Surgical criticality assisted assessment method based on combined machine learning
By combining machine learning methods with preoperative and intraoperative monitoring data, a critical illness assessment model was established, which solved the problem that existing technologies could not accurately predict critical illnesses. This enabled accurate assessment of the duration and severity of critical illnesses, assisting doctors in timely intervention.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHONGQING INST OF GREEN & INTELLIGENT TECH CHINESE ACAD OF SCI
- Filing Date
- 2022-06-12
- Publication Date
- 2026-05-05
AI Technical Summary
Current technology cannot accurately predict the timing and severity of critical illnesses based on monitoring data, making it impossible for doctors to intervene effectively in a timely manner.
A combined machine learning approach was adopted, combining historical monitoring and surveillance time-series data from preoperative and intraoperative procedures. Missing data were filled using a random forest regression algorithm based on decision trees, correlation analysis was performed, monitoring indicators affecting weights were extracted, and preoperative and intraoperative auxiliary assessment models were established. The monitoring cycle was optimized through clustering and loss functions, and the quantitative value of the severity of critical illness was calculated to assist doctors in assessment.
It enables accurate prediction of the duration and severity of critical illness, improves the model's generalization ability, and provides effective assessment of critical illness before and during surgery, assisting medical staff in timely intervention.
Smart Images
Figure CN115116605B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for auxiliary assessment of critically ill patients undergoing surgery based on combinatorial machine learning, belonging to the field of data mining, and is particularly applicable to such methods. Background Technology
[0002] Critical illness refers to the clinical signs of a severe, variable, and life-threatening emergency. Critically ill patients have unstable vital signs, rapidly changing conditions, and often exhibit one or more organ failures or dysfunctions. A comprehensive assessment can be made based on vital signs such as temperature, pulse, respiration, and blood pressure. Once critical illness occurs, the condition becomes severe and rapidly changes; even slight mishandling can lead to irreparable consequences. Therefore, doctors must be able to make accurate judgments and determine treatment measures within a short period.
[0003] The development of big data technology has made it possible to predict and assess patients' critical illness using monitoring data. However, judging the risk of critical illness based solely on a single moment or a short segment of online monitoring during surgery is extremely one-sided and inaccurate.
[0004] In fact, a patient diagnosed with a critical illness by doctors is not always in a critical condition. Rather, it's a situation where a suspected critical condition or a high probability of critical illness arises at a certain point in time, followed by effective intervention from doctors, which transforms the patient's condition into a non-critical one. Currently, existing research cannot accurately determine the timeframe and risk level of a critical illness through monitoring data, thus hindering medical personnel from providing timely and effective intervention. Summary of the Invention
[0005] In view of this, the present invention provides a surgical critical care auxiliary assessment method based on combinatorial machine learning, which aims to predict the accurate critical care cycle and the degree of critical care by combining historical monitoring and surveillance data from before and during surgery and through comprehensive analysis using combinatorial machine learning technology.
[0006] To achieve the above objectives, the present invention provides the following technical solution:
[0007] A surgical critical care auxiliary assessment method based on ensemble machine learning, combined with Figure 1 Its characteristic is that it includes the following steps:
[0008] S1: The historical monitoring and surveillance time series data are divided into preoperative historical monitoring and surveillance time series data and intraoperative historical monitoring and surveillance time series data according to the preoperative and intraoperative stages, and the critical illness and its cycle are artificially marked according to the doctor's diagnosis conclusion.
[0009] S2: For the time series data of historical monitoring and surveillance before and during the operation, the missing monitoring indicator data are filled in using the random forest regression algorithm based on decision trees.
[0010] S3: Based on the time-series data of preoperative historical monitoring and surveillance, conduct correlation analysis on common indicators of critical illness, and extract monitoring indicators with significant impact on critical illness as preoperative factors.
[0011] S4: Using preoperative factors as input to the model, a preoperative auxiliary assessment model adapted to different types of critical illnesses is established using combinatorial machine learning methods, and trained using time-series data of preoperative historical monitoring and surveillance.
[0012] S5: Input a patient's preoperative factors into a trained preoperative auxiliary assessment model for critical illnesses to assist doctors in assessing and analyzing preoperative critical illnesses.
