Evaluation of the Factors Influencing Blood Transfusion during Minimally Invasive Direct Coronary Artery Bypass Surgery

The shortage of blood resources has become a global dilemma. The global shortfall in blood resources is as much as 30 million units a year [1]. In China, we also have the same dilemma [2]. A study in the USA showed that the transfusion rate of allogeneic blood during adult heart surgery was as high as 50%, accounting for 10–15% of allogeneic blood use [3]. Compared with traditional surgical methods, minimally invasive surgery has the advantages of smaller trauma and faster recovery. The minimally invasive direct coronary artery bypass (MIDCAB) through a small incision in the left anterior chest can be completed under direct vision, with a clearer view, smaller trauma, and lower postoperative drainage volume than traditional coronary artery bypass grafting [4]. Although minimally invasive heart surgery has greatly reduced the allogenic blood transfusion rate [5], evaluating the possibility of different surgical blood usage and identifying factors that affect blood transfusion can help with scientific perioperative management and maximize patient safety.

Analyzing factors related to blood use in heart surgeries is mostly based on statistical models [6] but cannot engage in intelligent learning. AI can classify, predict data, and effectively analyze by learning from the data [7]. Using AI, more objective data results can be obtained [8]. AI can help identify useful clues for clinical diagnosis [9]. Using AI, we can provide data-supported blood transfusion plans to the clinic, which may help us achieve better patient blood management. A standardized diagnostic and treatment process exists for patients undergoing small incision coronary artery bypass grafting surgery [10]. Using AI to explore surgical-related risks helps doctors evaluate the blood needs during surgery and better manage patients’ blood. This study compared the accuracy of different AI models in predicting blood usage in MIDCAB and used the best model to explore the risk factors related to blood transfusion. The study investigated the advantages of AI in the perioperative management of MIDCAB patients.

This is a single-center, open, retrospective analysis aimed at patients who underwent MIDCAB operations, and no heart-lung machine was used in the Department of Cardiac Surgery at Peking University, Third Hospital, from January 2017 to September 2022. The analysis included the patient’s medical histories, general information, preoperative routine examination results, heart function classification, and surgical procedures. A model for predicting the use of allogeneic blood during surgery was established, using 80% of the data as the training set and 20% as the testing set. Additionally, the risk factors associated with allogeneic blood transfusion were analyzed.

A total of 766 patients who underwent left anterior thoracotomy MIDCAB from January 2017 to September 2022 were included as the study subjects. This study was approved by the Medical Ethics Committee of Peking University Third Hospital (Approval No. 2022 Medical Ethics Review No. 523-02) and was granted an exemption from informed consent. Exclusion criteria: patients without preoperative blood routine results; patients with a history of cardiac surgery (excluding interventional surgery); patients who converted to open chest surgery; patients who underwent secondary surgery; patients without blood routine results after surgery.

Medical history (pleural thickening, pleural fibrosis, tuberculosis, hyperlipidemia); diagnosis (extracted from the first admission note); Demographic characteristics: age, gender, BMI, blood volume (BV); cardiac function classification; surgical vessel (left anterior descending artery, circumflex artery, right coronary artery posterior descending branch, right coronary artery left ventricular posterior branch); myocardial injury detection: creatine kinase (CK), CK isoenzymes (CK-MB), hydroxybutyrate dehydrogenase, highly sensitive cardiac troponin T, N-terminal pro-brain natriuretic peptide (NT-proBNP); plasma lipids analysis: total cholesterol, triglycerides, high-density lipoprotein cholesterol, low-density lipoprotein cholesterol, apolipoprotein A1, apolipoprotein B, lipoprotein(a); coagulation function test: prothrombin time (PT), activated partial thromboplastin time, fibrinogen (Fib), international normalized ratio; blood routine before the surgery: hemoglobin, hematocrit (Hct), red blood cell count (RBC); platelets (PLT), white blood cell count; infection detection: C-reactive protein; liver function test: alanine aminotransferase, aspartate aminotransferase; kidney function test: creatinine (Cr), glomerular filtration rate (GFR), blood routine after the surgery: Hb, Hct; surgical characteristics (surgical vessels, surgical approach) and preoperative medication history.

