Ultrasound Assessment of Inferior Vena Cava Dynamics as a Predictor of Right Atrial Pressure

Right Atrial Pressure (RAP) represents a fundamental hemodynamic parameter in the evaluation of cardiac pathologies. However, the gold standard for assessing this parameter remains an invasive measurement. The invasiveness of such techniques limits their routine use, highlighting the need for reliable non-invasive alternatives. The diameter and Caval Index (CI) of the inferior vena cava (IVC), derived from ultrasound imaging, allow an indirect estimation of RAP through semi-quantitative criteria defined by current international guidelines, which classify RAP into three levels (low, intermediate, and high). While these recommendations offer a practical and straightforward approach, their application is often limited by suboptimal precision and reproducibility. Considering these limitations, a dedicated software tool named “VIPER” has been developed. This semi-automatic algorithm is capable of accurately tracking the borders of the IVC, from which it extracts key parameters: vessel diameter, CI, Respiratory (RCI) and Cardiac (CCI) Caval Index. In this thesis, two different algorithms of VIPER software, one developed on Matlab and one on Python, were employed to analyse and segment the IVC from ultrasound video recordings provided by expert clinicians at the Fondazione Toscana Gabriele Monasterio, Pisa. No statistically significant differences were observed between the two software tools for either diameter or CI measurements, confirming the consistency and reproducibility of the automated software. The IVC diameters measured by the operator had a mean value of 17.24 ± 4.03 mm, while the automated estimates obtained from the Python and Matlab algorithms reported mean values of 15.89 ± 5.20 mm and 16.13 ± 5.44 mm, respectively. Similarly, the CI computed from the Python software yielded a mean of 0.31 ± 0.16, compared to 0.28 ± 0.16 from the Matlab-based version. The study population consists of 37 patients (16 males and 21 females) with a mean age of 67.49 ± 16.81 years. For each subject, a comprehensive set of clinical, echocardiographic and invasive variables were available, along with additional features extracted by the two software tools. Notably, invasive measurements were excluded from the dataset to preserve the non-invasive nature of the prediction. A range of linear and non-linear classification models was implemented with the aim of assigning each patient to a RAP class. Among these, the best-performing approach was a meta-classifier trained using features extracted by the Matlab version of VIPER. This model integrates the predictions of two base classifiers: a linear Ridge model and a non-linear Support Vector Machine (SVM). The meta-classifier achieved an accuracy of 73%, as estimated through Leave-One-Out Cross-Validation (LOO-CV). The same meta-classifier architecture, when applied to features derived from the Python version of VIPER, achieved a LOO-CV accuracy of 65%. In comparison, when using the operator-measured diameter, the standard guideline-based assessment methods achieved a lower accuracy of 54%, suggesting that algorithmic approaches may offer a more reliable support for non-invasive RAP assessment and may represent a promising advancement in clinical practice.