Testing an inverse modeling approach with gradient boosting regression for stroke volume estimation using patient thermodilution data

Vasiliki (Vicky) Bikia ^1,2,3* Dionysios Adamopoulos ⁴ Marco Roffi ⁴ Georgios Rovas ² Stéphane Noble ⁴ François Mach ⁴ Nikolaos Stergiopulos ^2*

• ¹ Institute for Human-centered AI, Computer Science, Stanford University, Stanford, United States
• ² Department of Bioengineering, Swiss Federal Institute of Technology Lausanne, Lausanne, Vaud, Switzerland
• ³ Department of Biomedical Data Science, Stanford University, Stanford, California, United States
• ⁴ Hôpitaux universitaires de Genève (HUG), Genève, Geneva, Switzerland

Stroke volume (SV) is a major indicator of cardiovascular function, providing essential information about heart performance and blood flow adequacy. Accurate SV measurement is particularly important for assessing patients with heart failure, managing those undergoing major surgeries, and providing optimal care in critical settings. Traditional methods for estimating SV, such as thermodilution, are invasive and unsuitable for routine diagnostics.Non-invasive techniques, although safer and more accessible, often lack the precision and userfriendliness needed for continuous bedside monitoring. We developed a modified method for SV estimation that combines a validated 1-D model of the systemic circulation with machine learning. Our approach replaces the traditional optimization process developed in our previous work, with a regression method, utilizing an in silico-generated dataset of various hemodynamic profiles to create a gradient boosting regression-enabled SV estimator. This dataset accurately mimics the dynamic characteristics of the 1-D model, allowing for precise SV predictions without resource-intensive parameter adjustments. We evaluated our method against SV values derived from the gold standard thermodilution method in 24 patients. The results demonstrated that our approach provides a satisfactory agreement between the predicted and reference data, with a MAE of 16 mL, a normalized RMSE of 21%, a bias of -9.2 mL, and limits of agreement (LoA) of [-47,28] mL. A correlation value of r=0.7 (p<0.05) was reported, with the predicted SV slightly underestimated (68±23 mL) in comparison to the reference SV (77±26 mL). The significant reduction in computational time of our method for SV assessment should make it suitable for real-time clinical applications.