Research subject
This study adhered to tripod + AI (transparent reporting of multivariable predictive models for individual prognosis or diagnosis + artificial intelligence) and ensured comprehensive and transparent reporting. We conducted a large, multicenter retrospective cohort study at eight higher education hospitals in China. Patients who underwent mitral valve surgery between February 2016 and December 2018 were screened for qualification. Inclusion criteria were as follows: (1) Ages 18-75; (2) undergo mitral valve plasmography, mitral valve plasmography, or isolated mitral machine replacement. The exclusion criteria include: (2) Simultaneous aortic valve replacement or coronary bypass graft. (3) Secondary or emergency surgery. (4) Unable to provide the necessary clinical data.
A total of 1477 mitral valve procedures met eligibility and were performed by the core surgical team at each center during the study period. The distribution of cases was as follows: 2nd Xiangya Hospital (384 cases), 3rd Xiangya Hospital (112 cases), Beijing Aerospace General Hospital (92 cases), Qilu Hospital at Shandong University (192 cases), Fuwai Hospital (238 cases), Zejiangial People's Hospital of Southern Idiction Intigent of Hospital of Hospital of Hospital of Hospital of Southwisement Indiction (114 cases), Cardiovascular Hospital (176 cases). The cumulative surgical volumes for each core team over these three years ranged from 92 to 384 cases (median, 177 cases). All procedures were performed by the most volume surgical team at each center. Each team carried out more than 100 heart projects each year. All surgical surgeons had at least 10 years of dedicated experience in valve surgery. To minimize operator variability and ensure data quality, all centers followed standardized surgical protocols and quality control workflows.
The study was observational and therefore exempt from the written informed consent requirement. This study protocol received ethical approval from the Clinical Laboratory Ethics Committee of the 3rd Xiangya Hospital at Central South University (approval number 2019-S008). A flow chart depicting patient selection and grouping is shown in Figure 1.
Modeling procedures and research flow charts. (a) As shown in this figure, all variables including demographic information, clinical characteristics, clinical characteristics, clinical laboratory items, and diagnosis were extracted from the electronic medical records system of eight centres. In total, 30 preoperative variables were analyzed, of which 10 were screened. (b) Research flow chart.
Surgical procedures
All mitral valve surgeries were performed via median sternum resection. Cardiopulmonary bypass was performed under mild hypothermia and intraoperative esophageal echocardiography confirmed satisfactory valve function before weaning from the bypass. A standardized perioperative protocol was applied, including postoperative anticoagulation and restrictive transfusion strategies.
Cohort Construction and Data Collection
Based on the specified criteria, 1477 patients were enrolled in the study cohort. The transfusion group included patients who received RBC transfusions from the time of mitral valve surgery until discharge or end of hospital. The non-transfusion group consisted of patients who had not received an RBC transfusion from the start of mitral valve surgery until discharge or end of hospital in-hospital. Further validation of the proposed model was achieved by prospectively collecting data from 2nd Xiangya Hospital and 3rd Xiangya Hospital at Central South University for 35 patients undergoing mitral valve surgery between March 2022 and December 2023.
At each medical center, all data for the variables were extracted from the Electronic Health Records (EHR) system and the paper-based medical record system. The authors (such as YJW, JYZ, XJM) were granted access to the medical records systems of their respective affiliates to retrieve the data. Available preoperative variables known or suspected to be related to the need for RBC transfusion during mitral valve surgery were identified through consultation with medical professionals and literature review. For some variables requiring multiple measurements, data were assessed using the values closest to the start of the surgical procedure. Preoperative data collection includes demographic variables such as gender, age, height, weight, body mass index (BMI), and blood type. Treatment properties including left ventricular dilatation, atrial fibrillation, iron supplements and anemia drugs. Past medical history including cerebrovascular disease, hypertension, anemia, and diabetes. White blood cell (WBC) count, red blood cell (RBC) count, hemoglobin (HB) levels, hematocrit (HCT) levels, platelet (PLT) count, creatinine levels, total protein (TP), albumin, globulin, alanine aminotranserase (ALT), assartate aminotransease (AST), albumin, globulin, alanine aminotransparase (prothrombin time (PT), international normalized ratio (INR), left ventricular ejection fraction (LVEF), and type of surgery.
