Auxiliary evaluation system for acute and critical diseases based on machine learning

Abstract

The invention relates to the technical field of intelligent medical evaluation, in particular to an acute and critical disease auxiliary evaluation system based on machine learning, and the system comprises a data integration and feature extraction module which collects the medical data of an acute and critical disease patient, converts the medical data into a time sequence, analyzes the medical data change information in the time sequence, and outputs the medical data change information; and taking the medical data change information as medical indexes associated with the illness state of the patient, screening different medical indexes, and generating a key feature data set of the current critical patient according to a screening result. According to the method, the medical data is collected in real time and converted into the time sequence, the missing data is filled, the dynamic change of the medical indexes is analyzed, and the key feature data set of the patient can be effectively extracted. And more intuitive and meticulous abnormal data analysis is realized by combining the distribution state of the extreme data points in the time sequence and the record of the change interval of the extreme data points.

Description

Technical Field

[0001] The present invention relates to the field of intelligent medical assessment technology, and in particular to an emergency and critical illness auxiliary assessment system based on machine learning. Background Art

[0002] The emergency and critical care auxiliary assessment system refers to a system that provides condition assessment and diagnostic support for emergency and critically ill patients. It collects and integrates multi-source patient data in real time, including vital signs data, clinical examination results, medical history information, etc., and uses specific multi-parameter analysis methods to quantitatively assess the patient's condition, and combines specific diagnostic rules to judge possible changes in the condition.

[0003] Although the existing technology can provide a certain degree of support for disease assessment by collecting patient vital signs, clinical examinations and medical history data, it is difficult to accurately extract key disease characteristics in terms of data integration and dynamic change analysis. For example, in multi-source data analysis, there is a lack of systematic processing of missing data and abnormal fluctuations in time series, which makes it difficult to effectively capture key abnormal data points. For the processing of subjective symptoms, it mainly relies on the records of patient descriptions. Although it can be used as a reference, it lacks standardization and quantification methods, and it is difficult to deeply explore the correlation between subjective symptoms and objective indicators. For example, the subjective information of patients describing dyspnea cannot establish a clear connection with respiratory rate fluctuation data. In terms of disease trend assessment, fixed diagnostic rules are used to judge disease changes. Although static data can be processed, it is difficult to predict potential risks in real time in the face of dynamic and complex disease changes. For example, changes in patient oxygen saturation cannot effectively predict the risk of acute hypoxia that may occur. These problems will affect the accurate judgment of the development of the disease and limit the effectiveness of real-time monitoring and diagnostic support for critically ill patients. Summary of the invention

[0004] The purpose of the present invention is to solve the shortcomings of the prior art and propose an emergency and critical care auxiliary assessment system based on machine learning.

[0005] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme: The critical and severe disease auxiliary assessment system based on machine learning includes: The data integration and feature extraction module collects medical data of critically ill patients, converts the medical data into time series, analyzes the change information of medical data in the time series, and uses the change information of medical data as the medical indicator associated with the patient's condition. By screening different medical indicators, the key feature data set of the current critically ill patients is generated according to the screening results; The dynamic weighted quantile analysis module analyzes the distribution state of the medical indicators in the key feature data set of the current critically ill patient in the corresponding time series, determines the extreme data points according to the distribution state, obtains the medical indicator distribution record, extracts the change interval of the extreme data points in the time series from the medical indicator distribution record, and generates the patient abnormal data distribution record; The fuzzy symptom relationship analysis module classifies the extreme data points in the abnormal data distribution record of the patient, maps the current patient's subjective symptom text description into a standardized symptom category and performs encoding processing, quantifies the fuzzy relationship between the extreme data points and the symptom category, and generates a dynamic correlation feature set of medical indicators; The disease trend and risk dynamic analysis module uses the fuzzy relationship in the dynamic correlation feature set of the medical indicators as the input feature of the LSTM model, trains the input features, predicts the evolution pattern of extreme data points in the future target time window along the time series based on the training results, and establishes the disease trend auxiliary evaluation results based on the evolution pattern.

[0006] As a further solution of the present invention, the steps for obtaining the key feature data set of the current critically ill patient are specifically as follows: Collect medical data of critically ill patients in real time, including respiratory rate, blood oxygen saturation and mean arterial pressure, convert the medical data into time series, fill in missing values ​​in the time series, analyze the change information of medical data in the completed time series, including the change rate of respiratory rate and the fluctuation range of blood oxygen saturation and mean arterial pressure, and generate the current medical data set of the patient; The medical data change information in the patient's current medical data set is used as a medical indicator associated with the patient's condition, using the formula: Calculate the proportion of each medical indicator value that exceeds the standard range in the patient's current medical data set , ranging from 0 to 1, the ratio The larger the correlation, the stronger the correlation. Determine the degree of correlation and obtain the patient's condition-related information; in, is the number of times the medical indicators in the patient's current medical data set exceed the standard range. is the total number of data points of medical indicators in the patient's current medical data set; The medical indicators whose correlation degree is lower than a preset correlation threshold value in the patient's condition-related information are screened, and the screened medical indicators are organized into a structured format to generate a key feature data set of the current critically ill patients.

[0007] As a further solution of the present invention, the steps of obtaining the medical indicator distribution record are specifically as follows: Analyze the distribution status of the medical indicators in the key feature data set of the current critically ill patients in the corresponding time series, including determining the abnormal upper limit and abnormal lower limit of the medical indicators through a preset quantile threshold, using the formula: Calculate the lower limit of anomalies in the corresponding time series , and the corresponding abnormal upper limit in the time series ; in, It is the first quartile set according to the corresponding time series, that is, the lower quartile. is the third quartile set according to the corresponding time series, that is, the upper quartile. yes and The difference between , represents the interquartile range, It is the abnormal range adjustment coefficient set according to the discrete degree of the corresponding time series; According to the abnormal lower limit in the corresponding time series and the abnormal upper limit in the corresponding time series, extract extreme data points exceeding the abnormal upper limit and below the abnormal lower limit in the key feature data set of the current critically ill patient to generate an extreme data point set; By dynamically allocating weights of extreme data points in the extreme data point set, the distribution state of the medical indicator in the corresponding time series and the abnormal upper limit and the abnormal lower limit is highlighted, and the medical indicator distribution record is obtained.

