Development and Temporal Validation of Machine Learning Models for Hyponatremia Risk in Community-Dwelling Older...

Background/Objectives: Hyponatremia is a clinically important electrolyte disorder in older adults, yet early identification is hindered by complex, non-linear interactions between comorbidities and polypharmacy. This study aimed to develop and externally validate a machine learning (ML) prediction model for hyponatremia risk using nationwide claims data, focusing on medication patterns and clinical features. Methods: A retrospective cohort study was conducted using the South Korean Health Insurance Review and Assessment Service-Aged Patient Sample (HIRA-APS). Data from 2017 to 2018 were used for development, and 2019 data for temporal external validation (age ≥ 65). SHapley Additive exPlanations (SHAP)-based recursive feature elimination identified 33 high-impact predictors from 60 clinical features. Six ML algorithms, including LightGBM and CatBoost, were trained with 1:4 case-control matching and evaluated for discrimination and calibration. Results: The development and validation cohorts included 4810 and 648,586 patients, respectively. All models showed comparable discriminative performance (area under the receiver operating characteristic curve [AUROC] 0.741-0.746), with LightGBM achieving the highest (AUROC 0.746; 95% confidence interval (CI) 0.728-0.763). The models had very high negative predictive values (>0.999) for ruling out low-risk individuals. Tree-based ensemble matched linear models in discrimination but achieved better calibration. Conclusions: These validated, interpretable ML models can serve as clinical decision support tools that rule out low-risk patients and prioritize monitoring for high-risk individuals. Across sociodemographic subgroups, calibration was maintained after recalibration, whereas discrimination was lower in the oldest, most comorbid, frailest, highest-medication-burden, and lowest-socioeconomic groups-a gap to address before equitable deployment.