A method for predicting the risk of postoperative delirium of elderly abdominal surgery patients based on perioperative clinical characteristics

Abstract

The present application belongs to the technical field of medical artificial intelligence auxiliary decision-making, and particularly relates to a method for predicting postoperative delirium risk of elderly abdominal surgery patients based on perioperative clinical characteristics. The present application extracts the demographic characteristics, preoperative laboratory indicators and intraoperative anesthesia-related characteristics of elderly abdominal surgery patients based on an electronic medical record system, and constructs a full-feature data set for postoperative delirium risk prediction. Further, a machine learning algorithm is used to establish a postoperative delirium risk prediction model, and the target model is screened in combination with the model stability evaluation result. On this basis, the model is optimized by using SHAP feature contribution analysis and recursive feature elimination strategy, and a simplified prediction model with fewer variables and better generalization performance is obtained. The prediction model can realize early identification of high-risk groups of postoperative delirium of elderly abdominal surgery patients during the perioperative period, has high discrimination ability, stability and clinical applicability, and can provide a reliable tool for precise prevention and auxiliary decision-making during the perioperative period.

Description

Technical Field

[0001] This invention belongs to the field of medical artificial intelligence-assisted decision-making technology, specifically involving a method for predicting the risk of postoperative delirium in elderly patients undergoing abdominal surgery based on perioperative clinical characteristics. Background Technology

[0002] With the accelerating aging of the population, the number of elderly patients undergoing surgical treatment is increasing year by year. Especially with the continuous improvement in anesthesia techniques, perioperative monitoring, and surgical treatment levels, more and more elderly patients are eligible to undergo abdominal surgery. Abdominal surgeries mainly include liver and gallbladder, pancreas, and gastrointestinal surgeries, which are relatively common among elderly patients.

[0003] Postoperative delirium is a common perioperative neurocognitive complication in elderly patients undergoing abdominal surgery, typically manifesting as attention deficit, fluctuating levels of consciousness, disorientation, and acute changes in cognitive status. Postoperative delirium not only prolongs hospital stays but can also lead to persistent cognitive decline and increase the financial burden on patients' families and the healthcare system. Since there is currently no specific treatment for postoperative delirium, and interventions are limited once it occurs, early identification and warning of high-risk patients during the perioperative period are crucial for improving clinical outcomes.

[0004] Although the exact mechanisms underlying postoperative delirium are not fully understood, studies have shown that early screening, risk stratification, and targeted intervention in high-risk patients can reduce the incidence of postoperative delirium to some extent. Therefore, developing an accurate, stable, and clinically applicable risk prediction method based on perioperative clinical information has become an important research direction in this field.

[0005] Current technologies mainly rely on traditional statistical models, such as logistic regression and nomograms, to predict the risk of postoperative delirium, but they still have the following shortcomings in practical applications: (1) Traditional statistical prediction methods usually rely on a limited number of candidate variables to build models, which makes it difficult to fully integrate the diverse, high-dimensional and interrelated clinical information in perioperative electronic medical records, resulting in insufficient model representation of complex risk patterns. (2) Most existing methods assume that there is a linear relationship or a relatively simple association structure between each predictor variable and postoperative delirium outcome, which makes it difficult to characterize the potential nonlinear interaction between age, laboratory indicators, intraoperative anesthetic factors and perioperative management measures. (3) Some existing prediction models are mainly based on a small number of preoperative baseline factors or common risk factors reported in the literature. Their generalization ability and stability are insufficient in different time periods and different patient cohorts. They are prone to the problem that the internal dataset performs well while the time validation set performance declines. (4) With the continuous accumulation of perioperative data in the hospital's electronic medical record system, although machine learning methods have the potential to process high-dimensional data and mine complex patterns, there are still relatively few machine learning studies on the risk prediction of postoperative delirium in elderly patients undergoing abdominal surgery. Existing technologies generally lack a complete technical solution that takes into account prediction performance, time generalization ability, feature optimization ability and model interpretability.