[0008] As a further solution of the present invention, the steps of obtaining the abnormal data distribution record of the patient are specifically as follows: Extract extreme data points from the medical indicator distribution record, obtain continuous data points in the time range before and after the extreme data points in the corresponding time series, analyze the change interval of the extreme data points in the corresponding time series based on the continuous data points, and obtain change information of the extreme data points; Based on the change information of the extreme data points, the overall fluctuation trend of all data points within the change interval is analyzed, and the patient abnormal data distribution record is generated according to the analysis results.

[0009] As a further solution of the present invention, the step of mapping the current patient's subjective symptom text description into a standardized symptom category is specifically: Obtaining extreme data points in the abnormal data distribution record of the patient, extracting deviation values ​​between the extreme data points and the corresponding abnormal lower limit and abnormal upper limit, comparing the deviation values ​​with the span range of the abnormal lower limit and the abnormal upper limit, classifying the extreme data points according to the comparison results, and generating extreme data point classification results; Collect the symptom texts described subjectively by the current patients, parse the symptom texts, map the descriptions of each symptom text into standardized symptom categories, assign unique codes to the categories, and generate standardized symptom categories.

[0010] As a further solution of the present invention, the steps of acquiring the dynamic association feature set of medical indicators are specifically as follows: Based on the extreme data point classification results and the standardized symptom categories, the formula is adopted: Calculate the classification results of extreme data points The observed value of extreme data points Belongs to the standardized symptom categories Symptom category center value Membership ; in, It is the first in the standardized symptom categories The standardized width parameter for each symptom category, The weight index used to adjust the smoothness of the membership, yes and The absolute deviation of is the normalized result of absolute deviation; Different association ranges are divided according to the membership degree, and the fuzzy relationship between extreme data points and symptom categories is quantified with reference to the association ranges to generate a dynamic association feature set of medical indicators.

[0011] As a further solution of the present invention, the step of predicting the evolution pattern of extreme data points in the future target time window along with the time series is specifically: Arrange the fuzzy relationships in the dynamic association feature set of the medical indicators in the order of timestamps to form time series data, and divide them into input sequences of fixed time windows after normalization, use the input sequences as input features of the LSTM model, perform LSTM model training, and obtain LSTM model training results; Based on the LSTM model training results, the characteristic value of each time point in the future time window is predicted, and the evolution pattern of the extreme data points in the future target time window along the time series is analyzed to obtain the evolution data of the extreme data points.

[0012] As a further solution of the present invention, the steps for obtaining the auxiliary evaluation results of the disease trend are specifically as follows: According to the evolution data of the extreme data point, the state of the extreme data point in the future time window is determined, the state includes continuous abnormality and return to normal, and a change record of the extreme data point is generated; Based on the change records of the extreme data points, the reference range and corresponding time interval of the patient's condition trend are determined, and an auxiliary assessment result of the condition trend is established.

[0013] Compared with the prior art, the advantages and positive effects of the present invention are: In the present invention, by collecting medical data in real time and converting it into a time series, filling in missing data and analyzing the dynamic changes of medical indicators, the key feature data set of patients can be effectively extracted. Combined with the record and analysis of the distribution state of extreme data points in the time series and their change intervals, the abnormal distribution and change trend of medical indicators are weighted and structured, realizing a more intuitive and detailed abnormal data analysis. On this basis, combined with the standardization and coding of the patient's subjective symptom text, the fuzzy association method is used to quantify the correlation between symptom categories and extreme data points, and a multimodal feature set is constructed, thereby enhancing the comprehensive utilization of subjective and objective data. Through the deep learning model, the time series prediction of the disease trend can dynamically reflect the disease evolution trend within the future target time window, significantly improve the foresight of the disease change prediction, effectively reduce the omission of risk factors, and provide a clear basis for medical intervention. The overall solution significantly enhances the dynamic evaluation and prediction capabilities of the condition of critically ill patients through full-process and multi-level innovative analysis, and optimizes the efficiency and reliability of diagnostic support. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 is a system flow chart of the present invention; Figure 2 A flowchart for obtaining a key feature data set of a current critically ill patient for the present invention; Figure 3 A flow chart for obtaining medical indicator distribution records for the present invention; Figure 4 A flow chart for obtaining abnormal data distribution records of patients according to the present invention; Figure 5 A flowchart for mapping a current patient's subjective symptom text description into standardized symptom categories for the present invention; Figure 6 A flowchart of obtaining a dynamic correlation feature set of medical indicators according to the present invention; Figure 7 A flow chart of the present invention for predicting the evolution pattern of extreme data points in a future target time window along with the time series; Figure 8 The present invention is a flow chart for obtaining auxiliary evaluation results of disease trend. DETAILED DESCRIPTION

[0015] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0016] In the description of the present invention, it should be understood that the terms "length", "width", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, in the description of the present invention, "multiple" means two or more, unless otherwise clearly and specifically defined.

