Study For Perioperative Dynamic Risk Assessment And Prediction Method Of Geriatric Patients Based On Machine-Learning

Objective:We aimed to develop a scalable perioperative clinical data normalization pipeline for standardizing and integrating multi-source information about preoperative risk assessment,intraoperative records,postoperative follow-up through multidisciplinary cooperation;We also aimed to prospectively collect perioperative information of elderly patients older than 65 who underwent surgeries under general anesthesia in our institution from July 1,2019 to October 31,2020,and the data were used to train and internally validate machine learning algorithms for predicting postoperative 30-day mortality and unplanned readmission.Materials and Methods:1.Construction of scalable interactive platform for perioperative structured dataWe improved the content of preoperative risk assessment,intraoperative records,postoperative follow-up by consulting relevant domestic and foreign literature,guidelines,diagnostic standards and commonly used models,combined with the experts’suggestions.All the perioperative information were designed as structured fields and assigned a value.After the completion of the platform construction,we prospectively collected perioperative information of elderly patients older than 65who underwent surgeries under general anesthesia in our institution from July 1,2019to October 31,2020,establishing a preoperative-intraoperative-postoperative database.2.Development and validation of a machine learning model for prediction of postoperative 30-day mortalityWe extracted perioperative information of elderly patients older than 65 who underwent surgeries under general anesthesia in our institution from July 1,2019 to October 31,2020.The label"postoperative 30-day mortality"is defined as:all-cause death happened during surgery or within 30 days postoperatively.After data preprocessing including imputation for missing values and treatment for outliers,a total of 77 features were calculated or extracted,mainly including the patients’demographic characteristics,preoperative comorbidities and laboratory examinations;descriptive intraoperative vital signs,summary of drugs and fluids interventions;as well as patient operation descriptions such as type and duration of operation.The original data sets were randomly divided into training set（70%）and validation set（30%）,which were utilized for development and validation of machine learning models,respectively.A total of 7 machine learning algorithms were trained for predicting postoperative 30-day mortality,including Bagging classifier,Random Forest（RF）,e Xtreme Gradient Boosting（XGBoost）,Light Gradient Boosting Machine（Light GBM）,Logistic Regression（LR）,Support Vector Machine（SVM）and Multilayer Perceptron（MLP）.Model performance was assessed using the following metrics:Area Under ROC Curve（AUC）,Accuracy,Precision,Recall and F1 Score.In addition,the ROC curve and the Precision-Recall（P-R）curve were also drawn.Calibration of predictions were performed using Brier Score and Calibration Plots.A feature ablation analysis was performed and the change in AUC with the removal of each feature was then assessed to determine feature importance.3.Development and validation of a machine learning model for prediction of postoperative 30-day unplanned readmissionWe extracted perioperative information of elderly patients older than 65 who underwent surgeries under general anesthesia in our institution from July 1,2019 to October 31,2020.The label"postoperative 30-day unplanned readmission"is defined as:unexpected readmission due to the same surgical disease or postoperative complications within 30 days postoperatively.After data preprocessing including imputation for missing values and treatment for outliers,a total of 112 features were calculated or extracted,mainly including the patients’demographic characteristics,preoperative comorbidities and laboratory examinations;descriptive intraoperative vital signs,summary of drugs and fluids interventions;as well as patient operation descriptions such as type and duration of operation and postoperative complication.The original data sets were randomly divided into training set（70%）and validation set（30%）,which were utilized for development and validation of machine learning models,respectively.A total of 7 machine learning algorithms were trained for predicting postoperative 30-day unplanned readmission,including Bagging classifier,RF,XGBoost,Light GBM,LR,SVM and MLP.Model performance was assessed using the following metrics:AUC,Accuracy,Precision,Recall and F1 Score.In addition,the ROC curve and the P-R curve were also drawn.Calibration of predictions were performed using Brier Score and Calibration Plots.A feature ablation analysis was performed and the change in AUC with the removal of each feature was then assessed to determine feature importance.Results:1.Construction of scalable interactive platform for perioperative structured dataThe perioperative structured data interactive platform mainly includes preoperative risk assessment system,intraoperative information database,postoperative in-hospital follow-up system and postoperative out-of-hospital follow-up system.The preoperative risk assessment system mainly includes the patients’demographic characteristics such as age,gender;type of surgery,comorbidities and their severity of various organ systems,airway assessment information,past history,family history,special medication history,frailty assessment and laboratory examinations;the intraoperative information database mainly includes descriptive intraoperative vital signs,drugs and fluids interventions,blood transfusion,intraoperative special events,presence of an arterial or venous line and airway devices,etc;the postoperative in-hospital follow-up system mainly includes follow-up time,postoperative vital signs,postoperative complications of various organ systems and their severity,postoperative pain,postoperative outcome,etc;the postoperative out-of-hospital follow-up system is an automated follow-up platform which followed up patients by SMS or telephone 30 days after operation.Following up included:patient outcome,unplanned readmission or not,infection or not,postoperative pain,sleep quality,etc.2.Development and validation of a machine learning model for prediction of postoperative 30-day mortalityA total of 7467 elderly patients were included.The overall 30-day mortality was0.89%,and 30-day mortality in the training and test sets were 0.81%and 1.07%,respectively.The AUCs of the seven machine learning algorithms predicting postoperative 30-day mortality in the test set ranged from 0.743 to 0.973.The Light GBM classifier overall performed the best with an AUC of 0.959（95%CI,0.950-0.968）,accuracy of 0.996（95%CI,0.995-0.997）,precision of 0.871（95%CI,0.769-0.909）,recall of 0.620（95%CI,0.615-0.625）,and F₁score of 0.716（95%CI,0.690-0.741）;The brier scores of the seven machine learning algorithms predicting postoperative 30-day mortality in the test set ranged from 0.0039 to 0.0125,the brier score of the Light GBM model was 0.0039（95%CI,0.0037-0.0041）.The most five important features of the Light GBM model were Huaxi Perioperative Prognostic Index,body mass index（BMI）,creatinine,arrhythmia,ASA classification.3.Development and validation of a machine learning model for prediction of postoperative 30-day unplanned readmissionA total of 7467 elderly patients were included.The overall 30-day unplanned readmission rate was 3.79%,and 30-day unplanned readmission rate in the training and test sets were 3.73%and 3.93%,respectively.The AUCs of the seven machine learning algorithms predicting postoperative 30-day unplanned readmission in the test set ranged from 0.579 to 0.715.The RF classifier overall performed the best with an AUC of 0.711（95%CI,0.686-0.735）,accuracy of 0.962（95%CI,0.961-0.963）,precision of 0.500（95%CI,0.400-0.500）,recall of 0.012（95%CI,0.011-0.013）,and F1 score of 0.024（95%CI,0.023-0.025）;The brier scores of the seven machine learning algorithms predicting postoperative 30-day unplanned readmission in the test set ranged from 0.0377 to 0.0464,the brier score of the RF model was 0.0383（95%CI,0.0377-0.0388）.The most five important features of the RF model were total bilirubin concentration,operation duration,blood glucose concentration,white blood cell count,and BMI.Conclusion:1.The scalable interactive platform for perioperative structured data constructed in this study can effectively collect multi-dimensional,fine-grained perioperative clinical data of the patients,and realize the integrated management of the patients’perioperative data,which laid the foundation for medical risk management and clinical investigation.2.Machine learning-based models can accurately predict postoperative 30-day mortality of geriatric patients.3.Machine learning-based models can effectively predict postoperative 30-day unplanned readmission of geriatric patients.