Multivariate data mining-based emergency patient vital sign real-time monitoring method and system

Abstract

The application discloses a multiple data mining emergency patient vital sign real-time monitoring method and system, and relates to the technical field of vital sign monitoring. The method comprises the following steps: collecting multiple condition data of a target emergency patient; taking the multiple condition data as a constraint, taking a cerebral hemorrhage early warning as a warning target, determining a warning analysis time limit of the target emergency patient through a multiple data mining manner, and determining the contribution degree of each monitoring index in the preset monitoring index set to the warning target; taking the approximation of the warning analysis time limit as an optimization target, combining the contribution degree of each monitoring index and a preset monitoring frequency, performing a combination scheme optimization on the monitoring index in the preset monitoring index set, and determining an optimal monitoring index group; and performing a vital sign real-time monitoring on the target emergency patient based on an optimal monitoring scheme formed by the optimal monitoring index group. The application effectively improves the timeliness and accuracy of the emergency cerebral hemorrhage early warning.

Description

Technical Field

[0001] This invention relates to the field of vital sign monitoring technology, specifically to a method and system for real-time monitoring of vital signs in emergency patients using multivariate data mining. Background Technology

[0002] With the development of emergency medicine and the improvement of medical informatization, early warning of critical illnesses such as acute cerebral hemorrhage has become a key aspect of clinical treatment. Current technologies typically employ a fixed multi-parameter monitoring mode, collecting multiple physiological indicators such as electrocardiogram, blood pressure, and blood oxygen saturation, and combining them with preset alarm thresholds or general risk assessment models to determine changes in the patient's condition, providing basic data support for clinical practice.

[0003] However, existing monitoring technologies lack quantitative definitions of the early warning analysis timeframe, resulting in a lack of timeliness constraints and making it difficult to achieve effective early warnings within the golden rescue period. The widespread use of a full-indicator package-style monitoring model easily leads to information overload and alarm fatigue, failing to balance early warning accuracy and monitoring efficiency. Furthermore, the monitoring scheme cannot dynamically adjust the monitoring precision according to the complexity of the patient's condition and the urgency of the time limit, resulting in insufficient timeliness and accuracy of early warnings for cerebral hemorrhage in emergency patients, causing unnecessary waste of medical resources. Summary of the Invention

[0004] This invention provides a method and system for real-time monitoring of vital signs in emergency patients using multivariate data mining, aiming to solve the technical problem of insufficient timeliness and accuracy of existing technologies in early warning of cerebral hemorrhage in emergency patients.

[0005] In view of the above problems, the present invention provides a method and system for real-time monitoring of vital signs of emergency patients using multivariate data mining.

[0006] In a first aspect, the present invention provides a method for real-time monitoring of vital signs in emergency patients using multivariate data mining, including:

[0007] Collect diverse medical data of target emergency patients; Using the aforementioned multivariate medical data as constraints and cerebral hemorrhage warning as the warning target, the warning analysis time limit for the target emergency patients is determined through multivariate data mining, and the contribution of each monitoring indicator in the preset monitoring indicator set to the judgment of the warning target is determined. With the goal of approximating the early warning analysis time limit, and combining the contribution of each monitoring indicator with the preset monitoring frequency, the optimal monitoring indicator group is determined by optimizing the combination scheme of the monitoring indicators in the preset monitoring indicator set. During the optimization process, the candidate monitoring indicator group is evaluated by a preset monitoring scheme evaluation plugin. The quality evaluation is generated based on the prediction error probability of the early warning target. The optimal monitoring scheme, based on the optimal monitoring index group, is used to monitor the vital signs of the target emergency patient in real time.

[0008] Secondly, this invention provides a real-time monitoring system for vital signs of emergency patients based on multivariate data mining, including: The multi-data acquisition module is used to collect multi-dimensional medical data of the target emergency patients; The time limit and contribution determination module is used to determine the early warning analysis time limit for the target emergency patient by using the multivariate disease data as a constraint and cerebral hemorrhage early warning as the early warning target through multivariate data mining, and to determine the contribution of each monitoring indicator in the preset monitoring indicator set to the judgment of the early warning target. The combined scheme optimization module is used to optimize the combined scheme of the monitoring indicators in the preset monitoring indicator set by taking the approximation of the early warning analysis time limit as the optimization target, and combining the contribution of each monitoring indicator and the preset monitoring frequency to determine the optimal monitoring indicator group. In the optimization process, the candidate monitoring indicator group is evaluated by a preset monitoring scheme evaluation plugin. The quality evaluation is generated based on the prediction error probability of the early warning target. The vital signs monitoring module is used to monitor the vital signs of the target emergency patient in real time based on the optimal monitoring scheme composed of the optimal monitoring index group.

[0009] One or more technical solutions provided in this invention have at least the following technical effects or advantages: This invention provides a method and system for real-time monitoring of vital signs in emergency patients using multivariate data mining. By collecting diverse disease data from emergency patients, it provides comprehensive data support for individualized early warning analysis. By mining the early warning analysis time limit and quantifying the contribution of monitoring indicators, it achieves individualized definition of the early warning timeliness of cerebral hemorrhage and precise positioning of monitoring indicators. By combining contribution and early warning time limit to optimize indicator combinations and evaluating the scheme based on the prediction error probability, it can screen out the optimal monitoring indicator group that balances timeliness and accuracy, avoiding information overload and waste of medical resources. Based on the optimal monitoring scheme, real-time monitoring is implemented, ultimately ensuring that the early warning of cerebral hemorrhage is effectively triggered within the golden rescue period, improving the timeliness, accuracy, and clinical applicability of early warning for cerebral hemorrhage in emergency patients. Attached Figure Description

[0010] Figure 1 A flowchart illustrating the method for real-time monitoring of vital signs in emergency patients using multivariate data mining, as provided in an embodiment of the present invention. Figure 2 A schematic diagram of the structure of the real-time monitoring system for vital signs of emergency patients using multivariate data mining provided in an embodiment of the present invention; The components represented by each number in the attached diagram are explained below: Multi-source data acquisition module 11, time limit and contribution determination module 12, combination scheme optimization module 13, vital sign monitoring module 14. Detailed Implementation

[0011] This invention provides a method and system for real-time monitoring of vital signs in emergency patients using multivariate data mining, which addresses the technical problem of insufficient timeliness and accuracy of existing technologies in early warning of cerebral hemorrhage in emergency patients.

[0012] Example 1, as Figure 1 As shown, this invention provides a method for real-time monitoring of vital signs in emergency patients using multivariate data mining, the method comprising: S100: Collect multivariate medical data of target emergency patients.

[0013] In this embodiment of the invention, multidimensional medical data of target emergency patients are collected. The conditions of emergency cerebral hemorrhage patients exhibit significant individual differences; single-dimensional data cannot fully reflect the patient's underlying condition and real-time physiological state. Subsequent data mining stages, such as early warning analysis, time-limit mining, and matching with similar historical cases, require multidimensional standardized data as a constraint basis. To ensure the accuracy of subsequent case matching and the objectivity of core organ physiological state assessment, multidimensional medical data covering the patient's basic condition, real-time vital signs, and biochemical tests needs to be collected, providing comprehensive and accurate data support for subsequent multidimensional data mining and monitoring scheme optimization.

[0014] Step S100 in the method provided in this embodiment of the invention includes: Collect basic characteristic data of the target emergency patient, wherein the basic characteristic data includes at least age, gender, and history of cerebral hemorrhage; Collect preliminary vital sign data of the target emergency patient, wherein the preliminary vital sign data includes at least heart rate, blood pressure, respiratory rate and blood oxygen saturation; Collect bedside laboratory data from the target emergency patients, wherein the bedside laboratory data includes at least the international normalized ratio, activated partial thromboplastin time, and blood glucose concentration; The aforementioned basic characteristic data, initial vital sign data, and bedside test data are used as multivariate disease data.

[0015] First, basic characteristic data of the target emergency patient is collected. This basic characteristic data includes at least age, gender, and a history of cerebral hemorrhage. Basic characteristic data refers to static characteristic data reflecting the patient's congenital condition and past heart disease history, used as a basis for case matching in subsequent data mining. Information on the target emergency patient's age, gender, and history of cerebral hemorrhage is collected through the emergency department's medical information system or by interviewing the patient and their family on-site, and standardized data entry and storage are completed. For example, if the target emergency patient is a 58-year-old male with one history of cerebral hemorrhage, the collected basic characteristic data would be: age 58, male, and history of cerebral hemorrhage.

[0016] Secondly, preliminary vital sign data of the target emergency patient are collected. This preliminary vital sign data includes at least heart rate, blood pressure, respiratory rate, and blood oxygen saturation. Preliminary vital sign data refers to physiological indicators collected at the initial stage of hospital admission, directly reflecting the current perfusion status of the patient's core organs such as the heart and respiratory system. An emergency bedside vital sign monitor is used to monitor the patient in real time, collecting and recording the real-time values ​​of heart rate, blood pressure, respiratory rate, and blood oxygen saturation. For example, the preliminary vital sign data for the aforementioned 58-year-old male patient were: heart rate 102 beats / minute, systolic blood pressure 105 mmHg, diastolic blood pressure 65 mmHg, respiratory rate 24 breaths / minute, and blood oxygen saturation 92%.