[0017] See also Figure 1 , the critical care auxiliary assessment system based on machine learning includes: The data integration and feature extraction module collects medical data of critically ill patients, converts the medical data into time series, analyzes the change information of medical data in the time series, and uses the change information of medical data as the medical indicator associated with the patient's condition. By screening different medical indicators, the key feature data set of the current critically ill patients is generated according to the screening results; The dynamic weighted quantile analysis module analyzes the distribution status of medical indicators in the corresponding time series of the key feature data set of the current critically ill patients, determines the extreme data points according to the distribution status, obtains the medical indicator distribution record, extracts the change interval of the extreme data points in the time series from the medical indicator distribution record, and generates the patient abnormal data distribution record; The fuzzy symptom relationship analysis module classifies the extreme data points in the patient's abnormal data distribution records, maps the current patient's subjective symptom text description to standardized symptom categories and encodes them, quantifies the fuzzy relationship between extreme data points and symptom categories, and generates a dynamic correlation feature set of medical indicators; The disease trend and risk dynamic analysis module uses the fuzzy relationship in the dynamic correlation feature set of medical indicators as the input feature of the LSTM model, trains the input features, predicts the evolution pattern of extreme data points in the future target time window along the time series based on the training results, and establishes the disease trend auxiliary evaluation results based on the evolution pattern.

[0018] See also Figure 2 , the steps for obtaining the key feature data set of critically ill patients are as follows: Collect medical data of critically ill patients in real time, including respiratory rate, blood oxygen saturation and mean arterial pressure, convert the medical data into time series, fill in missing values ​​in the time series, analyze the change information of medical data in the completed time series, including the change rate of respiratory rate and the fluctuation range of blood oxygen saturation and mean arterial pressure, and generate the current medical data set of the patient; Medical data of critically ill patients were collected, including respiratory rate (unit: times / minute), blood oxygen saturation (unit: %) and mean arterial pressure (unit: mmHg). Medical data were recorded every minute in real time through bedside monitoring equipment. The collected data were sorted in chronological order and converted into time series format (unit time interval: every minute). For missing values ​​in the time series, linear interpolation was used to calculate and complete the missing values ​​with the data of adjacent time points before and after the missing values. The completed time series data were analyzed step by step to extract change information. The respiratory rate of change was obtained by calculating the difference in respiratory rate between adjacent time points in the time series and dividing it by the time interval (unit: change in times / minute). Among them, the respiratory rate was sampled in whole minutes, the fluctuation amplitude of blood oxygen saturation was obtained by extracting the difference between the maximum and minimum values ​​in the time series (unit: %), and the fluctuation amplitude of mean arterial pressure was determined by calculating the standard deviation of the time series data (unit: mmHg). Finally, the respiratory rate of change, the fluctuation amplitude of blood oxygen saturation and the fluctuation amplitude of mean arterial pressure were used as medical data change information.

[0019] The medical data change information in the patient's current medical data set is used as the medical indicator associated with the patient's condition, using the formula: Calculate the proportion of each medical indicator value that exceeds the standard range in the patient's current medical data set , ranging from 0 to 1, the ratio The larger the correlation, the stronger the correlation. Determine the degree of correlation and obtain the patient's condition-related information; in, It is the number of times the medical indicators in the patient's current medical data set exceed the standard range. The acquisition method is as follows: Standard range of respiratory rate change rate: determined through experiments, the normal respiratory rate change rate of critically ill patients is generally between 0 and 3 times / minute, that is, an increase or decrease of 0 to 3 times per minute. The experimental method is to use bedside respiratory monitoring equipment to record multiple patient data and count the range of changes. Standard range of blood oxygen saturation fluctuation amplitude: confirmed by the medical reference range (normal fluctuation range is 0 to 5%), standard range of mean arterial pressure fluctuation amplitude: obtained from literature or monitoring data analysis (normal fluctuation range is 0 to 10 mmHg), for each medical indicator, count the number of records exceeding the standard range in the time series, for example, directly filter the number of data points exceeding the standard range through MATLAB or Excel, It is the total number of data points of medical indicators in the patient's current medical data set. Data is imported from the continuous monitoring record file to count the number of all data points in the time series.

[0020] Assume that time series data records of three medical indicators are obtained: Respiratory rate change rate: total number of records , the number of times exceeding the standard range .

[0021] Blood oxygen saturation fluctuation range: total number of records , the number of times exceeding the standard range .

[0022] Mean arterial pressure fluctuation range: total number of records , the number of times exceeding the standard range .

[0023] For the respiratory rate change rate:

[0024] Fluctuation range of blood oxygen saturation:

[0025] For mean arterial pressure fluctuation:

[0026] The results showed that the correlation between the respiratory rate change rate , indicating that it has the greatest impact on the change of the condition. , indicating that it has the least impact on the disease. Correlation of mean arterial pressure fluctuation amplitude , indicating that its influence is moderate.

[0027] Screen the medical indicators whose correlation degree is lower than the preset correlation threshold in the patient's condition-related information, organize the screened medical indicators into a structured format, and generate a key feature data set for the current critically ill patients; According to the calculated correlation results, including the correlation of respiratory rate change rate of 0.2, the correlation of mean arterial pressure fluctuation amplitude of 0.1, and the correlation of blood oxygen saturation fluctuation amplitude of 0.05, the preset correlation threshold is set to 0.1, and the medical indicators with correlation lower than the threshold are screened out, namely, blood oxygen saturation fluctuation amplitude, which is excluded due to low correlation; after screening, the medical indicators with correlation higher than or equal to the threshold are retained, namely, respiratory rate change rate and mean arterial pressure fluctuation amplitude. The key feature data set of the current critically ill patients is generated for the two retained indicators, and their time series data is organized into the following structured format, including timestamp, medical indicator name and its corresponding data value. For example, in the recorded time series, the respiratory rate change rate at a certain moment is a change of 2 times / minute, and the mean arterial pressure fluctuation amplitude is 8 mmHg. The data at that moment is recorded as: timestamp information, and the medical indicators are "respiratory rate change rate" and "mean arterial pressure fluctuation amplitude", which correspond to the numerical values ​​of 2 times / minute and 8 mmHg respectively. The key medical indicator data of each time point are recorded in sequence and organized into a table or database form as the basic data set for subsequent patient condition evaluation.