[0006] Therefore, there is an urgent need to provide a technical solution for predicting postoperative delirium risk based on multidimensional clinical characteristics in the perioperative period, so as to achieve early identification of high-risk elderly patients undergoing abdominal surgery and provide reliable technical support for perioperative individualized intervention and clinical auxiliary decision-making. Summary of the Invention

[0007] To address the aforementioned issues, this invention aims to overcome the shortcomings of existing methods for predicting postoperative delirium risk in elderly patients undergoing abdominal surgery, such as insufficient integration of perioperative multidimensional clinical features, limited ability to characterize complex nonlinear risk relationships, unstable model generalization performance, and excessive feature redundancy. This invention proposes a method for predicting postoperative delirium risk in elderly patients undergoing abdominal surgery based on perioperative clinical features, which can be used to identify high-risk patients for postoperative delirium and assist in clinical decision-making.

[0008] The technical solution of the present invention is as follows Figure 1 As shown, it specifically includes: S1. Obtain perioperative clinical data of the subject to be evaluated, including demographic characteristics, preoperative laboratory indicators, and intraoperative anesthesia-related characteristics. S2. Preprocess and feature-construct the perioperative clinical data to obtain a feature dataset for predicting postoperative delirium risk, specifically including: S21. Perform data cleaning, mean or mode filling for missing values, outlier identification, and format standardization on the original clinical data to form structured input data; S22. Standardize the continuous variables in the structured input data. The standardization process refers to scaling the perioperative clinical data proportionally to make it fall into a small specific interval. Encode the categorical variables. The encoding process refers to converting non-numerical data into a numerical format that can be processed by the machine learning model. Based on correlation analysis, remove redundant variables with correlation coefficients greater than a preset threshold and construct a candidate feature set. S23. Based on the candidate feature set, a full feature dataset containing demographic features, preoperative laboratory indicators, and intraoperative anesthesia-related features is formed for subsequent postoperative delirium risk modeling. S3. Construct multiple candidate postoperative delirium risk prediction models based on the full feature dataset, and train and evaluate the performance of the candidate models, specifically including: S31. Divide the full feature dataset into a training set, a test set, and a time series validation set according to a preset ratio for model training, internal evaluation, and time generalization validation. S32. Modeling and training are performed using machine learning algorithms based on the training set. The machine learning algorithms include, but are not limited to, one or more of random forest, support vector machine, extreme gradient boosting, AdaBoost and k-nearest neighbor algorithm, to obtain multiple candidate prediction models. S33. Perform parameter optimization and cross-validation on the candidate prediction models, and compare each candidate model based on the discrimination results and calibration results of the test set and the time series validation set to select the basic prediction model with better performance. The discrimination performance indicators include area under the curve, accuracy, sensitivity, specificity, positive predicted value, negative predicted value, F1 value, and area under the precision-recall curve. The calibration performance indicators include calibration curve and Brier score. S4. Perform stability screening on the basic prediction model to obtain a target prediction model with better generalization performance, specifically including: S41. Obtain the discriminant performance metrics of each candidate model on the test set and on the time series validation set; S42. Calculate the model stability evaluation result based on the discriminant performance index to comprehensively reflect the predictive ability of the candidate model in the internal test set and its consistency performance in time-varying data; the stability evaluation result is characterized by the stability coefficient, which is used to comprehensively reflect the model's discriminant ability in the test set and its performance consistency with the time series validation set. S43. Select the target prediction model based on the stability evaluation results; S5. Based on feature contribution analysis, perform feature optimization on the target prediction model to obtain a simplified prediction model, specifically including: S51. The interpretability analysis method is used to quantify the contribution of each candidate variable in the full feature dataset to the model output, and the importance ranking of each feature is obtained. S52. Perform recursive feature elimination in order of feature importance from low to high. After each round of feature removal, retrain the model and re-evaluate the model performance. S53. Based on the model performance corresponding to different feature combinations during the optimization process, determine the optimal feature subset, and construct a simplified prediction model with better generalization performance based on the optimal feature subset. S6. Perform subgroup analysis to validate the simplified prediction model, specifically including: S61. Patients were divided into low ASA and high ASA groups according to the American Society of Anesthesiologists (ASA) classification. S62. Evaluate the discriminative performance of the simplified prediction model in each ASA subgroup on the test set and the time series validation set, respectively. S63. Based on the model performance results in each subgroup, determine the applicability and stability of the simplified prediction model in patients with different perioperative risk stratifications; S7. Input the perioperative clinical characteristics of the patient to be evaluated into the simplified prediction model, and output the probability of postoperative delirium and risk stratification results for early identification and clinical decision support of patients with high risk of postoperative delirium during the perioperative period.

