A New Machine Learning Approach for Actual Calcium Measurement

Area of Science:

• Biochemistry
• Medical Informatics
• Clinical Chemistry

Background:

• Total calcium measurement is common despite ionized calcium being a more accurate indicator.
• Existing calcium correction formulas show variable performance across populations.
• Machine learning offers potential for more accurate calcium level prediction by analyzing complex interactions.

Purpose of the Study:

• To evaluate the performance of Random Forest and XGBoost machine learning models.
• To compare machine learning models against conventional corrected calcium formulas.
• To assess prediction accuracy in an eastern Indian population.

Main Methods:

• Collected data on total calcium, ionized calcium, total protein, and albumin from 894 samples.
• Utilized a 65-15-20 data split for training, validation, and testing of machine learning models.
• Compared XGBoost and Random Forest models against six established corrected calcium formulas.

Main Results:

• The XGBoost model demonstrated the least deviation and a strong correlation (r=0.707) with actual calcium values.
• XGBoost outperformed conventional corrected calcium formulas in the study population.
• The XGBoost model also showed superior performance in patients with hypoalbuminemia.

Conclusions:

• The XGBoost machine learning model accurately predicts total calcium levels.
• XGBoost offers a more accurate alternative to conventional formulas when ionized calcium is not available.
• This approach can improve calcium status assessment in clinical settings.