[0028] See also Figure 3 ,The specific steps for obtaining medical indicator distribution records are: Analyze the distribution status of medical indicators in the corresponding time series in the key feature data set of current critically ill patients, including determining the abnormal upper and lower limits of medical indicators through preset quantile thresholds, using the formula: Calculate the lower limit of anomalies in the corresponding time series , and the corresponding abnormal upper limit in the time series ; in, It is the first quartile set according to the corresponding time series, that is, the lower quartile, which represents the 25% quantile value in the data. After sorting the time series data, the median value of the first 25% of the data is taken. It is the third quartile set according to the corresponding time series, that is, the upper quartile, which represents the 75% quantile value in the data. After sorting the time series data, the median value of the last 75% of the data is taken. yes and The difference represents the interquartile range, which is the span of the middle range of the data distribution. The calculation formula is , It is the abnormal range adjustment coefficient set according to the discrete degree of the corresponding time series. It indicates the proportional coefficient used to expand the upper and lower limits. It is often used for sensitivity adjustment. According to literature analysis or experimental statistics, the commonly used value is 1.5, but it can be adjusted according to the discrete degree of patient data. For example, in the data of critically ill patients, if the fluctuation range of the patient's medical indicators is larger, a higher value can be selected. To avoid normal values ​​being misjudged as abnormal, the degree of dispersion is determined as follows: The larger the value is, the higher the proportion is relative to the overall data range, and the greater the degree of data dispersion is. The gap with the total data range is small, indicating that the data is more concentrated. The formula is: gap ratio = ,The total data range refers to the difference between the maximum and minimum values ​​of the data points in the time series or data set, which is used to describe the overall distribution span of the data, highly concentrated: gap ratio , most of the data are concentrated near the median, the distribution is relatively dense, the dispersion is very low, and the concentration is medium: , the data distribution is relatively uniform, the middle 50% of the data accounts for a moderate proportion of the total range, the degree of dispersion is moderate, and it is relatively discrete: , the middle 50% of the data accounts for a large proportion of the total data range, the data distribution is relatively discrete, and the gap ratio is highly discrete: , the middle 50% of the data account for almost the entire range of the total data, indicating that the data distribution is highly discrete and the overall volatility is large.

[0029] If the original time series data of the respiratory rate change rate data is (Unit: change in times / minute), after sorting, divided into the first 25% and the last 75%, calculate and .

[0030] The original time series data of mean arterial pressure fluctuation amplitude data is (Unit: mmHg), after sorting, divided into the first 25% and the last 75%, calculate and .

[0031] Calculate quartiles and : Respiratory rate change rate: The first half of the data is , the median value is The second half of the data is , the median value is .

[0032] Mean arterial pressure fluctuation range: The first half of the data is , the median value is The second half of the data is , the median value is .

[0033] Calculate the interquartile range : Respiratory rate change rate: ; Mean arterial pressure fluctuation range: ; Respiratory rate change rate: Lower limit: ; Upper limit: ; Mean arterial pressure fluctuation range: Lower limit: ; Upper limit: ; The results show that the dynamic upper and lower limits of the respiratory rate change rate are , the standard range is The change in times / minute, the dynamic upper and lower limits are extended to The dynamic upper and lower limits of mean arterial pressure fluctuation are , the standard range is mmHg, dynamic upper and lower limits extended to . Since values ​​below and above the normal range are not necessarily considered extreme, they need to be further combined with distribution analysis. The range of extreme value determination is expanded through the upper and lower limits of the quantiles, which can dynamically adapt to the patient's personalized distribution characteristics, avoid misjudging medical normal values ​​as extreme values, and more accurately identify abnormal data points that exceed the patient's distribution pattern.

[0034] According to the abnormal lower limit and the abnormal upper limit in the corresponding time series, the extreme data points exceeding the abnormal upper limit and below the abnormal lower limit in the key feature data set of the current critically ill patients are extracted to generate an extreme data point set; Extract the extreme data points that exceed the abnormal upper limit and are lower than the abnormal lower limit from the key feature data set of the current critically ill patients. Combined with the dynamic upper and lower limit calculation results of paragraph 1, the dynamic upper and lower limits of the respiratory rate change rate are The dynamic upper and lower limits of mean arterial pressure fluctuation are , filter out extreme data points from time series data point by point. For example, the value below the lower limit is , which is equal to the lower limit is not considered abnormal; values ​​exceeding the upper limit are , which is equal to the upper limit are not considered abnormal; therefore, the extreme data points extracted are , recorded as: timestamp, indicator name (respiratory rate change rate), data value (change in 9 times / minute). For mean arterial pressure fluctuation amplitude data , the value below the lower limit is , the value exceeding the upper limit is , so the extreme data points extracted are , recorded as: timestamp, indicator name (mean arterial pressure fluctuation amplitude), data value (1mmHg and 15mmHg), and all extracted extreme data points are recorded and organized into a table format.

[0035] By dynamically allocating the weights of extreme data points in the extreme data point set, the distribution state of the medical indicators in the corresponding time series and the abnormal upper and lower limits is highlighted, and the distribution record of the medical indicators is obtained; Based on the extracted extreme data points, the weight of each data point is calculated. For example, its dynamic upper limit is , the deviation value is , the deviation value is compared with the range between the abnormal upper limit and the median Compare and calculate the weight as ; For extreme data points of mean arterial pressure fluctuation , the dynamic upper limit is , the deviation value is , the range is , the weight is calculated as Similarly, the extreme low value of the mean arterial pressure fluctuation Deviation from lower limit , the deviation value is , the weight is calculated as All calculated weight values ​​are attached to the extreme data point records and organized into a distribution record table, for example: timestamp, indicator name (respiratory rate change rate), data value (change of 9 times / minute), weight (0.5); timestamp, indicator name (mean arterial pressure fluctuation range), data value (15 mmHg), weight (0.167).

