessment information according to the N target weight values to obtain first target risk assessment information, and the reference formula (1) is as follows.

[0078] Formula (1)

[0079] Wherein, is the first target risk assessment information; is the target weight value of the i th risk assessment information; 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, according to formula (1), the first target risk assessment information is calculated, that is, 0.45*0.7+0.29*0.6+0.26*0.3=0.567.

[0081] According to the embodiment of the application, the first target risk assessment information can be used to judge the risk level by using a fixed threshold value: for example, a binary classification result, when ≥0.5 is determined as “high risk”, otherwise as “low risk”. It can also be a multi-threshold value grading judgment, for example, when ≥0.8 is the first risk level, when is the second risk level, and when For the third risk level, when <0.2 is the fourth risk level.

[0082] Figure 3 A structural block diagram of a calculation sub-module according to an embodiment of the present application is shown. Figure 4 A data flow diagram of determining a weight value of risk assessment information according to an embodiment of the present application is shown. The following will be described in detail in combination with Figure 3 and Figure 4 .

[0083] According to an embodiment of the present application, the calculation sub-module 210 comprises a first calculation unit 310, a second calculation unit 320 and a determination unit 330.

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

[0085] According to an embodiment of the present application, as Figure 4As shown, two independent performance indicators are calculated for each modality risk assessment model 401-1, 401-2, …, 401-N, respectively measuring the model's discrimination ability and precision-recall balance ability, ensuring that the reliability of the model can be comprehensively evaluated when subsequently weighted and fused. For example, the risk assessment model 401-1 can output a first evaluation value 402-1 and a second evaluation value 402-2; the risk assessment model 401-2 can output a first evaluation value 402-3 and a second evaluation value 402-4; and the risk assessment model 401-N can output a first evaluation value 402-(2N-1) and a second evaluation value 402-2N. For example, if the first evaluation value 402-1 of the risk assessment model 401-1 is AUC = 0.85, it means that when a positive sample and a negative sample are randomly selected, the risk assessment model 401-1 has an 85% probability of placing the positive sample before the negative sample. The evaluation model balances the ability between recall rate and precision rate, especially suitable for class imbalance scenarios such as critical patients in disaster rescue. If the second evaluation value 402-2 of the risk assessment model 401-1 is F1-score = 0.747, it means that the risk assessment model 401-1 has achieved a good balance between the "proportion of patients predicted as critical who are actually critical" (precision rate) and the "proportion of patients actually critical who are correctly predicted as critical" (recall rate). It neither overemphasizes "few missed diagnoses" and leads to a large number of misdiagnoses (high recall rate but low precision rate), nor overemphasizes "few misdiagnoses" and misses real critical patients (high precision rate but low recall rate), and in disaster rescue scenarios that require both identification accuracy and comprehensiveness, it can better balance the needs of "not misdiagnosis" and "not missed diagnosis".

[0086] The second calculation unit 320 is configured to calculate a target evaluation value of the risk assessment model according to the first evaluation value and the second evaluation value of the risk assessment model.

[0087] According to an embodiment of the present application, the target evaluation value can be a weighted combination of the first evaluation value (AUC) and the second evaluation value (F1-score), and the respective weights can be set according to business needs, for example, each accounting for 50%. For example: target evaluation value = 0.5 x AUC + 0.5 x F1-score. By integrating the two indicators, the limitations of a single indicator are avoided, for example, a model with high AUC but low F1-score may not perform well in actual application, thereby more comprehensively reflecting the accuracy of the model. 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 x 0.80 + 0.5 x 0.78 = 0.79.

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

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

[0090] For example, the biochemical detection modality uses XG Boost, which takes into account both model performance and feature interpretability; the proteomics modality selects a random forest model, which is suitable for its high-dimensional and small-sample characteristics; the metabolomics modality uses a light gradient boosting machine (Light GBM) to achieve efficient modeling and fast reasoning; the blood pressure modality selects a logistic regression model due to its small feature dimension and high deployment requirements; the electrocardiogram modality uses categorical boosting (Cat Boost) after time-domain feature extraction to improve the ability to distinguish between classes; and the electrical impedance imaging modality uses XG Boost after expressing statistical features to take into account accuracy and robustness. As shown in Table 1, each modality model shows good prediction performance.

[0091] The models selected by each modality described above are only exemplary and are not limited thereto. For example, the proteomics modality can use a graph neural network structure, the electrocardiogram modality can use a sequence model such as a long short-term memory (LSTM) if the collection accuracy is high and the complete electrocardiogram waveform sequence is included, and the electrical impedance imaging modality can use a model other than gradient boosting based on image feature vectors if the image features are obvious, such as a support vector machine.

