Tensor-based Multimodal Learning for Prediction of Pulmonary Arterial Wedge Pressure from Cardiac MRI

TL;DR

This work tackles non-invasive PAWP prediction by leveraging cardiac MRI data within a tensor-learning framework. It introduces MPCA-based tensor feature extraction, automated landmark detection with epistemic uncertainty, uncertainty-based sample binning, and multimodal data fusion of short-axis, four-chamber MRI, and cardiac measurements. On a cohort of $1346$ patients, the proposed tri-modal hybrid fusion achieves peak discrimination ($\mathrm{AUC}=0.8327$, $\mathrm{Accuracy}=0.8038$, $\mathrm{MCC}=0.5099$) with substantial gains over baselines ($Δ$AUC $=0.1027$, $Δ$Accuracy $=0.0628$, $Δ$MCC $=0.3917$), and decision-curve analysis confirms clinical utility for screening high-risk individuals. The method demonstrates the potential of tensor-based, uncertainty-aware, multimodal approaches to identify elevated PAWP non-invasively, though validation across multiple institutions remains a future priority.

Abstract

Heart failure is a serious and life-threatening condition that can lead to elevated pressure in the left ventricle. Pulmonary Arterial Wedge Pressure (PAWP) is an important surrogate marker indicating high pressure in the left ventricle. PAWP is determined by Right Heart Catheterization (RHC) but it is an invasive procedure. A non-invasive method is useful in quickly identifying high-risk patients from a large population. In this work, we develop a tensor learning-based pipeline for identifying PAWP from multimodal cardiac Magnetic Resonance Imaging (MRI). This pipeline extracts spatial and temporal features from high-dimensional scans. For quality control, we incorporate an epistemic uncertainty-based binning strategy to identify poor-quality training samples. To improve the performance, we learn complementary information by integrating features from multimodal data: cardiac MRI with short-axis and four-chamber views, and Electronic Health Records. The experimental analysis on a large cohort of $1346$ subjects who underwent the RHC procedure for PAWP estimation indicates that the proposed pipeline has a diagnostic value and can produce promising performance with significant improvement over the baseline in clinical practice (i.e., $Δ$AUC $=0.10$, $Δ$Accuracy $=0.06$, and $Δ$MCC $=0.39$). The decision curve analysis further confirms the clinical utility of our method.

Figures (5)

• Figure 1: The schematic overview of the PAWP prediction pipeline including preprocessing, tensor feature learning, and performance analysis. The blocks in gray color are explained in more detail in Section \ref{['s2']}.
• Figure 2: Performance comparison of removing a different number of bins of training data on $10$-fold cross-validation.
• Figure 3: The effect of combining CM features on short-axis and four-chamber. SA: Short-axis; FC: Four-chamber.
• Figure 4: The effect of combining CM features on the bi-modals including early and late fusion of four-chamber and short-axis. Early fusion: early fusion of short-axis and four-chamber; late fusion: late fusion of short-axis and four-chamber.
• Figure 5: Evaluating clinical utility of our method using Decision Curve Analysis (DCA) vickers2006decision."Treat All" means treating all patients, regardless of their actual disease status, while "Treat None" means treating no patients at all. Our predictive model's net benefit is compared with the net benefit of treating everyone or no one to determine its overall utility.