[0036] See also Figure 4 , the specific steps for obtaining the patient abnormal data distribution record are: Extract extreme data points from the medical indicator distribution records, obtain continuous data points in the time range before and after the extreme data points in the corresponding time series, analyze the change interval of the extreme data points in the corresponding time series based on the continuous data points, and obtain the change information of the extreme data points; The extreme data points of the respiratory rate change rate are (Unit: change in times / minute), the extreme data points of mean arterial pressure fluctuation are and (Unit: mmHg), extract the change range of these data points in the time series. For the respiratory rate change rate, the extreme data points The corresponding timestamp is timestamp 1, before and after Extract change data interval within minutes , the corresponding timestamp interval is the range before and after timestamp 1. For the fluctuation amplitude of mean arterial pressure, the extreme data points The corresponding timestamp is timestamp 2, and the change interval is , the corresponding timestamp interval is the range before and after timestamp 2; extreme data points The corresponding timestamp is timestamp 3, and the change interval is , the corresponding timestamp interval is the range before and after timestamp 3. These change interval data points and corresponding timestamps are recorded in a table, organized as: timestamp, indicator name (respiratory rate change rate or mean arterial pressure fluctuation range), change interval data (such as , units correspond to each indicator).

[0037] Based on the change information of extreme data points, analyze the overall fluctuation trend of all data points within the change interval, and generate abnormal data distribution records of patients according to the analysis results; For the respiratory rate change rate, the change range is (Unit: change in times / minute), the fluctuation trend is obtained by calculating the difference between the adjacent data points. The overall trend value is , indicating the overall upward trend of the respiratory rate change rate; for the fluctuation range of mean arterial pressure, the change range is and (Unit: mmHg), the difference is calculated as and The overall trend value is and , respectively, represent the upward trend and downward trend of the mean arterial pressure fluctuation amplitude. The trend data at each time point is recorded as: timestamp, indicator name (respiratory rate change rate or mean arterial pressure fluctuation amplitude), change trend (such as ), these trend records are organized into patient abnormal data distribution records.

[0038] See also Figure 5 The specific steps for mapping the current patient's subjective symptom text description into standardized symptom categories are: Obtain extreme data points in the patient's abnormal data distribution record, extract the deviation value between the extreme data point and the corresponding abnormal lower limit and abnormal upper limit, compare the deviation value with the span range of the abnormal lower limit and the abnormal upper limit, classify the extreme data points according to the comparison result, and generate the extreme data point classification result; The extreme data points in the patient's abnormal data distribution records are classified by obtaining the basic information of the extreme data points, including the name of the medical indicator, the upper and lower limit deviation values, and the corresponding timestamps. , the upper and lower limits are , the value that deviates from the upper limit is The change in beats / minute; the extreme data points for the fluctuation of mean arterial pressure , the upper and lower limits are , the value that deviates from the lower limit is mmHg; the other extreme data point of mean arterial pressure fluctuation The deviation from the upper limit is mmHg. Classification is performed by comparing the ratio between the deviation value and the span of the normal range of the corresponding indicator (the difference between the upper and lower limits), for example, the normal range span for the respiratory rate change rate Change in times / minute, deviation value Percentage of times / minute ; For the normal range of mean arterial pressure fluctuations mmHg, deviation mmHg percentage During the classification process, points with a deviation value of more than 50% are classified as "severely abnormal", such as the point with a fluctuation amplitude of mean arterial pressure. , which deviates from the lower limit mmHg accounts for a small proportion, no more than 50%, and is classified as "slightly abnormal". Data points with close timestamps are grouped by time period to ensure the integrity of the classification results.

[0039] Collect the symptom texts described by the current patients, parse the symptom texts, map the descriptions of each symptom text into standardized symptom categories, assign unique codes to the categories, and generate standardized symptom categories; First, by collecting the symptom texts described by the patients, such as keywords such as "dizziness", "chest tightness", and "rapid heartbeat", the medical terms corresponding to the symptom categories are extracted in combination with the commonly used medical standard manuals (such as the ICD-10 classification standards). For example, "dizziness" corresponds to "head discomfort", "chest tightness" corresponds to "dyspnea", and "rapid heartbeat" corresponds to "abnormal heart rate". Secondly, the extracted standardized categories are coded, and each category is assigned a unique numerical code, such as "head discomfort" is 001, "dyspnea" is 002, and "abnormal heart rate" is 003. The patient description text is processed by keyword matching technology and semantic parsing tools, and each text describing the symptoms is extracted as a keyword and mapped to the standardized category. At the same time, the accuracy of the symptom keywords is verified to avoid misclassification. Finally, the standardized symptom category and code value are recorded in correspondence with the timestamp. For example, the symptom "dizziness" corresponding to timestamp 1 is mapped to the category "head discomfort" and coded as 001. All results are organized into a record table, including timestamps, standardized symptom categories, and corresponding codes.