[0017] Next, bedside laboratory data were collected from the target emergency patients. This bedside data included at least the International Normalized Ratio (INR), activated partial thromboplastin time (APT), and blood glucose concentration. Bedside laboratory data refers to coagulation function and metabolic-related biochemical indicators obtained through rapid bedside testing equipment, serving as crucial reference data for early screening of cerebral hemorrhage. The INR is a standardized correction for prothrombin time (PT) in coagulation function testing. Its purpose is to eliminate the influence of different laboratories and testing reagents / instruments on prothrombin time test results, achieving a unified standard for coagulation function assessment. It is a core indicator for clinically evaluating the function of the body's extrinsic coagulation pathway and monitoring the effectiveness of anticoagulation therapy. A rapid bedside testing device was used to collect finger-prick or venous blood from the patient for rapid testing, collecting INR, APT, and blood glucose concentration values. For example, the bedside laboratory data obtained for the aforementioned 58-year-old male patient were: INR 1.2, APT 35s, and blood glucose concentration 6.8 mmol / L.

[0018] Finally, the basic characteristic data, initial vital sign examination data, and bedside test data are used as multidimensional disease data. Multidimensional disease data refers to multi-dimensional disease data integrated from basic characteristic data, initial vital sign examination data, and bedside test data. This data is used as a constraint for subsequent data mining to match historical cases with similar disease bases. For example, by summarizing and integrating the three types of data for the above patient, the complete multidimensional disease data for that patient can be obtained.

[0019] In this embodiment of the invention, by collecting multi-dimensional and diverse disease data, a precise basis is provided for subsequent historical case matching. This data can intuitively reflect the perfusion status of the patient's core organs and provides standardized constraints for subsequent data mining. This ensures the comprehensiveness and accuracy of the data source for subsequent early warning analysis, time limit mining, and contribution calculation of monitoring indicators, and avoids case matching bias caused by single data.

[0020] S200: Using the aforementioned multivariate medical data as constraints and cerebral hemorrhage warning as the warning target, the warning analysis time limit for the target emergency patient is determined through multivariate data mining, and the contribution of each monitoring indicator in the preset monitoring indicator set to the judgment of the warning target is determined.

[0021] In this embodiment of the invention, constrained by the multivariate medical data and with cerebral hemorrhage warning as the warning target, the warning analysis time limit for the target emergency patients is determined through multivariate data mining, and the contribution of each monitoring indicator in the preset monitoring indicator set to the judgment of the warning target is also determined. Emergency cerebral hemorrhage treatment has a strict golden rescue window. Patients of different ages, medical histories, and risk stratifications exhibit significant individual differences in the time span from abnormal physiological indicators to a confirmed diagnosis of cerebral hemorrhage. Using a uniform and fixed warning time limit would reduce the timeliness of the warning. Furthermore, cerebral hemorrhage warning involves numerous types of monitoring indicators, and the reference value of different indicators for judging cerebral hemorrhage is uneven. Failure to quantify the importance of indicators can lead to the omission of key indicators or redundant monitoring in subsequent monitoring plans. Therefore, it is necessary to use multivariate medical data as a constraint, and through multivariate data mining, determine the appropriate warning analysis time limit for the target patients individually, and accurately quantify the contribution of each monitoring indicator to the cerebral hemorrhage warning, providing time constraints and weighting basis for subsequent optimization of monitoring indicator combinations.

[0022] Step S200 in the method provided in this embodiment of the invention includes: The preset monitoring indicator set consists of multiple monitoring indicators suitable for early warning of cerebral hemorrhage. The preset monitoring indicator set includes at least bedside monitoring indicators and derived monitoring indicators. The bedside monitoring indicators include at least the peak systolic blood pressure, peak diastolic blood pressure, pulse pressure difference collected by invasive arterial blood pressure monitoring, and the absolute value of intracranial pressure collected by intracranial pressure monitoring. The derived monitoring indicators are generated in real time by a data processing module from the bedside monitoring indicators. The derived monitoring indicators include at least blood pressure variability, cerebral perfusion pressure, and intracranial pressure trend slope.

[0023] Specifically, using the aforementioned multivariate medical data as constraints and cerebral hemorrhage early warning as the early warning target, the early warning analysis time limit for the target emergency patients is determined through multivariate data mining, including: Using the aforementioned multivariate medical data as constraints, a historical case dataset with similar medical conditions to the target emergency patient is retrieved from the hospital information system. The similar medical conditions include at least the same age range and the same risk stratification for cerebral hemorrhage. In the historical case dataset, positive sample cases that were finally diagnosed as acute cerebral hemorrhage were marked, and continuous monitoring data of each positive sample case before the time of diagnosis were obtained; Analyze the continuous monitoring data of each positive sample case to find the time difference between the appearance of characteristic abnormal fluctuations in the monitoring data and the final diagnosis of cerebral hemorrhage. Perform statistical processing on the time differences of multiple positive sample cases and calculate the average of the multiple time differences as the early warning analysis time limit for the target emergency patient.

[0024] First, using the aforementioned multivariate medical data as constraints, a historical case dataset with similar medical conditions to the target emergency patient is retrieved from the hospital information system. This similarity in medical conditions includes at least the same age range and the same cerebral hemorrhage risk stratification. A similar medical condition base refers to a matching condition between the target emergency patient and historical cases that meets a preset similarity threshold in age range and cerebral hemorrhage risk stratification; this is the core criterion for selecting homogeneous historical samples. Using the multivariate medical data obtained in S100 as constraints, a case search is performed in the hospital information system to select historical cases within the same age range and with the same cerebral hemorrhage risk stratification as the target patient, and these are integrated to form a historical case dataset.

[0025] For example, the target emergency patient was a 58-year-old male, classified as high-risk for cerebral hemorrhage, with the corresponding age range being 55-65 years old and the cerebral hemorrhage risk stratification being high-risk. Based on this, a total of 1200 historical cases matching this similar medical condition were retrieved from the hospital information system.

[0026] Secondly, positive sample cases that were ultimately diagnosed with acute cerebral hemorrhage were identified in the historical case dataset, and continuous monitoring data for each positive sample case prior to the diagnosis were obtained. Positive sample cases refer to cases in the historical case dataset that were ultimately diagnosed with acute cerebral hemorrhage; continuous monitoring data refers to comprehensive monitoring data, including vital signs, continuously collected before the diagnosis of cerebral hemorrhage for positive sample cases. In the historical case dataset, positive sample cases that were ultimately diagnosed with acute cerebral hemorrhage were identified and filtered, while negative samples without a confirmed diagnosis of cerebral hemorrhage were removed. Complete continuous monitoring data prior to the diagnosis of each positive sample case was extracted. For example, in 1200 historical cases, 450 positive sample cases diagnosed with acute cerebral hemorrhage were identified, and continuous electrocardiogram and vital sign monitoring data for each of these 450 patients within 12 hours prior to diagnosis were extracted.

[0027] Furthermore, the continuous monitoring data of each positive sample case is analyzed to identify the time difference between the appearance of characteristic abnormal fluctuations in the monitoring data and the final diagnosis of cerebral hemorrhage. The time differences of multiple positive sample cases are statistically processed, and the average of these time differences is calculated as the early warning analysis time limit for the target emergency patient. Characteristic abnormal fluctuations refer to specific abnormal changes in monitoring data such as intracranial pressure, arterial blood pressure, and cerebral perfusion pressure that meet the diagnostic criteria for acute cerebral hemorrhage. The early warning analysis time limit refers to the average time difference between the appearance of characteristic abnormal fluctuations in the monitoring data and the final diagnosis of cerebral hemorrhage, and is the core time constraint for optimizing subsequent monitoring protocols. The continuous monitoring data of each positive sample case is analyzed one by one to locate the starting time of the characteristic abnormal fluctuations in the monitoring data, and the time difference from that time to the time of diagnosis is calculated. The arithmetic mean of the time differences of all positive samples is taken, and this average value is determined as the early warning analysis time limit for the target emergency patient.

[0028] For example, the time difference between the appearance of characteristic abnormal fluctuations in the monitoring data and the final diagnosis of cerebral hemorrhage in 450 positive sample cases were 18 minutes, 22 minutes, 20 minutes, 19 minutes, etc. The average of all time differences was calculated to be 20 minutes, that is, the warning analysis time limit for the target patient is 20 minutes.

