Heart operation postoperative complication prediction system

Abstract

The invention relates to the technical field of postoperative complication prediction, in particular to a cardiac surgery postoperative complication prediction system, and aims to improve prediction accuracy. The prediction system comprises: a data collection module, which is used for collecting perioperative characteristic sample data of a patient and a classification label indicating whether delirium occurs after an operation to obtain a sample data set; the feature screening module is used for screening out a plurality of important features from the perioperative period features according to the sample data set; the model construction module is used for evaluating N preset machine learning models based on the sample data of the important features and training an optimal model; the data acquisition module is used for acquiring important feature data of a to-be-predicted patient; the delirium prediction module is used for inputting the important feature data into the trained optimal model and outputting postoperative delirium occurrence categories of the patient to be predicted and the probability corresponding to each category, interference of redundant data is avoided, and prediction accuracy and reliability are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of postoperative complication prediction, and in particular to a postoperative complication prediction system for cardiac surgery. Background Art

[0002] Postoperative delirium after cardiac surgery is an acute and fluctuating change in mental state after anesthesia and surgery, manifested as confusion, disorientation, inattention and impaired cognitive function in patients. The incidence rate is 10%-50%, and its occurrence is usually related to multiple factors such as age, underlying diseases, drug use, and hypotension during surgery.

[0003] Studies have shown that intraoperative hemodynamic fluctuations, especially changes in mean arterial pressure (MAP), are closely related to the occurrence of postoperative delirium.

[0004] However, existing machine learning models for predicting delirium after cardiac surgery still make relatively extensive use of intraoperative MAP data and fail to fully tap the predictive value of its dynamic characteristics, which limits the accuracy of the final prediction results. Summary of the Invention

[0005] In order to solve the above problems in the prior art, the present invention proposes a cardiac surgery postoperative complication prediction system, which improves the accuracy of the prediction results.

[0006] The present invention provides a system for predicting complications after cardiac surgery, the system comprising:

[0007] A data collection module is used to collect sample data on perioperative characteristics of adult patients undergoing a specific type of cardiac surgery and classification labels for whether postoperative delirium occurs, thereby obtaining a sample data set;

[0008] A feature screening module, configured to screen the importance of the perioperative features according to the sample data set to obtain a number of important features;

[0009] A model building module, configured to evaluate N preset machine learning models based on the sample data of the important features and train the optimal model;

[0010] a data acquisition module, configured to acquire the important feature data of the patient to be predicted to undergo the specific type of cardiac surgery;

[0011] A delirium prediction module, configured to input the important feature data into the trained optimal model and output the postoperative delirium occurrence category of the patient to be predicted and the probability corresponding to each category;

[0012] in,

[0013] N is a preset value;

[0014] The perioperative characteristics include: the patient's preoperative basic information, past medical history, medication history, examination and test results, intraoperative information, and WTF score;

[0015] The WTF score is a comprehensive score of the degree of decrease in intraoperative MAP value (weight), cumulative time (time), and number of occurrences (frequency).

[0016] The prediction system of the present invention screens out important characteristics that have a greater impact on postoperative delirium from the perioperative characteristics and incorporates them into the prediction model, avoiding interference from redundant data; taking the WTF score as a comprehensive evaluation index for intraoperative MAP value fluctuations can comprehensively reflect the impact of the degree of intraoperative blood pressure decrease, cumulative time, and number of occurrences on postoperative delirium, and selects the optimal model from multiple machine learning models, thereby significantly improving the accuracy and reliability of the prediction.

[0017] Preferably, the system further includes:

[0018] A scoring module for calculating the WTF score.

[0019] The scoring module includes:

[0020] A MAP value acquisition unit for acquiring invasive arterial blood pressure data once every preset time interval throughout the patient's surgery; the invasive arterial blood pressure data includes: MAP (mean arterial pressure) value, systolic blood pressure value, and diastolic blood pressure value;

[0021] A first scoring unit for scoring the degree of blood pressure decrease of the patient in four preset segments when the MAP value is less than MAP4:

[0022] If MAP3 ≤ MAP < MAP4 and the duration exceeds 1 minute, then W = W1;

[0023] If MAP2 ≤ MAP < MAP3 and the duration exceeds 1 minute, then W = W2;

[0024] If MAP1 ≤ MAP < MAP2 and the duration exceeds 1 minute, then W = W3;

[0025] If MAP < MAP1 and the duration exceeds 1 minute, then W = W4;

[0026] Among them, MAP represents the mean arterial pressure, which is obtained by acquiring invasive arterial blood pressure data at each preset time interval. MAP1, MAP2, MAP3, and MAP4 are all preset mean arterial pressure thresholds. W represents the score of the degree of blood pressure drop, and W1, W2, W3, and W4 are all preset score values of the degree of blood pressure drop;

[0027] A second scoring unit for scoring the cumulative time of the patient's hypotension in each segment:

[0028] If t1 ≤ t < t2, then T = T1;

[0029] If t2 ≤ t < t3, then T = T2;

[0030] If t3 ≤ t < t4, then T = T3;

[0031] If t4 ≤ t < t5, then T = T4;

[0032] If t ≥ t5, then T = T5;