[0040] See also Figure 6 ,The specific steps for obtaining the dynamic association feature set of medical indicators are:

[0041] Based on the classification results of extreme data points and the standardized symptom categories, the formula is adopted:

[0042] Calculate the classification results of extreme data points The observed value of extreme data points Belongs to the standardized symptom categories Symptom category center value Membership ;

[0043] in, It is the classification result of extreme data points. extreme data points, such as the extreme data points of the respiratory rate change rate (change in times / minute), It is the first in the standardized symptom categories The central value of a symptom category represents the average value or typical value of a symptom category. The central value is calculated by statistically analyzing the historical data corresponding to the category. For example, the central value of the "dyspnea" symptom category is determined by analyzing the mean value of the respiratory rate change rate of patients with dyspnea. For example, for the dyspnea symptom category, the mean value of the respiratory rate change rate in all relevant cases is extracted. It is the first in the standardized symptom categories The standardized width parameter of each symptom category is used to measure the distribution range and sensitivity of the category. The standard deviation of the symptom category distribution is calculated based on historical data. For example, the standard deviation of the respiratory rate change rate of patients with dyspnea symptoms is calculated. The weight index used to adjust the smoothness of the membership, weight index The typical value range of , which is an empirical value obtained from a large amount of medical data modeling and classification practice. :It is suitable for situations where the data distribution is relatively concentrated and the category boundaries are clear. The change of membership is steeper and different categories can be quickly distinguished. : It is a commonly used neutral value, which is suitable for most practical scenarios. The change of membership degree is neither too sensitive nor too gentle. or higher: Applicable to situations where there is a lot of overlap between categories and blurred boundaries. The membership changes more smoothly and is suitable for multi-category association analysis. yes and The absolute deviation of It is the absolute deviation normalization result, which is used to measure the standardized deviation of the data point relative to the category width.

[0044] Respiratory frequency change rate raw time series data: (Unit: change in times / minute), in the data, Extremely high value, exceeding the dynamic upper limit The change in breaths / minute, assuming the symptom category is "dyspnea", the central value It is obtained through statistical analysis of historical data. Taking the mean of the respiratory rate change rate of patients with dyspnea in the case as the central value, assuming (change in times / minute), width parameter Determined by the standard deviation of the respiratory rate change rate of patients with "dyspnea" in historical data, assuming (Change in times / minute).

[0045] Substitute the formula to calculate the degree of membership:

[0046] Original time series data of mean arterial pressure fluctuation amplitude: (Unit: mmHg), in the data, Extremely low value (below the dynamic lower limit mmHg), Extremely high value (exceeding the dynamic upper limit mmHg), assuming the symptom category is "hypotension", the central value is the mean value of mean arterial pressure fluctuation in historical hypotension cases, assuming (mmHg), assuming the symptom category is "hypertension", the central value is the mean of historical hypertension cases, assuming (mmHg), width parameter mmHg (standard deviation in hypotensive cases), width parameter mmHg (standard deviation among cases of hypertension).

[0047] Substitute into the formula to calculate the membership degree (hypotension, ):

[0048] Substitute into the formula to calculate the membership degree (hypertension, ):

[0049] The results show that for extreme data points of respiratory rate change Membership , for extreme data points of mean arterial pressure fluctuation amplitude: Membership , Membership .

[0050] Different association ranges are divided according to the membership degree, and the fuzzy relationship between extreme data points and symptom categories is quantified with reference to the association range to generate a dynamic association feature set of medical indicators; The membership range is divided into: 0.0–0.3 indicates weak association, 0.3–0.6 indicates moderate association, and 0.6–1.0 indicates strong association. , calculate the membership , which falls into the medium correlation range, indicating that this data point has a medium degree of fuzzy relationship with the symptom category of "dyspnea". This shows that the respiratory rate change value The deviation from the normal range reflects that the change in the patient's respiratory rate may be related to "dyspnea", but it has not reached a strong correlation. , membership , falls into the weak correlation range, indicating that there is a weak fuzzy relationship between this data point and the "hypotension" symptom category, indicating that the patient's mean arterial pressure low value feature is weakly associated with "hypotension". On the other hand, the mean arterial pressure data point Membership , falls into the strong correlation range, indicating that there is a strong fuzzy relationship between it and the "hypertension" symptom category, indicating that the patient's mean arterial pressure is significantly higher and is highly correlated with the "hypertension" symptom. Through these membership values ​​and range divisions, the fuzzy relationship between extreme data points and symptom categories can be further quantified, providing a clearer reference for medical analysis.

[0051] See also Figure 7 , the steps to predict the evolution pattern of extreme data points in the future target time window along with the time series are as follows: The fuzzy relationships in the dynamic association feature set of medical indicators are arranged in the order of timestamps to form time series data, and after normalization, they are divided into input sequences of fixed time windows. The input sequences are used as input features of the LSTM model, and the LSTM model is trained to obtain the LSTM model training results. First, extract the input data of each time point from the dynamic association feature set, including membership, medical indicator values ​​at the corresponding time point (such as respiratory rate change rate and mean arterial pressure fluctuation amplitude), and symptom category, and construct the input feature vector. The feature vector is in the form of [timestamp, membership 1, membership 2, feature 1, feature 2, symptom category], where: Timestamp: refers to the time information when the data point is collected, which is used to record the temporal order of the data. Membership 1, Membership 2: respectively represent the membership values ​​of medical indicators (such as respiratory rate change rate and mean arterial pressure fluctuation amplitude) corresponding to specific symptom categories, for example Indicates the membership of the respiratory rate change rate to "dyspnea", Indicates the degree of membership of the mean arterial pressure fluctuation amplitude to "hypertension". Feature 1 and Feature 2: are specific medical indicator values, for example, feature 1 is the respiratory rate change rate (unit: change in times / minute), and feature 2 is the mean arterial pressure fluctuation amplitude (unit: mmHg). Symptom category: indicates the association of the data point with a specific symptom, for example, the respiratory rate change rate may correspond to "dyspnea", and the mean arterial pressure fluctuation amplitude may correspond to "hypertension". These features are sorted by time point to form time series input data to capture the trend of data changes over time. Subsequently, these feature values ​​are normalized to map all values ​​to the range of [0, 1] to reduce the impact of the difference in the numerical range between features on model training. The normalized data is divided into input sequences according to fixed time windows (such as 5 consecutive time points), and the features of the first 4 time points of the input sequence are used to predict the feature value of the 5th time point. During the training process, the LSTM model calculates the predicted value through forward propagation, uses the mean square error as the loss function, compares the error between the predicted value and the target value, and gradually adjusts the weight and bias parameters of the LSTM network through the backpropagation algorithm to reduce the error. After multiple rounds of iterative optimization, the model's predictive ability gradually improved and finally converged, completing the training process.