[0092] Table 1. Correspondence table of evaluation values and weight values of different modal types

[0093]

[0094] The biochemical detection modality has an AUC of 0.91 and an F1-score of 0.88, and has excellent overall performance. The proteomics modality has an AUC of 0.89 and an F1-score of 0.85, and the model still maintains a stable performance in a high-dimensional feature environment. The metabolomics modality has an AUC of 0.86 and an F1-score of 0.83, and takes into account the model accuracy and computational efficiency. The blood pressure modality has an AUC of 0.84 and an F1-score of 0.80. The electrocardiogram modality has an AUC of 0.81 and an F1-score of 0.77, and can better identify physiological fluctuations such as heart rate variability risks. The electrical impedance imaging modality has an AUC of 0.78 and an F1-score of 0.75. The product of the first evaluation value and the second evaluation value of the risk assessment model is calculated to obtain the target evaluation value of the risk assessment model. The weight value of any modality is the proportion of the target evaluation value in the cumulative evaluation value. For example, in Table 1, the cumulative evaluation values of the biochemical detection modality, the proteomics modality, the metabolomics modality, the blood pressure modality, the electrocardiogram modality, and the electrical impedance imaging modality are 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 value of the biochemical detection modality is 0.8008 / 4.1518=0.19. The proportions of other modalities are calculated in the same way, and are not described here.

[0095] According to an embodiment of the present application, the extrusion risk assessment device further comprises a second feature fusion module for performing feature fusion on the X sets of risk assessment information to obtain updated second target risk assessment information if any of the N sets of risk assessment information does not meet the predetermined condition.

[0096] According to an embodiment of the present application, in a multi-modal risk assessment scenario, considering the actual situation that the disaster site environment is complex, the equipment stability is insufficient, and part of the modal data may not be obtained in time. When a certain group of risk assessment information does not meet the predetermined standard, such as data missing, collection failure or model failure, the system will automatically exclude this group of information, and only the remaining X groups of valid information are subjected to feature fusion. The predetermined standard is a pre-set standard for judging whether a certain group of risk assessment information is valid, to ensure the reliability of the data input into the fusion model or the model output result. The predetermined standard can include the following cases. First, the data integrity standard: it is stipulated that the missing rate of a certain group of data should not exceed a threshold value, such as the single modal data missing rate ≤ 5%, or the key indicators must be complete, such as creatine kinase and blood potassium value in crush syndrome assessment. Second, the data quality standard: including the accuracy and rationality range of the data. For example, the normal reference range of blood potassium concentration in biochemical detection modal data is 3.5-5.5 mmol / L, if the collected data exceeds the extreme threshold value (such as > 8 mmol / L or < 2 mmol / L) and has no reasonable clinical explanation, it may be judged as collection error, which does not meet the standard; or the image data has serious noise and artifacts, which makes it impossible to extract effective features, and it will also be excluded. Third, the collection process standard: if the data collection process does not meet the pre-set specification, such as the sample collection time exceeding the window period, the operation process being wrong, etc., even if the data itself has no obvious abnormality, it may also be judged as invalid.

[0097] This dynamic adjustment mechanism generates updated second target risk assessment information by recalculating the weights and weighted summation, ensuring that the final result is not disturbed by low-quality data, and improving the reliability and robustness of the evaluation. For example, in disaster rescue, if the blood potassium data of a certain rapid detection device is unreliable due to operation error, the system will preferentially rely on the data of other modalities (such as electrocardiogram, image) to complete the comprehensive judgment, avoiding the influence of false information on the evaluation result.

[0098] According to an embodiment of the present application, the second feature fusion module includes an updating submodule and a second obtaining submodule.

[0099] The updating submodule is configured to update the respective weight values of the X groups of risk assessment information to obtain updated X weight values in the case that any one group of risk assessment information does not meet the predetermined condition; and the second obtaining submodule is configured 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 application, if a certain modality cannot output the prediction probability due to data missing, collection failure or model failure , the system will automatically identify this modality as an invalid modality and exclude it from the fusion process. At this time, the remaining X valid modalities constitute a set To keep the sum of fusion weights as 1, the system proportionally normalizes the weights of the remaining valid modalities. The updated weight values are calculated as , referring to formula (2):

[0101] Formula (2)

[0102] wherein, is the target weight value of the jth risk assessment information.