[0029] The determination of the contribution of each monitoring indicator in the preset monitoring indicator set to the judgment of the early warning target includes: A supervised learning model is constructed to evaluate the importance of indicators, wherein the input of the supervised learning model is all the monitoring indicators in the preset monitoring indicator set, and the output of the supervised learning model is a predicted label of whether cerebral hemorrhage has occurred. The supervised learning model is trained using labeled historical case data. During the training process, the supervised learning model automatically learns the correlation strength between each input monitoring indicator and the output prediction label. After the supervised learning model training converges, the feature importance quantification value corresponding to each input monitoring indicator is extracted by the model interpreter, and the feature importance quantification value is determined as the contribution of the corresponding monitoring indicator to the judgment of the early warning target, wherein the value of the contribution is proportional to the importance of the monitoring indicator in the early warning of cerebral hemorrhage.

[0030] First, a pre-defined set of monitoring indicators for subsequent contribution calculations is defined. This set of indicators is specifically adapted to the early warning scenario of cerebral hemorrhage and consists of multiple monitoring indicators suitable for early warning of cerebral hemorrhage. Specifically, it is divided into bedside monitoring indicators and derived monitoring indicators. The bedside monitoring indicators include at least the peak systolic blood pressure, peak diastolic blood pressure, pulse pressure difference collected by invasive arterial blood pressure monitoring, and the absolute value of intracranial pressure collected by intracranial pressure monitoring. The derived monitoring indicators are generated in real time by the data processing module based on the bedside monitoring indicators. The derived monitoring indicators include at least blood pressure variability, cerebral perfusion pressure, and intracranial pressure trend slope.

[0031] Based on this, a supervised learning model is constructed to evaluate the importance of indicators. The input of the supervised learning model is all monitoring indicators in the preset monitoring indicator set, and the output of the supervised learning model is a predicted label indicating whether cerebral hemorrhage has occurred. Specifically, the supervised learning model refers to a multilayer perceptron (MLP) classification model constructed to quantify the importance of cerebral hemorrhage early warning monitoring indicators. It takes preset monitoring indicators as input and cerebral hemorrhage diagnosis results as output, and is used to learn the correlation between indicators and early warning targets.

[0032] Specifically, a 3-layer fully connected multilayer perceptron (MLP) structure is adopted. The number of neurons in the input layer is 7, consistent with the total number of preset monitoring indicators. The first hidden layer has 64 neurons, and the second hidden layer has 32 neurons. The activation function for the hidden layers is ReLU. The output layer has 1 neuron, and the activation function is Sigmoid. The output is mapped to the 0-1 interval, corresponding to the binary classification of "no cerebral hemorrhage" or "cerebral hemorrhage has occurred". The model input is the quantified value of all 7 indicators in the preset monitoring indicator set, and the output is the predicted value in the 0-1 interval. ≥0.5 is judged as cerebral hemorrhage has occurred, and <0.5 is judged as no cerebral hemorrhage has occurred, i.e., the binary classification prediction label of whether cerebral hemorrhage has occurred.

[0033] Secondly, the supervised learning model is trained using labeled historical case data. During training, the supervised learning model automatically learns the correlation strength between each input monitoring indicator and the output predicted label. Labeled historical case data refers to emergency case data from the hospital information system that are similar to the target patient's condition. Each case is labeled with a true label indicating whether acute cerebral hemorrhage was diagnosed. True labels are assigned to all historical data obtained above: 1 for diagnosed cerebral hemorrhage and 0 for undiagnosed cerebral hemorrhage. Then, all historical datasets are divided into training, validation, and test sets in a 7:2:1 ratio. Mini-batch gradient descent is used for training, with a batch size of 32 and the Adam optimizer selected. A binary cross-entropy loss function is used to quantify the deviation between the model's predicted label and the true label. The training set data is input into the model, and the model parameters are updated through backpropagation. Each iteration of the complete training set constitutes one round, while the validation set is used to evaluate the model performance. This iteration continues until the convergence condition is met.

[0034] For example, 1200 historical data points are divided into a 7:2:1 ratio: 840 training samples, 240 validation samples, and 120 test samples. The 840 training samples are input into the MLP model described above, with a batch size of 32 and an Adam optimizer learning rate of 0.001. Iterative training is performed using a binary cross-entropy loss function. The validation loss value is calculated after each epoch. During training, the model automatically learns the correlation strength between each input monitoring indicator and the output predicted label.

[0035] Furthermore, after the supervised learning model training converges, the feature importance quantification value corresponding to each input monitoring indicator is extracted through the model interpreter. This quantification value is then determined as the contribution of the corresponding monitoring indicator to the judgment of the early warning target. The magnitude of the contribution value is proportional to the importance of the monitoring indicator in the early warning of cerebral hemorrhage. Model convergence refers to the model training reaching a stable performance state, specifically defined as a decrease in the validation set loss function value of <0.001 for 10 consecutive rounds, and a validation set prediction accuracy ≥92%. The model interpreter uses the SHAP value interpretation module to quantify the contribution of each input indicator to the model output. The feature importance quantification value is a quantified value calculated by the model interpreter, reflecting the importance of the indicator in the early warning of cerebral hemorrhage, and its value ranges from 0 to 1. The contribution value is equivalent to the feature importance quantification value; a larger value indicates a higher importance of the indicator in the early warning judgment of cerebral hemorrhage.

[0036] Specifically, the convergence criteria are as follows: The model is considered converged when the binary cross-entropy loss value on the validation set decreases from the initial 0.65 to 0.12, and the loss value fluctuation is less than 0.001 for 10 consecutive epochs, and the validation set accuracy reaches 93%. Feature importance extraction: The SHAP value interpretation module is called, and the validation set data and the converged model are input to calculate the SHAP mean of each input monitoring indicator, which is the quantitative value of feature importance. Contribution determination: The SHAP mean of each indicator is directly determined as the contribution of the corresponding monitoring indicator to the judgment of the cerebral hemorrhage early warning target, and the value is proportional to the importance of the indicator's early warning.

[0037] For example, after the supervised learning model converges, the feature importance values ​​of seven indicators are extracted using the SHAP interpreter: absolute intracranial pressure contribution (0.32), cerebral perfusion pressure contribution (0.25), peak systolic blood pressure contribution (0.18), blood pressure variability contribution (0.12), peak diastolic blood pressure contribution (0.08), pulse pressure contribution (0.03), and intracranial pressure trend slope contribution (0.02). These values ​​represent the contribution of each indicator, with absolute intracranial pressure contributing the most and being the most crucial monitoring indicator for early warning of cerebral hemorrhage.

[0038] In this embodiment of the invention, by mining historical cases with similar conditions, the early warning analysis time limit for target patients is determined individually, thereby achieving precise constraints on the early warning time of cerebral hemorrhage and meeting the treatment requirements of the golden rescue period. By supervising learning and model interpretation, the early warning contribution of each monitoring indicator is quantified, the weight of the core early warning indicator is clarified, and a scientific basis is provided for the optimization of subsequent monitoring indicator combinations, avoiding the blindness of indicator selection and improving the pertinence and rationality of subsequent monitoring plans.

[0039] S300: With the goal of approximating the early warning analysis time limit, and combining the contribution of each monitoring indicator with the preset monitoring frequency, the optimal monitoring indicator group is determined by optimizing the combination scheme of the monitoring indicators in the preset monitoring indicator set. During the optimization process, the candidate monitoring indicator group is evaluated for quality through a preset monitoring scheme evaluation plugin. The quality evaluation is generated based on the prediction error probability of the early warning target.

[0040] In this embodiment of the invention, with the goal of approximating the early warning analysis time limit, the optimal monitoring indicator group is determined by combining the contribution of each monitoring indicator with the preset monitoring frequency and performing a combination scheme optimization on the monitoring indicators in the preset monitoring indicator set. During the optimization process, a preset monitoring scheme evaluation plugin is used to perform a quality assessment on the candidate monitoring indicator groups. This quality assessment is generated based on the prediction error probability of the early warning target. Existing emergency cerebral hemorrhage early warning monitoring schemes mostly adopt a unified monitoring mode for all indicators, without combining the indicator contribution with the early warning time limit for personalized combinations. This easily leads to two types of problems: first, full indicator monitoring causes data collection and processing time to exceed the early warning analysis time limit of the golden rescue period, resulting in a loss of early warning timeliness; second, blindly simplifying indicators leads to an increased probability of early warning errors, increasing the risk of missed or misjudged cases. Furthermore, under different disease complexities and sample sizes, the evaluation accuracy of the monitoring scheme lacks a dynamic adaptation mechanism, failing to balance early warning accuracy and timeliness constraints. Therefore, the optimization goal should be to approach the early warning analysis time limit. Combined optimization should be carried out by considering the contribution of indicators and the monitoring frequency. The error probability of the evaluation plug-in should be quantified by dynamically adapted monitoring schemes. Finally, the optimal monitoring indicator group with the lowest error probability and the time consumption ≤ the early warning time limit should be selected.