[0033] Among them, t represents the cumulative time corresponding to the patient's blood pressure being hypotensive in one or more segments during the operation. t1, t2, t3, t4, and t5 are all preset cumulative time thresholds. T represents the score of the cumulative time, and T1, T2, T3, T4, and T5 are all preset score values of the cumulative time;

[0034] A third scoring unit for scoring the cumulative occurrence times of the patient's hypotension in each segment:

[0035] If n ≤ n1, then F = F1;

[0036] If n1 < n ≤ n2, then F = F2;

[0037] If n2 < n ≤ n3, then F = F3;

[0038] If n ≥ n3, then F = F4;

[0039] Among them, n represents the cumulative occurrence times corresponding to the patient's blood pressure being hypotensive in one or more segments during the operation. n1, n2, and n3 are all preset cumulative occurrence times thresholds, F represents the score of the cumulative occurrence times, and F1, F2, F3, and F4 are all preset score values of the cumulative occurrence times;

[0040] A summing unit for calculating the total score according to the W, T, and F scores corresponding to the patient's hypotension in each segment to obtain the WTF score.

[0041] The WTF score proposed in this invention quantifies the severity, cumulative time, and number of occurrences of intraoperative hypotension in multiple dimensions, which not only improves the accuracy of postoperative delirium risk assessment, but also provides a scientific tool for real-time intraoperative intervention and mechanism research.

[0042] Preferably, the types of cardiac surgery include: coronary artery bypass grafting, heart valve surgery, congenital heart disease correction, great vessel surgery, heart transplantation, mechanical assist device implantation and arrhythmia surgery;

[0043] The specific type is one of the types of cardiac surgery;

[0044] The scoring module also includes:

[0045] The MAP value correction unit is used to determine whether each of the obtained MAP values is abnormal according to a preset condition, and replace the abnormal MAP value by linear interpolation.

[0046] The present invention accurately determines whether the MAP value is abnormal through preset conditions and replaces abnormal values with linear interpolation, effectively eliminating noise data caused by measurement errors or interference, ensuring the integrity and accuracy of perioperative MAP data, and providing a high-quality data foundation for subsequent MAP data-based analysis (such as WTF score calculation, postoperative delirium prediction, etc.).

[0047] Optionally, the type of cardiac surgery is off-pump coronary artery bypass grafting; the feature screening module includes:

[0048] A first screening unit is used to screen the importance of various perioperative features using the Boruta algorithm to obtain a preliminary selection set;

[0049] The second screening unit is used to determine the correlation between the features marked as "Confirmed" or "Tentative" in the preliminary set through the variance inflation factor and the Pearson correlation coefficient matrix, remove strongly correlated features, and obtain the set of important features.

[0050] Preferably, the model building module includes:

[0051] a data acquisition unit, configured to acquire sample data of the important features from the sample data set to form a first historical data set;

[0052] a preprocessing unit, configured to normalize the values in the first historical data set and use a SMOTE (Synthetic Minority Over-sampling Technique) oversampling technique to address the class imbalance problem, thereby obtaining a second historical data set;

[0053] a model selection unit, configured to evaluate the normalization performance of each of the N preset machine learning models using a 10-fold cross-validation method based on the second historical data set, and then select the optimal model according to the evaluation index;

[0054] A parameter setting unit, configured to set hyperparameters, decision thresholds, and penalty coefficients for the optimal model;

[0055] a training unit, configured to train the optimal model based on the second historical data set;

[0056] in,

[0057] The evaluation indicators include: area under the receiver operating characteristic curve, accuracy, precision, recall, area under the precision-recall curve and F1 score.

[0058] Preferably, the delirium occurrence categories include: delirium occurrence and no delirium occurrence;

[0059] The system further comprises:

[0060] The alarm module is used to issue an alarm message when the probability of delirium exceeds a preset probability threshold.

[0061] Preferably, the system further comprises:

[0062] An evaluation module is used to analyze the clinical usefulness of the optimal model using a decision curve and generate a calibration curve to evaluate the accuracy of risk prediction.

[0063] The global explanation module is used to perform a global explanation on the important features input into the optimal model through SHAP (SHapley Additive exPlanation) values, and give an importance ranking of the important features to the prediction results.

[0064] The local explanation module is used to use the ICE (Individual Conditional Expectation) mean curve and the PDP (Partial Dependence Plot) curve to explain the impact of each important feature on the prediction result under different values.

[0065] Through the above-mentioned evaluation and explanation, the prediction system of the present invention can enable medical staff to understand the impact of each of the important features on the prediction results, thereby helping medical staff to formulate more accurate intervention measures.

[0066] Preferably, the system further comprises:

[0067] A first label replacement module is configured to replace the classification label in the sample data set with whether acute kidney injury occurs after surgery before the feature screening module, the model building module, and the data acquisition module are run;

[0068] The renal injury prediction module is used to input the important feature data into the trained optimal model and output the occurrence category of postoperative acute renal injury of the patient to be predicted and the probability corresponding to each category.

[0069] Preferably, the system further comprises:

[0070] A second label replacement module is configured to replace the classification label in the sample data set with whether myocardial injury occurs after surgery before the feature screening module, the model building module, and the data acquisition module are run;

[0071] The myocardial injury prediction module is used to input the important feature data into the trained optimal model and output the postoperative myocardial injury occurrence category of the patient to be predicted and the probability corresponding to each category.

