One-day-ahead prediction of severe hyponatremia in patients with cancer using medication and diagnostic trajectories without electrolyte indicators.

INTRO: Hyponatremia is a common complication in patients with cancer and is associated with poor prognosis; however, early detection remains challenging. Existing models often rely on laboratory parameters directly related to sodium levels, limiting early risk identification before electrolyte testing. This study developed a machine learning model to predict severe hyponatremia 1 day before its occurrence using longitudinal prescription and diagnostic data without relying on serum electrolyte variables.

METHODS

This retrospective single-center study included 33,160 adult patients (2010-2021). De-identified data were transformed into the Observational Medical Outcomes Partnership (OMOP) Common Data Model. Severe hyponatremia was defined as serum sodium ≤ 125 mmol/L. Using 30-day time-series data, the eXtreme Gradient Boosting (XGBoost) and Transformer models were compared across three feature sets: a baseline set (Case 1), an information-rich set that included expanded laboratory variables (Case 2), and a medication-focused set (Case 3).

RESULTS

Among 33,160 patients, 2,916 (8.8 %) developed severe hyponatremia. XGBoost consistently outperformed the Transformer model. In Case 3, XGBoost achieved an area under the curve (AUC) of 0.853 and an F1-score of 0.411, comparable to those in Case 2 (AUC 0.862 and F1-score 0.426), which included laboratory data. Shapley Additive exPlanations (SHAP) and attention analyses identified cisplatin, lung cancer, male sex, dexamethasone, and trimethoprim/sulfamethoxazole as key predictors.

CONCLUSION

Severe hyponatremia can be predicted 1 day in advance with clinically meaningful predictive performance using 30-day medication and diagnostic trajectories without relying on electrolyte-related laboratory data. This finding suggests that high-risk patients could be identified early even when electrolyte-related laboratory variables are unavailable, such as during cancer treatment planning or when routine electrolyte monitoring is not performed, supporting more targeted electrolyte testing and earlier clinician-directed intervention pending external validation.