[0103] Further, the fusion model only calculates the weighted probability for the valid modalities. The second target risk assessment information is calculated according to formula (3):

[0104] Formula (3)

[0105] When the fusion probability satisfies the following condition, the system outputs the second target risk assessment information: ≥ 0.5 is determined as “high risk”, otherwise as “low risk”.

[0106] wherein, the evaluation results of different modalities can be fused in a “soft voting weighting” manner. Specifically, fixed weights are pre-allocated according to the overall performance (such as accuracy, reliability, etc.) of each modality model, and the weights are adjusted to keep the sum as 1, so that the prediction result can be stably output even if some modalities are missing. In order to make this fusion method more adaptable to different situations, “stacked fusion, hybrid fusion” technology in “ensemble learning” can be used to automatically learn the nonlinear relationship between the outputs of different modalities in a higher level model, so that the expression ability of the fusion model is stronger; or a “fusion strategy based on attention mechanism” can be used to dynamically adjust the importance weights of each modality according to the specific situation of each wounded person, so as to adapt to the differences in the role of each modality in different samples.

[0107] According to the embodiments of the present application, the fault-tolerant mechanism supports the case of missing any number of modalities, and the system can automatically identify valid modalities and dynamically adjust the fusion strategy to ensure that the prediction function does not interrupt. The mechanism does not depend on a fixed upper limit of the number of modalities, and theoretically supports prediction calculation under the premise that at least one valid modality exists. In the verification experiment of simulating modality missing, the maximum decrease of AUC is not more than 2%, the maximum decrease of F1-score is not more than 1.5%, and the overall performance is still better than any single modality model, showing good robustness and actual deployment adaptability.

[0108] According to the embodiment of the present application, in the aspect of adapting to the incomplete scene data collection of disasters, the present application designs a soft voting fusion strategy based on performance weighting, introduces a modal effectiveness identification mechanism and weight normalization logic, and even when a single mode or multiple modes are missing, the system can automatically adjust the fusion strategy to ensure uninterrupted prediction. The experimental results show that the AUC and F1-score indicators of the fusion model decrease by no more than 2%, improving the robustness. In addition, it can not only output the overall risk probability of crush syndrome, but also support individualized assessment of complications such as acute kidney injury and death risk, providing support for medical personnel.

[0109] The crush risk assessment method of the present application can also be used to simultaneously evaluate the crush syndrome risk probability, acute kidney injury risk probability and short-term death risk probability, that is, in the training phase of the model, the historical detection data of the historical object are taken as sample data, and the crush syndrome evaluation information of the historical object is taken as a label to train the crush syndrome risk assessment model; the same historical detection data of the same historical object are taken as sample data, and the acute kidney injury risk assessment information of the historical object is taken as a label to train the acute kidney injury risk assessment model; the same historical detection data of the same historical object are taken as sample data, and the short-term death risk assessment information of the historical object is taken as a label to train the short-term death risk assessment model. When actually needed to evaluate the risk of the to-be-evaluated object, the to-be-detected data of the to-be-evaluated object 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 to-be-evaluated object having crush syndrome, the risk probability of the to-be-evaluated object having acute kidney injury and the risk probability of the to-be-evaluated object having short-term death respectively.

[0110] The output of the first target risk assessment information is structured data, which can be integrated into an electronic medical record system, a disaster scene mobile terminal or a remote command platform through a standard format. Moreover, it is shown which modalities and models are used to complete the evaluation work.

[0111] It should be noted that the first target risk assessment information or the second target risk assessment information output by the above model is only used as reference information for medical personnel when rescuing on-site wounded personnel. In order to obtain processing information parameters for preferential treatment of high-risk wounded personnel, it is not used as a direct diagnosis result. Moreover, the entire process of the information processing method of all steps of the above method does not require the participation of medical personnel, and the direct purpose is not to obtain a diagnosis result or a health condition. The obtained information itself cannot directly obtain a diagnosis result or a health condition of a disease.

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

[0113] As Figure 5As shown, the extrusion risk assessment method of the embodiment includes steps S510-S540.

[0114] In step S510, N groups of to-be-detected data of at least one to-be-evaluated object in a target scene are acquired, wherein the N groups of to-be-detected data are related to a target marker, N is a positive integer greater than 2, and the N groups of to-be-detected data are data of different modal types.

[0115] In step S520, feature extractors corresponding to the modal types are used to respectively perform feature extraction on the N groups of to-be-detected data, to obtain N groups of risk features.

[0116] In step S530, the N groups of risk features are respectively input into risk assessment models corresponding to the modal types, to generate N groups of risk assessment information.

