Key predictors of mortality in profound hyponatremia beyond the correction rate

Abstract Background Current guidelines for hyponatremia recommend slow sodium correction; however, recent large-scale observational studies have associated slow correction with increased mortality. The association between correction rate and mortality can be influenced by numerous factors, including comorbidities, overall condition, and treatment interventions, but this complex interplay remains unclear. We aimed to clarify this association by developing interpretable machine learning models using detailed clinical features. Methods We included 546 patients with serum sodium ≤120 mEq/L, collected clinical features through chart review, and developed four machine learning models to predict in-hospital mortality. The best-performing model, selected by the area under the receiver operating characteristic curve (ROC-AUC), was interpreted using SHapley Additive exPlanations (SHAP) to quantify each feature’s contribution to mortality prediction. Results In-hospital mortality was 18%. The random forest model demonstrated the best predictive performance (ROC-AUC = 0.907; 95% CI, 0.832–0.965). SHAP analysis revealed that the most influential predictors were baseline characteristics reflecting underlying illness severity: serum albumin, C-reactive protein, Charlson Comorbidity Index, and metastatic malignancy; their predictive effects were consistent across correction rates. Among treatment-related features, intravenous fluid choice and sodium monitoring frequency had a greater predictive impact than correction rate. Although slower correction was associated with higher mortality, the correction rate ranked 14th among all 59 features in predictive importance. Conclusions In profound hyponatremia, multiple key predictors of mortality exist beyond the correction rate. These findings suggest that the observed association between slow correction and higher mortality may not be causal but rather an epiphenomenon driven by underlying illness severity and treatment intensity.