A geriatric nsc emergency risk stratification method, apparatus, and medium

Abstract

The application relates to the field of clinical diagnostics, and discloses an old NSC emergency risk stratification method, equipment and medium, the method comprising the following steps: S1: obtaining multi-dimensional evaluation data of an old NSC patient; S2: converting the multi-dimensional evaluation data into single risk values of a unified scale by using a preset risk mapping rule; S3: dividing the single risk values into multiple risk sub-models, and calculating the dimension risk scores of each risk sub-model; S4: integrating the risk sub-models and the interaction effects of the single risk values based on a preset fusion strategy, and calculating the total risk score of the patient; and S5: determining the risk level of the patient according to the total risk score, and outputting a clinical diagnosis and treatment suggestion. The application can more comprehensively reflect the potential pathological state of the patient by integrating various biomarkers and key clinical parameters, and constructing a multi-dimensional joint risk evaluation model.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of clinical diagnostics, and in particular to an elderly NSC emergency risk stratification method, device and medium. BACKGROUND

[0002] The elderly patient group with non-specific complaints (NSC) appearing in the emergency department is well known but not well defined (simple complaints of feeling unwell, dizziness, etc.), and this type of patient often has difficulty in diagnosis, which is a great challenge to emergency physicians. Studies have shown that up to 20% of elderly patients in the emergency department have non-specific complaints, and about 50% of them are eventually found to have acute medical problems. In addition, elderly patients themselves belong to a high-risk group that is more prone to functional decline, dependence, and death, and other adverse outcomes.

[0003] In the emergency environment, a variety of biomarkers have the potential for risk assessment. suPAR (a broad-spectrum inflammatory activity indicator) is a non-specific biomarker that reflects the systemic inflammatory state and disease prognosis, and has been shown to be an important indicator for risk stratification of emergency patients. Studies have shown that the introduction of suPAR can improve the performance of predicting short-term mortality in emergency patients; in addition, MR-proADM (midregional proadrenomedullin) and D-dimer also show excellent effects in predicting the risk of death in elderly emergency patients, and the odds ratio (OR) of 30-day mortality is as high as 11.12 (about 7.65 after multivariate correction) when the level of MR-proADM increases in univariate logistic regression analysis. The corresponding area under the curve (AUC) is about 0.86; and the AUC of D-dimer at the standard cutoff value is also more than 0.85. These results suggest that a multi-marker model can significantly improve the prediction accuracy.

[0004] In summary, there is still a lack of comprehensive solutions for risk assessment of elderly NSC patients, and the marker selection and algorithm design of existing methods need to be optimized and innovated. SUMMARY

[0005] Therefore, the present application provides an elderly NSC emergency risk stratification method, device and medium to solve the above problems.

[0006] To solve the above technical problems, the present application provides an elderly NSC emergency risk stratification method, comprising:

[0007] S1: obtaining multi-dimensional evaluation data of elderly NSC patients, and standardizing the multi-dimensional evaluation data;

[0008] S2: For each item of evaluation data after standardization, a preset risk mapping rule is used to convert it into a single risk value of a unified scale; the risk mapping rule includes setting according to the pathological significance of each item of evaluation data and the physiological characteristics of the elderly NSC patients;

[0009] S3: The single risk value is divided into multiple risk sub-models, and the dimension risk score of each risk sub-model is calculated;

[0010] S4: Based on the preset fusion strategy, the interaction effect between the dimension risk scores of multiple risk sub-models and the single risk value is integrated, and the total risk score of the patient is calculated;

[0011] S5: Determine the risk level of the patient according to the total risk score, and output the corresponding clinical diagnosis and treatment path suggestion based on the risk level.

[0012] As an optional way, the multi-dimensional evaluation data at least includes basic clinical information, laboratory biomarker indicators and vital sign indicators;

[0013] The laboratory biomarker indicators include at least one of soluble urokinase-type plasminogen activator receptor, adrenal medulla precursor middle segment, D-dimer, vasopressin C-terminal precursor and peroxide reductase-4.

[0014] As an optional way, the standardization process uses Z-score algorithm to realize standardization by eliminating the differences in measurement units and dimensions of different evaluation data, which includes:

[0015] Based on the mean and standard deviation of each item of evaluation data in historical training data, the original value of the current patient's evaluation data is converted and unified.

[0016] As an optional way, the preset risk mapping rule includes a nonlinear mapping function, and adopts a dedicated function form for different evaluation data pathological mechanisms;

[0017] Among them, piecewise quadratic function is used for indicators reflecting broad-spectrum inflammatory activity, Sigmoid function is used for indicators reflecting organ perfusion state, power function is used for indicators reflecting the degree of functional decline, and nonlinear function based on normal interval is used for indicators reflecting heart rate.

