Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Manifold-Aware Diffusion-Augmented Contrastive Learning for Noise-Robust Biosignal Representation

Rami Zewail

TL;DR

Physiological time-series data suffer from limited annotations and class imbalance. This work introduces Diffusion-Augmented Contrastive Learning (DACL), which uses forward diffusion-based augmentation within a context-aware Scattering Transformer latent space to learn noise-robust, discriminative biosignal representations. DACL achieves a 0.9741 AUROC on PhysioNet 2017 ECG AF detection, rivaling supervised state-of-the-art while enabling single-pass inference and efficient online augmentation. The study demonstrates that early-stage diffusion acts as a local manifold explorer, providing precise, manifold-consistent perturbations that improve both robustness and efficiency for real-time clinical monitoring.

Abstract

Learning robust representations for physiological time-series signals continues to pose a substantial challenge in developing efficient few-shot learning applications. This difficulty is largely due to the complex pathological variations in biosignals. In this context, this paper introduces a manifold-aware Diffusion-Augmented Contrastive Learning (DACL) framework, which efficiently leverages the generative structure of latent diffusion models with the discriminative power of supervised contrastive learning. The proposed framework operates within a contextualized scattering latent space derived from Scattering Transformer (ST) features. Within a contrastive learning framework, we employ a forward diffusion process in the scattering latent space as a structured manifold-aware feature augmentation technique. We assessed the proposed framework using the PhysioNet 2017 ECG benchmark dataset. The proposed method achieved a competitive AUROC of 0.9741 in the task of detecting atrial fibrillation from a single-lead ECG signal. The proposed framework achieved performance on par with relevant state-of-the-art related works. In-depth evaluation findings suggest that early-stage diffusion serves as an ideal "local manifold explorer," producing embeddings with greater precision than typical augmentation methods while preserving inference efficiency.

Manifold-Aware Diffusion-Augmented Contrastive Learning for Noise-Robust Biosignal Representation

TL;DR

Physiological time-series data suffer from limited annotations and class imbalance. This work introduces Diffusion-Augmented Contrastive Learning (DACL), which uses forward diffusion-based augmentation within a context-aware Scattering Transformer latent space to learn noise-robust, discriminative biosignal representations. DACL achieves a 0.9741 AUROC on PhysioNet 2017 ECG AF detection, rivaling supervised state-of-the-art while enabling single-pass inference and efficient online augmentation. The study demonstrates that early-stage diffusion acts as a local manifold explorer, providing precise, manifold-consistent perturbations that improve both robustness and efficiency for real-time clinical monitoring.

Abstract

Learning robust representations for physiological time-series signals continues to pose a substantial challenge in developing efficient few-shot learning applications. This difficulty is largely due to the complex pathological variations in biosignals. In this context, this paper introduces a manifold-aware Diffusion-Augmented Contrastive Learning (DACL) framework, which efficiently leverages the generative structure of latent diffusion models with the discriminative power of supervised contrastive learning. The proposed framework operates within a contextualized scattering latent space derived from Scattering Transformer (ST) features. Within a contrastive learning framework, we employ a forward diffusion process in the scattering latent space as a structured manifold-aware feature augmentation technique. We assessed the proposed framework using the PhysioNet 2017 ECG benchmark dataset. The proposed method achieved a competitive AUROC of 0.9741 in the task of detecting atrial fibrillation from a single-lead ECG signal. The proposed framework achieved performance on par with relevant state-of-the-art related works. In-depth evaluation findings suggest that early-stage diffusion serves as an ideal "local manifold explorer," producing embeddings with greater precision than typical augmentation methods while preserving inference efficiency.

Paper Structure

This paper contains 24 sections, 5 equations, 4 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Generative Data Augmentation (Offline Synthesis)
  4. Reconstruction-based Anomaly Detection
  5. Diffusion-Guided Representation Learning
  6. Methodology
  7. Framework Overview
  8. Latent Space Construction
  9. Diffusion-Based Latent Augmentation
  10. UNet-like Encoder Architecture
  11. Combined Contrastive Objective
  12. Inference Strategy
  13. Dataset and Preprocessing
  14. Evaluation Protocol and Baselines
  15. Internal Benchmarking Results
  16. ...and 9 more sections

Figures (4)

  • Figure 1: Conceptual diagram of the DACL framework. Stage I involves training a VAE on all ST features to create a general-purpose latent space. In Stage II, this latent space is used as the main training loop. A triplet of latent vectors is sampled, and noisy views are generated via the forward diffusion process. These views are passed through a shared U-Net encoder, normalized to the unit hypersphere, and trained via a supervised contrastive objective to produce noise-robust, class-discriminative embeddings.
  • Figure 2: Patient-Level ROC Curve for the DACL model utilizing the Combined Contrastive Loss on PhysioNet 2017. The curve demonstrates robust performance with an AUROC of 0.9741, confirming high sensitivity at low false-positive rates.
  • Figure 3: t-SNE visualization of the test embeddings generated by the DACL encoder. The clusters show markedly improved separation compared to baselines, with the Normal class (Red) forming a compact, well-defined manifold distinct from anomalies.
  • Figure 4: Results from the diffusion timestep ablation study. Performance clearly peaks during Early timesteps ($t \le 16$) and degrades as noise intensity increases, confirming that local manifold exploration is superior to heavy corruption for this task.