Domain-independent PPG signal quality assessment framework via representation learning

Fine-tunes a pretrained PPG transformer (PPG-PT) with triplet loss; quality = embedding distance to reference high-quality segments, so no retraining for new devices/sites. Evaluated on finger (CSL), wrist (WCS) and neck (Apnea & Artifacts) PPG.

• Form factor: any

Photoplethysmography (PPG) is a widely used non-invasive sensing technique for monitoring cardiovascular and respiratory parameters, however, its measurements are highly susceptible to artifacts, which can degrade signal quality and, in turn, the accuracy of derived vital signs. Reliable PPG signal quality assessment (SQA) is therefore critical, yet existing SQA methods often rely on handcrafted features or ad-hoc thresholds and tend to be dataset-specific, limiting their generalizability across different sensor locations and users. To address these gaps, this paper proposes a novel SQA approach for PPG that is real-time compatible and sensor-location invariant. The method leverages deep representation learning by fine-tuning a pre-trained PPG transformer (PPG-PT) with a triplet loss to learn an embedding space where PPG segments are meaningfully partitioned based on their morphology. Signal quality is then quantified by computing each window’s embedding distance to a set of reference embeddings representing high-quality PPG segments, yielding a flexible reference-based quality score without requiring any model retraining when adapting to new devices or domains. The proposed framework was evaluated on three PPG datasets: the Complex System Laboratory (CSL) dataset providing finger PPG, the Wearable and Clinical Devices (WCS) for wrist PPG signals, and the Apnea and Artifacts (AA) dataset representing neck PPG signals, encompassing a wide range of artifact types and recording conditions. The model achieved strong agreement with expert annotations on the CSL dataset, with an F1-score of

~

90%. It also demonstrated robust generalization on the WCS wearable dataset: using a generalized reference set drawn from multiple subjects, the approach attained a high F1-score of about

~

95% on wrist PPG data, highlighting the method’s potential for population-level deployment. Performance on the AA neck PPG dataset was moderately acceptable with

~

70% F1-score, likely due to the complexity and varying executions of the different artifacts presented in the dataset. The proposed methodology provides a deep learning SQA framework capable of generalizing across different PPG sensor locations and devices.

Verbatim from the publisher's DOI record (opens in a new tab) · Publisher licence: https://creativecommons.org/licenses/by/4.0

No results from this paper on ChatPPG leaderboards yet. Browse the leaderboards.