[0018] As an optional way, the risk mapping rule also includes an age dynamic correction threshold, which is used to distinguish the normal and abnormal risk intervals of D-dimer.

[0019] As an optional mode, the uniform scale single-item risk value output range is provided with a preset interval, and a value in the preset interval is used to realize risk probability expression of the single-item evaluation data, and the key parameters of each risk mapping rule are automatically optimized based on statistical distribution of historical training data.

[0020] As an optional mode, the plurality of risk sub-models include sub-models of five pathological mechanism dimensions, which are respectively: an inflammation-immunity sub-model, a coagulation-circulation sub-model, an organ function sub-model, a clinical perception sub-model and a basic risk adjustment sub-model.

[0021] The inflammation-immunity sub-model is used to integrate single-item risk values related to inflammation and immunity; the coagulation-circulation sub-model is used to integrate single-item risk values related to coagulation and cardiovascular system; the organ function sub-model is used to integrate single-item risk values related to organ perfusion and respiratory function; the clinical perception sub-model is used to integrate single-item risk values related to functional decline degree and doctor's preliminary impression; and the basic risk adjustment sub-model is constructed based on patient age.

[0022] As an optional mode, the dimension risk score of each risk sub-model is a weighted linear combination of all single-item risk values under the sub-model.

[0023] The weight of each single-item risk value is calculated based on historical training data, and the sum of the weights of all single-item risk values in each sub-model is 1.

[0024] As an optional mode, the preset fusion strategy includes:

[0025] The first fusion layer is set as global weighting of the sub-models, and the dimension risk scores of the five risk sub-models are linearly aggregated through global weights;

[0026] The second fusion layer is set as an interaction term penalty, and an interaction weight is used to correct pairs of single-item risk values that have a synergistic deterioration effect, the pairs of single-item risk values that have the synergistic deterioration effect including one or more of suPAR and CRP, MR-proADM and respiratory rate, and blood oxygen saturation and systolic pressure.

[0027] As an optional mode, the risk levels include three levels of low risk, medium risk and high risk, each level is provided with a different total risk score interval, and the threshold of the score interval is dynamically updated based on hospital local historical data, and the updating mode is to determine the optimal threshold by calculating the Youden index through ROC curve analysis every quarter.

[0028] On the other hand, the application also provides an electronic device, which comprises a memory for storing a computer program and a processor for executing the computer program to realize the steps of the method for risk stratification of elderly NSC emergency.

[0029] In another aspect, the application also provides a storage medium, wherein the storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the method for emergency risk stratification of elderly NSC.

[0030] The application has the following beneficial effects:

[0031] The application can more comprehensively reflect the potential pathological state of a patient by integrating multiple biomarkers and key clinical parameters to construct a multi-dimensional joint risk assessment model. Compared with traditional methods that rely on a single indicator or scoring system, the model has stronger discrimination ability in identifying the potential severe risk of elderly patients with non-specific complaints, especially in the early stage of atypical clinical manifestations. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 is a flowchart of the method for emergency risk stratification of elderly NSC. DETAILED DESCRIPTION

[0033] To make the objectives, technical solutions, and advantages of the embodiments of the application clearer, the various embodiments of the application will be described in detail below with reference to the accompanying drawings. However, those skilled in the art can understand that, in the various embodiments of the application, many technical details are presented to make the readers better understand the application. However, the technical solutions claimed by the application can be implemented even without these technical details and various changes and modifications based on the following embodiments.

[0034] As Figure 1 shown, the present detailed description is used to describe the technical implementation process of the method and system for emergency risk stratification of elderly NSC based on multi-parameter biomarker fusion, which is suitable for emergency patients aged 65 and above with non-specific complaints. Such patients often have symptoms such as feeling unwell and dizziness without clear direction. The present embodiment can realize accurate prediction and stratified management of short-term serious adverse outcomes of patients through multi-dimensional data acquisition, standardized processing, risk mapping, sub-model construction, total risk fusion, and diagnosis and treatment path linkage. The present embodiment is realized as follows:

[0035] S1: obtaining multi-dimensional evaluation data of elderly NSC patients and performing standardized processing on the multi-dimensional evaluation data;

[0036] S2: converting each item of evaluation data after standardized processing into a single risk value of a uniform scale by using a preset risk mapping rule; the risk mapping rule includes setting according to the pathological significance of each item of evaluation data and the physiological characteristics of elderly NSC patients;

[0037] S3: dividing the single risk value into multiple risk sub-models, and calculating the dimension risk score of each risk sub-model;

[0038] S4: based on the preset fusion strategy, integrating the dimension risk scores of multiple risk sub-models and the interaction effect between the single risk value, to calculate the total risk score of the patient;

[0039] S5: determining the risk level of the patient according to the total risk score, and outputting the corresponding clinical diagnosis and treatment path suggestion based on the risk level.