[0117] In step S540, the N groups of risk assessment information are fused to obtain first target risk assessment information, so as to generate a processing scheme suitable for the to-be-evaluated object according to the first target risk assessment information.

[0118] According to the embodiment of the present application, N groups of different modal data are collected to capture risk information related to the target marker from multiple dimensions. For example, in disaster rescue, multi-modal data of a crush victim are simultaneously acquired to avoid missed or incorrect judgments due to one-sided information. Further, since the characteristics of different modal data differ greatly, a corresponding feature extractor and risk assessment model are matched for each modal to more accurately retain the core risk features of each modal data. The N groups of risk assessment information are integrated into unified first target risk assessment information through feature fusion technology to avoid the deviation of single modal assessment. The multi-modal fusion of the correlation relationship not only achieves comprehensive coverage of information, but also improves the time-sensitive, causally interpretable, and robustness of abnormal recognition of risk assessment by mining the internal relationship between modalities, more accurately captures the risk evolution trend, and reduces the delay risk caused by assessment lag or deviation. Therefore, the comprehensiveness and accuracy of risk assessment are improved, which is especially suitable for scenes such as disaster rescue with limited resources and time constraints, can quickly obtain a processing priority according to the first target risk assessment information, reduces the subjectivity and delay of human judgment, and provides a more practical processing scheme for the to-be-evaluated object.

[0119] The above-described specific embodiments further illustrate the purpose, technical solutions, and advantages of the present application. It should be understood that the above-described specific embodiments are merely examples of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A device for assessing extrusion risk, characterized in that: include: an acquisition module for acquiring N sets of to-be-detected data for at least one subject to be evaluated in a target scenario, wherein the N sets of to-be-detected data correspond to target markers of N modal types, 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 association relationships exists between the N modal types: a causal relationship characterizing that an abnormal marker of a first modal type is a cause of an abnormal marker of a second modal type; a temporal sequence relationship characterizing that an abnormal marker of a third modal type occurs earlier than an abnormal marker of a fourth modal type; and a synergistic change relationship characterizing that a marker of a fifth modal type and a marker of a sixth modal type change synchronously with a preset trend; An extraction module, configured to perform feature extraction on the N groups of to-be-detected data using feature extractors corresponding to the N modal types, respectively, to obtain N groups of risk features; a generating module, configured to input the N groups of risk features into risk assessment models corresponding to the N modality types, respectively, to generate N groups of risk assessment information; The first feature fusion module is used to perform feature fusion on the N groups of risk assessment information based on the association relationship between the N modal types to obtain first target risk assessment information, so as to generate a processing solution applicable to the at least one object to be assessed based on the first target risk assessment information.

2. The device according to claim 1, characterized in that The N groups of data to be detected include at least data to be detected in a proteomics modality, data to be detected in a metabolomics modality, data to be detected in a biochemical detection modality, data to be detected in a blood pressure modality, data to be detected in an electrocardiogram modality, and data to be detected in an electrical impedance imaging modality. The proteomics data acquisition unit is used to acquire the data to be detected of the proteomics modality of the object to be evaluated, wherein the target markers of the proteomics modality include at least: phosphoglycerate mutase 2, lactate dehydrogenase A, malate dehydrogenase 1, fructose-1,6-bisphosphate aldolase, pyruvate kinase M type, glyceraldehyde-3-phosphate dehydrogenase, phosphoglycerate kinase, phosphoglucomutase and carbonic anhydrase 3; The metabolomics data acquisition unit is used to acquire the data to be detected of the metabolomics modality of the subject to be evaluated, wherein the target marker of the metabolomics modality includes 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 malonate metabolism data; The biochemical test data acquisition unit is used to acquire the test data of the biochemical test modality of the subject to be evaluated, and the target markers of the biochemical test 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, corpuscular volume distribution width, platelet count, mean platelet volume, platelet distribution width, platelet hematocrit, arterial blood pH, carbon dioxide partial pressure, oxygen partial pressure, 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 blood pressure modality data to be detected of the subject to be evaluated, wherein the target markers of the blood pressure modality include at least: systolic pressure, diastolic pressure, mean arterial pressure, pulse rate, pulsating pressure and blood pressure variability; The electrocardiogram data acquisition unit is used to acquire the test data of the electrocardiogram modality of the subject to be evaluated, wherein the target markers of the electrocardiogram modality include at least: heart rate, the time interval between two adjacent ventricular peaks, the duration of the ventricular depolarization wave group, the time interval from the start of ventricular depolarization to the end of repolarization, the corrected total ventricular repolarization time, the change of the potential at the early stage of ventricular repolarization, and the waveform characteristics of the late stage of ventricular repolarization; The electrical impedance imaging data acquisition unit is used to acquire the test data of the electrical impedance imaging modality of the object to be evaluated, and the target markers of the electrical impedance imaging modality include at least: global conductivity change, conductivity change range, conductivity change standard deviation and conductivity change, regional weighted average conductivity, time slope of conductivity change, multi-frequency electrical impedance response characteristics, local area area ratio, conductivity distribution center offset, and symmetry index.