[0041] Step S300 in the method provided in this embodiment of the invention includes: Sort all monitoring indicators in the preset monitoring indicator set in descending order of their contribution values ​​to generate a contribution ranking list. Starting with the monitoring indicator with the highest contribution from the contribution ranking list, monitoring indicators are added one by one in sequence to construct multiple candidate monitoring indicator groups composed of different numbers of monitoring indicators. For each constructed candidate monitoring indicator group, the prediction error probability is calculated through the monitoring scheme evaluation plugin. At the same time, the total time required for the candidate monitoring indicator group to complete a complete data collection and processing is estimated based on the preset monitoring frequency of each monitoring indicator. The estimated total time is compared with the early warning analysis time limit. All candidate monitoring indicator groups whose total time is less than or equal to the early warning analysis time limit are selected. The candidate monitoring indicator group with the lowest prediction error probability is selected from the selected candidate monitoring indicator groups as the optimal monitoring indicator group.

[0042] First, we built a plugin for evaluating monitoring schemes.

[0043] The construction steps of the monitoring scheme evaluation plugin include: Using the aforementioned multivariate disease data as constraints, several sample monitoring indicator groups are collected based on the historical data of emergency patients' cerebral hemorrhage warnings. The proportion of historical warning error events in different sample monitoring indicator groups within the historical time range is used as the sample prediction error probability to obtain several sample prediction error probabilities. The proportion of historical warning error events is the sum of the proportion of historical misjudgment events and the proportion of historical missed judgment events. The aforementioned sample monitoring index groups and the predicted error probabilities of several samples are used as training data, and P-fold cross-partitioning with replacement is performed to obtain P sample training sets. Deep learning models are trained to convergence using the P sample training sets to obtain P monitoring scheme evaluation units, which are then integrated and constructed according to the mean fusion strategy to build the monitoring scheme evaluation plugin.

[0044] First, using the aforementioned multivariate disease data as constraints, several sample monitoring indicator groups are collected based on the historical data of emergency patients' cerebral hemorrhage warnings. The proportion of historical warning error events in different sample monitoring indicator groups within the historical time range is calculated as the sample prediction error probability, resulting in several sample prediction error probabilities. The proportion of historical warning error events is the sum of the proportion of historical misjudgment events and the proportion of historical missed judgment events.

[0045] The sample monitoring indicator group refers to a sample of monitoring schemes with different combinations of indicators extracted from the historical records of emergency cerebral hemorrhage warnings, constrained by the target patient's multivariate disease data. The sample prediction error probability is the percentage of historical warning error events corresponding to a certain sample monitoring indicator group, which is the sum of the percentage of misjudged events and the percentage of missed judgment events in that group. A misjudgment means that a non-cerebral hemorrhage patient is warned as having cerebral hemorrhage, and a missed judgment means that a cerebral hemorrhage patient is not warned. The multivariate disease data constraint means that only historical data with similar disease backgrounds to the target patient are collected.

[0046] Specifically, constrained by the diverse disease data of the target patients, historical data on cerebral hemorrhage warnings of similar emergency patients are retrieved from the hospital information system; different combinations of monitoring indicators are extracted from the historical data to form several sample monitoring indicator groups; for each sample monitoring indicator group, the number of misjudged events and the number of missed events within the historical time range are counted, and the historical warning error event ratio = (number of misjudged events + number of missed events) / total number of warning events × 100%, and the result is used as the sample prediction error probability of that group.

[0047] For example, 50 different sample monitoring indicator groups were extracted from the historical records of 1200 high-risk stratified patients aged 55-65 years with intracerebral hemorrhage. Among them, the intracranial pressure absolute value group alone generated 800 warning events, with 80 misjudgments and 40 missed judgments, and its sample prediction error probability = (80+40) / 800×100%=15%; the intracranial pressure absolute value + cerebral perfusion pressure group generated 900 warning events, with 36 misjudgments and 36 missed judgments, and the sample prediction error probability = (36+36) / 900×100%=8%.

[0048] Secondly, the aforementioned sample monitoring index groups and sample prediction error probabilities are used as training data, and a P-fold cross-partition with replacement is performed to obtain a P-part training set. P-fold cross-partition refers to dividing the training data into P equal parts and constructing the training set through random selection with replacement. Here, P is an integer ≥ 10. Random selection with replacement means that each time data parts are selected, previously selected parts are not excluded and can be repeatedly selected into the training set. The sample training set refers to the dataset consisting of the P data parts selected with replacement, used to train a single evaluation unit.

[0049] For example, 50 sets of training data are equally divided into 10 parts, each with 5 groups, numbered 1-10; in the first round, 10 random selections are made with replacement, and the selected parts are [1,3,5,2,10,3,7,8,5,9]. These parts are then merged to obtain the first training set; in the second round, [2,4,6,1,9,4,8,7,6,10] are selected and merged to obtain the second training set; and so on, until 10 training sets are finally generated.

[0050] Furthermore, deep learning models are trained to convergence using the P sample training sets, resulting in P monitoring scheme evaluation units. These units are then integrated and constructed as a monitoring scheme evaluation plugin using a mean fusion strategy. Each monitoring scheme evaluation unit is a deep learning model trained to convergence using a single sample training set, capable of independently outputting the prediction error probability of candidate indicator groups. The model convergence condition is that the loss function value of the deep learning model decreases by less than 0.001 over 10 consecutive rounds, and the prediction error on the validation set is less than 5%. The mean fusion strategy refers to taking the arithmetic mean of the prediction results from multiple evaluation units as the final output of the plugin.

[0051] Specifically, a lightweight fully connected deep learning model is selected, with the number of neurons in the input layer equal to the feature dimension of the monitoring index group, two hidden layers (64 or 32 neurons respectively) with ReLU activation, and one neuron in the output layer. The model is trained with P sample training sets until all models reach convergence, resulting in P independent monitoring scheme evaluation units. Based on the mean fusion strategy, the P evaluation units are encapsulated into a unified monitoring scheme evaluation plugin. After the plugin is input into the candidate monitoring index group, it calls the error probabilities of all evaluation units and takes the arithmetic mean as the final prediction error probability of the candidate group.

[0052] For example, 10 fully connected deep learning models are trained using 10 training sets, each trained until the loss function decreases from the initial 0.8 to 0.1 and converges. The 10 fully connected deep learning models are integrated into a plugin. When the absolute value of intracranial pressure + cerebral perfusion pressure of the candidate group is input, the error probabilities output by the 10 units are 7.8%, 8.2%, 7.9%, 8.1%, 8.0%, 7.7%, 8.3%, 8.0%, 7.9%, and 8.1%, respectively. The final prediction error probability output by the plugin is (7.8 + 8.2 + 7.9 + 8.1 + 8.0 + 7.7 + 8.3 + 8.0 + 7.9 + 8.1) / 10 = 8.0%.

[0053] Based on this, all monitoring indicators in the preset monitoring indicator set are sorted in descending order of their contribution values ​​to generate a contribution ranking list. The contribution ranking list is formed by sorting the indicators in the preset monitoring indicator set in descending order of their contribution values, with the contribution value being directly proportional to the importance of the indicator's early warning function for cerebral hemorrhage. The contribution values ​​of each monitoring indicator determined in S200 are extracted, and the seven preset monitoring indicators are sorted in descending order to generate a structured contribution ranking list.

[0054] For example, based on the contribution of the indicators for the target patient: absolute intracranial pressure contribution 0.32, cerebral perfusion pressure contribution 0.25, peak systolic blood pressure contribution 0.18, blood pressure variability contribution 0.12, peak diastolic blood pressure contribution 0.08, pulse pressure contribution 0.03, and intracranial pressure trend slope contribution 0.02, a sorted list is generated: absolute intracranial pressure, cerebral perfusion pressure, peak systolic blood pressure, blood pressure variability, peak diastolic blood pressure, pulse pressure, and intracranial pressure trend slope.

[0055] Secondly, starting with the monitoring indicator with the highest contribution in the contribution ranking list, monitoring indicators are added sequentially one by one to construct multiple candidate monitoring indicator groups composed of different numbers of monitoring indicators. A candidate monitoring indicator group refers to a set of monitoring schemes with different combinations of indicators, constructed by progressively adding indicators from high to low in the contribution ranking list. Starting with the indicator at the top of the contribution ranking list, the next indicator is added sequentially to construct candidate groups containing different numbers of indicators, with only one new indicator added to each group, until all monitoring indicators are included.

[0056] For example, seven candidate indicator groups are constructed according to the sorting results: Candidate group 1: Absolute value of intracranial pressure; Candidate group 2: Absolute value of intracranial pressure and cerebral perfusion pressure; Candidate group 3: Absolute value of intracranial pressure, cerebral perfusion pressure and peak systolic blood pressure; ...; Candidate group 7: All seven indicators.

[0057] Furthermore, for each constructed candidate monitoring indicator group, the prediction error probability is calculated using the monitoring scheme evaluation plugin, and the total time required for the candidate monitoring indicator group to complete one full data collection and processing is estimated based on the preset monitoring frequency of each monitoring indicator.