[0040] In the above S1, the multi-dimensional evaluation data needs to be automatically collected from the hospital information system, the inspection information system and the electronic medical record, to ensure the integrity and consistency of the data. The core data collected is divided into three categories. The basic clinical information index includes age, ADL score and doctor's first impression score. The age needs to be 65 years old and above, which is the core of the risk baseline adjustment; the ADL score is used to evaluate the degree of functional decline, ranging from 0 to 1, 0 indicating no functional decline, and 1 indicating complete loss of function; the doctor's first impression score is the subjective evaluation of the patient's risk by the clinician based on the initial diagnosis, also ranging from 0 to 1.

[0041] The laboratory biomarker index includes soluble urokinase-type plasminogen activator receptor, adrenal medulla precursor middle segment fragment, D-dimer, vasopressin C-terminal precursor, peroxide reductase-4, C-reactive protein and procalcitonin. The soluble urokinase-type plasminogen activator receptor reflects the systemic inflammatory activity, the adrenal medulla precursor middle segment fragment reflects the organ perfusion state, the D-dimer reflects the coagulation activation state, the vasopressin C-terminal precursor reflects the neuroendocrine stress response, the peroxide reductase-4 reflects the oxidative stress level, the C-reactive protein and the procalcitonin both reflect the degree of inflammation, and the procalcitonin is also related to the inflammation associated with bacterial infection. The vital signs index includes heart rate, systolic blood pressure, blood oxygen saturation, respiratory rate and body temperature.

[0042] In order to eliminate the differences in measurement units and dimensions of different indicators, such as the unit of soluble urokinase-type plasminogen activator receptor is ng / ml and the unit of heart rate is bpm, the embodiment adopts Z-score algorithm to standardize all original data. The specific way of standardization is to convert the original measurement value of the current patient for each index based on the mean and standard deviation of the historical training data. The historical training data needs to include at least 1000 cases of elderly NSC patients' emergency treatment data, and these data need to have 30-day outcome labels. Through this conversion, all index data have unified model fusion adaptability, avoiding distortion of risk contribution due to dimensional differences.

[0043] In S2, the embodiment converts each index into a risk value in the range of 0 to 1 by a preset risk mapping rule, where 0 represents no risk and 1 represents extremely high risk. The risk mapping rule is designed according to the pathological significance of the index and the physiological characteristics of the elderly patients.

[0044] In the embodiment, a dedicated nonlinear function form is used for different pathological mechanisms of the index.

[0045] For the soluble urokinase-type plasminogen activator receptor reflecting the activity of broad-spectrum inflammation, a piecewise quadratic function is used:

[0046]

[0047] When the index is lower than 4.0 ng / ml, the risk value is 0; when the index is between 4.0 and 6.5 ng / ml, the risk value increases quadratically with the value; and when the index exceeds 6.5 ng / ml, the risk value is fixed at 1.0, thereby reflecting the risk difference in different concentration intervals.

[0048] For D-dimer reflecting the activation state of blood coagulation, an age dynamic correction threshold is introduced, which is the value of the patient's age divided by 100. For example, the threshold for a 75-year-old patient is 0.75 μg / ml, and a logarithmic function is used for mapping:

[0049]

[0050] The more the value exceeds the threshold, the more significant the risk value increases, which is more in line with the risk characteristics of blood coagulation abnormalities in elderly patients.

[0051] For the adrenal medulla precursor middle segment fragment reflecting the organ perfusion state, a Sigmoid function is used for setting:

[0052]

[0053] With 1.0 nmol / L as the risk inflection point, the risk value increases sharply after exceeding this value, which is in line with the feature that the increase of the index is strongly related to the risk of organ failure in clinical practice.

[0054] For the ADL score reflecting the degree of functional decline, a power function is used to amplify the risk of patients with high functional decline by the power of 1.5:

[0055]

[0056] For example, when the ADL score is 0.6, the risk value is about 0.465, and when the ADL score is 0.8, the risk value is about 0.715, highlighting the impact of functional decline on risk.

[0057] For heart rate reflecting the function of the circulatory system, a nonlinear function based on the normal interval is used:

[0058]

[0059] Wherein, 60 to 100 bpm is the normal interval, the risk value is 0 in this interval, below 60 bpm or above 100 bpm, the risk value increases nonlinearly with the degree of deviation, which conforms to the influence law of abnormal heart rate on the circulatory stability of elderly patients.