Preoperative test results are within 72 h before surgery. The allogeneic red blood cell transfusion volume is recorded based on the transfusion management system. The chief surgeon and the anesthesiologist decide the intraoperative allogeneic blood transfusion based on the patient’s bleeding volume and rate, as well as the patient’s underlying conditions. In our cardiovascular surgery department, the routine indications for allogeneic blood transfusion are intraoperative blood loss greater than 800 mL and/or Hb level lower than 80 g/L. Clinical physicians will choose between restrictive and nonrestrictive transfusion strategies based on their specific surgical circumstances. For patients with the following conditions: low venous saturation, low cardiac output syndrome, lactate accumulation or metabolic acidosis, high catecholamine dose, hypovolemia, we adopt a more liberal transfusion threshold. Collect the postoperative examination within 72 h according to the following principles: for patients who received blood transfusions within 20 h of the start of surgery, collect the results that are closest to the most recent blood transfusion; for patients who did not receive blood transfusions within 20 h of the start of surgery, collect the results that are closest to the time of the end of the surgery.

To predict whether a patient will receive an allogeneic blood transfusion during surgery, we made models using Extreme Gradient Boosting, Gradient Boosting Classifier, Extra Trees Classifier, Logistic Regression, CatBoost Classifier, AdaBoost Classifier, Linear Discriminant Analysis, Random Forest Classifier, Decision Tree Classifier, and K Neighbors Classifier. We compared the result and chose the best model using 80% of the data as the training set to train the model and 20% as the test set to evaluate the model using the ROC curves. Additionally, we used the SHapley Additive exPlanations (SHAP) tool to analyze and demonstrate the factors that have the highest importance in decision-making. We also visualized the specific impact of important features using a local dependence plot.

Statistical analysis was conducted using SPSS26.0 software. Continuous variables with normal distribution were expressed as means, while those with non-normal distribution were expressed as medians (interquartile range). Categorical variables were expressed as frequency or (and) percentage. T-tests for two independent samples were used for normally distributed data, and nonparametric t-tests for two independent samples were used for non-normally distributed data. The χ² or Fisher’s exact test was used to count data, and p < 0.05 was considered statistically significant.

The 766 surgical patients were included, and 107 (13.97%) cases received allogeneic blood transfusion during surgery. The distribution of postoperative Hb levels in transfused patients is shown (Fig. 1). The average postoperative Hb level in transfused patients was 98.77 g/L, and there were 22 cases with a postoperative Hb level ≥120 g/L, accounting for 20.56% of transfused patients. Among the total number of patients, 576 had applicated for red blood cells of 4 units or more, accounting for 75.20%. The number of transfused patients with a red blood cell application of 4 units or more was 77, and the transfusion rate was 13.37% for patients requiring more than 4 units of red blood cells.

First, perform a univariate analysis of preoperative data based on whether allogeneic blood was transfused or not (Table 1). The results showed that advanced age, male gender, low BMI, low blood volume, tachycardia, elevated CK isoenzyme, increased high-sensitivity cardiac troponin T, high N-terminal pro-brain natriuretic peptide, low Hb, low Hct, low red blood cell count, increased Cr, low GFR, and after cardiovascular stent implantation were all risk factors for allogeneic blood transfusion, and the differences were statistically significant (all p < 0.05). The result of CK in the transfusion group is lower than the non-transfusion group (p < 0.05) because CK is not sensitive in determining heart damage.