[0072] Preferably, the system further comprises:

[0073] a third label replacement module, configured to replace the classification label in the sample data set with whether new-onset atrial fibrillation occurs after surgery before the feature screening module, the model building module, and the data acquisition module are run;

[0074] The atrial fibrillation prediction module is used to input the important feature data into the trained optimal model and output the occurrence category of new atrial fibrillation after surgery of the patient to be predicted and the probability corresponding to each category.

[0075] The prediction system of the present invention only needs to perform a simple label replacement operation to use the sample data set collected during the previous delirium prediction to build an optimal model for other complications such as acute kidney injury, myocardial injury or new-onset atrial fibrillation, thereby realizing the prediction of other complications such as acute kidney injury, myocardial injury or new-onset atrial fibrillation. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 This is a schematic diagram of the main structure of the first embodiment of the cardiac surgery postoperative complication prediction system of the present invention;

[0077] Figure 2 This is a schematic diagram of the main structure of the second embodiment of the cardiac surgery postoperative complication prediction system of the present invention;

[0078] Figure 3 This is a schematic diagram of the main structure of Example 3 of the cardiac surgery postoperative complication prediction system of the present invention;

[0079] Figure 4 is the SHAP value ranking of important features used in delirium prediction in the embodiment of the present invention;

[0080] Figure 5(a)-Figure 5(e) They are the ICE mean curve and PDP curve of preoperative lactate dehydrogenase, preoperative hemoglobin, preoperative plasma albumin, patient age and intraoperative WTF score in the embodiment of the present invention. DETAILED DESCRIPTION

[0081] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these embodiments are only used to explain the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0082] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0083] It should be noted that, in the description of the present invention, the terms "first" and "second" are merely for the convenience of description, and do not indicate or imply the relative importance of the devices, elements or parameters, and therefore should not be understood as limiting the present invention. In addition, the term "and / or" in the present invention is merely a description of the association relationship between associated objects, indicating that three relationships may exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this document, unless otherwise specified, generally indicates that the associated objects are in an "or" relationship.

[0084] The present invention is mainly aimed at predicting the risk of delirium after cardiac surgery, and can also be used to predict other complications such as postoperative acute kidney injury, myocardial injury and atrial fibrillation.

[0085] Figure 1 This is a schematic diagram of the main structure of the embodiment 1 of the cardiac surgery postoperative complications prediction system of the present invention. Figure 1 As shown, the prediction system of this embodiment includes: a data collection module 101, a feature screening module 102, a model construction module 103, a data acquisition module 104 and a delirium prediction module 105.

[0086] The data collection module 101 is used to collect sample data of perioperative characteristics of adult patients undergoing a specific type of cardiac surgery and classification labels of whether postoperative delirium occurs, to obtain a sample data set.

[0087] Among them, the types of cardiac surgery include: coronary artery bypass grafting, heart valve surgery, congenital heart disease correction, great vessel surgery, heart transplantation, mechanical assist device implantation, and arrhythmia surgery; the "specific type" in the present invention refers to one of the above-mentioned types of cardiac surgery. Perioperative characteristics include: the patient's preoperative basic information (age, height, weight, education level, etc.), medical history, medication history, examination and test results, intraoperative information, and WTF score. The WTF score is a score obtained by comprehensively evaluating the degree of decrease in the intraoperative MAP value, the cumulative time, and the number of occurrences.

[0088] In this example, the specific type is OPCABG (off-pump coronary artery bypass grafting). Therefore, we collected perioperative data from 2,081 adult patients who underwent OPCABG at a Class A tertiary hospital between June 1, 2021, and April 30, 2023. We excluded all patients who did not undergo continuous intraoperative arterial blood pressure data collection and those who died after surgery. Features with missing values exceeding 10% were removed, and missing categorical data were imputed using the mode, while missing continuous data were imputed using the mean.

[0089] OPCABG completes vascular anastomosis without stopping the heart, avoiding non-physiological perfusion of extracorporeal circulation and cardiac arrest.

[0090] In this embodiment, the delirium occurrence categories include: delirium occurs and delirium does not occur. When adding classification labels, the diagnosis and exclusion criteria for postoperative delirium are:

[0091] (1) Diagnostic criteria: The mental status of all patients was assessed preoperatively and postoperatively using the Mini-Mental State Examination (MMSE), with a maximum score of 30 points. Scores between 27 and 30 were considered normal, and scores <27 were considered cognitive impairment.

[0092] (2) Exclusion criteria: Patients were excluded if any of the following conditions were present: i. MMSE score < 27 before surgery; ii. Severe renal insufficiency or abnormal liver function (transaminase greater than 1.5 times the normal concentration) before surgery.

[0093] The feature screening module 102 is used to screen the importance of perioperative features according to the sample data set to obtain several important features.

[0094] Specifically, for different cardiac surgery types or different sample data, the important feature screening algorithm may be different. For example, when the cardiac surgery type is off-pump coronary artery bypass grafting, the feature screening module 102 may include: a first screening unit and a second screening unit.

[0095] Among them, the first screening unit is used to use the Boruta algorithm to screen the importance of various perioperative features to obtain a preliminary set; the second screening unit is used to determine the correlation between features marked as "Confirmed" or "Tentative" in the preliminary set through the variance inflation factor and Pearson correlation coefficient matrix, remove strongly correlated features, and obtain a set of important features.

