Machine learning model predicts cardiac surgery complications from routine blood tests

Researchers in China have built and tested an interpretable machine learning model that predicts major adverse cardiac and cerebrovascular events (MACCE) after cardiac surgery using routine laboratory data. The multicenter retrospective cohort study, published in BMC Cardiovascular Disorders, drew on patients who underwent cardiac surgery with cardiopulmonary bypass (CPB) at four tertiary cardiac centers between 2019 and 2023. Rather than relying on specialized or expensive biomarkers, the model was trained on information already collected in everyday clinical practice: baseline patient characteristics, perioperative blood counts and routine coagulation parameters.

The study included a total of 27,847 patients, of whom 26,687 were assigned to the training set and 1,160 to the validation set. Multiple machine learning models were developed and evaluated on this data. The best performer was an Explainable Boosting Machine, or EBM, which achieved an area under the receiver operating characteristic curve of 0.789, with a 95% confidence interval of 0.724 to 0.847. That figure places the model’s discriminative ability in a moderate range, meaning it can help flag patients at higher risk but cannot identify every patient who will experience an event.

The clinical rationale behind the model rests on the inflammation–coagulation axis. Cardiac surgery under CPB can trigger a systemic inflammatory response and coagulation disorder, processes that the study’s authors note may increase the risk of MACCE after the operation. Because both inflammation and clotting are reflected in standard blood tests, the researchers reasoned that these routine parameters could carry predictive signal. The model’s top important predictors bore this out: postoperative lymphocyte percentage, preoperative thrombin time, postoperative international normalized ratio, lymphocyte count, red blood cell distribution width–standard deviation, fibrinogen, prothrombin time and plateletcrit all ranked among the most influential variables.

What distinguishes the EBM from many black-box algorithms is its interpretability. The researchers selected it specifically for global and local interpretability analyses, alongside sensitivity analyses. Global interpretability shows which variables the model weighs most heavily across the whole population, while local interpretability explains how individual factors contributed to a prediction for a single patient. This transparency matters in a clinical setting, where clinicians need to understand why a risk score was generated before acting on it. The prominence of lymphocyte measures and coagulation markers in the model also gives researchers a concrete, testable picture of how postoperative inflammation and clotting disturbances relate to adverse outcomes.

The model’s intended role is early postoperative risk stratification. According to the study’s conclusions, the EBM effectively predicts MACCE after cardiac surgery and identifies key biomarkers related to the inflammation–coagulation axis, positioning it as an early postoperative risk-stratification tool. In practice, such a tool could help care teams decide which patients warrant closer monitoring in the hours and days following an operation, using data that has already been drawn for routine testing rather than additional procedures.

The study’s design shapes how its findings should be read. As a retrospective cohort study, it analyzed records of patients who had already undergone surgery, rather than prospectively enrolling patients and tracking outcomes forward. The population was limited to patients at four tertiary cardiac centers in China undergoing CPB cardiac surgery during a five-year window, so performance in other health systems, surgical settings or patient populations remains to be established. The authors themselves state that further validation in larger multicenter prospective studies is warranted, and the reported AUC, while the best among the models tested, indicates room for improvement in discrimination.

Even with those caveats, the work illustrates a practical direction for predictive modeling in cardiac surgery. Routine blood counts and coagulation panels are inexpensive, widely available and already part of standard perioperative care, which makes a model built on them easier to deploy than one requiring novel assays. The combination of moderate predictive performance and built-in interpretability also addresses two common obstacles to clinical adoption of machine learning: cost of inputs and opacity of outputs. Whether the model improves outcomes when embedded in real-world workflows is a question for the prospective studies the authors call for, but the present results establish that routine laboratory data, processed through an explainable algorithm, can carry meaningful signal about postoperative risk.

Subject of Research: Development of an interpretable machine learning model using routine blood tests to predict major adverse cardiac and cerebrovascular events after cardiac surgery.

Article Title: Development and validation of an interpretable machine learning model with multicenter routine tests to predict major adverse cardiac and cerebrovascular events after cardiac surgery

Article References: Tan, J., Yang, H., Wang, Y., Wei, X., Chen, C., Du, K., Meng, G., Liu, Y., Zhou, C., & Du, L. (2026). Development and validation of an interpretable machine learning model with multicenter routine tests to predict major adverse cardiac and cerebrovascular events after cardiac surgery. BMC Cardiovascular Disorders. https://doi.org/10.1186/s12872-026-06741-5

Image Credits: AI Generated

DOI: 10.1186/s12872-026-06741-5

Keywords: machine learning, cardiac surgery, MACCE, cardiopulmonary bypass, inflammation, coagulation, risk stratification, interpretable AI, blood tests, retrospective cohort study, Development, validation