[0060] The key parameters of all mapping functions, including function inflection point, growth rate, and saturation value, are automatically optimized based on the statistical distribution of historical training data, with the optimization target being to maximize the area under the curve of 30-day adverse outcome prediction of a single indicator, ensuring that the parameters have clinical rationality and data-driven scientificity. At the same time, all mapping results are reduced to the interval of 0 to 1, achieving risk probability unified expression.

[0061] Taking a 75-year-old female patient as an example, the original values and risk mapping results of some indicators are as follows: soluble urokinase-type plasminogen activator receptor is 5.8 ng / ml, the risk mapping value is 0.324; D-dimer is 1.5 μg / ml, the risk mapping value is 0.084; adrenomedullin precursor mid-segment fragment is 1.2 nmol / L, the risk mapping value is 0.645; ADL score is 0.6, the risk mapping value is 0.465; heart rate is 110 bpm, the risk mapping value is 0.477, through these values, the risk situation of single indicator can be intuitively understood.

[0062] The risk values obtained in S2 will be input into the next step of risk sub-model construction, which is used for multi-dimensional disease mechanism aggregation analysis.

[0063] For S3, the single risk value obtained in step S2 is divided into five risk sub-models according to the pathological mechanism dimension, and the dimension risk score of each sub-model is calculated by weighted linear combination, the dimension risk score range is between 0 and 1, which realizes the structural attribution of risk source and improves the explainability of risk assessment.

[0064] The five risk sub-models are classified according to the pathological mechanism of biomarkers. The inflammation-immunity sub-model integrates single risk values related to inflammation and immunity, and the core indicators include soluble urokinase plasminogen activator receptor, C-reactive protein, procalcitonin, and peroxidase-4, mainly reflecting the patient's systemic inflammation and immune activation state. The coagulation-circulation sub-model integrates single risk values related to coagulation and cardiovascular function, and the core indicators include D-dimer, systolic blood pressure, and heart rate, reflecting the patient's coagulation abnormalities and circulatory stability. The organ function sub-model integrates single risk values related to organ perfusion and respiratory function, and the core indicators include mid-segment fragment of adrenomedullin precursor, blood oxygen saturation, and respiratory rate, reflecting the patient's organ damage and risk of functional failure. The clinical perception sub-model integrates single risk values related to clinical subjective evaluation and functional status, and the core indicators include the doctor's first impression score and ADL score, reflecting the clinical intuitive risk and functional reserve. The basic risk adjustment sub-model takes age as the core indicator, reflecting the basic risk of elderly patients due to physiological degradation, and the risk increases quadratically after the age of 65, and the risk value is 0 before the age of 65.

[0065] The weight of each indicator in each sub-model is calculated based on historical training data, and the calculation is based on the area under the curve of single indicators, the coefficients of logistic regression or the feature importance of gradient boosting trees, and the sum of the weights of all indicators in each sub-model is 1, ensuring the rationality and comparability of the weights.

[0066] In an implementable scenario, in the inflammation-immunity sub-model, the weight of soluble urokinase plasminogen activator receptor is 0.35, the weight of C-reactive protein is 0.30, the weight of procalcitonin is 0.20, and the weight of peroxidase-4 is 0.15;

[0067] In the coagulation-circulation sub-model, the weight of D-dimer is 0.50, the weight of systolic blood pressure is 0.20, and the weight of heart rate is 0.30; in the organ function sub-model, the weight of mid-segment fragment of adrenomedullin precursor is 0.55, the weight of blood oxygen saturation is 0.25, and the weight of respiratory rate is 0.20;

[0068] In the clinical perception sub-model, the weight of the doctor's first impression score is 0.55, and the weight of the ADL score is 0.45;

[0069] In the basic risk adjustment sub-model, the weight of age is 1.00, and the dimension score is calculated separately.

[0070] The dimension risk score of each sub-model is the weighted sum of all single risk values and corresponding weights in the sub-model. Taking the inflammation-immunity sub-model as an example, the dimension score is

[0071]

[0072] The dimension score of the coagulation-circulation sub-model is:

[0073]

[0074] The organ function sub-model dimension score is:

[0075]

[0076] The clinical perception sub-model dimension score is obtained by aggregating the following indicators ADL decline degree and physician preliminary impression score DOI:

[0077]

[0078] The basic risk adjustment sub-model is set as:

[0079]

[0080] In the above model, all Risk mapping results from S2;

[0081] All Sub-model internal indicator weights obtained for training;

[0082] Sub-model output values Indicate the relative risk contribution degree under the pathological mechanism dimension.