[0058] Specifically, for each constructed candidate monitoring indicator group, the prediction error probability is calculated using the monitoring scheme evaluation plugin, including: The disease complexity index of the target emergency patient is calculated based on the multivariate disease data, wherein the disease complexity index is comprehensively evaluated based on at least the age value, the number of comorbidities, and the variability of the initial vital signs data. Obtain the preset baseline disease complexity index and the preset baseline early warning analysis time limit; The ratio of the disease complexity index to the baseline disease complexity index is calculated and used as the compensation coefficient for the first unit. Calculate the ratio of the baseline early warning analysis time limit to the early warning analysis time limit, and use it as the compensation coefficient for the second unit; A weighted summation operation is performed on the compensation coefficients selected by the first unit and the second unit to generate the compensation coefficients selected by the adaptation unit. The product of the adaptation unit selection compensation coefficient and the initial unit selection quantity K is rounded down to obtain Q. Q monitoring scheme evaluation units are randomly selected from the P monitoring scheme evaluation units to evaluate the quality of the candidate monitoring index group. Here, K is 1 / 3 of P. If Q is less than 1, Q is set to 1. If Q is greater than P, Q is set to P.

[0059] First, the complexity index of the target emergency patient's condition is calculated based on the multivariate medical data. This complexity index is comprehensively assessed based on at least age, number of comorbidities, and the variability of initial vital sign data. The complexity index is a numerical value that comprehensively quantifies the complexity of the target patient's condition, calculated by weighting age, number of comorbidities, and the variability of initial vital sign data; a higher value indicates a more complex condition. The variability of initial vital sign data refers to the dispersion of initial data such as heart rate, blood pressure, respiratory rate, and blood oxygen saturation in the target patient. Variation = standard deviation / mean, reflecting the stability of vital signs. Multivariate medical data is extracted, and the standard deviation and mean of each vital sign indicator are calculated separately. The average value is taken as the overall variability. The complexity index is calculated using the weighted formula: Complexity Index = 0.3 × (age / 100) + 0.4 × number of comorbidities + 0.3 × variability of initial vital sign data. The weighting is based on clinical guidelines: age accounts for 30%, comorbidities for 40%, and vital sign stability for 30%.

[0060] For example, the target patient was a 58-year-old male with two comorbidities: hypertension and diabetes. The initial vital signs data were as follows: mean heart rate 102 beats / min, standard deviation 5 beats / min; mean systolic blood pressure 105 mmHg, standard deviation 8 mmHg; mean respiratory rate 24 breaths / min, standard deviation 3 breaths / min; mean blood oxygen saturation 92%, standard deviation 2%. Variation of each indicator: Heart rate variation = 5 / 102≈0.049, Systolic blood pressure variation = 8 / 105≈0.076, Respiratory rate variation = 3 / 24=0.125, Blood oxygen saturation variation = 2 / 92≈0.022; Overall variation = (0.049+0.076+0.125+0.022) / 4≈0.068; Complexity index of the condition = 0.3×(58 / 100)+0.4×2+0.3×0.068=0.174+0.8+0.0204=0.9944.

[0061] Secondly, obtain the preset baseline disease complexity index and the preset baseline warning analysis time limit. The baseline disease complexity index is a preset standardized reference value for disease complexity, obtained based on statistics from a large number of emergency cerebral hemorrhage patients, with a default value of 1.0, representing a moderately complex cerebral hemorrhage patient. The baseline warning analysis time limit refers to the preset average warning time limit for similar emergency conditions, without needing to be limited to the same condition as the target patient. It is used as a reference benchmark for the urgency of the warning time limit and can be customized according to the clinical scenario. For example, if the retrieved baseline disease complexity index = 1.0 and baseline warning analysis time limit = 30 minutes; and S200 indicates that the warning analysis time limit for the target patient is 20 minutes, it means that their condition is deteriorating faster than the baseline level.

[0062] Subsequently, the ratio of the disease complexity index to the baseline disease complexity index is calculated and used as the compensation coefficient for the first unit selection. The first unit selection compensation coefficient reflects the multiple of the target patient's disease complexity relative to the baseline level. A value > 1 indicates a more complex disease, requiring more evaluation units to ensure assessment accuracy; a value < 1 indicates a simpler disease, allowing for fewer evaluation units. It is calculated using the formula: First Unit Selection Compensation Coefficient = Disease Complexity Index / Baseline Disease Complexity Index. For example, a first unit selection compensation coefficient = 0.9944 / 1.0 = 0.9944 ≈ 1, indicating that the target patient's disease complexity is basically consistent with the baseline level.

[0063] Next, the ratio of the baseline early warning analysis time limit to the target patient's early warning analysis time limit is calculated and used as the compensation coefficient for the second unit selection. The compensation coefficient for the second unit selection reflects the multiple of the urgency of the target patient's early warning time limit relative to the baseline level. A larger value indicates a shorter early warning time limit, meaning a faster deterioration of the condition, requiring higher assessment timeliness and accuracy, and necessitating more assessment units; a smaller value indicates a longer early warning time limit, allowing for fewer assessment units and saving resources. The formula is: Compensation coefficient for the second unit selection = Baseline early warning analysis time limit / Target patient early warning analysis time limit. For example, a compensation coefficient for the second unit selection of 30 minutes / 20 minutes = 1.5 indicates that the target patient's early warning time limit is shorter than the baseline, the condition is deteriorating faster, and more assessment units are needed to ensure assessment accuracy.

[0064] Then, a weighted summation operation is performed on the compensation coefficients selected by the first unit and the second unit to generate the compensation coefficients selected by the adaptation unit.

[0065] The process of generating the adaptation unit selection compensation coefficients involves performing a weighted summation operation on the compensation coefficients selected by the first unit and the second unit, including: The total amount of historical sample data used to mine the early warning analysis time limit is statistically analyzed, wherein the total amount of historical sample data is the number of positive sample cases diagnosed with cerebral hemorrhage; The ratio of the total amount of historical sample data to the preset sample size threshold is used as the second weight compensation coefficient to compensate and adjust the second initial weight, thereby obtaining the second adaptive weight. The first adaptation weight is obtained by subtracting the second adaptation weight from 1. Based on the first adaptation weight and the second adaptation weight, a weighted summation operation is performed on the compensation coefficients selected by the first unit and the second unit to generate the compensation coefficients selected by the adaptation unit.

[0066] First, the total amount of historical sample data used for mining the aforementioned early warning analysis timeframe is calculated. This total amount of historical sample data represents the number of positive sample cases diagnosed with cerebral hemorrhage. The total amount of historical sample data refers to the number of historical positive sample cases used for mining the early warning analysis timeframe for the target patient, i.e., the number of cases diagnosed with acute cerebral hemorrhage. A larger value indicates more accurate early warning analysis timeframe mining results. The historical case dataset previously used for mining the early warning analysis timeframe is retrieved from the hospital information system, and the number of positive sample cases diagnosed with acute cerebral hemorrhage is counted. For example, the number of historical positive sample cases retrieved for mining the early warning analysis timeframe is 450.

[0067] Secondly, the ratio of the total historical sample data to the preset sample size threshold is used as the second weight compensation coefficient to compensate and adjust the second initial weight, resulting in the second adaptive weight. The second weight compensation coefficient is a coefficient used to adjust the second initial weight, with a value range limited to 0.7~1.3, determined by the ratio of the total historical sample data to the preset sample size threshold. The second initial weight is the preset initial weight of the second unit selection compensation coefficient, with a default value of 0.5. The second adaptive weight is the final weight after adjustment by the second weight compensation coefficient; a larger value indicates a higher proportion of the second unit selection compensation coefficient in the total compensation coefficient. The preset sample size threshold is set to 500 cases, and the second initial weight is set to 0.5; the initial second weight compensation coefficient = total historical sample data / preset sample size threshold; if the initial second weight compensation coefficient < 0.7, then 0.7 is used; if the initial second weight compensation coefficient > 1.3, then 1.3 is used; the second adaptive weight = second initial weight × constrained second weight compensation coefficient.

[0068] For example, the initial second weight compensation coefficient = 450 / 500 = 0.9, which is in the range of 0.7 to 1.3 and does not need to be adjusted; the second adaptation weight = 0.5 × 0.9 = 0.45, indicating that because the sample size is slightly lower than the threshold, the warning time limit weight is slightly lower than the initial level.

[0069] Next, the first adaptation weight is obtained by subtracting the second adaptation weight from 1. The first adaptation weight refers to the final weight of the compensation coefficient selected in the first unit, obtained by subtracting the second adaptation weight from 1, ensuring that the sum of the two weights is 1; the larger the sample size, the higher the second adaptation weight and the lower the first adaptation weight. It is calculated using the formula: First adaptation weight = 1 - Second adaptation weight. For example, the first adaptation weight = 1 - 0.45 = 0.55, meaning the complexity of the condition accounts for 55% and the urgency of the warning time limit accounts for 45%.

