Q-PART: Quasi-Periodic Adaptive Regression with Test-time Training for Pediatric Left Ventricular Ejection Fraction Regression

TL;DR

Q-PART tackles the challenge of adaptive pediatric LVEF regression under distribution shifts by introducing a Quasi-Period Network that explicitly separates periodic cardiac motion from aperiodic variations, modeled with Neural CDEs. At test time, it leverages domain-specific augmentations and a variance minimization objective, with differential learning rates for periodic and aperiodic components, and provides a theoretical bound linking variance minimization to regression error. Empirical evaluation on EchoNet-Dynamic and EchoNet-Pediatric across age groups shows state-of-the-art LVEF prediction and strong clinical screening performance (mAUROC up to 0.9747) while exhibiting gender-fair behavior. The approach also demonstrates ablation-supported gains from each component and robustness to acquisition-related quality issues, suggesting practical applicability in pediatric echocardiography workflows.

Abstract

In this work, we address the challenge of adaptive pediatric Left Ventricular Ejection Fraction (LVEF) assessment. While Test-time Training (TTT) approaches show promise for this task, they suffer from two significant limitations. Existing TTT works are primarily designed for classification tasks rather than continuous value regression, and they lack mechanisms to handle the quasi-periodic nature of cardiac signals. To tackle these issues, we propose a novel \textbf{Q}uasi-\textbf{P}eriodic \textbf{A}daptive \textbf{R}egression with \textbf{T}est-time Training (Q-PART) framework. In the training stage, the proposed Quasi-Period Network decomposes the echocardiogram into periodic and aperiodic components within latent space by combining parameterized helix trajectories with Neural Controlled Differential Equations. During inference, our framework further employs a variance minimization strategy across image augmentations that simulate common quality issues in echocardiogram acquisition, along with differential adaptation rates for periodic and aperiodic components. Theoretical analysis is provided to demonstrate that our variance minimization objective effectively bounds the regression error under mild conditions. Furthermore, extensive experiments across three pediatric age groups demonstrate that Q-PART not only significantly outperforms existing approaches in pediatric LVEF prediction, but also exhibits strong clinical screening capability with high mAUROC scores (up to 0.9747) and maintains gender-fair performance across all metrics, validating its robustness and practical utility in pediatric echocardiography analysis.

Figures (6)

• Figure 1: (a) Adult and Pediatric Echocardiogram Visualization. These two kinds of echocardiograms exhibit anatomical and physiological variations, including distinct differences in heart size and shape Moreover, We demonstrate two quality issue, i.e., acoustic shadow and speckle noise. (b) Quasi-Period Signal. Electrocardiogram (ECG) signals can be decomposed into periodic components originating from regular cardiac cycles and aperiodic components induced by physiological variations and noise.
• Figure 2: Overview of Q-PART Framework. The Quasi-Period Network decomposes input echocardiogram sequences into periodic and aperiodic components. The periodic component is modeled through a parameterized helix trajectory with frequency $\boldsymbol{f}$, phase shift $\boldsymbol{\phi}$, offset $\boldsymbol{b}$, and velocity $\boldsymbol{v}$. The aperiodic component is captured by Neural CDE. The model is jointly optimized with regression loss $\mathcal{L}_{reg}$ for LVEF prediction and reconstruction loss $\mathcal{L}_{rec}$ for feature learning.
• Figure 3: Variance Minimization during Test. we generate multiple augmented views of the test sequence through domain-specific transformations. The model adapts to each test case by minimizing prediction variance $\mathcal{L}_{var}$ across augmented views and maintaining reconstruction consistency $\mathcal{L}_{rec}$, with different learning rates applied to periodic and aperiodic components.
• Figure 4: ROC Analysis at Multiple Clinical Thresholds. Please zoom in for better view. Receiver Operating Characteristic (ROC) curves for LVEF regression with four clinically significant thresholds: 35% (blue), 40% (green), 45% (red), and 50% (cyan). These thresholds represent different degrees of cardiac dysfunction severity. The x-axis shows 1-specificity (false positive rate) and the y-axis shows sensitivity (true positive rate). The mAUROC is shown in gray box. A high mAUROC value indicates strong clinical reliability. Higher sensitivity ensures early detection of cardiac dysfunction, while higher specificity reduces unnecessary follow-up examinations and patient anxiety, collectively supporting more accurate clinical decision-making and resource allocation.
• Figure 5: Impact of $K$ Augmentation Number on Model Performance. The plot shows how performance varies with different numbers of augmentation number across three age cohorts.