In this study, we development six ML models for predicting the risk of GV occurrence within the first 24 hours postoperatively following CPB. Among these models, the XGBoost model demonstrated the best predictive performance. Consequently, the XGBoost model was chosen as the optimal model for the early identification of GV risk in patients post-CPB surgery. The XGBoost model showed acceptable accuracy in predicting glycemic variability risk. This allows for proactive interventions to improve patient outcomes.
Over the past decade, the application of machine learning techniques to redict anomalous blood glucose levels has grown extensively. Previously, various methodologies were employed, including neural network 44,45, linear model with multiple input variables 24, and mathematical model utilizing both the first and second derivatives of the continuous glucose monitoring (CGM) data 46 or incorporating constant endogenous glucose production along with other physiological parameters for real-time application to develop prediction models 47. The choice of how we define the research question and select the study population significantly impacts a model's generalizability and usefulness in clinical settings. At present, there is a lack of predictive models for GV specifically in non-diabetic patients undergoing CPB.
In this study, the SHAP values were used to interpret the outputs of the XGBoost model and identified several important variables associated GV in non-diabetic patients undergoing cardiac surgery with CPB. This approach helps in understanding how each feature in the dataset influences the model’s predictions, providing a clearer insight into the decision-making process of the machine learning model. Postoperative insulin use has been recognized as the most important variable in the model. The results indicated that postoperative insulin use increased the risk of GV. The univariate analysis (Tabe S1) also supports the finding that patients who used insulin postoperative had higher SDBG. For postoperative patients with blood sugar exceeding 10.0 mmol/L, intravenous insulin infusion is the preferred treatment. Unlike injections, intravenous delivery sends insulin directly into the bloodstream, allowing it to reach tissues quickly and lower blood sugar levels. However, overuse of insulin can lead to hypoglycemia 48. Strong glucose control methods can lead to hypoglycemia and wider glucose fluctuations. The study by Sanjay OP et al. 49 found that attempting to maintain normoglycemia with insulin during CPB might lead to postoperative hypoglycemia. Currently, expert opinions from various countries differ slightly regarding the requirements for blood glucose control targets during the perioperative period. Based on the available literature 50,51, it is generally accepted that perioperative target blood glucose should be controlled to a range of 7.8 to 10 mmol/L. The international guidelines also supplement that the minimum target range for blood glucose can be set from 4.4 to 8 mmol/L 51,52. However, the optimal approach to glucose regulation remains unclear. Future studies with more patients and comprehensive glucose monitoring data are necessary to solidify the link between glycemic variability and insulin administration.
BMI was also identified as an important variable. This study was consistent with the findings by Wang et al. 53, which demonstrated that lower BMI correlated with increased GV. This may be ascribed to the possibility that individuals with lower BMI have poorer beta-cell function compared to those who are overweight or obese. Moreover, high BMI was associated with insulin resistance. In contrast, lower BMI was primarily associated with insulin deficiency, which makes controlling blood glucose with medication or insulin more difficult. Therefore, we propose monitoring of blood glucose in patients with lower BMI to prevent GV.
Meanwhile, we observed that higher intraoperative mean glucose levels were associated with an increased risk of GV, confirming the results obtained by Cornelia Knaak et al 54. In the study, they found a significant rise in both mean (p < 0.001) and maximum BG levels (p = 0.001) postoperatively in patients experiencing intraoperative dysglycemia. It is widely recognized that surgical procedures induce a stress response, which in turn prompts the secretion of both catecholamines and cortisol. These hormones can cause temporary insulin resistance, leading to a condition called stress hyperglycemia 54. Despite this, the incidence of intraoperative hyperglycemia is often underestimated in non-diabetic patients. Therefore, ideal clinical blood glucose management should focus not only on postoperative blood glucose levels and complications, but also on the trends of intraoperative blood glucose change to reduce the postoperative GV.
Our study also identified the duration of CPB as a significant predictor of GV. Patients undergoing longer CPB durations experienced a greater increase in GV. This association likely relates to the stress response that invariably occurs during cardiac surgery with CPB. The longer the duration of CPB, the stronger the stress response in the body, which increases systemic inflammatory response and insulin resistance. This decreases sensitivity of body tissues to insulin, triggering increased insulin usage. In addition, patients undergoing CPB are routinely exposed to hypothermia. Cueni-Villoz N et al. 55 reported that hypothermia increased the levels of blood glucose concentrations, elevated GV, and enhanced insulin requirement, which was consistent with our findings. With the improvements in surgical techniques, the operative time, CPB time and cross-clamp time are expected to be significantly reduced. This may effectively improve the body’s internal environment and decrease perioperative GV.
This study has several advantages. Firstly, our study focused on non-diabetic patients. According to Krinsley JS 56 and Hao-ming Hestudy et al. 57, GV was significantly associated with mortality and had the poorest prognosisin in the non-diabetic population. Another study 58 also supports this finding, indicating that an increase in GV is associated with a higher risk of mortality in non-diabetic patients, but this is not necessarily the case for diabetic patients. Thus, it appears that non-diabetes patients exhibit less tolerance to high glucose variation compared to those with diabetes. Currently, there are no reliable predictive models for GV in non-diabetes patients. Thus, this study can improve the management of postoperation glycemia in non-diabetes patients. Secondly, we established the XGBoost technique, with can rapid computation, good generalization and excellent predictive capabilities 59–61. XGBoost, unlike traditional machine learning models, employs an ensemble of decision trees. The outputs of all decision trees are combined to create the final output of the XGBoost model 62. XGBoost has enabled the development of new generation medical applications, from accurate diagnosis to personalized patient management. These applications hold immense potential to improve real-world healthcare outcomes 63–65. Here, we found that XGBoost could predict GV in non-diabetic patients undergoing cardiac surgery with CPB. The third advantage of this study is that we employed SHAP values to visually explain the selected variables. Despite the high accuracy of ML algorithms, it is limited by difficult interpretability, known as "black box"66. The SHAP algorithm showed good ability to address this problem. It ranked each feature’s importance within the model and illustrated how different variables contribute to predictive outcomes.
This study has some limitations. Firstly, the data were collected from a single institution with a relatively small sample size. To reduce overfitting and maximize the models’ advantages, larger sample sizes should be enrolled in future studies. Secondly, to improve the generalization of our models, further studies are needed to test the generalization of the models in independent external validation dataset from multiple institutions. Thirdly, no gold standard inclusion or exclusion criteria have been proposed for GV. Fourthly, we monitored postoperative glucose levels only during the initial 24 hours, neglecting potential impacts from subsequent fluctuations. Moreover, we excluded 58.38% of the initial participants, which might affect the generalizability of our findings.