[0096] The Boruta algorithm uses a random forest model to calculate the importance of each feature related to the outcome and rank them. Features marked as "Confirmed" have the strongest correlation with the outcome and are considered important features. Features marked as "Rejected" have low importance and are considered irrelevant. Features whose importance cannot be clearly determined are marked as "Tentative."

[0097] In this embodiment, the "Rejected" feature is discarded, and the "Confirmed" and "Tentative" features are again judged by the variance inflation factor (VIF) and Pearson correlation coefficient matrix to determine the correlation between the features. If the VIF between the features is greater than 5 or the correlation coefficient is greater than 0.5, only the features with high importance are retained. After screening, the test results of preoperative lactate dehydrogenase, preoperative plasma albumin, preoperative hemoglobin, patient age and intraoperative hypotension WTF score are used for delirium prediction. The perioperative feature screening process for other postoperative complications such as acute kidney injury, myocardial injury and atrial fibrillation mentioned in the optional embodiments below is consistent with the screening method for delirium.

[0098] The model building module 103 is used to evaluate N preset machine learning models based on sample data of important features and train the optimal model.

[0099] Where N is a preset value, and sample data of important features can be obtained from the sample dataset. In this embodiment, seven machine learning models are preset, including adaptive boosting (Adaboost), decision tree (DT), gradient boosting machine (GBM), logistic regression (LR), random forest (RF), support vector machine (SVM), and extreme gradient boosting (XGBoost).

[0100] Specifically, the model building module 103 may include: a data acquisition unit, a preprocessing unit, a model selection unit, a parameter setting unit and a training unit.

[0101] Among them, the data acquisition unit is used to obtain sample data of important features from the sample data set to form a first historical data set; the preprocessing unit is used to standardize the numerical values in the first historical data set, and use SMOTE (Synthetic Minority Over-sampling Technique) oversampling technology to solve the class imbalance problem to obtain a second historical data set; the model selection unit is used to evaluate the normalization performance of N preset machine learning models based on the second historical data set using a 10-fold cross-validation method, and then select the optimal model according to the evaluation index; the parameter setting unit is used to set hyperparameters, decision thresholds and penalty coefficients for the optimal model; the training unit is used to train the optimal model based on the second historical data set.

[0102] SMOTE is a data augmentation technique used to address class imbalance. Its core idea is to increase the number of minority class samples by synthesizing new minority class samples, thereby improving the model's prediction ability for the minority class.

[0103] In this embodiment, 10-fold cross validation is used in the model selection unit. The sample data of important features are randomly divided into 10 parts without repeated sampling. Each time, 9 parts are selected as training sets and the remaining 1 part is used as test sets. This is repeated 10 times. Each time, the best model is fitted on the training set, and then prediction is performed on the test set. The evaluation index of the model is calculated and saved. Finally, the average values of the indexes of the 10 test sets of the 7 machine learning models are calculated and filled in Table 1:

[0104] Table 1 Evaluation metrics of each machine learning prediction model for postoperative delirium

[0105]

[0106] As shown in Table 1, the evaluation metrics include: area under the receiver operating characteristic curve (AUC-ROC), accuracy, precision, recall, area under the precision-recall curve (PR-AUC), and F1 score. Table 1 shows that the XGBoost model performs best, with an AUC-ROC of 0.873, indicating that the model is well able to distinguish between postoperative delirium and non-postoperative delirium. The accuracy of 0.759 indicates that the model has good ability to correctly predict delirium. The precision of 0.902 indicates that the model has high ability to predict delirium. The recall of 0.564, the area under the precision-recall curve (PR-AUC) of 0.880, and the F1 score of 0.692 indicate that the model has few misjudgments when predicting delirium and can more comprehensively capture postoperative delirium. Therefore, XGBoost is selected as the optimal model.

[0107] The data acquisition module 104 is used to acquire important feature data of the patient to be predicted who will undergo a specific type of cardiac surgery.

[0108] The delirium prediction module 105 is used to input the important feature data into the trained optimal model, and output the postoperative delirium occurrence category of the patient to be predicted and the probability corresponding to each category.

[0109] Figure 2 This is a schematic diagram of the main structure of the second embodiment of the cardiac surgery postoperative complications prediction system of the present invention. Figure 2 As shown, the prediction system of this embodiment includes: a data collection module 201, a feature screening module 202, a model construction module 203, a data acquisition module 204, a delirium prediction module 205, a scoring module 206 and an alarm module 207.

[0110] Among them, the data collection module 201, the feature screening module 202, the model construction module 203, the data acquisition module 204 and the delirium prediction module 205 correspond to the data collection module 101, the feature screening module 102, the model construction module 103, the data acquisition module 104 and the delirium prediction module 105 in the above-mentioned embodiment 1, respectively, and are not repeated here.

[0111] In this embodiment, the scoring module 206 is used to calculate the WTF score.

[0112] Specifically, the scoring module 206 may include: a MAP value acquiring unit, a MAP value correcting unit, a first scoring unit, a second scoring unit, a third scoring unit, and a summing unit.

