TinyArtifactNet: A Resource-Efficient Deep Learning Model for On-Device Motion Artifact Detection in PPG

TinyArtifactNet is a lightweight hybrid multi-scale temporal CNN for on-device PPG motion-artifact detection. Authors report DaLiA accuracy 88.17% and Dice 88.77% at 21.3K parameters / 0.16 MB, with cross-dataset TROIKA/WESAD numbers and STM32+MAX30102 real-time accuracy 87.2%. No public code or weights located as of 2026-11-06.

• Form factor: wrist

Photoplethysmogram (PPG) signals are widely used in wearable devices for continuous monitoring of cardiovascular vital signs. However, PPG signals are highly susceptible to motion artifacts, which severely degrade signal quality and lead to inaccurate physiological measurements. Consequently, any vital signs derived from contaminated PPG lead to wrong clinical or health-related decisions. To address this challenge, this study proposes TinyArtifactNet, a lightweight hybrid multi-scale temporal deep learning model for efficient and accurate motion artifact detection in PPG signals. The proposed architecture integrates ghost modules, multi-scale dilated convolutions, and squeeze-and-excitation mechanisms to achieve a favorable trade-off between performance and computational efficiency. The effectiveness of TinyArtifactNet was evaluated using three publicly available datasets and one self-curated private dataset. This work presents three key contributions: the design of a lightweight architecture, its deployment on an embedded platform, and real-time validation, all combining within a unified framework. On the PPG-DaLiA dataset, the proposed model achieved state-of-the-art performance among embedded-deployable lightweight models with an accuracy of 88.17% and a Dice score of 88.77%, while requiring only 21.3K parameters and 0.16 MB of memory. Furthermore, cross-dataset evaluation on two completely unseen datasets, TROIKA and WESAD, yielded accuracies of 81.84% and 90.34%, respectively, and Dice scores of 82.93% and 90.05%, respectively, demonstrating strong generalization. In addition, real-time evaluation using raw PPG signals acquired from a low-cost MAX30102 sensor and deployed on an STM32 microcontroller achieved an accuracy of 87.2%. These results confirm the suitability of TinyArtifactNet for real-time motion artifact detection on low-cost, resource-constrained platforms, making it a promising solution for wearable healthcare systems.

Verbatim from OpenAlex (opens in a new tab) · Abstract via OpenAlex (publisher-deposited); © the authors/publisher

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

• IEEE Signal Processing Cup 2015 (TROIKA)wrist PPG during intensive exerciseEvaluation12 subjects · PPG · ECG · IMU
• PPG-DaLiAPPG dataset for motion compensation / HRTrainingEvaluation15 subjects · PPG · ECG · IMU
• WESADWearable Stress and Affect DetectionEvaluation15 subjects · PPG · ECG · IMU