Comparing conventional correction formulas and machine learning-based prediction of ionized calcium

Accurate assessment of ionized calcium (Ca++) is critical in clinical settings but remains technically and logistically challenging in many healthcare facilities. This study aimed to evaluate the performance of machine learning (ML) models in predicting Ca++ levels measured by blood gas analysis, using routinely available biochemical parameters-total calcium (TotCa), total protein, and albumin-and to compare them with values obtained through direct measurement and established correction formulas.

A retrospective analysis was conducted on 84,410 patients aged 20-70 years (43,863 men, 40,547 women). Whole-blood Ca++, serum TotCa, albumin, and total protein levels were retrieved from hospital records. Three ML algorithms-Support Vector Machine (SVM), Random Forest (RF), and Gradient Boosting (GB)-were trained and validated using 5-fold cross-validation. Their performance was benchmarked against three conventional correction formulas: Hanna, Zeisler, and Butler.

Among the conventional formulas, the Hanna method showed the highest mean absolute error (MAE = 0.3626), while Zeisler (MAE = 0.0719) and Butler (MAE = 0.0988) performed more closely to measured Ca++. The ML models outperformed all formula-based methods, with GB (R² = 0.6742), SVM (R² = 0.6732), and RF (R² = 0.6730) achieving the highest explained variance. In contrast, Butler and Zeisler yielded R² values of 0.2684 and 0.4879, respectively.

ML models demonstrate superior predictive accuracy for Ca++ compared with conventional correction formulas when using routine biochemical parameters. These findings support the potential integration of ML-based tools into clinical decision support systems. Future research should address model interpretability, pH incorporation, and prospective external validation.