[0009] The technical solution provided by this invention has the following technical advantages: (1) The present invention constructs a postoperative delirium risk prediction model based on demographic characteristics, preoperative laboratory indicators and intraoperative anesthesia-related characteristics that are routinely available in the electronic medical record system. It does not require additional complex testing costs and has strong clinical feasibility and promotion value. (2) This invention can make comprehensive use of multidimensional perioperative clinical information, break through the limitations of traditional statistical models on the number of variables and linear relationships, and is more conducive to exploring complex risk patterns related to postoperative delirium, thereby improving the accuracy of prediction; (3) By introducing a model stability assessment and time series validation mechanism, this invention can effectively screen candidate models with better generalization ability, reduce the performance fluctuation of the model in data at different time periods, and improve the reliability of clinical application. (4) This invention combines feature contribution analysis and recursive feature optimization strategy to reduce the number of redundant variables while maintaining high predictive performance, thereby obtaining a more concise, interpretable and clinically deployable predictive model. (5) This invention can identify patients at high risk of postoperative delirium in the early perioperative period, providing an auxiliary basis for clinicians to formulate stratified management, precise prevention and intervention strategies. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating the overall process of the postoperative delirium risk prediction method involved in this invention.

[0011] Figure 2 This is a comparison chart of the ROC values ​​of different candidate models in the test set in this invention.

[0012] Figure 3 This is a comparison chart of ROC values ​​for different candidate models in the time series validation set in this invention.

[0013] Figure 4This is a schematic diagram illustrating the feature contribution analysis of the target prediction model in this invention.

[0014] Figure 5 This is a schematic diagram of the recursive feature optimization process and the selection of the optimal feature subset in this invention. Detailed Implementation

