Abstract
Physiological signals from wearables and clinical monitors are often corrupted by noise and artifacts, posing a challenge for standard foundation models. We present SiamQuality, a CNN-based foundation model trained on 36 million signal pairs using a novel quality-pairing contrastive learning strategy. By leveraging temporal proximity to pair clean and noisy segments from the same patient, SiamQuality learns robust, quality-invariant representations. Our model achieves state-of-the-art performance in respiration rate estimation and atrial fibrillation detection, demonstrating significant resilience to data imperfections.
Introduction
Foundation models (FMs) have transformed natural language processing and computer vision by training on massive datasets. However, applying FMs to physiological signals—such as photoplethysmography (PPG) from wearables or ICU monitors—is difficult because real-world data is rarely "clean." Motion artifacts, sensor displacement, and environmental noise frequently degrade signal quality. Existing biosignal FMs often rely on complex Transformer architectures that struggle with these imperfections. SiamQuality addresses this gap by introducing a foundation model inherently robust to poor data quality, bridging the divide between ideal training conditions and clinical reality.
Methods
SiamQuality utilizes a Convolutional Neural Network (CNN) backbone, specifically ResNet, rather than the more common Transformer architecture. The core of the approach is the SimQuality self-supervised task, which employs a "quality-pairing" strategy.
Quality-Pairing and SimSiam Architecture
Instead of standard data augmentations, the model leverages the temporal consistency of medical data. It identifies a "good quality" anchor signal and pairs it with a "bad quality" segment from the same patient within a 5-minute window. Since the underlying physiological state (e.g., heart rhythm) is unlikely to change drastically in such a short period, the model is forced to produce similar representations for both clean and noisy versions of the same state. This process teaches the model to ignore noise and focus on biological features.
The training uses a Siamese network structure that minimizes the distance between these pairs without requiring negative samples. Furthermore, a curriculum learning approach is applied, where the model is progressively exposed to pairs with increasing levels of noise.
Datasets and Evaluation
The model was pre-trained on a massive dataset from the UCSF Medical Center, comprising over 36 million 30-second PPG pairs (approx. 600,000 hours) from 21,000 patients. Downstream evaluation was conducted on several public benchmarks:
- Heart Rate & Blood Pressure: TROIKA, Dalia, WESAD, and PulseDB.
- Respiration Rate (RR): BIDMC PPG and Respiration Dataset.
- Atrial Fibrillation (AF) Detection: Stanford AF dataset.
Results
SiamQuality established new state-of-the-art benchmarks across six downstream tasks. Notably, it achieved a Mean Absolute Error (MAE) of 0.89 for Respiration Rate, a 43% improvement over previous methods. In AF detection, the model reached an F1 score of 0.71 with 100% artifact coverage.
A key contribution is the introduction of the Artifact Tolerance Curve (AT-Curve), which visualizes performance degradation as signal quality decreases. SiamQuality demonstrated remarkable resilience; on the noisy DALIA dataset, it maintained high accuracy even when signals contained over 60% artifacts, whereas competing models failed. Additionally, the model exhibited clear scaling laws, with performance improving consistently as the ResNet depth increased from 50 to 152 layers.
Discussion and Limitations
While SiamQuality represents a significant step forward, several areas for future work remain. The model currently requires fine-tuning for specific tasks and does not support zero-shot execution. It also faces challenges with class imbalance in rare events like Premature Ventricular Contractions. Future iterations could benefit from more generalized quality assessment tools applicable to other modalities like ECG or EEG. Finally, clinical integration will require further engineering and validation through clinical trials.
Summary updated