[0113] The MAP value acquisition unit is used to acquire invasive arterial blood pressure data once every preset time interval during the entire surgical process of the patient; the invasive arterial blood pressure data includes: MAP value, systolic blood pressure value, and diastolic blood pressure value.

[0114] Among them, the MAP (mean arterial pressure) value is the average value of arterial blood pressure in one cardiac cycle, called mean arterial pressure, and is calculated according to the following formula (1):

[0115] MAP = diastolic blood pressure + 1 / 3 (systolic blood pressure - diastolic blood pressure) (1)

[0116] The MAP value can comprehensively reflect the functions and states of the heart and peripheral blood vessels.

[0117] In this embodiment, invasive arterial blood pressure data is collected by a LiDCO hemodynamic monitor during the operation, and the acquisition frequency is 1 Hz (that is, the preset time interval is 1 second).

[0118] The MAP value correction unit is used to determine whether each acquired MAP value is abnormal according to preset conditions, and replace the abnormal MAP values by linear interpolation.

[0119] In this embodiment, the preset condition for determining that the MAP value is abnormal is: systolic blood pressure < 60 mmHg, or systolic blood pressure > 180 mmHg, or diastolic blood pressure < 30 mmHg, or diastolic blood pressure > 110 mmHg.

[0120] The first scoring unit is used to score the degree of blood pressure drop of the patient according to four preset segments when the MAP value is less than MAP4: [[ID=2,22]]

[0121] If MAP3 ≤ MAP < MAP4 and the duration exceeds 1 minute, then W = W1;

[0122] If MAP2 ≤ MAP < MAP3 and the duration exceeds 1 minute, then W = W2;

[0123] If MAP1 ≤ MAP < MAP2 and the duration exceeds 1 minute, then W = W3;

[0124] If MAP < MAP1 and the duration exceeds 1 minute, then W = W4;

[0125] Among them, MAP represents mean arterial pressure, which is obtained by acquiring invasive arterial blood pressure data at each preset time interval. MAP1, MAP2, MAP3, and MAP4 are all preset mean arterial pressure thresholds, W represents the score of blood pressure decline degree, and W1, W2, W3, and W4 are all preset blood pressure decline degree score values. In this embodiment, to predict the probability of delirium after off - pump coronary artery bypass grafting, MAP1, MAP2, MAP3, and MAP4 are 50 mmHg, 55 mmHg, 60 mmHg, and 65 mmHg respectively, and W1, W2, W3, and W4 are 1, 2, 3, and 4 respectively.

[0126] A second scoring unit for scoring the cumulative time of the patient's hypotension in each segment:

[0127] If t1 ≤ t < t2, then T = T1;

[0128] If t2 ≤ t < t3, then T = T2;

[0129] If t3 ≤ t < t4, then T = T3;

[0130] If t4 ≤ t < t5, then T = T4;

[0131] If t ≥ t5, then T = T5;

[0132] Among them, t represents the cumulative time corresponding to the patient's blood pressure being in a certain segment of hypotension once or multiple times during the operation. t1, t2, t3, t4, and t5 are all preset cumulative time thresholds, T represents the score of the cumulative time, and T1, T2, T3, T4, and T5 are all preset cumulative time score values. In this embodiment, t1, t2, t3, t4, and t5 are 1 min, 2 min, 5 min, 10 min, and 20 min respectively, and T1, T2, T3, T4, and T5 are 1, 2, 3, 4, and 5 respectively.

[0133] A third scoring unit for scoring the cumulative occurrence times of the patient's hypotension in each segment:

[0134] If n ≤ n1, then F = F1;

[0135] If n1 < n ≤ n2, then F = F2;

[0136] If n2 < n ≤ n3, then F = F3;

[0137] If n ≥ n3, then F = F4;

[0138] Where n represents the cumulative number of times a patient's blood pressure experiences one or more episodes of hypotension in a certain segment during surgery, n1, n2, and n3 are preset thresholds for the cumulative number of occurrences, F represents a score for the cumulative number of occurrences, and F1, F2, F3, and F4 are preset scores for the cumulative number of occurrences. In this embodiment, n1, n2, and n3 are 2, 5, and 10, respectively, and F1, F2, F3, and F4 are 1, 2, 3, and 4, respectively.

[0139] The summing unit is used to calculate the total score according to the W, T and F scores corresponding to each segment of hypotension of the patient to obtain the WTF score.

[0140] For example, the invasive arterial blood pressure data for a patient are as follows:

[0141] The MAP value was between 60 and 65 mmHg for 424 seconds and occurred 3 times; therefore, in this stage, the W score was 1, the T score was 3, and the F score was 2;

[0142] The MAP value was between 55 and 60 mmHg for 77 seconds and occurred 1 time; therefore, the W score was 2, the T score was 1, and the F score was 1 during this period;

[0143] The time that the MAP value was between 50 and 55 mmHg was 335 seconds, and the number of occurrences was 4. Therefore, in this stage, the W score was 3, the T score was 3, and the F score was 2;

[0144] The time between MAP values < 50 mmHg was 770 seconds and the number of occurrences was 5, so W score was 4, T score was 4, and F score was 2 during this period.

[0145] Calculate the sum of each part above: 1+3+2+2+1+1+3+3+2+4+4+2=28, that is, the patient's WTF score is 28.