[0015] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. Example

[0016] As a first embodiment of the present invention, a method for predicting the risk of postoperative delirium in elderly patients undergoing abdominal surgery based on perioperative clinical characteristics is provided. The method includes: S1. Obtain perioperative clinical data of the patients to be evaluated. The patients to be evaluated are elderly patients aged ≥60 years undergoing abdominal surgery, including one or more of gastrointestinal surgery, hepatobiliary surgery, pancreatic surgery, small bowel surgery, and colorectal surgery. The development cohort includes 379 patients, with 130 positive cases, and the time-validation cohort includes 82 patients, with 27 positive cases. The perioperative clinical data are obtained from the hospital's electronic medical record system and anesthesia information system, and the data includes demographic information, preoperative laboratory indicators, and intraoperative anesthesia-related indicators. S2. Preprocess and feature-construct the perioperative clinical data to obtain a feature dataset for predicting postoperative delirium risk, specifically including: S21. Preprocess the raw clinical data, including performing Z-score processing on continuous variables and 0 / 1 encoding on categorical variables to form structured input data; S22. Based on correlation analysis, redundant variables are eliminated, Spearman correlation coefficients are calculated between variables, and the variable with the greatest clinical significance is retained among multiple variables with a correlation coefficient greater than 0.8 to construct a candidate feature set; S23. Based on the candidate feature set, a full feature dataset containing demographic features, preoperative laboratory indicators, and intraoperative anesthesia-related features is formed for subsequent postoperative delirium risk modeling. S3. Construct multiple candidate postoperative delirium risk prediction models based on the full feature dataset, and train and evaluate the performance of the candidate models, specifically including: S31. Divide the full feature dataset into a development queue and a time series validation set according to time sequence. The development queue is divided into a training set and a test set in an 8:2 ratio for model training and internal evaluation. The time series validation set is used for time generalization validation. S32. Modeling and training are performed using machine learning algorithms based on the training set. The machine learning algorithms include random forest, support vector machine, extreme gradient boosting, AdaBoost and k-nearest neighbor algorithm to obtain multiple candidate prediction models. S33. A grid search combined with five-fold cross-validation method is used to optimize the hyperparameters of each candidate prediction model on the training set; and the discriminative performance and calibration performance of each candidate prediction model are evaluated on the test set and time series validation set, respectively; wherein, the discriminative performance indicators include the area under the receiver operating characteristic curve, the area under the precision-recall curve, accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and F1 score; the 95% confidence interval of each indicator is estimated by 2000 bootstrap resampling; the classification threshold is determined by maximizing the Youden index; the calibration performance is characterized by the calibration curve and Brier score; in this embodiment, the ROC comparison of different candidate models on the test set and time series validation set is as follows: Figure 2 and 3 As shown; S4. Perform stability screening on the candidate prediction models to obtain a target prediction model with better generalization performance, specifically including: S41. Obtain the area under the receiver operating characteristic curve (AUC) of each candidate prediction model on the test set and the area under the AUC of each candidate prediction model on the time series validation set. S42. Based on the area under the receiver operating characteristic (ROC) curves in the test set and time series validation set, calculate the stability coefficient of the candidate prediction model. The formula for calculating the stability coefficient is as follows: Stability coefficient = AUROC test × (1 - |AUROC test - AUROC temporal |); Among them, AUROC test AUROC represents the area under the receiver operating characteristic curve of the candidate prediction model on the test set. temporal The stability coefficient represents the area under the receiver operating characteristic curve of the candidate prediction model on the time series validation set; the stability coefficient is used to comprehensively reflect the model's discriminative performance on the internal test set and its performance consistency in time-varying data. S43. Sort the candidate prediction models according to the stability coefficient and select the target prediction model with a higher stability coefficient and more consistent performance across datasets. In this embodiment, the random forest model built based on the full feature dataset is determined as the target prediction model for subsequent feature optimization. S5. Based on feature contribution analysis, perform feature optimization on the target prediction model to obtain a simplified prediction model, specifically including: S51. The SHApley Additive Explanations (SHAP) interpretability analysis method is used to quantify the contribution of each feature in the target prediction model to the model output. In this embodiment, KernelExplainer is used to calculate the SHAP value of each feature, and the global feature importance ranking is determined based on the average absolute SHAP value of each feature. In this embodiment, the average absolute SHAP value of each feature is as follows: Figure 4 As shown; S52. Based on the importance ranking of the features, perform SHAP-guided backward stepwise feature elimination; in each iteration, delete the feature with the smallest average absolute SHAP value, and retrain the target prediction model based on the remaining features, while re-evaluating the model performance on the test set. S53. Repeat the feature deletion, model reconstruction, and performance evaluation process until the number of remaining features is 1; take the feature combination corresponding to the highest area under the subject operating characteristic curve in the test set as the optimal feature subset, and construct a simplified prediction model based on the optimal feature subset; in this embodiment, the recursive feature optimization process and the optimal feature subset selection are illustrated as follows: Figure 5 As shown; S6. After constructing the simplified prediction model, a subgroup analysis is performed on the simplified prediction model to verify its predictive stability and applicability in patients with different perioperative baseline risk levels. Given the strong correlation between ASA classification and the risk of postoperative delirium, patients are stratified according to the ASA classification, with ASA I-II patients defined as the low ASA group and ASA III-V patients defined as the high ASA group. The predictive performance of the simplified prediction model in each subgroup is evaluated in the test set and time-series validation set, respectively. S7. Input the perioperative clinical characteristics of the patient to be evaluated into the simplified prediction model, output the risk probability of postoperative delirium, and give a high-risk or low-risk classification result according to the optimal classification threshold; wherein, when the risk probability is higher than the optimal classification threshold, the patient is determined to be a high-risk patient for postoperative delirium, so as to be used for early perioperative warning, stratified management and auxiliary clinical intervention decision-making.

