A Hybrid Learning Framework for Predicting Post-Treatment Serum Sodium in Patients with Hyponatremia

Ghanahshyam B. Kshirsagar, Robert Hayden, Prasheen Shah, Christopher El Mouhayyar, Sahir Kalim, Sagar Nigwekar, Raimond Winslow, Qingchu Jin

Confirmation: I have read and agree with the IEEE BHI 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: Hyponatremia, Predictive Modeling, Serum Sodium Correction, Physiological Equations, Machine Learning

TL;DR: We developed a hybrid mechanistic-ML model to predict serum sodium correction in hyponatremia. Validated on two clinical cohorts, it improves generalizability for data with heterogenous risk and robustness for handling missing or delayed inputs

Abstract:

Hyponatremia (serum sodium [S-Na⁺] <135 mEq/L) is common in hospitalized patients and linked to adverse outcomes. Management of severe hyponatremia ([S-Na⁺] <120 mEq/L) remains controversial and challenging, requiring a controlled rate of correction—typically a rise in [S-Na⁺] of 6 mEq/L over 24 hours—to reduce the risk of complications in high-risk patients. Rapid correction (>8 mEq/L/day) has been associated with a higher risk of developing osmotic demyelination syndrome, while overly slow correction has been associated with longer hospital stays. Achieving the target correction often requires 3% hypertonic saline (3% NaCl) and sometimes desmopressin acetate (DDAVP). The required amount of 3% NaCl varies with patient-specific factors and clinician judgment, often resulting in suboptimal correction due to the complexity of predicting individual responses.
Predictive modeling of post-treatment [S-Na⁺] may offer a promising solution. While mechanistic models may simulate fluid-electrolyte dynamics, and machine learning (ML) models may offer additional data-driven insights, existing models are limited by non-individualized predictions and small, homogeneous datasets, thus limiting their generalizability.
To address these gaps, we developed a hybrid modeling framework that integrates mechanistic physiology with ML-based prediction. We developed our approach using two internal cohorts drawn from Massachusetts General Hospital and Brigham and Women’s Hospital. Cohort 1 (n=144) included heterogeneous-risk patients, and Cohort 2 (n=73) was a high-risk subset of DDAVP-treated patients from Cohort 1 (primary data), which was then externally validated on MIMIC-IV. Our hybrid model outperformed existing models, particularly in the high-risk Cohort 2. We submit that this hybrid modeling framework can be applied to other clinical challenges with small and complex clinical datasets.

Track: 4. Clinical Informatics

Registration Id: R8NKRN7VHTG

Submission Number: 349