[0070] Subsequently, based on the first and second adaptation weights, a weighted summation operation is performed on the first unit selection compensation coefficient and the second unit selection compensation coefficient to generate the adaptation unit selection compensation coefficient. The adaptation unit selection compensation coefficient is the final compensation coefficient that comprehensively considers the complexity of the condition and the urgency of the warning time limit, used to determine the number of evaluation units to be selected. It is calculated using the weighted summation formula: Adaptation unit selection compensation coefficient = First adaptation weight × First unit selection compensation coefficient + Second adaptation weight × Second unit selection compensation coefficient. For example, the adaptation unit selection compensation coefficient = 0.55 × 0.9944 + 0.45 × 1.5 ≈ 1.22.

[0071] Further, the product of the adaptation unit selection compensation coefficient and the initial unit selection quantity K is rounded down to obtain Q. Then, Q monitoring scheme evaluation units are randomly selected from the P monitoring scheme evaluation units to perform quality assessment on the candidate monitoring indicator group. Here, K is 1 / 3 of P; if Q is less than 1, Q is set to 1; if Q is greater than P, Q is set to P. The initial unit selection quantity K is a preset initial selection quantity of evaluation units, which is 1 / 3 of P. Q is the final number of selected evaluation units, obtained by rounding down the adaptation unit selection compensation coefficient × K, and is limited to Q ≥ 1 and Q ≤ P. A monitoring scheme evaluation unit refers to an independent deep learning model constituting the monitoring scheme evaluation plugin, and its number is P.

[0072] For example, P=10, K=P / 3≈3.33; Q=rounded down (1.2219×3.33)≈rounded down (4.069)=4; Q=4 is in the range of 1~10 and does not need to be adjusted; therefore, 4 are randomly selected from the 10 monitoring scheme evaluation units to calculate the prediction error probability of the candidate indicator group.

[0073] Furthermore, the total time required for a candidate monitoring indicator group to complete one full data collection and processing cycle is estimated based on the preset monitoring frequency for each monitoring indicator. The preset monitoring frequency, determined by clinical monitoring guidelines, is the pre-defined collection frequency for each monitoring indicator to ensure the timeliness of the monitoring data. The total time refers to the cumulative time for a candidate monitoring indicator group to complete one full data collection, transmission, and processing cycle, which must meet the minimum requirements of the preset monitoring frequencies for all indicators in that group. The preset monitoring frequencies and individual indicator times for each monitoring indicator clarify the collection frequency for each indicator in the preset monitoring indicator group, as well as the time required for each individual indicator to complete one full data collection, transmission, and processing cycle. For each candidate monitoring indicator group, the individual indicator times for all monitoring indicators within the group are summed to obtain the total time required for that candidate group to complete one full data collection and processing cycle.

[0074] For example, the preset time for each monitoring indicator is as follows: Absolute value of intracranial pressure: frequency 1 time / 1 minute, single time 2 minutes; Cerebral perfusion pressure: frequency 1 time / 2 minutes, single time 1.5 minutes; Peak systolic blood pressure: frequency 1 time / 3 minutes, single time 1 minute; The remaining 4 indicators: frequency 1 time / 5 minutes, single time 1 minute each; The total time for each candidate group is calculated as follows: Candidate group 1: 2 minutes; Candidate group 2: 2+1.5=3.5 minutes; Candidate group 3: 2+1.5+1=4.5 minutes; ...; Candidate group 7: 2+1.5+1+1+1+1+1=8.5 minutes.

[0075] Finally, the estimated total time is compared with the warning analysis time limit. All candidate monitoring indicator groups whose total time is less than or equal to the warning analysis time limit are selected, and the candidate monitoring indicator group with the lowest prediction error probability is chosen as the optimal monitoring indicator group. The warning analysis time limit refers to the average time from abnormal indicators to diagnosis of cerebral hemorrhage for the target patient, obtained through historical sample mining, and is a constraint threshold for the total time. The optimal monitoring indicator group is the candidate monitoring indicator group that satisfies the condition of total time ≤ warning analysis time limit and has the lowest prediction error probability, balancing the timeliness and accuracy of the warning. The total time of each candidate monitoring indicator group is compared with the warning analysis time limit for the target patient, and all candidate monitoring indicator groups with a total time ≤ warning analysis time limit are selected, while candidate groups with a total time exceeding the time limit are eliminated. Among the selected candidate monitoring indicator groups, the prediction error probability of each group is compared, and the candidate monitoring indicator group with the lowest prediction error probability is selected as the final optimal monitoring indicator group.

[0076] For example, the target patient early warning analysis time limit is 20 minutes, and the total time for all candidate groups is ≤20 minutes, so all of them pass the screening; the prediction error probability of each group is as follows: candidate group 1 prediction error probability 15%, candidate group 2 prediction error probability 8%, candidate group 3 prediction error probability 6%, ..., candidate group 7 prediction error probability 3%; the candidate group 7 with the lowest prediction error probability is selected as the optimal monitoring indicator group; if the clinic needs to take into account the minimum necessary indicators, the candidate group with the fewest indicators can be selected on the premise that the error probability meets the clinical requirements.

[0077] In this embodiment of the invention, a dynamically adaptable monitoring scheme evaluation plugin is constructed to accurately assess the probability of prediction errors under different disease complexities and sample sizes. Candidate indicator groups are constructed by ranking contributions, avoiding the blind inclusion of non-core indicators. The total time consumption is screened using the early warning analysis time limit as a constraint, ensuring the timeliness of the monitoring scheme. Simultaneously, the group with the lowest error probability is selected, balancing the requirements of early warning accuracy and timeliness. The ultimately determined optimal monitoring indicator group not only meets the time limit requirement of the golden rescue period for cerebral hemorrhage but also minimizes the risk of missed or misjudged early warnings, while avoiding resource waste and information overload caused by full indicator monitoring.

[0078] S400: Real-time monitoring of vital signs of the target emergency patient is performed based on the optimal monitoring scheme composed of the optimal monitoring index group.

[0079] In this embodiment of the invention, the target emergency patient's vital signs are monitored in real time based on the optimal monitoring scheme composed of the optimal monitoring indicator group. An optimal monitoring scheme is constructed based on the determined optimal monitoring indicator group. Continuous, real-time data collection is performed on the target emergency patient according to the preset monitoring frequency of each monitoring indicator in the optimal monitoring scheme. The collected indicator data is processed and analyzed in real time by the data processing module. The analyzed indicator data is compared in real time with a preset cerebral hemorrhage warning threshold to continuously determine the risk of cerebral hemorrhage. Once the indicator data triggers the warning threshold, a cerebral hemorrhage warning signal is immediately generated and pushed to the emergency medical terminal, achieving closed-loop real-time monitoring of vital signs.

[0080] For example, taking a 58-year-old high-risk cerebral hemorrhage emergency patient as an example, the optimal monitoring indicator group consists of seven indicators: absolute intracranial pressure, cerebral perfusion pressure, peak systolic blood pressure, blood pressure variability, peak diastolic blood pressure, pulse pressure difference, and intracranial pressure trend slope. Data is collected in real time according to the preset monitoring frequency of each indicator. The data processing module calculates the values ​​of derived monitoring indicators in real time and compares the real-time values ​​of each indicator with the cerebral hemorrhage warning threshold. For example, when the absolute intracranial pressure exceeds 20 mmHg and the blood pressure variability is greater than 15%, a high-risk cerebral hemorrhage warning is immediately issued to the medical terminal, completing the entire process of real-time monitoring and warning triggering.

[0081] In this embodiment of the invention, individualized real-time monitoring is implemented through an optimal monitoring scheme. While ensuring the accuracy of cerebral hemorrhage early warning, it strictly meets the timeliness requirements of early warning analysis, avoids information overload and waste of medical resources caused by full-indicator monitoring, and achieves early and real-time early warning of cerebral hemorrhage risk, thus gaining golden rescue time for emergency cerebral hemorrhage patients and effectively improving the timeliness, accuracy and clinical applicability of emergency cerebral hemorrhage monitoring.

[0082] Through the specific implementation methods described above, the embodiments of the present invention achieve the following technical effects: This invention provides a method and system for real-time monitoring of vital signs in emergency patients using multivariate data mining. By collecting multivariate patient data step-by-step, it provides comprehensive and accurate data support for subsequent analysis. Based on data mining, it individually determines the early warning analysis timeframe and quantifies the contribution of monitoring indicators, achieving precise constraints on the timeliness of early warning for cerebral hemorrhage and scientific positioning of core indicators. Combining contribution ranking with early warning timeframes, it optimizes indicator combinations and completes quality assessment through a monitoring scheme evaluation plugin that dynamically adapts to the complexity of the patient's condition, early warning timeframe, and sample size, selecting the optimal monitoring indicator set that balances timeliness and accuracy. This effectively solves the problems of information overload, resource waste, and insufficient early warning accuracy. Finally, based on the optimal monitoring scheme, it implements real-time monitoring of vital signs, strictly adhering to the golden rescue window for cerebral hemorrhage while achieving high-precision adaptive monitoring for high-risk patients and high-efficiency monitoring for low-risk patients. This comprehensively improves the timeliness, accuracy, and clinical applicability of early warning for emergency cerebral hemorrhage, providing reliable technical support for the early treatment of patients with emergency cerebral hemorrhage.

