[論文レビュー] Aortic Valve Disease Detection from PPG via Physiology-Informed Self-Supervised Learning

Jiaze Wang, Qinghao Zhao+5 more|arXiv (Cornell University)|Feb 4, 2026

Cardiac Valve Diseases and Treatments被引用数 0

ひとこと要約

The paper introduces Physiology-Guided Self-Supervised Learning (PG-SSL) using unlabeled PPG data to screen Aortic Stenosis (AS) and Aortic Regurgitation (AR), achieving improved AUC over supervised baselines with strong prognostic value.

ABSTRACT

Traditional diagnosis of aortic valve disease relies on echocardiography, but its cost and required expertise limit its use in large-scale early screening. Photoplethysmography (PPG) has emerged as a promising screening modality due to its widespread availability in wearable devices and its ability to reflect underlying hemodynamic dynamics. However, the extreme scarcity of gold-standard labeled PPG data severely constrains the effectiveness of data-driven approaches. To address this challenge, we propose and validate a new paradigm, Physiology-Guided Self-Supervised Learning (PG-SSL), aimed at unlocking the value of large-scale unlabeled PPG data for efficient screening of Aortic Stenosis (AS) and Aortic Regurgitation (AR). Using over 170,000 unlabeled PPG samples from the UK Biobank, we formalize clinical knowledge into a set of PPG morphological phenotypes and construct a pulse pattern recognition proxy task for self-supervised pre-training. A dual-branch, gated-fusion architecture is then employed for efficient fine-tuning on a small labeled subset. The proposed PG-SSL framework achieves AUCs of 0.765 and 0.776 for AS and AR screening, respectively, significantly outperforming supervised baselines trained on limited labeled data. Multivariable analysis further validates the model output as an independent digital biomarker with sustained prognostic value after adjustment for standard clinical risk factors. This study demonstrates that PG-SSL provides an effective, domain knowledge-driven solution to label scarcity in medical artificial intelligence and shows strong potential for enabling low-cost, large-scale early screening of aortic valve disease.

研究の動機と目的

• Motivate scalable, low-cost screening for AVD beyond echocardiography constraints.
• Leverage large-scale unlabeled PPG data to learn disease-relevant morphology without direct labels.
• Translate physiological knowledge of AS/AR into a proxy self-supervised task for pre-training.
• Fine-tune on a small labeled AS/AR subset with an efficient dual-branch architecture.
• Demonstrate prognostic value and independence from standard clinical risk factors.

提案手法

• Formulate clinical AVD physiology into computable PPG morphological phenotypes to create a proxy pretext task.
• Pre-train on 170,702 unlabeled PPG samples from UK Biobank to learn domain-specific features.
• Use a dual-branch gated-fusion architecture to fine-tune on a small labeled AS/AR subset.
• Compare PG-SSL against supervised baselines and generic SSL approaches (SimCLR, reconstruction, K-means).
• Visualize model focus with Grad-CAM to verify physiological relevance of features.
• Evaluate prognostic independence via multivariable Cox regression and propensity score matching.

Figure 1 : Performance evaluation of the model on the independent test set. (A, B) ROC curves for AS and AR detection. The model achieves an AUC of 0.765 for AS and 0.776 for AR. (C) Calibration curves. The axes focus on the [0, 0.45] interval, showing high consistency between predicted probabilitie

実験結果

リサーチクエスチョン

• RQ1 Can physiology-informed self-supervised pre-training on unlabeled PPG improve AS/AR screening with limited labeled data?
• RQ2 Do the learned features provide independent prognostic value beyond standard clinical risk factors?
• RQ3 How does PG-SSL compare to generic SSL methods and traditional supervised models on AS/AR tasks?
• RQ4 Is the model attention aligned with known hemodynamic mechanisms of AS and AR?
• RQ5 Can early time-to-diagnosis windows be exploited for improved screening performance?

主な発見

Method	AUROC	S@60Sp	BalAcc
ResNet1D (Baseline)	0.6985	0.6531	0.6613
TimesNet 45	0.6977	0.7347	0.6821
Attn-LRCN 8	0.7020	0.7551	0.7041
SimCLR	0.7175	0.6939	0.6939
Reconstruction	0.7216	0.7143	0.6922
K-Means	0.7000	0.6735	0.6576
Feature-based Clustering	0.7201	0.6735	0.6613
PiLA (Ours)	0.7645	0.7755	0.7072
ResNet1D (Baseline)	0.7013	0.7143	0.6786
TimesNet 45	0.6793	0.6486	0.6494
Attn-LRCN 8	0.7236	0.7297	0.7025
SimCLR	0.7252	0.6905	0.6819
Reconstruction	0.7508	0.7619	0.7033
K-Means	0.7199	0.7619	0.7029
Feature-based Clustering	0.6783	0.6786	0.6289
PiLA (Ours)	0.7756	0.7857	0.7490

• PG-SSL achieved AUROC of 0.765 for AS and 0.776 for AR on the independent test set.
• At 60% specificity, S@60Sp was 77.6% for AS and 78.6% for AR.
• PG-SSL outperformed supervised baselines and several SSL baselines (e.g., SimCLR, reconstruction).
• Model predictions remained prognostically meaningful after adjusting for clinical risk factors in Cox models.
• Grad-CAM visualizations showed physiologically plausible attention patterns corresponding to AS and AR features.
• Handcrafted features fused with deep features reduced performance, indicating high completeness of PiLA features.

Figure 2 : Grad-CAM visualization of model attention across different groups. (Left) Healthy controls show diffuse attention, verifying holistic waveform integrity. (Center) AS patients show focused attention on the delayed systolic upstroke and dicrotic notch. (Right) AR patients exhibit broad cove