Subsequent analysis mentioned these variables using abbreviations and using brevity and clarity. For example, the number of red blood cells (RBCs) was called RBCs, and hemoglobin (HB) levels were called hemoglobin.
Of the 30 variables, one hospital had no height data. Therefore, prior to modeling, we implemented the K-nearest Neighbors (KNN) strategy to address missing data. Data collected by multiple medical centers were standardized. Ordinary variables were assigned to three surgical types: mechanical replacement of isolated mitral valves, mitral valve replacement, and mitral valve plasification.
Building a predictive model
Maximum Association and Minimum Redundancy (MRMR) Algorithm15 It was used to identify important variables from 30 preoperative variables. The MRMR algorithm served as a complete filtering feature selection model that allowed us to find the most relevant features to the target outcomes, while creating the most redundant features of the selected features.16. The MRMR algorithm first selected the most appropriate feature by identifying important mutual information in the target value. Second, additional features were gradually added depending on their correlation with existing features, ensuring minimal redundancy within the feature set. This approach was able to determine the most representative features and showed strong correlation with target values while being highly independent of each other. Furthermore, MRMR technology is not affected by assessing functionality during the feature selection process without building a model, reducing the risk of method modeling and overimplantation17. As a result, the MRMR algorithm was suitable for subsequent modeling and evaluation models.
This study employs a typical machine learning model to construct a predictive model of category outcomes. These models included an optical gradient boost machine (LightGBM), an extreme gradient boost (xgboost), an adaptive boost (adaboost), a multilayer perceptron (MLP), a support vector machine (SVM), complement naive Bayes (CNB), and logistic regression (LR).
Each model had advantages and disadvantages in a diverse dataset environment. There were no absolute advantages or disadvantages. After developing and verifying the above models through a 5x cross-validation technique, the area under the receiver operating characteristic curve (AUROC) for each model was then calculated in both the training and validation datasets. Additionally, F1 scores, kappa values, accuracy, sensitivity, specificity, positive predictive values (PPV), negative predictive values (NPV), and cutoff values were evaluated to identify the optimal model to predict the need for intraoperative and postoperative RBC transfusions during intraoperative valve surgery. Additionally, we used Brier scores and net clinical benefits to assess and compare model performance. The Brier score is an indicator of the accuracy of the predicted probability. The lower the score, the better the model18. Net clinical benefits demonstrate a trade-off between the benefits and harms of choosing different thresholds, providing insight into the practical application of the model in a clinical setting.
The prediction of the optimal model was elucidated using the Shapley Additive Description (SHAP) method after comparing the model's performance. By involving cooperative game theory19,SHAP package converted the results of raw nonlinear ML to the sum of attribution effects for all variables. This process quantified the contribution of each feature to the final prediction of the model. We approximated the risks associated with obtaining relevant outcomes for each patient. Therefore, the SHAP conversion results provided a simple means of identifying the effects of preoperative variables on the final outcome.
Statistical analysis
Median interquartile ranges were applied to represent continuous variables, but percentages were used to describe categorical variables. The T-TEST or MANN-WHITNEY U test (subject to circumstances) was used to compare continuous variables between the transfusion and non-transfusion groups. Chi-square tests were employed to analyze differences in categorical variables. Statistical significance was determined as p <0.05 using 2 tails pvalue. Predictive models were constructed with feature sets of key variables selected by the MRMR algorithm. By comparing the predictive models, the optimal predictive model was identified. The average area of the receiver operating characteristic (ROC) curve (AUC) from 5x cross-validation served as an evaluation metric to compare the accuracy of the predictive model. Higher AUC values indicate the superior predictive power of the model. Python was used for all statistical analysis of baseline characteristics.
