Machine learning-based prediction of intraoperative blood transfusion in major surgery: exploiting clinical variables and systemic inflammatory indices.

BACKGROUND

Intraoperative blood transfusion is common in major surgery. Predicting transfusion risk may improve perioperative management, optimize blood use, and enhance surgical planning. Machine Learning (ML) offers promising tools for individualized risk prediction.

AIM: To develop and validate an ML model predicting intraoperative transfusion risk in major surgery using clinical variables and inflammatory indices.

MATERIAL AND

METHODS

1858 adult patients who underwent major surgical procedures at AUSL-IRCCS di Reggio Emilia between September 2021 and December 2023 were retrospectively analyzed. The dataset was splitted into training (60%), internal validation (18%) and internal test (22%) sets. The primary outcome was the intraoperative red blood cell transfusion. Thirty-three candidate variables (29 clinical and laboratory parameters and 4 calculated indices) were evaluated. Highly correlated variables were excluded and key predictors were identified using CatBoost. Predictors distributions were compared by transfusion status and oncological diagnosis. Model performance was assessed by AUC, sensitivity, specificity, NPV, PPV, and F1-score. SHAP graph were used to interpret feature contributions.

RESULTS

The CatBoost model demonstrated strong predictive performance (AUC 0.88 training, 0.80 test). and a high NPV (0.89 training, 0.87 test), reliably identifyinglow-risk patients. Key predictors included type of surgery, preoperative haemoglobin, age, BMI, MCV RDW, PT and inflammatory indices such as MLR, PLR, NLR.

CONCLUSIONS

This ML-driven model accurately predicts intraoperative transfusion risk and identifies clinical and laboratory predictors, supporting perioperative management and more efficient resource allocation.