SiamQuality

SiamQuality is a self-supervised foundation model for photoplethysmography (PPG), the optical pulse waveform captured by pulse oximeters and consumer wearables. PPG is cheap and ubiquitous but notoriously noisy: motion, poor perfusion, and sensor contact routinely corrupt the signal, and most prior deep-learning pipelines cope by simply discarding low-quality segments. This both wastes data and biases models toward clean conditions that rarely hold in real ICU or ambulatory monitoring. SiamQuality instead treats signal quality as the central learning objective, building representations that remain stable across the full range of waveform quality.

Developed by Cheng Ding and colleagues at the Georgia Institute of Technology and Emory University, with collaborators at Duke University and the University of California, San Francisco, the model was published in Physiological Measurement in August 2024. Its key idea is a quality-aware contrastive scheme: rather than augmenting a clean signal to create a positive pair, SiamQuality pairs a high-quality PPG segment with a temporally adjacent low-quality segment from the same patient, which shares the underlying physiological state but differs in artifact level. The network is then trained to map both to similar embeddings.

By forcing the encoder to ignore quality-driven nuisance variation while preserving physiology, SiamQuality produces a frozen backbone that transfers strongly to clinical downstream tasks, positioning it among the first PPG-specific foundation models aimed at robust, deployable cardiovascular monitoring.

#Key Features

• Quality-pairing contrastive objective: Positive pairs are formed from real high- and low-quality PPG segments captured close in time (within short windows from the same patient), so the model learns physiology-preserving, artifact-invariant representations rather than augmentation-invariant ones.
• SimSiam framework, no negatives: Built on the SimSiam architecture with encoder, projector, and predictor heads trained via a cosine-similarity loss, avoiding negative sampling and large memory banks.
• ConvNet backbone: Uses ResNet-family encoders (ResNet-50/101/152 variants evaluated) rather than a transformer, keeping the model efficient for 1D physiological signals.
• Curriculum learning over artifacts: Training progressively increases the difficulty of paired low-quality signals, stabilizing convergence under heavy noise.
• Frozen-encoder transfer: The pretrained encoder is reused as a fixed feature extractor across four distinct clinical tasks with only lightweight task heads trained.

#Technical Details

SiamQuality was pretrained on more than 36 million 30-second PPG signal pairs drawn from roughly 21,000 ICU patients at UCSF Medical Center, totaling over 600,000 hours of continuous recordings (acquired at 240 Hz, downsampled to 40 Hz). The self-supervised model is then frozen and evaluated on four downstream tasks across six public datasets: heart rate estimation (TROIKA, DaLiA, WESAD), respiratory rate estimation (BIDMC), blood pressure estimation (PulseDB, combining MIMIC-III and VitalDB), and atrial fibrillation detection (Stanford AF). Across these benchmarks SiamQuality reported substantial gains over prior task-specific baselines—for example, lower mean absolute error on respiratory rate from the BIDMC set and improved F1 for AF detection—demonstrating that a single quality-robust encoder can support both regression and classification on PPG. Reported comparisons carry the usual caveats, since some prior methods filtered out poor-quality signals and are therefore not strictly comparable.

#Applications

SiamQuality targets vital-sign monitoring from PPG in settings where signal quality cannot be guaranteed: ICU bedside monitors, telemetry, and wrist- or finger-worn wearables. Because the encoder is pretrained once and frozen, researchers and device developers can adapt it to new tasks—heart rate, respiratory rate, blood pressure, or arrhythmia screening—by training small heads on modest labeled datasets, rather than collecting large clean-signal corpora. This is especially useful for ambulatory and continuous-monitoring applications, where motion artifact is pervasive and discarding noisy data would discard most of the recording.

#Impact

SiamQuality is an early demonstration that the foundation-model paradigm can be adapted to noisy physiological waveforms by making signal quality the learning target rather than a filtering criterion, and that frozen self-supervised PPG features transfer across heterogeneous clinical tasks. Its quality-pairing strategy offers a template for other biosignal modalities (such as ECG) that face the same clean-data bias. Adoption is constrained, however, by openness limitations: the public GitHub repository provides training and inference code but ships no declared license, and the pretrained weights are not released—the repository contains only an empty weight placeholder—so reproducing the foundation model requires access to the proprietary UCSF ICU dataset. The absence of released checkpoints and an explicit license currently limits straightforward reuse despite the method's promise.