[0146] In this embodiment, the alarm module 207 is configured to issue an alarm message when the probability of delirium exceeds a preset probability threshold.

[0147] Figure 3 This is a schematic diagram of the main structure of the third embodiment of the cardiac surgery postoperative complications prediction system of the present invention. Figure 3 As shown, the prediction system of this embodiment includes: a data collection module 301, a feature screening module 302, a model building module 303, a data acquisition module 304, a delirium prediction module 305, a scoring module 306, an alarm module 307, an evaluation module 308, a global interpretation module 309 and a local interpretation module 310.

[0148] Among them, the data collection module 301, the feature screening module 302, the model construction module 303, the data acquisition module 304, the delirium prediction module 305, the scoring module 306 and the alarm module 307 correspond to the data collection module 201, the feature screening module 202, the model construction module 203, the data acquisition module 204, the delirium prediction module 205, the scoring module 206 and the alarm module 207 in the above-mentioned embodiment 2, and are not repeated here.

[0149] The evaluation module 308 is used to analyze the clinical usefulness of the optimal model using a decision curve and generate a calibration curve to evaluate the accuracy of risk prediction.

[0150] In this example, the XGBoost model showed a greater net gain in threshold probability compared with the other six machine learning models according to the decision curve, indicating that the XGBoost model has good clinical utility.

[0151] The global explanation module 309 is used to perform a global explanation on the important features of the input optimal model through SHAP values, and give an importance ranking of the important features to the prediction results.

[0152] Figure 4 is the SHAP value ranking of the important features used in delirium prediction in the embodiment of the present invention. Figure 4 As shown in the figure, the vertical axis represents the important features, and the horizontal axis represents the average absolute SHAP value. In the prediction of postoperative delirium, the importance of each important feature is as follows from high to low: age, preoperative plasma albumin, preoperative lactate dehydrogenase, intraoperative WTF score, and preoperative hemoglobin.

[0153] The local explanation module 310 is used to use the ICE (Individual Conditional Expectation) mean curve and the PDP (Partial Dependence Plot) curve to explain the impact of each important feature on the prediction result under different values.

[0154] Figure 5(a)-Figure 5(e) They are the ICE mean curve and PDP curve of preoperative lactate dehydrogenase, preoperative hemoglobin, preoperative plasma albumin, patient age and intraoperative WTF score in the embodiment of the present invention.

[0155] The ICE curve can reflect how the prediction results of a single sample patient change with the value of a certain important feature. In this embodiment, the average value of the ICE curves of all sample patients is plotted. Figure 5(a)-Figure 5(e)The medium gray ICE mean curve reflects the influence of the important feature values of most patients on the output (predicted probability) of the delirium prediction model. The 95% confidence interval of the ICE mean curve is estimated based on the variance of the statistical model output, which is shown as the gray shaded part in the figure, reflecting the stability and reliability of the prediction results. The black curve (Smoothed PDP) is the curve obtained by spline smoothing the mean of ICE (ie PDP) of all sample patients, which aims to present the overall trend more intuitively. ICE / PDP calculations and their graphics are implemented using PyCharm 2024.2.1. As shown in Figure 5(e), the WTF PDP curve decreases in the first half because the clinicians manage the patients' blood pressure well during surgery, resulting in a large base of people with low WTF scores. However, in the second half, WTF of more than 30 points significantly increases the probability of postoperative delirium. Therefore, high intraoperative WTF scores are one of the important reasons for the increased probability of postoperative delirium. According to Figure 5(a)-Figure 5(e) It can be seen that high preoperative lactate dehydrogenase level, high preoperative hemoglobin level, high intraoperative WTF score, high age and low preoperative plasma albumin level can lead to the occurrence of postoperative delirium.

[0156] In an optional embodiment, the prediction system may further include: a first label replacement module and a renal injury prediction module.

[0157] The first label replacement module is used to replace the classification labels in the sample dataset with whether or not postoperative acute kidney injury has occurred before the feature screening module 302, model building module 303, and data acquisition module 304 are executed. The kidney injury prediction module is used to input the important feature data into the trained optimal model after the execution of these three modules, and output the postoperative acute kidney injury category and the corresponding probability of each category for the patient to be predicted. In this optional embodiment, when predicting acute kidney injury, the delirium prediction module is no longer executed, and the alarm module can be modified to issue an alarm message if the probability of acute kidney injury exceeds a preset value.

[0158] At the time of label replacement, the diagnosis and exclusion criteria for postoperative acute kidney injury were:

[0159] (1) Diagnostic criteria: The definition of AKI was based on the Kidney Disease Improving Global Outcomes (KDIGO) criteria 3 based on serum creatinine (Cr), with an absolute increase of Cr ≥26.5 μmol / L (≥0.3 mg / dL) or ≥1.5-fold compared with baseline within 7 days after surgery;

[0160] (2) Exclusion criteria: Patients were excluded if any of the following conditions were present: i. A history of kidney-related disease or nephrectomy; ii. Preoperative renal insufficiency or need for renal replacement therapy; iii. Missing preoperative or postoperative Cr data; iv. Death within 7 days after surgery.

[0161] In another optional embodiment, the prediction system may further include: a second label replacement module and a myocardial injury prediction module.