[0083] Example 2, as Figure 2 As shown, this invention provides a real-time monitoring system for vital signs of emergency patients based on multivariate data mining, the system comprising: The multi-dimensional data acquisition module 11 is used to collect multi-dimensional disease data of the target emergency patients; The time limit and contribution determination module 12 is used to determine the early warning analysis time limit of the target emergency patient by means of multivariate data mining, with the multivariate disease data as constraints and the early warning target of cerebral hemorrhage as the early warning target, and to determine the contribution of each monitoring indicator in the preset monitoring indicator set to the judgment of the early warning target. The combination scheme optimization module 13 is used to optimize the combination scheme of the monitoring indicators in the preset monitoring indicator set by taking the approximation of the early warning analysis time limit as the optimization target, and combining the contribution of each monitoring indicator and the preset monitoring frequency to determine the optimal monitoring indicator group. In the optimization process, the candidate monitoring indicator group is evaluated by a preset monitoring scheme evaluation plugin. The quality evaluation is generated based on the prediction error probability of the early warning target. The vital signs monitoring module 14 is used to monitor the vital signs of the target emergency patient in real time based on the optimal monitoring scheme composed of the optimal monitoring index group.

[0084] In one embodiment, the multi-source data acquisition module 11 is further configured to: Collect basic characteristic data of the target emergency patient, wherein the basic characteristic data includes at least age, gender, and history of cerebral hemorrhage; Collect preliminary vital sign data of the target emergency patient, wherein the preliminary vital sign data includes at least heart rate, blood pressure, respiratory rate and blood oxygen saturation; Collect bedside laboratory data from the target emergency patients, wherein the bedside laboratory data includes at least the international normalized ratio, activated partial thromboplastin time, and blood glucose concentration; The aforementioned basic characteristic data, initial vital sign data, and bedside test data are used as multivariate disease data.

[0085] In one embodiment, the time limit and contribution determination module 12 is further configured to: The preset monitoring indicator set comprises multiple monitoring indicators suitable for early warning of cerebral hemorrhage. This set includes at least bedside monitoring indicators and derived monitoring indicators. The bedside monitoring indicators at least include peak systolic blood pressure, peak diastolic blood pressure, pulse pressure difference collected via invasive arterial blood pressure monitoring, and absolute intracranial pressure collected via intracranial pressure monitoring. The derived monitoring indicators are generated in real-time from the bedside monitoring indicators through a data processing module, and at least include blood pressure variability, cerebral perfusion pressure, and intracranial pressure trend slope.

[0086] Specifically, using the aforementioned multivariate medical data as constraints and cerebral hemorrhage early warning as the early warning target, the early warning analysis time limit for the target emergency patients is determined through multivariate data mining, including: Using the aforementioned multivariate medical data as constraints, a historical case dataset with similar medical conditions to the target emergency patient is retrieved from the hospital information system. The similar medical conditions include at least the same age range and the same risk stratification for cerebral hemorrhage. In the historical case dataset, positive sample cases that were finally diagnosed as acute cerebral hemorrhage are marked, and continuous monitoring data of each positive sample case before the time of diagnosis are obtained; Analyze the continuous monitoring data of each positive sample case to find the time difference between the appearance of characteristic abnormal fluctuations in the monitoring data and the final diagnosis of cerebral hemorrhage. Perform statistical processing on the time differences of multiple positive sample cases and calculate the average of the multiple time differences as the early warning analysis time limit for the target emergency patient.

[0087] The determination of the contribution of each monitoring indicator in the preset monitoring indicator set to the judgment of the early warning target includes: A supervised learning model is constructed to evaluate the importance of indicators, wherein the input of the supervised learning model is all the monitoring indicators in the preset monitoring indicator set, and the output of the supervised learning model is a predicted label of whether cerebral hemorrhage has occurred. The supervised learning model is trained using labeled historical case data. During the training process, the supervised learning model automatically learns the correlation strength between each input monitoring indicator and the output prediction label. After the supervised learning model training converges, the feature importance quantification value corresponding to each input monitoring indicator is extracted by the model interpreter, and the feature importance quantification value is determined as the contribution of the corresponding monitoring indicator to the judgment of the early warning target, wherein the value of the contribution is proportional to the importance of the monitoring indicator in the early warning of cerebral hemorrhage.

[0088] In one embodiment, the combination scheme optimization module 13 is further configured to: Sort all monitoring indicators in the preset monitoring indicator set in descending order of their contribution values ​​to generate a contribution ranking list. Starting with the monitoring indicator with the highest contribution from the contribution ranking list, monitoring indicators are added one by one in sequence to construct multiple candidate monitoring indicator groups composed of different numbers of monitoring indicators. For each constructed candidate monitoring indicator group, the prediction error probability is calculated through the monitoring scheme evaluation plugin. At the same time, the total time required for the candidate monitoring indicator group to complete a full data collection and processing is estimated based on the preset monitoring frequency of each monitoring indicator. The estimated total time is compared with the early warning analysis time limit. All candidate monitoring indicator groups whose total time is less than or equal to the early warning analysis time limit are selected. The candidate monitoring indicator group with the lowest prediction error probability is selected from the selected candidate monitoring indicator groups as the optimal monitoring indicator group.

[0089] The construction steps of the monitoring scheme evaluation plugin include: Using the aforementioned multivariate disease data as constraints, several sample monitoring indicator groups are collected based on the historical data of emergency patients' cerebral hemorrhage warnings. The proportion of historical warning error events in different sample monitoring indicator groups within the historical time range is used as the sample prediction error probability to obtain several sample prediction error probabilities. The proportion of historical warning error events is the sum of the proportion of historical misjudgment events and the proportion of historical missed judgment events. The aforementioned sample monitoring index groups and the predicted error probabilities of several samples are used as training data, and P-fold cross-partitioning with replacement is performed to obtain P sample training sets. Deep learning models are trained to convergence using the P sample training sets to obtain P monitoring scheme evaluation units, which are then integrated and constructed according to the mean fusion strategy to build the monitoring scheme evaluation plugin.

[0090] Specifically, for each constructed candidate monitoring indicator group, the prediction error probability is calculated using the monitoring scheme evaluation plugin, including: The disease complexity index of the target emergency patient is calculated based on the multivariate disease data, wherein the disease complexity index is comprehensively evaluated based on at least the age value, the number of comorbidities, and the variability of the initial vital signs data. Obtain the preset baseline disease complexity index and the preset baseline early warning analysis time limit; The ratio of the disease complexity index to the baseline disease complexity index is calculated and used as the compensation coefficient for the first unit. Calculate the ratio of the baseline early warning analysis time limit to the early warning analysis time limit, and use it as the compensation coefficient for the second unit; A weighted summation operation is performed on the compensation coefficients selected by the first unit and the second unit to generate the compensation coefficients selected by the adaptation unit. The product of the adaptation unit selection compensation coefficient and the initial unit selection quantity K is rounded down to obtain Q. Q monitoring scheme evaluation units are randomly selected from the P monitoring scheme evaluation units to evaluate the quality of the candidate monitoring index group. Here, K is 1 / 3 of P. If Q is less than 1, Q is set to 1. If Q is greater than P, Q is set to P.

[0091] The process of generating the adaptation unit selection compensation coefficients involves performing a weighted summation operation on the compensation coefficients selected by the first unit and the second unit, including: The total amount of historical sample data used to mine the early warning analysis time limit is statistically analyzed, wherein the total amount of historical sample data is the number of positive sample cases diagnosed with cerebral hemorrhage; The ratio of the total amount of historical sample data to the preset sample size threshold is used as the second weight compensation coefficient to compensate and adjust the second initial weight, thereby obtaining the second adaptive weight. The first adaptation weight is obtained by subtracting the second adaptation weight from 1. Based on the first adaptation weight and the second adaptation weight, a weighted summation operation is performed on the compensation coefficients selected by the first unit and the second unit to generate the compensation coefficients selected by the adaptation unit.

[0092] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for real-time monitoring of vital signs in emergency patients using multivariate data mining, characterized in that, The methods include: Collect diverse medical data of target emergency patients; Using the aforementioned multivariate medical data as constraints and cerebral hemorrhage warning as the warning target, the warning analysis time limit for the target emergency patients is determined through multivariate data mining, and the contribution of each monitoring indicator in the preset monitoring indicator set to the judgment of the warning target is determined. With the goal of approximating the early warning analysis time limit, and combining the contribution of each monitoring indicator with the preset monitoring frequency, the optimal monitoring indicator group is determined by optimizing the combination scheme of the monitoring indicators in the preset monitoring indicator set. During the optimization process, the candidate monitoring indicator group is evaluated by a preset monitoring scheme evaluation plugin. The quality evaluation is generated based on the prediction error probability of the early warning target. The optimal monitoring scheme, based on the optimal monitoring index group, is used to monitor the vital signs of the target emergency patient in real time.

