Unveiling the Gaps: Machine Learning Models for Unmeasured Ions.

Area of Science:

• Critical Care Medicine
• Physiology
• Biochemistry

Background:

• Unmeasured ions (UIs) significantly impact acid-base balance in critically ill patients.
• Current estimation methods (AGc, SIG, BEGap) have uncertain optimal utility.
• Accurate UI assessment is crucial for managing acid-base disturbances.

Purpose of the Study:

• To compare the explanatory performance of traditional, Stewart, and partitioned base excess (BE) approaches for estimating UIs.
• To evaluate the utility of BEGap, SIG, and AGc as determinants of arterial pH.
• To assess the generalizability of analytical models using machine learning.

Main Methods:

• Retrospective cohort study with development (2274 patients) and validation (1202 patients) cohorts.
• Evaluation of traditional, Stewart, and partitioned BE approaches using multivariable linear regression and machine learning (RF, XGBoost, SVR).
• Assessment of model performance via adjusted R², RMSE, MAE, and variable importance metrics (SHAP, permutation).

Main Results:

• The partitioned BE approach demonstrated the highest explanatory performance in MLR (adjusted R² = 0.949) and ML analyses (R² up to 0.989).
• BEGap consistently emerged as a strong, independent determinant of arterial pH, outperforming SIG and AGc across all models.
• XGBoost exhibited the most stable and accurate performance among machine learning models.

Conclusions:

• BEGap is a practical, physiologically informative, and bedside-applicable parameter for assessing UIs.
• BEGap outperforms both AGc and SIG in linear and non-linear analytical models for critically ill patients.
• The partitioned BE approach, particularly BEGap, offers superior insights into acid-base disturbances driven by UIs.