[0013] S6: Calculate intraoperative feature values for data marked as critical illness cycles in the intraoperative historical monitoring and surveillance time series data. Similarly, calculate intraoperative feature values for patient data marked as non-critical illness cycles in the intraoperative historical monitoring and surveillance time series data. Then, cluster the data according to the intraoperative feature values to obtain critical illness cluster centers and non-critical illness cluster centers.
[0014] S7: Initially set a monitoring cycle, calculate intraoperative feature values based on the time series data of historical monitoring and surveillance during the operation according to the monitoring cycle, classify according to the distance between the intraoperative feature values and the cluster centers of critical and non-critical cases, and calculate the optimal monitoring cycle using the loss function;
[0015] S8: Calculate the quantitative value of the severity of critical illness using the time-series data of intraoperative historical monitoring and surveillance within the optimal monitoring period;
[0016] S9: Using preoperative factors, intraoperative feature values, and critical illness severity quantification values as inputs to the model, a combined machine learning method is used to establish an intraoperative auxiliary assessment model that adapts to different types of critical illnesses, and the model is trained using time-series data of historical monitoring and surveillance before and during the operation.
[0017] S10: Input a patient's preoperative factors, intraoperative characteristic values, and criticality severity quantification values into a trained intraoperative auxiliary assessment model for critical illnesses to assist doctors in assessing and analyzing intraoperative critical illnesses.
[0018] Furthermore, the aforementioned cycle includes non-critical illness cycles and critical illness cycles; the monitoring cycle is a period of time for monitoring, which may be a non-critical illness cycle, a critical illness cycle, or a combination of both.
[0019] Furthermore, the correlation analysis techniques in step S3 mainly include: covariance, correlation coefficient, etc. In application, the monitoring indicators are ranked according to their correlation, and the monitoring indicators whose total weight meets the set threshold are used as preoperative factors.
[0020] Furthermore, the combined machine learning method described in steps S4 and S9 is as follows: First, an ensemble learning method is used to learn from artificial deep neural networks and the XGBoost regression method (see reference). https: / / arxiv.org / pdf / 1603.02754.pdf From the perspectives of support vector machines and random forest regression methods, a regression analysis model is established from input to critical illness determination result as output; then, all regression analysis models are evaluated by indicators, and the optimal regression analysis model is selected for the final determination of critical illness result.
[0021] Furthermore, the ensemble learning method (see reference) http: / / wfwei.github.io / posts / ensemble-methods / The available algorithms include: Bootstrap Aggregating, Boosting, and Random Forest; the evaluation metrics include mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R-squared).
[0022] Furthermore, the intraoperative characteristic value mentioned in step S6 is a four-dimensional vector consisting of the first moment (mean), the second central moment (variance), the third central moment (skewness), and the fourth central moment (kurtosis) of the data within the period.
[0023] Furthermore, the clustering method described in step S6 is a binary classification method that uses the K-means method to classify critically ill patients into cluster centers and non-critically ill patients into cluster centers.
[0024] Furthermore, the loss function described in step S7 is:
[0025]
[0026] Among them, c w c n , , represent the cluster centers for critically ill patients and those for non-critically ill patients, respectively, and p represents the time-series data of historical monitoring and surveillance of a single patient before and during surgery.
[0027] Furthermore, step S7 specifically includes:
[0028] S701: Set the sampling period of the time series data of the historical monitoring and surveillance during the operation to the minimum step size. Initially, the number of steps in one monitoring cycle is set to 2.
[0029] Specifically, for the time-series data of historical monitoring and surveillance of patients marked as non-critical, the intraoperative characteristic value is calculated on a cycle of the entire monitoring period, and then the process directly jumps to step S8.
[0030] S702: Calculate intraoperative characteristic values based on the time series data of historical monitoring and surveillance during the operation according to the monitoring cycle, and classify them according to their distance from the critical illness cluster center and the non-critical illness cluster center;
[0031] S703: Calculate the loss function based on the classification results and intraoperative feature values within the monitoring period;
[0032] S704: Increase the monitoring cycle step by 1, and repeat steps S702 to S703 until N times, selecting the monitoring cycle with the smallest loss function as the optimal monitoring cycle. Ideally, an integer N should be chosen reasonably while satisfying real-time performance and accuracy requirements.
[0033] Furthermore, the calculation method for the critical illness severity quantification value y in step S8 is as follows: when the classification result is critical illness, Where, r w Let y be the envelope radius of the critical illness cluster, and ||·|| be the F-norm; when the classification result is non-critical illness, y = 0.
