Estimation of hemoglobin concentration at the initiation of cardiopulmonary bypass using support vector regression.

BACKGROUND

Hemodilution during cardiopulmonary bypass (CPB) is a standard perfusion strategy used to reduce blood viscosity and enhance microcirculatory flow. The hemodilution rate, expressed as hemoglobin (Hb) concentration, is a key control index in CPB and is currently estimated from total blood volume (TBV). The objective of this study was to propose a novel formula to accurately predict Hb concentration at the initiation of CPB (HbCPB) by incorporating circulating blood volume, laboratory data, physical measurements, and patient history.

METHODS

We retrospectively analyzed 577 adult patients who underwent elective CPB at Fujita Health University Hospital from January 2016 to December 2020. Thirty-six preoperative variables - including demographics, laboratory data, circuit parameters, and indices such as TBV and ideal weight - were standardized. Categorical variables underwent one-hot encoding. We compared generalized linear models (GLM), support vector regression (SVR), and multilayer perceptron (MLP). Model performance was evaluated using the coefficient of determination (R2), mean squared error (MSE), and Bland-Altman analysis (bias and 95% limits of agreement [LoA]). Predictions from two conventional TBV-based formulas were used as benchmarks.

RESULTS

Of 993 screened cases, 577 met inclusion criteria (447 males, mean age 66.8 ± 11.7 years; 130 females, 69.5 ± 10.6 years). SVR on standardized predictors achieved the highest accuracy (R2 = 0.498, MSE = 0.517), outperforming GLM (R2 = 0.429, MSE = 0.797) and MLP (R2 = 0.332, MSE = 0.669). Conventional formulas showed lower performance (R2 = 0.325, MSE ≥ 1.48). Bland-Altman analysis for SVR demonstrated minimal bias (-0.0028 g/dL) and narrower LoA (-1.42 to 1.41 g/dL) than conventional methods (bias -1.33 g/dL; LoA -3.49 to 0.83 g/dL).

CONCLUSION

These findings suggest that an SVR-based model improves prediction of HbCPB over conventional approaches, supporting optimized transfusion management and reduced hemodilution-related risks.