2. The method for real-time monitoring of vital signs in emergency patients using multivariate data mining according to claim 1, characterized in that, Collect diverse medical data on the target emergency patients, including: Collect basic characteristic data of the target emergency patient, wherein the basic characteristic data includes at least age, gender, and history of cerebral hemorrhage; Collect preliminary vital sign data of the target emergency patient, wherein the preliminary vital sign data includes at least heart rate, blood pressure, respiratory rate and blood oxygen saturation; Collect bedside laboratory data from the target emergency patients, wherein the bedside laboratory data includes at least the international normalized ratio, activated partial thromboplastin time, and blood glucose concentration; The aforementioned basic characteristic data, initial vital sign data, and bedside test data are used as multivariate disease data.

3. The method for real-time monitoring of vital signs in emergency patients using multivariate data mining according to claim 1, characterized in that, The preset monitoring indicator set consists of multiple monitoring indicators suitable for early warning of cerebral hemorrhage. The preset monitoring indicator set includes at least bedside monitoring indicators and derived monitoring indicators. The bedside monitoring indicators include at least the peak systolic blood pressure, peak diastolic blood pressure, pulse pressure difference collected by invasive arterial blood pressure monitoring, and the absolute value of intracranial pressure collected by intracranial pressure monitoring. The derived monitoring indicators are generated in real time by a data processing module from the bedside monitoring indicators. The derived monitoring indicators include at least blood pressure variability, cerebral perfusion pressure, and intracranial pressure trend slope.

4. The method for real-time monitoring of vital signs in emergency patients using multivariate data mining according to claim 1, characterized in that, Using the aforementioned multivariate medical data as constraints and cerebral hemorrhage warning as the early warning target, the early warning analysis time limit for the target emergency patients is determined through multivariate data mining, including: Using the aforementioned multivariate medical data as constraints, retrieve historical case datasets from the hospital information system that have similar medical conditions to the target emergency patient; In the historical case dataset, positive sample cases that were finally diagnosed as acute cerebral hemorrhage were marked, and continuous monitoring data of each positive sample case before the time of diagnosis were obtained; Analyze the continuous monitoring data of each positive sample case to find the time difference between the appearance of characteristic abnormal fluctuations in the monitoring data and the final diagnosis of cerebral hemorrhage. Perform statistical processing on the time differences of multiple positive sample cases and calculate the average of the multiple time differences as the early warning analysis time limit for the target emergency patient.

5. The method for real-time monitoring of vital signs of emergency patients using multivariate data mining according to claim 1, characterized in that, Determine the contribution of each monitoring indicator in the preset monitoring indicator set to the judgment of the early warning target, including: A supervised learning model is constructed to evaluate the importance of indicators, wherein the input of the supervised learning model is all the monitoring indicators in the preset monitoring indicator set, and the output of the supervised learning model is a predicted label of whether cerebral hemorrhage has occurred. The supervised learning model is trained using labeled historical case data. During the training process, the supervised learning model automatically learns the correlation strength between each input monitoring indicator and the output prediction label. After the supervised learning model training converges, the feature importance quantification value corresponding to each input monitoring indicator is extracted by the model interpreter, and the feature importance quantification value is determined as the contribution of the corresponding monitoring indicator to the judgment of the early warning target, wherein the value of the contribution is proportional to the importance of the monitoring indicator in the early warning of cerebral hemorrhage.

6. The method for real-time monitoring of vital signs of emergency patients using multivariate data mining according to claim 1, characterized in that, With the goal of approximating the aforementioned early warning analysis time limit, and combining the contribution of each monitoring indicator with a preset monitoring frequency, the optimal combination scheme of monitoring indicators in the preset monitoring indicator set is determined, including: Sort all monitoring indicators in the preset monitoring indicator set in descending order of their contribution values ​​to generate a contribution ranking list. Starting with the monitoring indicator with the highest contribution from the contribution ranking list, monitoring indicators are added one by one in sequence to construct multiple candidate monitoring indicator groups composed of different numbers of monitoring indicators. For each constructed candidate monitoring indicator group, the prediction error probability is calculated through the monitoring scheme evaluation plugin. At the same time, the total time required for the candidate monitoring indicator group to complete a complete data collection and processing is estimated based on the preset monitoring frequency of each monitoring indicator. The estimated total time is compared with the early warning analysis time limit. All candidate monitoring indicator groups whose total time is less than or equal to the early warning analysis time limit are selected. The candidate monitoring indicator group with the lowest prediction error probability is selected from the selected candidate monitoring indicator groups as the optimal monitoring indicator group.

7. The method for real-time monitoring of vital signs of emergency patients using multivariate data mining according to claim 6, characterized in that, The construction steps of the monitoring scheme evaluation plugin include: Using the aforementioned multivariate disease data as constraints, several sample monitoring indicator groups are collected based on the historical data of emergency patients' cerebral hemorrhage warnings. The proportion of historical warning error events in different sample monitoring indicator groups within the historical time range is used as the sample prediction error probability to obtain several sample prediction error probabilities. The proportion of historical warning error events is the sum of the proportion of historical misjudgment events and the proportion of historical missed judgment events. The aforementioned sample monitoring index groups and the predicted error probabilities of several samples are used as training data, and P-fold cross-partitioning with replacement is performed to obtain P sample training sets. Deep learning models are trained to convergence using the P sample training sets to obtain P monitoring scheme evaluation units, which are then integrated and constructed according to the mean fusion strategy to build the monitoring scheme evaluation plugin.

8. The method for real-time monitoring of vital signs of emergency patients using multivariate data mining according to claim 7, characterized in that, For each constructed candidate monitoring indicator group, the prediction error probability is calculated using the monitoring scheme evaluation plugin, including: The disease complexity index of the target emergency patient is calculated based on the multivariate disease data, wherein the disease complexity index is comprehensively evaluated based on at least the age value, the number of comorbidities, and the variability of the initial vital signs data. Obtain the preset baseline disease complexity index and the preset baseline early warning analysis time limit; The ratio of the disease complexity index to the benchmark disease complexity index is calculated and used as the compensation coefficient for the first unit. Calculate the ratio of the baseline early warning analysis time limit to the early warning analysis time limit, and use it as the compensation coefficient for the second unit; A weighted summation operation is performed on the compensation coefficients selected by the first unit and the second unit to generate the compensation coefficients selected by the adaptation unit. The product of the adaptation unit selection compensation coefficient and the initial unit selection quantity K is rounded down to obtain Q. Q monitoring scheme evaluation units are randomly selected from the P monitoring scheme evaluation units to evaluate the quality of the candidate monitoring index group. Here, K is 1 / 3 of P. If Q is less than 1, Q is set to 1. If Q is greater than P, Q is set to P.

9. The method for real-time monitoring of vital signs in emergency patients using multivariate data mining according to claim 8, characterized in that, A weighted summation operation is performed on the compensation coefficients selected by the first unit and the second unit to generate the compensation coefficients selected by the adaptation unit, including: The total amount of historical sample data used to mine the early warning analysis time limit is statistically analyzed, wherein the total amount of historical sample data is the number of positive sample cases diagnosed with cerebral hemorrhage; The ratio of the total amount of historical sample data to the preset sample size threshold is used as the second weight compensation coefficient to compensate and adjust the second initial weight, thereby obtaining the second adaptive weight. The first adaptation weight is obtained by subtracting the second adaptation weight from 1. Based on the first adaptation weight and the second adaptation weight, a weighted summation operation is performed on the compensation coefficients selected by the first unit and the second unit to generate the compensation coefficients selected by the adaptation unit.

10. A real-time monitoring system for vital signs of emergency patients based on multivariate data mining, characterized in that: The system for implementing the multivariate data mining method for real-time monitoring of vital signs in emergency patients according to any one of claims 1-9, the system comprising: The multi-data acquisition module is used to collect multi-dimensional medical data of the target emergency patients; The time limit and contribution determination module is used to determine the early warning analysis time limit for the target emergency patient by using the multivariate disease data as a constraint and cerebral hemorrhage early warning as the early warning target through multivariate data mining, and to determine the contribution of each monitoring indicator in the preset monitoring indicator set to the judgment of the early warning target. The combined scheme optimization module is used to optimize the combined scheme of the monitoring indicators in the preset monitoring indicator set by taking the approximation of the early warning analysis time limit as the optimization target, and combining the contribution of each monitoring indicator and the preset monitoring frequency to determine the optimal monitoring indicator group. In the optimization process, the candidate monitoring indicator group is evaluated by a preset monitoring scheme evaluation plugin. The quality evaluation is generated based on the prediction error probability of the early warning target. The vital signs monitoring module is used to monitor the vital signs of the target emergency patient in real time based on the optimal monitoring scheme composed of the optimal monitoring index group.

Images