[0052] Based on the LSTM model training results, the characteristic value of each time point in the future time window is predicted, and the evolution pattern of extreme data points in the future target time window with the time series is analyzed to obtain the evolution data of extreme data points; According to the training results, the evolution pattern of extreme data points in the future target time window with the time series is predicted. The prediction process is based on the trained LSTM model, and the latest time window data in the dynamic association feature set is used as the starting input data. The input feature vector is in the form of [timestamp, membership 1, membership 2, feature 1, feature 2, symptom category], such as the feature values ​​in the last 5 time points, including membership, medical indicator values ​​(such as respiratory rate change rate and mean arterial pressure fluctuation amplitude) and corresponding timestamps. The meaning of the input feature vector is consistent with the training process. The model receives these input data and gradually calculates the feature values ​​of the future target time points through forward propagation. In the prediction process, a recursive iteration method is used, and the predicted output value of each time point will be used as the input of the next time point. For example, the feature data of the last 5 time points are input into the model, and the feature value of the 6th time point is calculated. Then the predicted value of the 6th time point is combined with the data of the previous 4 time points to form a new input for predicting the value of the 7th time point. Through this recursive iteration method, the time evolution sequence of all characteristic values ​​in the target time window is gradually generated, and finally the prediction process of the evolution pattern of the extreme data point time series is completed.

[0053] See also Figure 8 The specific steps for obtaining the auxiliary evaluation results of the disease trend are as follows: According to the evolution data of extreme data points, the state of extreme data points in the future time window is determined, including continued abnormality and return to normal, and the change record of extreme data points is generated; The LSTM model is used to predict the respiratory rate change rate and mean arterial pressure fluctuation amplitude in the future time window, and gradually observe whether these extreme data points continue to be abnormal or return to the normal range. If the predicted value at a certain time point still belongs to the extreme data point range (such as the respiratory rate change rate continues to be higher than the upper limit of the normal range), it is marked as "continuously abnormal" and the timestamp and corresponding value are recorded; if the predicted value at a certain time point gradually returns to the normal range, it is marked as "return to normal" and the risk mark for its subsequent evolution is stopped. By gradually judging each extreme data point in the future time series, the information containing the timestamp, extreme data point value and state change is extracted, and the dynamic evolution pattern of these extreme data points is recorded, such as the continuous change trend of extreme data points in different time periods and their end time points, and finally a time series information record for disease trend analysis is generated.

[0054] Based on the change records of extreme data points, determine the reference range and corresponding time interval of the patient's condition trend, and establish auxiliary assessment results of the condition trend; First, all extreme data points marked as “persistent abnormality” or “return to normal” within the future target time window are arranged in chronological order, such as the respiratory rate change rate at time point Persistent exception (value is The change in beats / minute) at time point Return to normal range (value is Then, based on the information at these time points, analyze the fluctuation trend of the disease within a specific time period, such as the change rate of respiratory rate from time point arrive The value of gradually decreases, showing a recovery trend. For extreme data points with persistent abnormalities (such as the fluctuation amplitude of mean arterial pressure at time point If the patient's blood pressure is continuously above the extreme range, it is marked as "risk increase" and the predicted duration is used to determine the time period when the condition may worsen, such as from Start to Then, these trend change data are combined with the fluctuation range of specific medical indicators to generate a reference range for the condition change in the future time period. For example, the respiratory rate change rate may be The mean arterial pressure fluctuates in the range of mmHg range changes. Based on these reference ranges, auxiliary assessment results of disease trends are generated, including: 1) description of disease trends in future time periods (such as "recovery" or "increased risk"); 2) time intervals corresponding to each disease trend; 3) dynamic ranges of different medical indicators. This information can be further used to guide decisions on medical interventions and provide a quantitative reference for disease trend changes.

[0055] The above are only preferred embodiments of the present invention and are not intended to limit the present invention in other forms. Any technician familiar with the profession may use the technical contents disclosed above to change or modify them into equivalent embodiments with equivalent changes and apply them to other fields. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. An emergency and critical care auxiliary assessment system based on machine learning, characterized in that: The system comprises: The data integration and feature extraction module collects medical data of critically ill patients, converts the medical data into time series, analyzes the change information of medical data in the time series, and uses the change information of medical data as the medical indicator associated with the patient's condition. By screening different medical indicators, the key feature data set of the current critically ill patients is generated according to the screening results; The dynamic weighted quantile analysis module analyzes the distribution state of the medical indicators in the key feature data set of the current critically ill patient in the corresponding time series, determines the extreme data points according to the distribution state, obtains the medical indicator distribution record, extracts the change interval of the extreme data points in the time series from the medical indicator distribution record, and generates the patient abnormal data distribution record; The fuzzy symptom relationship analysis module classifies the extreme data points in the abnormal data distribution record of the patient, maps the current patient's subjective symptom text description into a standardized symptom category and performs encoding processing, quantifies the fuzzy relationship between the extreme data points and the symptom category, and generates a dynamic correlation feature set of medical indicators; The disease trend and risk dynamic analysis module uses the fuzzy relationship in the dynamic correlation feature set of the medical indicators as the input feature of the LSTM model, trains the input features, predicts the evolution pattern of extreme data points in the future target time window along the time series based on the training results, and establishes the disease trend auxiliary evaluation results based on the evolution pattern.