[0034] Ideally, the severity of illness can be graded according to the magnitude of the quantitative value y of the severity of illness. The specifics require doctors to develop industry standards and corresponding intervention plans based on their experience.
[0035] The beneficial effects of this invention are as follows: This invention provides an auxiliary assessment method for critical surgical conditions based on combinatorial machine learning. First, it optimizes the combinatorial machine learning method using an ensemble approach and establishes preoperative and intraoperative auxiliary assessment models, thereby improving the generalization ability of the models. Second, it uses correlation analysis to analyze preoperative factors and trains the preoperative auxiliary assessment model. Then, it uses intraoperative feature values to determine the degree of critical illness and, combined with the time-series data of historical monitoring and surveillance before and during surgery, trains the intraoperative auxiliary assessment model, which can comprehensively and accurately predict the cycle and degree of critical illness, thus helping medical personnel to make effective interventions. Attached Figure Description
[0036] To illustrate the objectives and technical solutions of this invention, the following figures are provided:
[0037] Figure 1 This is a flowchart of the method of the present invention;
[0038] Figure 2 This is an architecture diagram of the ensemble learning method according to an embodiment of the present invention;
[0039] Figure 3This is a framework diagram for intraoperative auxiliary evaluation in an embodiment of the present invention. Detailed Implementation
[0040] Example: A hospital provides historical patient monitoring data from 2014 to 2019, containing eight common indicators for different critical illnesses (diastolic blood pressure, systolic blood pressure, heart rate, body temperature, respiratory rate, arterial oxygen partial pressure, central venous pressure, blood pH, etc.). This data has been segmented into preoperative and intraoperative time-series monitoring data, and further artificially labeled according to physician diagnoses, with negative critical illnesses marked as 0 and positive critical illnesses marked as 1. Further, for model training, all data is divided into a training set (2014-2018) and a test set (2019). To accurately predict the duration and severity of critical illness, this invention proposes a "surgical critical illness auxiliary assessment method based on combinatorial machine learning."
[0041] The following will be combined with the appendix Figure 1 The preferred embodiments of the present invention will be described in detail below.
[0042] Step 1: For the time series data of historical monitoring and surveillance before and during the operation, the missing monitoring indicator data are filled in using the random forest regression algorithm based on decision trees.
[0043] Make full use of effective data in decision-making, use random forest for regression processing, and minimize the regression loss by adjusting the parameters of the random forest algorithm.
[0044] Step 2: Based on the time-series data of preoperative historical monitoring and surveillance, calculate the correlation coefficient of common indicators for critically ill patients, sort the monitoring indicators according to their correlation degree, and take the monitoring indicators with the highest weight total of 99% as preoperative factors.
[0045] Step 3: Using preoperative factors as input to the model, a preoperative auxiliary assessment model adapted to different types of critical illnesses is established using ensemble machine learning methods. The model is trained using time-series data of preoperative historical monitoring and surveillance from the training set, and then tested using data from the test set.
[0046] The aforementioned combined machine learning method is as follows: First, an ensemble learning approach is used to establish regression analysis models from input to critical illness assessment results, employing XGBoost regression, Support Vector Machine regression (SVR), and Random Forest regression methods respectively. Then, all regression analysis models are evaluated using metrics, and the optimal model is selected for the final critical illness assessment. Figure 2The architecture adopted is as follows: three parallel regression analysis models are connected in series with an indicator evaluation model. The function of the indicator evaluation model is to select the output of the regression analysis model with the best evaluation effect of a certain type of indicator among the three regression analysis models as the final output.
[0047] The ensemble learning method is a boosting algorithm; the evaluation metrics include mean squared error (MSE) and mean absolute error (MAE).
[0048] Step 4: Input a patient's preoperative factors into a trained preoperative auxiliary assessment model for critical illnesses to assist doctors in assessing and analyzing preoperative critical illnesses.
[0049] Step 5: Calculate intraoperative feature values for the data marked as critical illness cycles in the intraoperative historical monitoring and surveillance time series data. Similarly, calculate intraoperative feature values for the patient data marked as non-critical illness cycles in the intraoperative historical monitoring and surveillance time series data. Then, cluster the data according to the intraoperative feature values to obtain critical illness cluster centers and non-critical illness cluster centers.