Claims

1. A method for predicting the risk of postoperative delirium in elderly patients undergoing abdominal surgery based on perioperative clinical characteristics, characterized in that, Includes the following steps: S1. Obtain perioperative clinical data of the patient to be evaluated, including demographic characteristics, preoperative laboratory indicators, and intraoperative anesthesia-related characteristics. S2. Preprocess and feature construct the perioperative clinical data to obtain a feature dataset for predicting the risk of postoperative delirium; S3. Construct and train multiple candidate postoperative delirium risk prediction models based on the feature dataset, and evaluate the performance of the candidate postoperative delirium risk prediction models. S4. Based on the performance of each candidate postoperative delirium risk prediction model on the test set and time series validation set, the stability of the candidate postoperative delirium risk prediction models is screened to obtain the target prediction model. S5. Based on feature contribution analysis, the target prediction model is optimized to obtain a simplified prediction model; S6. Perform ASA subgroup analysis to validate the simplified prediction model; S7. Input the perioperative clinical characteristics of the patient to be evaluated into the simplified prediction model and output the postoperative delirium risk prediction results.

2. The method according to claim 1, characterized in that, The patients to be evaluated are elderly patients aged 60 years or older who have undergone abdominal surgery, including one or more of the following: gastrointestinal surgery, hepatobiliary surgery, pancreatic surgery, small bowel surgery, and colorectal surgery.

3. The method according to claim 1, characterized in that, Step S2 includes: S21. Perform data cleaning, missing value imputation, outlier identification, and format standardization on the original perioperative clinical data to form structured input data; S22. Standardize the continuous variables in the structured input data, encode the categorical variables, and remove redundant variables based on correlation analysis to construct a candidate feature set. S23. Based on the candidate feature set, form a feature dataset containing demographic features, preoperative laboratory indicators, and intraoperative anesthesia-related features.

4. The method according to claim 1, characterized in that, Step S3 includes: S31. Divide the feature dataset into a training set, a test set, and a time series validation set; S32. Modeling and training are performed using machine learning algorithms based on the training set. The machine learning algorithms include, but are not limited to, one or more of random forest, support vector machine, extreme gradient boosting, AdaBoost and k-nearest neighbor algorithm, to obtain multiple candidate postoperative delirium risk prediction models. S33. Perform parameter optimization, cross-validation, and performance evaluation on the candidate postoperative delirium risk prediction model.

5. The method according to claim 1, characterized in that, The performance evaluation includes discriminant performance evaluation and calibration performance evaluation; the discriminant performance evaluation metrics include area under the receiver operating characteristic curve, accuracy, sensitivity, specificity, positive predictive value, negative predictive value, F1 score, and area under the precision-recall curve; the calibration performance evaluation metrics include calibration curve and Brier score.

6. The method according to claim 1, characterized in that, S4 includes: S41. Obtain the first discriminant performance index of each candidate postoperative delirium risk prediction model on the test set and the second discriminant performance index on the time series validation set; S42. Calculate the stability coefficient based on the first and second discrimination performance indices; S43. Select the target prediction model based on the stability coefficient.

7. The method according to claim 1, characterized in that, S5 includes: S51. The contribution of each feature to the output of the target prediction model is quantified using the interpretability analysis method to obtain the feature importance ranking; S52. Perform recursive feature elimination in order of feature importance from low to high. After each round of feature removal, retrain the model and re-evaluate the model performance. S53. Determine the optimal feature subset based on the model performance corresponding to different feature combinations, and construct the simplified prediction model based on the optimal feature subset.

8. The method according to claim 1, characterized in that, S6 includes: S61. Patients were stratified according to the American Association of Anesthesiologists classification, with ASA I-II patients defined as the low ASA group and ASA III-V patients defined as the high ASA group. S62. Evaluate the prediction performance of the simplified prediction model in the low ASA group and the high ASA group in the test set and the time series validation set, respectively; S63. Based on the prediction performance results in each ASA subgroup, verify the applicability and stability of the simplified prediction model in patients with different perioperative risk stratifications.

9. The method according to claim 15, characterized in that, The predictive performance in S62 includes the area under the receiver operating characteristic curve, accuracy, sensitivity, specificity, positive predictive value, negative predictive value, F1 score, and area under the precision-recall curve.

10. The method according to claim 1, characterized in that, The postoperative delirium risk prediction results output in S7 include the probability of postoperative delirium and the result of high or low risk level.

Images