3. The device according to claim 1, characterized in that The first feature fusion module includes: a calculation submodule, configured to calculate an initial weight value of any one of the N groups of risk assessment information, to obtain an initial weight value of each of the N groups of risk assessment information; an adjustment submodule, configured to adjust the N initial weight values ​​according to the association relationship between the N modal types to obtain N target weight values, wherein, for the first modal type and the second modal type having the 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 having the 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 modal type and the sixth modal type having the collaborative 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 configured to perform weighted summation on the N groups of risk assessment information according to the N target weight values ​​to obtain the first target risk assessment information.

4. The device according to claim 3, characterized in that The calculation submodule includes: a first calculation unit, configured to calculate, for each risk assessment model used for the risk assessment information, a first evaluation value and a second evaluation value of the risk assessment model, 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 strike a balance between precision and recall; a second calculation unit, configured to calculate a target evaluation value of the risk assessment model according to the first evaluation value and the second evaluation value of the risk assessment model; A determination unit is used to determine the weight value of the risk assessment model according to the proportion of the target assessment value of the risk assessment model to the cumulative assessment value, wherein the cumulative assessment value is determined according to the target assessment value of each of N risk assessment models.

5. The device according to claim 1, characterized in that The device further comprises: The second feature fusion module is configured to perform feature fusion on X groups of risk assessment information to obtain updated second target risk assessment information if any group of risk assessment information among the N groups of risk assessment information does not meet a predetermined condition, where X<N and X is a positive integer greater than or equal to 0.

6. The device according to claim 5, characterized in that The second feature fusion module includes: an updating submodule, configured to update the respective weight values ​​of the X groups of risk assessment information if any of the groups of risk assessment information does not meet a predetermined condition, to obtain X updated weight values; The second obtaining submodule is configured 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.

7. The device according to claim 1, characterized in that The device further comprises: A first determination module is used to identify missing values ​​for the N groups 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 device according to claim 7, characterized in that The filling module includes: a first filling submodule, configured to fill in the missing data based on historical test data of a historical subject, when the modality type of the missing data is at least one of a biochemical test modality, a blood pressure modality, an electrocardiogram modality, and an electrical impedance imaging modality, wherein the historical subject is similar to the subject to be evaluated; A second filling submodule is used to fill the missing data using a mean interpolation method when the missing data is a proteomics modality; The third filling submodule is used to fill the missing data using a multiple interpolation method when the missing data is a metabolomics modality.

9. The device according to claim 1, characterized in that The target scenario is an emergency scenario, and the extrusion risk assessment device further includes: a second determining module, configured to determine a plurality of initial markers associated with the target scene; The screening module is used to screen the multiple target markers from the multiple initial markers according to the multiple initial marker data of the historical objects and the historical risk assessment information of the historical objects.

10. A method for assessing extrusion risk, characterized in that: The method comprises: Obtaining N sets of data to be tested for at least one object to be evaluated in a target scenario, wherein the N sets of data to be tested correspond to target markers of N modal types, and N is a positive integer greater than 2, the N sets of data to be tested 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, and at least one of the following association relationships exists between the N modal types: a causal relationship characterizing that an abnormality in a marker of a first modal type is a cause of an abnormality in a marker of a second modal type, a temporal sequence relationship characterizing that an abnormality in a marker of a third modal type occurs earlier than an abnormality in a marker of a fourth modal type, and a synergistic change relationship characterizing that a marker of a fifth modal type and a marker of a sixth modal type change synchronously with a preset trend; Using feature extractors corresponding to the N modal types, respectively extract features from the N groups of data to be detected to obtain N groups of risk features; Inputting the N groups of risk features into risk assessment models corresponding to the N modality types, respectively, to generate N groups of risk assessment information; According to the association relationship between the N modality types, feature fusion is performed on the N groups of risk assessment information to obtain first target risk assessment information, so as to generate a treatment plan applicable to the at least one object to be assessed based on the first target risk assessment information.

Images