[0050] The intraoperative characteristic values are: a four-dimensional vector consisting of the first moment (mean), the second central moment (variance), the third central moment (skewness), and the fourth central moment (kurtosis) of the data within the period.
[0051] The clustering method described is a binary classification method that uses K-means to classify cluster centers of critically ill patients and cluster centers of non-critically ill patients.
[0052] Step 6: Combining Figure 3 An initial monitoring cycle is set. Intraoperative historical monitoring and time-series data are used to calculate intraoperative feature values according to the monitoring cycle. The data are then categorized based on the distance between the intraoperative feature values and the cluster centers for critical and non-critical cases. Finally, a loss function is used to calculate the optimal monitoring cycle. Specifically:
[0053] (1) Set the sampling period of the time series data of the historical monitoring and surveillance during the operation to the minimum step size, and initially set the number of steps of a monitoring cycle to 2;
[0054] (2) Calculate intraoperative characteristic values for the historical monitoring and surveillance data during the operation according to the monitoring cycle, and classify them according to their distance from the critical illness cluster center and the non-critical illness cluster center;
[0055] Specifically, for the time-series data of historical monitoring and surveillance of patients marked as non-critical, intraoperative characteristic values are calculated on a periodic basis over the entire monitoring period, and then the process is directly skipped to step 7.
[0056] (3) Calculate the loss function based on the classification results and the intraoperative characteristic values within the monitoring period;
[0057] (4) Increase the number of monitoring cycle steps by 1 and repeat steps S702 to S703 until 20 times. Select the monitoring cycle with the smallest loss function in 20 iterations as the optimal monitoring cycle.
[0058] The loss function is:
[0059]
[0060] Among them, c w c n , , represent the cluster centers for critically ill patients and those for non-critically ill patients, respectively, and p represents the time-series data of historical monitoring and surveillance of a single patient before and during surgery.
[0061] Step 7: Calculate the quantitative value of the severity of critical illness using the time-series data of intraoperative historical monitoring and surveillance within the optimal monitoring period.
[0062] The method for calculating the critical illness severity quantification value y is as follows: when the classification result is critical illness, Where, r w y is the envelope radius for the critical illness cluster; when the classification result is non-critical illness, y = 0.
[0063] Specifically, the severity of illness can be graded according to the magnitude of the quantitative value 'y'. Specific criteria and corresponding intervention plans need to be developed by doctors based on their experience and industry standards. For example, 0 < y < 0.5 is considered Grade 1 critical illness, and 0.5 ≤ y < 0.7 is considered Grade 2 critical illness.
[0064] Step 8: Using preoperative factors, intraoperative feature values, and criticality severity quantification values as inputs to the model, a combined machine learning method is used to establish an intraoperative auxiliary assessment model adapted to different types of critical illnesses. The model is trained using time-series data of preoperative and intraoperative historical monitoring and surveillance from the training set, and then tested using the data.
[0065] Step 9: Input a patient's preoperative factors, intraoperative characteristic values, and criticality severity quantification values into a trained intraoperative auxiliary assessment model for critical illnesses to assist doctors in assessing and analyzing intraoperative critical illnesses.
[0066] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made to it in form and detail without departing from the scope defined by the claims of the present invention.