A decision model for allogeneic blood transfusion was established, and the effects of different models were compared (Table 2). XGBoost was chosen as the prediction model based on a comprehensive evaluation. After training, SHAP was used to display the feature priorities (Fig. 2). It can be seen that preoperative anemia, elevated PT, low BMI and weight, multiple types of diseases, history of percutaneous coronary intervention (PCI), increased heart rate, low Fib and PLT, and elevated Cr are the top ten risk factors affecting blood transfusion during surgery. The relationship dependence diagram was drawn for the top six influential factors of SHAP values (Fig. 3), which showed that preoperative Hb less than 120 g/L, PT greater than 13.75, BMI less than 22.7 kg/m², diagnose length longer than 4 letters, the history of PCI, weight lower than 67 kg were the main risk factors for blood transfusion. The learning curve shows the validation accuracy of the XGBoost (Fig. 4). It shows that accuracy is relatively high when the training set has more than 400 samples.

In recent years, with increasing awareness and attention to the risks of blood transfusions, more and more doctors have begun to focus on patient blood management and adopt restrictive transfusion strategies [11]. Studies have shown that in hospitals where patient blood management is not implemented, 25% of trauma patients receive blood products even when their Hb levels are greater than 104 g/L [12]. While some doctors acknowledge the restrictive transfusion threshold of 70 g/L for surgical patients, in actual clinical practice, 84–92% of patients are not strictly adhering to the restrictive transfusion strategy [13]. Therefore, it is necessary to strengthen education on transfusion practice, and also, we can practice patient blood management by reducing bleeding, blood conservation, and rational use of allogeneic transfusions. Early screening for the cause of blooding has proven beneficial. This screening should be tailored to different diseases.

This study investigates patients who underwent MIDCAB surgery to find the risk factors of bleeding with the goal of preventing RBC transfusion. We construct 10 machine learning models to preliminarily identify that preoperative anemia, coagulation indicators (PT, Fib, PLT), low BMI and weight, multiple types of diseases, history of PCI, increased heart rate, elevated Cr have relatively high weights in predicting allogeneic blood transfusions during surgery.

It has been proved that the incorporation of common factors into clinical pathways can better reduce perioperative risks and economic costs for patients [14, 15]. Currently, there are few predictions and assessments related to blood transfusion, and clinicians often rely on clinical experience to judge the infusion and evaluation of blood products [16]. Guidelines on blood transfusion only provide rough descriptions based on Hb, coagulation function, and platelet values, and there are no standards for preoperative evaluation or management of specific diseases. Currently, most predictions related to blood transfusion in cardiac surgery are based on linear regression results, which cannot provide a quantitative relationship between risk factors and prediction results. We used SHAP analysis to identify risk factors and determine the critical values of risk through local relationship dependence graphs and found that preoperative Hb less than 120 g/L, PT greater than 13.75 s, BMI less than 22.7 kg/m², weight lower than 67 kg are consistent with the results of previous studies [17‒21]. Diagnose length longer than 4 letters and history of PCI have a high risk of transfusion [22]. Coronary heart disease was the primarily diagnosis in our study. In Chinese, coronary heart disease is 3 characters. Our results show that the risk of bleeding increases for patients whose diagnosis length is greater than 4, which means that in addition to coronary heart disease, these patients also have other comorbidities. The risk of blood transfusion in patients undergoing venous bypass grafting is greater than in other surgical methods, which is consistent with the risk factors identified by our univariate analysis [23]. N-terminal pro-brain natriuretic peptide and CK isoenzyme, which assess myocardial damage, are also indicators of blood transfusion. Some studies have indicated that the use of clopidogrel and platelet function are also factors for predicting transfusion risk [24], while others have shown that continuing the original anticoagulation regimen does not increase the risk of intraoperative bleeding in cardiac patients undergoing minimally invasive surgery [25]. In our hospital, we kept aspirin on before the surgery, which did not increase the risk of allogeneic blood transfusion in the risk factor analysis model.