[0083] This step still takes a 75-year-old female patient as an example, assuming that the C-reactive protein risk value is 0.6, the procalcitonin risk value is 0.5, the peroxide reductase-4 risk value is 0.4, the systolic blood pressure risk value is 0.3, the blood oxygen saturation risk value is 0.2, the respiratory rate risk value is 0.3, and the physician first impression score risk value is 0.7, combined with the previous single risk value, the dimension scores of each sub-model can be calculated: the inflammation-immunity sub-model dimension score is 0.35x0.324+0.30x0.6+0.20x0.5+0.15x0.4=0.6275; the coagulation-circulation sub-model dimension score is 0.50x0.084+0.20x0.3+0.30x0.477=0.69; the organ function sub-model dimension score is 0.55x0.645+0.25x0.2+0.20x0.3=0.64; the clinical perception sub-model dimension score is 0.55x0.7+0.45x0.465=0.655; and the basic risk adjustment sub-model dimension score is 0.02x(75-65)²=3.38.

[0084] In S4, the embodiment is provided with a double-layer weighted fusion mechanism, which integrates the dimension scores of the five types of sub-models and the interaction effect of the single risk value to obtain the final total risk score.

[0085] Specifically, first, the total score function is set as:

[0086]

[0087] wherein is the score result of five sub-models;

[0088] is the global weight of the kkkth sub-model (derived from training optimization);

[0089] is the standardized single risk indicator participating in scoring;

[0090] is the penalty weight between the index pairs with known interaction reinforcement effect;

[0091] is the interaction factor adjustment intensity (hyperparameter, adjusts the influence degree of interaction term weight on F(X)).

[0092] Sigmoid risk score function:

[0093]

[0094] The output range of RiskScore in this function is (0, 1), representing the overall risk probability, which can be used for three-stage classification (low / medium / high risk).

[0095] Based on the above preconditions, the first layer of the fusion strategy in this embodiment is the global weighting of sub-models. According to the prediction contribution of each sub-model to the 30-day adverse outcome, a global weight is given to each sub-model. The global weight is obtained by training the complete model through a logistic regression or gradient boosting tree algorithm, and is normalized based on the prediction importance of the sub-model. The sum of the global weights of all sub-models is 1. For example, the global weight of the inflammation-immunity sub-model is 0.25, the global weight of the coagulation-circulation sub-model is 0.20, the global weight of the organ function sub-model is 0.20, the global weight of the clinical perception sub-model is 0.15, and the global weight of the basic risk adjustment sub-model is 0.20. These weights reflect the contribution differences of different sub-models to the total risk.

[0096] The second layer of the fusion strategy is the interaction penalty, which corrects the single-item risk value pairs that have synergistic deterioration effects to avoid underestimation of risk. Synergistic deterioration effect refers to the risk level when some indicators increase simultaneously, which is much higher than the simple superposition of individual increase, for example, when soluble urokinase-type plasminogen activator receptor and C-reactive protein increase simultaneously, the risk of inflammation is significantly higher than that of individual increase. In addition to the above-mentioned combinations, other index pairs with clinically proven synergistic effects include mid-regional proadrenomedullin and respiratory rate, oxygen saturation and systolic blood pressure. The interaction weight is determined by interaction variable regression or SHAP interaction analysis, with a value range of-1 to 1. A positive number indicates synergistic risk enhancement, and a negative number indicates synergistic risk reduction. The default high-risk interaction pair, such as soluble urokinase-type plasminogen activator receptor and C-reactive protein, has an interaction weight of 0.25, and the weakly related interaction pair, such as oxygen saturation and systolic blood pressure, has an interaction weight of 0.10. At the same time, the interaction factor adjustment strength is set, which is a hyperparameter with a default value of 1.0, used to control the influence of the interaction term on the total score.

[0097] In the implementation process of the present embodiment, the calculation of the total risk score is divided into three steps. The first step is to calculate the linear weighted sum of the five types of sub-model dimension scores, that is, to multiply each sub-model dimension score by its corresponding global weight and then sum them up. The second step is to calculate the interaction correction value, which is the product of the interaction factor adjustment strength, the interaction weight of each interaction pair, and the sum of the products of the two single-item risk values in the interaction pair. Only the top several interaction pairs that significantly contribute to risk prediction are included in the calculation. The third step is to normalize the sum of the linear weighted sum and the interaction correction value by using the Sigmoid function to obtain the total risk score. This function can map numerical values to the range of 0 to 1, eliminate numerical deviations in different calculation steps, and ensure the stability and comparability of the total risk score.