[0162] Among them, the second label replacement module is used to replace the classification label in the sample data set with whether myocardial damage occurs after surgery before the feature screening module 302, the model building module 303 and the data acquisition module 304 are run; the myocardial damage prediction module is used to input important feature data into the trained optimal model after the above three modules are run, and output the category of postoperative myocardial damage in the patient to be predicted and the corresponding probability of each category.

[0163] When label replacement is performed, the diagnostic and exclusion criteria for postoperative myocardial injury are:

[0164] (1) Diagnostic criteria: Myocardial injury can be defined as an increase in serum creatinine troponin cTnT level to more than 99% of the upper reference limit (0.2 μg / L) 24 hours after surgery, and cTnT < 0.1 μg / L is considered normal.

[0165] (2) Exclusion criteria: Patients were excluded if any of the following conditions were present: i. Preoperative or postoperative cTnT deficiency; ii. Preoperative cTnT abnormality, greater than 0.2 μg / L; iii. Preoperative left ventricular ejection fraction <50%; iv. Previous cardiac surgery or preoperative NYHA assessment class III-IV.

[0166] The above outcomes excluded all patients who did not undergo continuous arterial blood pressure data collection during surgery and those who died after surgery.

[0167] In yet another optional embodiment, the prediction system may further include: a third label replacement module and a new-onset atrial fibrillation prediction module.

[0168] Among them, the third label replacement module is used to replace the classification label in the sample data set with whether new-onset atrial fibrillation occurs after surgery before the feature screening module 302, the model building module 303 and the data acquisition module 304 are run; the atrial fibrillation prediction module is used to input important feature data into the trained optimal model after the above three modules are run, and output the category of new-onset atrial fibrillation after surgery for the patient to be predicted and the corresponding probability of each category.

[0169] When label replacement is performed, the diagnostic and exclusion criteria for new-onset atrial fibrillation after surgery are:

[0170] (1) Diagnostic criteria: Postoperative electrocardiogram assessment of whether atrial fibrillation occurs;

[0171] (2) Exclusion criteria: patients with a history of atrial fibrillation or other arrhythmias before surgery.

[0172] In the embodiment of the present invention, the evaluation indicators calculated by the optimal model for delirium, acute kidney injury (AKI), myocardial injury (MI), and new-onset atrial fibrillation (AF) after OPCABG are shown in Table 2:

[0173] Table 2 Evaluation metrics of the optimal machine learning prediction model for each complication

[0174]

[0175] Those skilled in the art should be able to appreciate that the method steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of electronic hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in electronic hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0176] Thus far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is clearly not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art may make equivalent modifications or substitutions to the relevant technical features, and the technical solutions after such modifications or substitutions will fall within the scope of protection of the present invention.

Claims

1. A cardiac surgery postoperative complication prediction system, characterized in that: The system includes: A data collection module, which is used to collect sample data of the perioperative characteristics of adult patients undergoing a specific type of cardiac surgery and classification labels indicating whether delirium occurs after surgery, so as to obtain a sample data set; A feature screening module, which is used to screen the importance of the perioperative characteristics according to the sample data set to obtain several important features; A model construction module, which is used to evaluate N preset machine learning models based on the sample data of the important features and train the optimal model; A data acquisition module, which is used to acquire the important feature data of the patient to be predicted undergoing the specific type of cardiac surgery; A delirium prediction module, which is used to input the important feature data into the trained optimal model and output the postoperative delirium occurrence category of the patient to be predicted and the probability corresponding to each category; Wherein, N is a preset value; The perioperative characteristics include: the patient's preoperative basic information, past medical history, medication history, examination and test results, intraoperative information, and WTF score; The WTF score is a comprehensive score of the degree of decrease in intraoperative MAP value, cumulative time, and occurrence frequency.

2. The cardiac surgery postoperative complication prediction system according to claim 1, characterized in that: The system further includes: A scoring module, which is used to calculate the WTF score; The scoring module includes: A MAP value acquisition unit, which is used to acquire invasive arterial blood pressure data once every preset time interval during the entire operation of the patient; the invasive arterial blood pressure data includes: MAP value, systolic blood pressure value, and diastolic blood pressure value; A first scoring unit, which is used to score the degree of blood pressure decrease of the patient in four preset segments when the MAP value is less than MAP4: If MAP3 ≤ MAP < MAP4 and the duration exceeds 1 minute, then W = W1; If MAP2 ≤ MAP < MAP3 and the duration exceeds 1 minute, then W = W2; If MAP1 ≤ MAP < MAP2 and the duration exceeds 1 minute, then W = W3; If MAP < MAP1 and the duration exceeds 1 minute, then W = W4; Wherein, MAP represents the mean arterial pressure, which is obtained by acquiring invasive arterial blood pressure data once every preset time interval, MAP1, MAP2, MAP3, and MAP4 are all preset mean arterial pressure thresholds, W represents the score of the degree of blood pressure decrease, and W1, W2, W3, and W4 are all preset score values of the degree of blood pressure decrease; A second scoring unit, which is used to score the cumulative time of the patient's hypotension in each segment: If t1 ≤ t < t2, then T = T1; If t2 ≤ t < t3, then T = T2; If t3 ≤ t < t4, then T = T3; If t4 ≤ t < t5, then T = T4; If t ≥ t5, then T = T5; Wherein, t represents the cumulative time corresponding to the patient's blood pressure being in a certain segment of hypotension one or more times during the operation, t1, t2, t3, t4, and t5 are all preset cumulative time thresholds, T represents the score of the cumulative time, and T1, T2, T3, T4, and T5 are all preset score values of the cumulative time; A third scoring unit, which is used to score the cumulative occurrence frequency of the patient's hypotension in each segment: If n ≤ n1, then F = F1; If n1 < n ≤ n2, then F = F2; If \(n_2 \lt n \leq n_3\), then \(F = F_3\); If \(n \geq n_3\), then \(F = F_4\); Where \(n\) represents the cumulative occurrence times corresponding to the patient's blood pressure being in a certain segment of hypotension once or multiple times during the operation, \(n_1\), \(n_2\), and \(n_3\) are all preset cumulative occurrence times thresholds, \(F\) represents the score of the cumulative occurrence times, and \(F_1\), \(F_2\), \(F_3\), and \(F_4\) are all preset cumulative occurrence times score values; A summation unit, configured to calculate the total score according to the \(W\), \(T\), and \(F\) scores corresponding to the patient at each segment of hypotension, and obtain the \(WTF\) score.