Multiple-center randomized controlled studies have shown that simply strengthening the auditing and evaluation of blood transfusion does not improve the quality control of blood transfusion. It is suggested that a learning software system should be used to improve the rationality of blood transfusion [26]. Previous studies have shown that intelligent software-based blood transfusion interventions can help reduce the use of red blood cells [27, 28]. Therefore, the effectiveness of the blood transfusion model can be verified through the clinical application, and optimization can continue during the application process. Our hospital’s evaluation of the use of blood products based on doctors’ experience shows that the Hb levels of patients who receive red blood cell transfusions after surgery are generally between 80 and 100 g/L. Clinical physicians will choose between restrictive and nonrestrictive transfusion strategies based on their specific surgical circumstances. According to the recommendations of the AABB [29], the indications for a liberal transfusion threshold are Hb levels between 90 g/L and 100 g/L. Therefore, using this more liberal transfusion threshold to transfuse patients, the postoperative Hb levels would not exceed 120 g/L, but there were also 22 (20.56%) cases of postoperative Hb levels ≥120 g/L among patients who received red blood cell transfusions. The process of red blood cell application or evaluation indicators for these patients still needs to be improved. The intelligent blood transfusion model that has been constructed will be embedded in the clinical medical record system and will be automatically triggered when doctors submit a blood transfusion application. The model will use preoperative variables to analyze and predict the risk level of surgical blood usage and related risk factors [30]. The predicting model can help doctors to reduce the risk of intraoperative blood use and play a role in patient blood management.

Studies have shown that reasonable preoperative blood preparation is an important measure for blood inventory management and can also reduce medical expenses [31]. For surgeries with a low risk of transfusion reactions, two units of red blood cells are routinely prepared before surgery to deal with the risk of unexpected intraoperative bleeding. To eliminate the interference of routine blood preparation operations, RBC applications lower than two units were removed, and a total of 576 cases applicated RBC of 4 units or more with an experience evaluation. There were 77 cases (13.7%) transfused among RBC applications of 4 units or more. Blood preparation volumes that exceed the maximum blood preparation list not only increase medical expenses due to excessive testing but also pose difficulties for blood inventory management. The blood transfusion department can use the intelligent blood transfusion model to manage blood inventory in a reasonable manner.

Accumulating evidence supports our results that our blood transfusion model can help reduce the transfusion rate and economic costs. Although many studies have already confirmed the role of intelligent models in different diseases, AI for blood transfusion prediction studies are still in the early stages and need more exploration. Future research could discover models that have high accuracy.

This study has some limitations that must be highlighted. First, this study is a single-center retrospective study, and the transfusion probability provided by the machine model only represents the current level of medical care in our hospital. Second, the accuracy of the software prediction needs to be further corrected through prospective experiments.

This model only predicts the transfusion rate of a single type of disease, and in the future, more types of diseases can be added for analysis to improve the evaluation of preoperative blood demand and better allocate blood products. Alternatively, dynamic blood inventory management can be carried out using artificial intelligence to achieve reasonable blood allocation in a blood shortage situation. Based on the factors affecting transfusion, we can make different processing flows and therapeutic schemes to provide better medical decisions, aiming to achieve intelligent blood management for patients before, during, and after surgery.

The result reveals that the XGBoost model can accurately predict transfusion or not transfusion in MIDCAB. The risk factors discovered by the model can help doctors perform effective patient blood management.

This study was approved by the Medical Ethics Committee of Peking University Third Hospital (Approval No. 2017 Medical Ethics Review No. 523-02) and was granted an exemption from informed consent.

The authors declare that there is no conflict of interest regarding the publication of this paper.

This research was funded by National Natural Science Foundation of China (Grant No. 81971012).

Zhenmin Sun: conceptualization, methodology, formal analysis, investigation, resources, writing – original draft, funding acquisition, and writing – review and editing; Zhongqi Cui: methodology, software, validation, formal analysis, investigation, writing – original draft, writing – review and editing, and project administration; Yan Xie: methodology; investigation; data curation; visualization; validation; and writing – original draft; Lei Wang: investigation, data curation, visualization, validation, and writing – original draft; Zhengqian Li, Xiaoyu Yang, and Xiaoqing Zhang: investigation, data curation, and writing – original draft; and Jun Wang: supervision, methodology, formal analysis, writing – review and editing.

The data that support the findings of this study are not publicly available due to their containing information that could compromise the privacy of research participants but are available from the corresponding author (J.W.) upon reasonable request.