[0098] The embodiment continues to take a 75-year-old female patient as an example, the first step calculates the linear weighted sum: 0.25x0.6275+0.20x0.69+0.20x0.64+0.15x0.655+0.20x3.38=1.19715. The second step calculates the interaction correction value, assuming that the interaction factor adjusts the intensity to 1.0, and selecting the soluble urokinase-type plasminogen activator receptor and C-reactive protein, the adrenal medulla precursor middle fragment and respiratory rate, and the blood oxygen saturation and systolic pressure three groups of interaction pairs, the interaction weights are 0.25, 0.15 and 0.10 respectively, and the corresponding single risk value products are 0.324x0.6, 0.645x0.3 and 0.2x0.3 respectively, and the interaction correction value is 1.0x(0.25x0.324x0.6+0.15x0.645x0.3+0.10x0.2x0.3)=0.1875. The third step calculates the total risk score: 1 ÷(1+e^(-(1.19715+0.1875)))≈0.80, the total risk score of the patient is 0.80, which belongs to the high risk category.

[0099] Through S1-S4 described above, the embodiment obtains the total risk score of the patient, so as to implement S5, the total risk score is converted into an operable clinical risk level, and the emergency information system is linked to push the diagnosis and treatment path suggestion, so as to realize the closed-loop management from risk assessment to clinical intervention, and improve the emergency treatment efficiency and patient prognosis.

[0100] As an optional way, the risk level adopts a double-threshold three-section grading method, and the basic threshold is determined based on the ROC curve analysis of historical data. The low risk corresponds to a total risk score less than 0.35, and the probability of adverse outcomes within 30 days of the patients in this level is usually less than 5%; the medium risk corresponds to a total risk score between 0.35 and 0.65, and the probability of adverse outcomes within 30 days is usually between 5% and 20%; the high risk corresponds to a total risk score greater than or equal to 0.65, and the probability of adverse outcomes within 30 days is usually more than 20%.

[0101] In order to adapt to the patient population characteristics of different hospitals, such as the difference in patient risk distribution between community hospitals and third-grade hospitals, the model needs to be retrained based on the real data of old NSC patients in the hospital and the 30-day outcome every quarter, and the grading threshold is updated. The update process is as follows: first, collect at least 200 evaluation data and 30-day outcomes of old NSC patients in the hospital every quarter, the outcomes include death, hospitalization and ICU treatment; then use the existing model structure to re-fit the new data and calculate the ROC curve; then select the score when the Youden index is maximum as the new threshold, the Youden index is calculated by sensitivity plus specificity minus 1, which can balance sensitivity and specificity; finally, the new threshold is loaded into the system through the model configuration file, and it can take effect without restarting the system, ensuring the timeliness and adaptability of the threshold.

[0102] Based on the risk level, the corresponding clinical diagnosis and treatment path suggestion is output, and the suggestions for different risk levels have different focuses. The low-risk patients are pushed to outpatient assessment or home follow-up suggestion, and the system automatically generates a follow-up template, which includes 72-hour telephone follow-up reminders and medical guidance when symptoms worsen. Such patients do not need to be hospitalized or observed, reducing unnecessary occupation of medical resources. The medium-risk patients are pushed to observation re-examination suggestions, and are prompted to recheck core biomarkers such as soluble urokinase-type plasminogen activator receptor and adrenal medulla precursor mid-fragment after 6 to 12 hours, while linking with secondary departments such as geriatric department for initial diagnosis and closely monitoring patient condition changes. The high-risk patients are pushed to red alerts, and the system automatically triggers multiple tasks: patient information is displayed at the top of the on-duty doctor terminal to ensure that the doctor pays attention first; links with the ICU consultation system to put the patient into the priority queue for consultation, shortening the waiting time for consultation; interfaces with the bed management system to pre-allocate ICU or emergency observation beds to avoid delays caused by bed shortages; generates an emergency admission process sheet, which includes examination items that need to be completed first, such as blood gas analysis, chest CT, etc., to speed up the admission processing speed.

[0103] Taking a 75-year-old female patient as an example, her total risk score is 0.80, which belongs to high risk, and the system triggered linkage operations include: the emergency information system value display window pops up a high-risk patient prompt on the main control console, suggesting immediate ICU consultation; automatically sends a short message to the ICU on-duty doctor, with the patient's core index information attached, such as soluble urokinase-type plasminogen activator receptor 5.8 ng / ml and adrenal medulla precursor mid-fragment 1.2 nmol / L; the bed system displays one idle ICU bed and locks the bed for 30 minutes, reserving time for the patient; after the doctor confirms, the system automatically generates an ICU admission evaluation sheet, reducing manual document processing time and improving overall processing efficiency.

[0104] In addition, to ensure the generalization ability and clinical adaptability of the model, all parameters of the model are trained and optimized through historical data in this embodiment, including sub-model index weight, sub-model global weight, and interaction term weight. The training process focuses on the combination of data-driven and clinical rationality.

