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

Area of Science:

• Biochemistry and Clinical Chemistry
• Artificial Intelligence in Medicine
• Health Informatics

Background:

• Accurate ionized calcium (Ca++) measurement is crucial but challenging in clinical settings.
• Existing methods for estimating Ca++ using routine biochemical parameters have limitations.
• Machine learning (ML) offers a potential solution for more accurate Ca++ prediction.

Purpose of the Study:

• To evaluate the performance of ML models in predicting Ca++ levels.
• To compare ML model predictions with direct Ca++ measurements and established correction formulas.
• To assess the utility of routinely available biochemical parameters (total calcium, total protein, albumin) for Ca++ prediction.

Main Methods:

• Retrospective analysis of 84,410 adult patients.
• Training and validation of Support Vector Machine (SVM), Random Forest (RF), and Gradient Boosting (GB) ML algorithms.
• Benchmarking ML models against Hanna, Zeisler, and Butler correction formulas.

Main Results:

• ML models significantly outperformed all conventional formulas in predicting Ca++ levels.
• Gradient Boosting (GB) achieved the highest explained variance (R² = 0.6742), followed by SVM and RF.
• Conventional formulas like Zeisler and Butler showed lower predictive accuracy (R² = 0.4879 and 0.2684, respectively).

Conclusions:

• ML models demonstrate superior accuracy for predicting ionized calcium (Ca++) compared to traditional formulas.
• These ML-based tools show potential for integration into clinical decision support systems.
• Future research should focus on model interpretability, incorporating pH, and external validation.