Research map / Probabilistic and scientific modeling / Molecules and proteins
Q-PART: Quasi-Periodic Adaptive Regression with Test-time Training for Pediatric Left Ventricular Ejection Fraction Regression
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Q-PART, a test-time training framework for pediatric LVEF regression that decomposes echocardiogram latents into periodic (helix trajectory) and aperiodic (Neural CDE) parts and minimizes prediction variance across ultrasound-specific augmentations at test time.
Problem
Adapting LVEF regression models to pediatric echocardiograms at test time, where existing TTT methods are designed for classification (entropy-based) and ignore the quasi-periodic nature of cardiac signals.
Approach
A Quasi-Period Network splits latent echo features into a periodic helix-trajectory component and an aperiodic Neural CDE component, and test-time adaptation uses a variance-minimization loss over augmented views with different BN learning rates for the two components.
Key results
- On the Pre-School cohort Q-PART reaches MAE 7.235, MAPE 0.1611, RMSE 9.290, beating the best TTT baseline NC-TTT (MAE 8.352) and TTT++ (8.053).
- Clinical screening mAUROC reaches 0.9362, 0.9747, and 0.9083 for Pre-School, School Age, and Adolescence cohorts.
- Ablation on Pre-School shows the full model (MAE 7.235) improving over the baseline without components (MAE 8.129).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add QP-Net (ablation) | Q-PART · Pre-School | MAE: 8.129 -> 7.949 | Helped |
| Add variance minimization to QP-Net (ablation) | Q-PART · Pre-School | MAE: 7.949 -> 7.283 | Helped |
| Add the differential learning-rate strategy to QP-Net and VM (ablation) | Q-PART · Pre-School | MAE: 7.283 -> 7.235 | Helped |
| Use Q-PART instead of TTT++ | pediatric LVEF regression · Pre-School | MAE: 8.053 -> 7.235 | Helped |
| Use Q-PART instead of NC-TTT | pediatric LVEF regression · School Age | MAE: 7.285 -> 6.706 | Helped |
| Use Q-PART instead of NC-TTT | pediatric LVEF regression · Adolescence | MAE: 7.422 -> 6.980 | Helped |
Surprising findings
- Using K=2 augmented views for variance minimization performs worse than no variance minimization (MAE 8.036 vs 7.909 on Pre-School).
Limitations
- Theoretical assumptions (unbiased augmentation, zero cross-covariance across augmentations) are idealized and only mildly justified.
- Only apical-4-chamber echo videos from two datasets were evaluated, with a single adult-to-pediatric transfer setting.
Experimental setup
Datasets: EchoNet-Dynamic EchoNet-Pediatric
Benchmarks: EchoNet-Pediatric
Metrics: MAE RMSE MAPE mAUROC
Baselines: EchoNet MemSAM EchoCLIP Source-only TTT++ Clust3 NC-TTT
Builds on: Test-Time Training (TTT) TTT++ Tent Neural CDE EchoNet-Dynamic
Closest papers by meaning
- EchoXFlow: A Beamspace Echocardiography Dataset for Cardiac Motion, Flow, and Function (NeurIPS 2026 · similarity 0.83)
- ECHOPulse: ECG Controlled Echocardio-gram Video Generation (ICLR 2025 · similarity 0.83)
- CardioLens: Revealing the Clinical Reality Gap of MLLMs via Multi-Sequence Cardiac MRI Evaluations (NeurIPS 2026 · similarity 0.83)
- Self-supervised Learning of Echocardiographic Video Representations via Online Cluster Distillation (NeurIPS 2025 · similarity 0.83)
- Med-Real2Sim: Non-Invasive Medical Digital Twins using Physics-Informed Self-Supervised Learning (NeurIPS 2024 · similarity 0.82)
- CMR-RD: Long-Tailed Adaptive VLM for Explainable CMR Diagnosis (CVPR 2026 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.