[0105] In an alternative scenario, the training data needs to meet certain requirements to ensure the training effect. In terms of sample size, at least 1000 elderly NSC emergency patients aged 65 and above are included, and the sample needs to cover different genders, different underlying diseases such as hypertension and diabetes, and ensure the representativeness of the sample. In terms of data integrity, each sample needs to include all core indicators in step S1, including original values and standardized values, to avoid affecting model training due to missing indicators. In terms of outcome labels, each sample needs to have a clear 30-day outcome label, including at least one of whether to die, whether to be hospitalized, and whether to enter ICU, to provide a clear training target for the model. In terms of data sources, historical case data is extracted from hospital information systems, laboratory information systems, and electronic medical records, and samples with a missing rate of more than 20% are excluded to ensure data quality.

[0106] Parameter training adopts targeted strategies. The training of sub-model weights is modeled separately for each sub-model, and can optionally use logistic regression or gradient boosting tree algorithm. Logistic regression uses regression coefficient size as weight basis, and the larger the coefficient, the greater the contribution of the indicator to risk prediction. Gradient boosting tree uses feature importance as a basis, and the higher the feature importance, the more critical the indicator is in the model. Then the original weight is normalized to ensure that the sum of all indicator weights in each sub-model is 1. The training of sub-model global weights is to construct a logistic regression model to predict 30-day adverse outcomes using five types of sub-model dimension scores as input variables. The regression coefficients of the dimension scores of each sub-model in the model are normalized to obtain global weights, reflecting the overall contribution of each sub-model to total risk prediction. The training of interaction term weights is to add the indicator pair as a new variable to the model through interaction variable regression, and the regression coefficient of the new variable is used as the interaction term weight. At the same time, the weight value range is constrained between -1 and 1, only retaining the interaction pairs with a P value less than 0.05 in statistical tests to ensure that the contribution of the interaction pair to risk prediction has statistical significance.

[0107] In the model evaluation and optimization link, the embodiment adopts multiple indexes to comprehensively evaluate the model performance, avoiding the limitation of a single index. The area under the curve is used to evaluate the discrimination ability of the model, and the area under the curve exceeding 0.85 indicates that the model has excellent discrimination ability; the F1 score takes into account the precision and recall, and the F1 score exceeding 0.8 indicates that the model performs excellently in terms of precision and recall; the Brier score is used to evaluate the calibration degree of the risk probability, and the Brier score less than 0.1 indicates that the model has excellent calibration degree. In order to avoid model overfitting, 5-fold cross-validation is adopted, and the training data is divided into 5 parts, each time 4 parts are used as the training set and 1 part is used as the validation set, and the average performance index is taken after repeating 5 times. In terms of optimization algorithm, the grid search and Early stopping strategy are adopted to optimize the hyperparameters such as interaction factor adjustment strength, the grid search finds the optimal value by traversing the preset parameter combination, the Early stopping stops training when the model performance no longer improves, preventing overfitting; at the same time, the L1 regularization, i.e. LASSO, is used to filter redundant indexes, for example, if the area under the curve of peroxide reductase-4 is less than 0.65, it is excluded, which simplifies the model structure while ensuring the prediction performance.

[0108] Through the above scheme, the embodiment realizes the precise risk stratification of elderly NSC emergency patients. Compared with the prior art, the embodiment covers multiple dimensions such as inflammation, coagulation and organ function through multi-marker fusion, integrates new markers such as soluble urokinase-type plasminogen activator receptor and mid-fragment of adrenomedullin precursor, and improves the prediction accuracy; through age correction threshold, age-based risk sub-model and other ways, the physiological characteristics of elderly patients are fully adapted; it has strong clinical operability, system supports automatic data access and path pushing, reduces manual intervention, and improves emergency disposal efficiency; it has dynamic adaptability, supports hospitals to update thresholds and weights based on local data, and ensures the applicability of the model in different clinical scenarios. It can be directly applied to emergency geriatric medicine of hospitals at all levels, helping doctors to early identify high-risk patients, optimize medical resource allocation, and improve patient prognosis.

[0109] The embodiment also provides an electronic device, including: a memory for storing a computer program; a processor for executing the computer program to realize the steps of the elderly NSC emergency risk stratification method.

[0110] The embodiment also provides a storage medium, the storage medium stores a computer program, and the computer program is executed by a processor to realize the steps of the elderly NSC emergency risk stratification method.

[0111] The above has carried out the detailed introduction to the embodiment of the application. The progressive way is adopted in the description of each embodiment, and each embodiment mainly explains the difference from other embodiments. The same and similar parts of each embodiment can be referred to each other. For the device disclosed by the embodiment, since it corresponds to the method disclosed by the embodiment, the description is relatively simple, and the related part can be referred to the method part. It should be pointed out that, for ordinary skilled in the art, without departing from the principle of the application, the application can be improved and modified, and these improvements and modifications also fall within the protection scope of the claims of the application.