2. The machine learning-based emergency and critical illness auxiliary assessment system according to claim 1, characterized in that: The steps for obtaining the key feature data set of the current critically ill patients are specifically as follows: Collect medical data of critically ill patients in real time, including respiratory rate, blood oxygen saturation and mean arterial pressure, convert the medical data into time series, fill in missing values ​​in the time series, analyze the change information of medical data in the completed time series, including the change rate of respiratory rate and the fluctuation range of blood oxygen saturation and mean arterial pressure, and generate the current medical data set of the patient; The medical data change information in the patient's current medical data set is used as a medical indicator associated with the patient's condition, using the formula: Calculate the proportion of each medical indicator value that exceeds the standard range in the patient's current medical data set , ranging from 0 to 1, the ratio The larger the correlation, the stronger the correlation. Determine the degree of correlation and obtain the patient's condition-related information; in, is the number of times the medical indicators in the patient's current medical data set exceed the standard range. is the total number of data points of medical indicators in the patient's current medical data set; The medical indicators whose correlation degree is lower than a preset correlation threshold value in the patient's condition-related information are screened, and the screened medical indicators are organized into a structured format to generate a key feature data set of the current critically ill patients.

3. The critical illness auxiliary assessment system based on machine learning according to claim 1 is characterized in that: The steps for obtaining the medical indicator distribution record are specifically as follows: Analyze the distribution status of the medical indicators in the key feature data set of the current critically ill patients in the corresponding time series, including determining the abnormal upper limit and abnormal lower limit of the medical indicators through a preset quantile threshold, using the formula: Calculate the lower limit of anomalies in the corresponding time series , and the corresponding abnormal upper limit in the time series ; in, It is the first quartile set according to the corresponding time series, that is, the lower quartile. is the third quartile set according to the corresponding time series, that is, the upper quartile. yes and The difference between , represents the interquartile range, It is the abnormal range adjustment coefficient set according to the discrete degree of the corresponding time series; According to the abnormal lower limit in the corresponding time series and the abnormal upper limit in the corresponding time series, extract extreme data points exceeding the abnormal upper limit and below the abnormal lower limit in the key feature data set of the current critically ill patient to generate an extreme data point set; By dynamically allocating weights of extreme data points in the extreme data point set, the distribution state of the medical indicator in the corresponding time series and the abnormal upper limit and the abnormal lower limit is highlighted, and the medical indicator distribution record is obtained.

4. The machine learning-based emergency and critical illness auxiliary assessment system according to claim 1, characterized in that: The steps for obtaining the abnormal data distribution record of the patient are specifically as follows: Extract extreme data points from the medical indicator distribution record, obtain continuous data points in the time range before and after the extreme data points in the corresponding time series, analyze the change interval of the extreme data points in the corresponding time series based on the continuous data points, and obtain change information of the extreme data points; Based on the change information of the extreme data points, the overall fluctuation trend of all data points within the change interval is analyzed, and the patient abnormal data distribution record is generated according to the analysis results.

5. The machine learning-based emergency and critical illness auxiliary assessment system according to claim 3, characterized in that: The step of mapping the current patient's subjective symptom text description into a standardized symptom category is specifically as follows: Obtaining extreme data points in the abnormal data distribution record of the patient, extracting deviation values ​​between the extreme data points and the corresponding abnormal lower limit and abnormal upper limit, comparing the deviation values ​​with the span range of the abnormal lower limit and the abnormal upper limit, classifying the extreme data points according to the comparison results, and generating extreme data point classification results; Collect the symptom texts described subjectively by the current patients, parse the symptom texts, map the descriptions of each symptom text into standardized symptom categories, assign unique codes to the categories, and generate standardized symptom categories.

6. The machine learning-based emergency and critical illness auxiliary assessment system according to claim 5, characterized in that: The steps for obtaining the dynamic association feature set of medical indicators are specifically as follows: Based on the extreme data point classification results and the standardized symptom categories, the formula is adopted: Calculate the classification results of extreme data points The observed value of extreme data points Belongs to the standardized symptom categories Symptom category center value Membership ; in, It is the first in the standardized symptom categories The standardized width parameter for each symptom category, The weight index used to adjust the smoothness of the membership, yes and The absolute deviation of is the normalized result of absolute deviation; Different association ranges are divided according to the membership degree, and the fuzzy relationship between extreme data points and symptom categories is quantified with reference to the association ranges to generate a dynamic association feature set of medical indicators.

7. The machine learning-based emergency and critical illness auxiliary assessment system according to claim 1, characterized in that: The steps of predicting the evolution pattern of extreme data points in the future target time window along with the time series are specifically as follows: Arrange the fuzzy relationships in the dynamic association feature set of the medical indicators in the order of timestamps to form time series data, and divide them into input sequences of fixed time windows after normalization, use the input sequences as input features of the LSTM model, perform LSTM model training, and obtain LSTM model training results; Based on the LSTM model training results, the characteristic value of each time point in the future time window is predicted, and the evolution pattern of the extreme data points in the future target time window along the time series is analyzed to obtain the evolution data of the extreme data points.

8. The machine learning-based emergency and critical illness auxiliary assessment system according to claim 7, characterized in that: The steps for obtaining the auxiliary evaluation results of the disease trend are specifically as follows: According to the evolution data of the extreme data point, the state of the extreme data point in the future time window is determined, the state includes continuous abnormality and return to normal, and a change record of the extreme data point is generated; Based on the change records of the extreme data points, the reference range and corresponding time interval of the patient's condition trend are determined, and an auxiliary assessment result of the condition trend is established.

Images