Predicting dyscalcemia using machine learning models based on routine laboratory data

Background: Dyscalcemia is associated with significant clinical complications, and the demand for calcium testing is rising disproportionately. Objectives: This study aimed to develop machine learning (ML) models to predict total calcium abnormalities using routinely available demographic and laboratory data to optimize laboratory test utilization. Materials and methods: This retrospective study analyzed data from 7,951 patients at Siriraj Hospital between April 2023 and March 2024. Feature selection identified ten predictors, including patient status, absolute lymphocyte count, albumin, bicarbonate, hematocrit, hemoglobin, platelet count, potassium, red blood cell count, and total protein. Seven ML algorithms were developed and validated using an 80:20 training-testing split with 10-fold cross-validation and hyperparameter tuning. The Synthetic Minority Oversampling Technique was applied to address class imbalance. Additionally, model calibration was evaluated using Brier scores, and Platt scaling was applied to ensure the reliability of predicted probabilities. Results: XGBoost achieved the highest AUC (0.84), indicating slightly better discriminative performance, and the highest specificity (0.86), reflecting a stronger ability to correctly identify normal cases. Random forest yielded the highest F1 score (0.71) and recall (0.76), indicating a better balance between precision and recall, with higher sensitivity for detecting abnormal cases. Following Platt scaling, the calibrated models achieved robust Brier scores (e.g., 0.1496 for random forest and 0.1497 for XGBoost), demonstrating highly accurate probability risk estimation. SHAP analysis identified albumin as the most influential feature, followed by total protein and hemoglobin. Conclusion: ML models utilizing demographic and laboratory data can accurately predict plasma total calcium status, and probability calibration ensures reliable risk estimates. XGBoost and random forest demonstrated robust performance with complementary strengths. These models show promise as screening tools for identifying high-risk patients and optimizing rational laboratory utilization.