Artificial intelligence-based characterization of multi-organ ultrasound congestion across the heart failure Spectrum

TL;DR

This study uses artificial intelligence to analyze ultrasound signs of congestion in heart failure patients, revealing a multidimensional congestion phenotype across different stages of the disease.

Contribution

The novel use of AI to integrate multi-organ ultrasound findings with clinical and echocardiographic data to characterize congestion in heart failure.

Findings

AI models identified key predictors of multi-organ congestion, including pulmonary artery pressures and medication use.

Multi-organ congestion was observed in 274 patients with ≥2 ultrasound signs of congestion.

Congestion features clustered into four domains: medical history, biohumoral variables, left heart function, and right heart/pulmonary circulation.

Abstract

To investigate, using artificial intelligence (AI), the relationships between ultrasound (US)-defined systemic congestion and demographic, echocardiographic, and biohumoral parameters across the heart failure (HF) spectrum.

A total of 1588 subjects (651 Stage A–B, 376 HF with reduced left ventricular ejection fraction [HFrEF, <50%], and 561 HF with preserved ejection fraction [HFpEF, ≥50%]) underwent comprehensive clinical evaluation, laboratory testing, echocardiography, and US assessment of congestion, including inferior vena cava (IVC), lung ultrasound (LUS), renal venous flow (RVF), portal venous flow (PVF), and hepatic venous flow (HVF). Assessment of IVC, LUS, and RVF was available in the entire cohort, whereas HVF and PVF were performed in 359 and 289 patients, respectively. Overall, 856 patients had no US signs of congestion, 458 had one US sign, and 274 had ≥2 US signs (multi-organ congestion). AI-based predictive models were developed for each site of congestion and for multi-organ congestion using a 3-item model (IVC, LUS, RVF). Congestion-related features clustered into four domains: medical history, biohumoral variables, left heart morphology and function, and right heart and pulmonary circulation. The 3-item model identified mitral annular systolic velocity, systolic and diastolic pulmonary artery pressure, triglycerides, left atrial volume index, diabetes, and treatment with furosemide or angiotensin-converting enzyme inhibitors/angiotensin II receptor blockers as key predictors of multi-organ congestion (area under the curve = 0.79).

AI-assisted integration of multi-organ US characterizes congestion as a multidimensional phenotype beyond conventional clinical assessment and biomarkers across the HF spectrum.

Graphical AbstractA model for each site of congestion (one-site models for IVC, LUS, RVF, HVF, PVF) and a 3-item model (IVC, LUS and RVF) were built integrating routine clinical, laboratory, and echocardiographic to predict one-site and multi-organ (≥2 US signs) congestion respectively. The minimal set of input features was selected for the 3-item model. Shapley additive explanations violin plot illustrates the direction and magnitude of each variable’s impact on the prediction of multisite congestion. ACEi/ARB: angiotensin-converting enzyme inhibitors/angiotensin II receptor blockers; AI: artificial intelligence; AUC = area under the receiver operating characteristic curve; dPAP = diastolic pulmonary artery pressure; HF = heart failure; HFpEF = heart failure with preserved ejection fraction; HFrEF = heart failure with reduced ejection fraction; HPV = hepatic venous flow; IVC = inferior vena cava; LAVi = left atrium volume index; LUS = lung ultrasound; LV S′ = left ventricle systolic mitral annulus tissue velocity; PVF = portal venous flow; RVF = renal venous flow; sPAP = systolic pulmonary artery pressure, US = ultrasound; XBG = extreme gradient boosting.For image description, please refer to the figure legend and surrounding text.

A model for each site of congestion (one-site models for IVC, LUS, RVF, HVF, PVF) and a 3-item model (IVC, LUS and RVF) were built integrating routine clinical, laboratory, and echocardiographic to predict one-site and multi-organ (≥2 US signs) congestion respectively. The minimal set of input features was selected for the 3-item model. Shapley additive explanations violin plot illustrates the direction and magnitude of each variable’s impact on the prediction of multisite congestion. ACEi/ARB: angiotensin-converting enzyme inhibitors/angiotensin II receptor blockers; AI: artificial intelligence; AUC = area under the receiver operating characteristic curve; dPAP = diastolic pulmonary artery pressure; HF = heart failure; HFpEF = heart failure with preserved ejection fraction; HFrEF = heart failure with reduced ejection fraction; HPV = hepatic venous flow; IVC = inferior vena cava; LAVi = left atrium volume index; LUS = lung ultrasound; LV S′ = left ventricle systolic mitral annulus tissue velocity; PVF = portal venous flow; RVF = renal venous flow; sPAP = systolic pulmonary artery pressure, US = ultrasound; XBG = extreme gradient boosting.

Figures