Evaluating Hyponatremia Through Inferior Vena Cava Ultrasound: A Non-Invasive Approach to Volume Status Assessment

Hyponatremia, defined as serum sodium concentration of less than 135 mEq/L, is a common electrolyte disorder observed in hospitalized patients. Its diagnosis is clinically relevant due to the associated risk of increased morbidity and mortality. A critical step in managing hyponatremia is the correct assessment of the patient’s volume status (hypovolemic, euvolemic, or hypervolemic), which guides therapeutic decisions. However, current diagnostic approaches, which are based on clinical evaluation, laboratory data, and static ultrasound assessments, often lack sensitivity and specificity, leading to frequent misclassifications. This thesis explores the integration of a novel software solution, developed by VIPER s.r.l, into current diagnostic methods for assessing volume status. The software enables semi-automated tracking of the inferior vena cava (IVC) walls in both transverse and longitudinal ultrasound views, to extract key parameters such as diameter, caval index (CI), respiratory caval index (RCI) and cardiac caval index (CCI). The primary objective was to develop and validate an accurate and reproducible automated classification model for volume status. The secondary aim was to identify a minimal subset of features capable of achieving reliable classification performance. The study enrolled 33 hyponatremic patients (21 females, 12 males; mean age 81.45 ± 12.78 years) who were classified as hypovolemic (6), euvolemic (14), or hypervolemic (13). Three machine learning classifiers were evaluated using 3-fold cross-validation. For all models, the inclusion of the full set of VIPER-derived features, in both short (SA) and long-axis (LA) views, significantly enhanced performance compared to using clinical variables alone. To identify the most predictive subset of features, a grid search was performed, restricted to all VIPER-derived features and two key clinical variables: edema and central venous pressure (CVP). During this feature selection phase, the Random Forest model consistently outperformed the other classifiers. As a result, the best-performing feature subset identified by Random Forest, comprising SA CCI, LA CCI, SA maximum diameter, and edema, was selected for further evaluation. The final model, trained on this reduced feature set, was then evaluated with leave-one-out cross-validation (leaving out one patient per fold) to ensure an unbiased performance estimation. Concurrently, an intra-operator repeatability analysis of VIPER measurements was performed by calculating Intraclass Correlation Coefficients (ICC). The optimal Random Forest model, based on the reduced feature set, achieved a balanced accuracy of 82%. The repeatability analysis confirmed excellent consistency for measurements acquired in the short-axis view, with ICCs of 0.97 for diameter, 0.92 for RCI, and 0.90 for CCI. Measurements from the long-axis view demonstrated lower reproducibility (ICC 0.62-0.89). These findings suggest that integrating innovative VIPER features into machine learning models offers a non-invasive and reliable tool to support the differential diagnosis of volume status in hyponatremia. Future studies should focus on developing a fully automated classification system that relies exclusively on VIPER features.