Claims
1. A surgical critical care auxiliary assessment method based on ensemble machine learning, characterized in that, Includes the following steps: S1: The historical monitoring and surveillance time series data are divided into preoperative historical monitoring and surveillance time series data and intraoperative historical monitoring and surveillance time series data according to the preoperative and intraoperative stages, and the critical illness and its cycle are artificially marked according to the doctor's diagnosis conclusion. S2: For the time series data of historical monitoring and surveillance before and during the operation, the missing monitoring indicator data are filled in using the random forest regression algorithm based on decision trees. S3: Based on the time-series data of preoperative historical monitoring and surveillance, conduct correlation analysis on common indicators of critical illness, and extract the monitoring indicators with the greatest impact on critical illness as preoperative factors. S4: Using preoperative factors as input to the model, a preoperative auxiliary assessment model adapted to different types of critical illnesses is established using combinatorial machine learning methods, and trained using time-series data of preoperative historical monitoring and surveillance. S5: Input a patient's preoperative factors into a trained preoperative auxiliary assessment model for critical illnesses to assist doctors in assessing and analyzing preoperative critical illnesses. S6: Calculate intraoperative feature values for data marked as critical illness cycles in the intraoperative historical monitoring and surveillance time series data. Similarly, calculate intraoperative feature values for patient data marked as non-critical illness cycles in the intraoperative historical monitoring and surveillance time series data. Then, cluster the data according to the intraoperative feature values to obtain critical illness cluster centers and non-critical illness cluster centers. S7: Initially set a monitoring cycle, calculate intraoperative feature values based on the time series data of historical monitoring and surveillance during the operation according to the monitoring cycle, classify according to the distance between the intraoperative feature values and the cluster centers of critical and non-critical cases, and calculate the optimal monitoring cycle using the loss function; S8: Calculate the quantitative value of the severity of critical illness using the time-series data of intraoperative historical monitoring and surveillance within the optimal monitoring period; S9: Using preoperative factors, intraoperative feature values, and critical illness severity quantification values as inputs to the model, a combined machine learning method is used to establish an intraoperative auxiliary assessment model that adapts to different types of critical illnesses, and the model is trained using time-series data of historical monitoring and surveillance before and during the operation. S10: Input a patient's preoperative factors, intraoperative characteristic values, and criticality severity quantification values into a trained intraoperative auxiliary assessment model for critical illnesses to assist doctors in assessing and analyzing intraoperative critical illnesses.
2. The surgical critical care auxiliary assessment method based on combinatorial machine learning according to claim 1, characterized in that, The combined machine learning method described in steps S4 and S9 is as follows: First, an ensemble learning method is used to establish regression analysis models from the perspectives of artificial deep neural networks, XGBOOST regression method, support vector machine, and random forest regression method, respectively, to output the critical illness judgment result; then, all regression analysis models are evaluated by indicators, and the optimal regression analysis model is selected for the final judgment of the critical illness result.
3. The surgical critical care auxiliary assessment method based on combinatorial machine learning according to claim 2, characterized in that, The ensemble learning methods mentioned include Bootstrap Aggregating, Boosting, and Random Forest; the evaluation metrics include Mean Squared Error (MSE), Mean Absolute Error (MAE), and Coefficient of Determinism (R-squared).
4. The surgical critical care auxiliary assessment method based on combinatorial machine learning according to claim 1, characterized in that, The intraoperative characteristic value mentioned in step S6 is a four-dimensional vector consisting of the first moment (mean), second central moment (variance), third central moment (skewness), and fourth central moment (kurtosis) of the data within the period.
5. The surgical critical care auxiliary assessment method based on ensemble machine learning according to claim 1, characterized in that, The clustering method described in step S6 is a binary classification method that uses the K-means method to classify critically ill patients into cluster centers and non-critically ill patients into cluster centers.
6. The surgical critical care auxiliary assessment method based on combinatorial machine learning according to claim 1, characterized in that, The loss function mentioned in step S7 is: in, , They are designated as cluster centers for critically ill patients and non-critically ill patients, respectively. This is time-series data of historical monitoring and surveillance of a single patient before and during surgery.
7. The surgical critical care auxiliary assessment method based on combinatorial machine learning according to claim 1, characterized in that, The specific steps of step S7 are as follows: S701: Set the sampling period of the time series data of the historical monitoring and surveillance during the operation to the minimum step size. Initially, the number of steps in one monitoring cycle is set to 2. S702: Calculate intraoperative characteristic values based on the time series data of historical monitoring and surveillance during the operation according to the monitoring cycle, and classify them according to their distance from the critical illness cluster center and the non-critical illness cluster center; S703: Calculate the loss function based on the classification results and intraoperative feature values within the monitoring period; S704: Increase the number of monitoring cycle steps by 1, and repeat steps S702~S703 until N times, and select the monitoring cycle with the smallest loss function as the optimal monitoring cycle.
8. The surgical critical care auxiliary assessment method based on combinatorial machine learning according to claim 1, characterized in that, The quantification value of the severity of illness described in step S8 The calculation method is as follows: when the classification result is critical illness, ,in, This represents the envelope radius for the critical illness cluster; when the classification result is non-critical illness... .
Citation Information
Patent Citations
Device monitoring system
EP3839975A1
Postoperative complication prediction method, postoperative complication prediction program, and postoperative complication prediction device
JP2020057053A