Tugay Solar

Background/Objectives: This study aimed to develop and evaluate machine learning models for predicting ionized calcium levels in patients with hypoalbuminemia, compare their performance with classical correction formulas and unadjusted total calcium levels, and assess their transportability through external validation. Methods: In a retrospective development cohort of 14,745 adults with hypoalbuminemia, XGBoost and ordinary least squares (OLS) models used total calcium, albumin, total protein, and eGFR and were evaluated with nested cross-validation. Frozen models were then applied without recalibration to an independent external cohort of 2751 patients using the same analyzer platforms. Results: Development XGBoost performance was R² = 0.500 and AUC-ROC = 0.798 (95% CI 0.788–0.807), with 98.5% sensitivity and 26.0% specificity. External R² was 0.397 for XGBoost and 0.363 for OLS. External AUC-ROC was 0.818 (95% CI 0.791–0.845) for XGBoost, 0.823 (0.797–0.850) for OLS, and 0.819 (0.791–0.845) for total calcium. Threshold-matched and exploratory decision-curve analyses showed no clear incremental XGBoost advantage over total calcium. Conclusions: Machine learning models provided useful continuous ionized-calcium estimates and high sensitivity for hypocalcemia detection, although specificity was low. In independent external validation, hypocalcemia discrimination was comparable to unadjusted total calcium. The models should be considered complementary rather than substitutes for measured ionized calcium, and further multicenter and prospective evaluation is warranted.