Claims

1. A risk stratification method for elderly patients with NSC emergency care, characterized in that, include: S1: Obtain multidimensional assessment data of elderly NSC patients and standardize the multidimensional assessment data; The multidimensional assessment data includes at least basic clinical information, laboratory biomarker indicators, and vital sign indicators; the laboratory biomarker indicators include at least one of soluble urokinase plasminogen activator receptor, adrenal medullary precursor mid-terminus, D-dimer, angiotensin C-terminal precursor, and peroxidase-4. S2: For each standardized assessment data point, a pre-defined risk mapping rule is used to convert it into a single risk value with a uniform scale; The risk mapping rules are set based on the pathological significance of each assessment data and the physiological characteristics of elderly NSC patients. The preset risk mapping rule includes a nonlinear mapping function, and adopts a specific function form for the pathological mechanism of different assessment data; specifically, a piecewise quadratic function is used for indicators reflecting broad-spectrum inflammatory activity, a sigmoid function is used for indicators reflecting organ perfusion status, a power function is used for indicators reflecting the degree of functional decline, and a nonlinear function based on the normal range is used for indicators reflecting heart rate; the risk mapping rule also includes an age-dynamic correction threshold, which is used to distinguish between normal and abnormal risk ranges for D-dimer; S3: Divide the single risk value into multiple risk sub-models and calculate the dimensional risk score for each risk sub-model; the multiple risk sub-models include sub-models of five pathological mechanism dimensions, namely: inflammation-immunity sub-model, coagulation-circulation sub-model, organ function sub-model, clinical perception sub-model, and basic risk regulation sub-model; among them, the inflammation-immunity sub-model is used to integrate single risk values ​​related to inflammation and immunity; the coagulation-circulation sub-model is used to integrate single risk values ​​related to coagulation and cardiovascular function; the organ function sub-model is used to integrate single risk values ​​related to organ perfusion and respiratory function; the clinical perception sub-model is used to integrate single risk values ​​related to the degree of functional decline and the doctor's initial impression; the basic risk regulation sub-model is constructed based on the patient's age; S4: Based on a preset fusion strategy, the dimensional risk scores of multiple risk sub-models and the interaction effects between the individual risk values ​​are integrated to calculate the patient's total risk score. The preset fusion strategy includes: setting the first fusion layer as a global weighted sub-model, and linearly aggregating the dimensional risk scores of the five risk sub-models through global weights; setting the second fusion layer as an interaction term penalty, and correcting single risk value pairs with synergistic deterioration effects through interaction weights, wherein the single risk value pairs with synergistic deterioration effects include one or more of suPAR and CRP, MR-proADM and respiratory rate, and blood oxygen saturation and systolic blood pressure; S5: Determine the patient's risk level based on the total risk score, and output corresponding clinical treatment pathway recommendations based on the risk level.

2. The method for risk stratification in emergency care for elderly patients with NSC according to claim 1, characterized in that, The standardization process employs the Z-score algorithm, which achieves standardization by eliminating differences in measurement units and dimensions among different evaluation data. This includes: Based on the mean and standard deviation of each assessment data in the historical training data, the original values ​​of the current patient's assessment data are transformed and standardized.

3. The method for risk stratification in emergency care for elderly patients with NSC according to claim 1, characterized in that, The output range of the single risk value of the unified scale is set with a preset interval. The values ​​within the preset interval are used to realize the probabilistic expression of the risk of the single assessment data, and the key parameters of each risk mapping rule are automatically optimized based on the statistical distribution of historical training data.

4. The method for risk stratification in emergency care for elderly patients with NSC according to claim 1, characterized in that, The dimensional risk score of each risk sub-model is a weighted linear combination of all individual risk values ​​under that sub-model; The weight of each individual risk value is calculated based on historical training data, and the sum of the weights of all individual risk values ​​within each sub-model is 1.

5. The method for risk stratification in emergency care for elderly patients with NSC according to claim 1, characterized in that, The risk levels include three levels: low risk, medium risk, and high risk. Each level has a different total risk score range. The threshold of the score range is dynamically updated based on the hospital's local historical data. The update method is to determine the optimal threshold by calculating the Youden index through ROC curve analysis every quarter.

6. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of a method for risk stratification in emergency care for elderly patients with NSC as described in any one of claims 1-5 when executing the computer program.

7. A storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the method for risk stratification in emergency care for elderly patients with NSC as described in any one of claims 1-5.

Images