3. The postoperative complication prediction system for cardiac surgery according to claim 2, wherein The types of the cardiac surgery include: coronary artery bypass grafting, cardiac valve surgery, congenital heart disease correction surgery, great vessel surgery, heart transplantation, mechanical assist device implantation, and arrhythmia surgery; The specific type is one of the types of the cardiac surgery; The scoring module further includes: A MAP value correction unit, configured to determine whether each obtained MAP value is abnormal according to a preset condition, and replace the abnormal MAP value by linear interpolation.

4. The cardiac surgery postoperative complication prediction system according to claim 3, characterized in that: The type of the cardiac surgery is off - pump coronary artery bypass grafting; The feature screening module includes: A first screening unit, configured to perform importance screening on various peri - operative features by using the Boruta algorithm to obtain a primary selection set; A second screening unit, configured to, for the features marked as "Confirmed" or "Tentative" in the primary selection set, judge the correlation between features through the variance inflation factor and the Pearson correlation coefficient matrix, and remove strongly correlated features to obtain the set of important features.

5. The cardiac surgery postoperative complication prediction system according to claim 1, characterized in that: The model construction module includes: A data acquisition unit, configured to acquire the sample data of the important features from the sample data set to form a first historical data set; A pre - processing unit, configured to perform normalization processing on the numerical values in the first historical data set, and use the SMOTE oversampling technique to solve the class imbalance problem to obtain a second historical data set; A model selection unit, configured to, based on the second historical data set, use the 10 - fold cross - validation method to evaluate the generalization performance of \(N\) preset machine learning models respectively, and then select the optimal model according to the evaluation indexes; A parameter setting unit, configured to set hyperparameters, decision thresholds, and penalty coefficients for the optimal model; A training unit, configured to train the optimal model based on the second historical data set; Where The evaluation indexes include: the area under the receiver operating characteristic curve, accuracy, precision, recall, the area under the precision - recall curve, and the F1 score.

6. The cardiac surgery postoperative complication prediction system according to claim 1, characterized in that: The delirium occurrence categories include: having delirium and not having delirium; The system further includes: An alarm module, configured to send an alarm message when the probability of having delirium exceeds a preset probability threshold.

7. The cardiac surgery postoperative complication prediction system according to claim 1, characterized in that: The system further includes: An evaluation module, configured to use decision curve analysis to analyze the clinical usefulness of the optimal model and generate a calibration curve to evaluate the accuracy of risk prediction. A global interpretation module is used to perform a global interpretation of the important features input into the optimal model through SHAP values, and give an importance ranking of the important features to the prediction results; The local explanation module is used to use the ICE mean curve and the PDP curve to explain the impact of each important feature on the prediction result under different values.

8. The cardiac surgery postoperative complication prediction system according to claim 1, characterized in that: The system further comprises: A first label replacement module is configured to replace the classification label in the sample data set with whether acute kidney injury occurs after surgery before the feature screening module, the model building module, and the data acquisition module are run; The renal injury prediction module is used to input the important feature data into the trained optimal model and output the occurrence category of postoperative acute renal injury of the patient to be predicted and the probability corresponding to each category.

9. The cardiac surgery postoperative complication prediction system according to claim 1, characterized in that: The system further comprises: A second label replacement module is configured to replace the classification label in the sample data set with whether myocardial injury occurs after surgery before the feature screening module, the model building module, and the data acquisition module are run; The myocardial injury prediction module is used to input the important feature data into the trained optimal model and output the postoperative myocardial injury occurrence category of the patient to be predicted and the probability corresponding to each category.

10. The cardiac surgery postoperative complication prediction system according to claim 1, characterized in that: The system further comprises: a third label replacement module, configured to replace the classification label in the sample data set with whether new-onset atrial fibrillation occurs after surgery before the feature screening module, the model building module, and the data acquisition module are run; The atrial fibrillation prediction module is used to input the important feature data into the trained optimal model and output the occurrence category of new atrial fibrillation after surgery of the patient to be predicted and the probability corresponding to each category.

Images