EnhancePPG: Improving PPG-based Heart Rate Estimation with Self-Supervision and Augmentation
Luca Benfenati 1 , Sofia Belloni 1 , Alessio Burrello 1 5 , Panagiotis Kasnesis 2 3 , Xiaying Wang 4 , Luca Benini 4 5 , Massimo Poncino 1 , Enrico Macii 6 , Daniele Jahier Pagliari 1 · 2024
Read EnhancePPG: Improving PPG-based Heart Rate Estimation with Self-Supervision and Augmentation by Luca Benfenati 1 , Sofia Belloni 1 , Alessio Burrello 1 5 , Panagiotis Kasnesis 2 3 , Xiaying Wang 4 , Luca Benini 4 5 , Massimo Poncino 1 , Enrico Macii 6 , Daniele Jahier Pagliari 1 (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-17860.txt.
Introduction
Luca Benfenati and a team of co-authors improve heart rate reading from wrist light sensors. Photoplethysmography, the green-light method in wearables, struggles when motion muddies the signal. Deep learning helps, but it needs lots of labeled data.
Their method, EnhancePPG, pairs self-supervised pre-training with data augmentation. The model first learns to rebuild raw signals from unlabeled data, using a U-Net-like autoencoder shape. That training needs only small changes to existing top models and adds little delay at run time. On the PPG-DaLiA benchmark, error fell by about twelve percent, from just over four beats per minute to about three and a half. The authors stress the gain comes from training alone.
Catalog record
| CREATOR | Luca Benfenati 1 , Sofia Belloni 1 , Alessio Burrello 1 5 , Panagiotis Kasnesis 2 3 , Xiaying Wang 4 , Luca Benini 4 5 , Massimo Poncino 1 , Enrico Macii 6 , Daniele Jahier Pagliari 1 |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 22 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2412-17860
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
EnhancePPG: Improving PPG-based Heart Rate Estimation with Self-Supervision and Augmentation Luca Benfenati 1 , Sofia Belloni 1 , Alessio Burrello 1 5 , Panagiotis Kasnesis 2 3 , Xiaying Wang 4 , Luca Benini 4 5 , Massimo Poncino 1 , Enrico Macii 6 , Daniele Jahier Pagliari 1 Abstract Heart rate (HR) estimation from photoplethysmography (PPG) signals is a key feature of modern wearable devices for health and wellness monitoring. While deep learning models show promise, their performance relies on the availability of large datasets. We present EnhancePPG, a method that enhances state-of-the-art models by integrating self-supervised learning with data augmentation (DA). Our approach combines self-supervised pre-training with DA, allowing the model to learn more generalizable features, without needing more labelled data. Inspired by a U-Net-like autoencoder architecture, we utilize unsupervised PPG signal reconstruction, taking advantage of large amounts of unlabeled data during the pre-training phase combined with data augmentation, to improve state-of-the-art models’ performance. Thanks to our approach and minimal modification to the state-of-the-art model, we improve the best HR estimation by 12.2%, lowering from 4.03 Beats-Per-Minute (BPM) to 3.54 BPM the error on PPG-DaLiA. Importantly, our EnhancePPG approach focuses exclusively on the training of the selected deep learning model, without significantly increasing its inference latency. ©2024 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. I Introduction & Related Works Wearable health monitoring devices have rapidly advanced in the health and wellness Internet-of-Things (IoT) era, integrating sophisticated sensors, efficient computational platforms, and cutting-edge algorithms [ 1 , 2 , 3 ] . Heart Rate (HR) is a vital physiological parameter that offers insights into physical activity, stress levels, and potential health concerns [ 4 ] . Although HR can be assessed intermittently in clinical or home environments, continuous monitoring is essential for the early detection of health issues. One way to continuously monitor HR is wrist-worn Photoplethysmographic (PPG) sensors. These sensors have become popular with the rise of affordable fitness trackers like Apple Watch [ 5 ] and Fitbit [ 6 ] . They measure HR by detecting light from pulsating light-emitting diodes (LEDs) absorbed or reflected by blood vessels [ 7 ] . Despite their user comfort, PPG sensors face accuracy challenges due to Motion Artifacts (MA) from variations in sensor pressure, external light interference, and arm movements, particularly during intense physical activity when blood flow variability complicates HR estimation. To address the challenges posed by MA, initial approaches in the literature proposed filtering techniques, where the correlation between acceleration data and PPG signals was leveraged to mitigate MA noise on PPG signal [ 8 ] , [ 9 ] . However, these methods often suffer from poor generalization on unseen data due to the high number of tunable hyperparameters. Deep learning techniques have been proposed to overcome this limitation, showing promising results on various public datasets [ 10 ] . In [ 11 , 12 ] , the authors address the challenges of poor generalization and the high complexity of deep learning by integrating Neural Architecture Search (NAS) strategies with Temporal Convolutional Networks (TCNs) for HR monitoring. Despite the progress made, most deep learning solutions fail to achieve a sufficiently low Mean Absolute Error (MAE) in HR estimation for practical deployment in real-world scenarios, where inter-subject variability makes accurate HR monitoring particularly difficult. In this work, we propose an innovative approach that integrates self-supervised learning and data augmentation techniques [ 13 , 14 ] to enhance HR estimation performance of deep learning models. While our method is mostly orthogonal to the specific deep neural network selected, we assess its performance using PULSE ( P pg and im U signa L fu S ion for HR E stimation [ 15 ] ), the state-of-the-art deep learning model for HR estimation from PPG signals on PPG-DaLiA [ 10 ] , the biggest public dataset available. By introducing a pre-training phase combined with augmentation techniques, we improve the robustness and the accuracy of PULSE without significantly impacting inference latency, ensuring suitability for deployment on low-power edge devices like smartwatches. Our contributions are as follows: • Self-supervised learning (SSL) for PPG data : we are the first to introduce a self-supervised pre-training step by restructuring the selected deep learning model into an autoencoder-like framework, inspired by U-Net [ 16 ] . This allowed us to pre-train the model on two additional unlabelled datasets: WESAD [ 17 ] and a new dataset provided by West Attica University. The pre-training task is the PPG signal reconstruction. This approach allows for the use of more readily accessible data, as they can be collected without needing any labelling. • Data Augmentation (DA) Integration : we leverage DA techniques, inspired by [ 13 ] , to expand the pre-training dataset with a focus on increasing HR variability, thereby enhancing the model’s generalization and robustness to noisy inputs. Additionally, we evaluate how our SSL framework supports the use of DA without altering the HR labels, ensuring that the model can learn broader features without compromising its ability to generalize to unseen data. Ours is the first attempt to exploit large unlabelled datasets combined with data augmentation in a self-supervised way for HR estimation, paving the way to utilising large unlabelled datasets to improve the generalization of this vital task. Importantly, our EnhancePPG approach, which combines self-supervised pre-training with data augmentation, solely affects the training phase, leaving the inference process unchanged. To benchmark our method’s accuracy, we used a modified PULSE architecture derived from the original one to reduce training time and improve performance. These architectural changes alone lowered the Mean Absolute Error (MAE) from 4.03 BPM to 3.73 BPM. Applying EnhancePPG, we brought the MAE of the architecture down to 3.54 BPM, achieving a further 5.09% reduction. Notably, we achieved significant reductions in MAE for the two most critical patients, S5 and S8, reducing the MAE achieved by the original PULSE model by 18.9% and 36.8%, respectively. Our modified PULSE model can be deployed on an STM32 NUCLEO-H743ZI2 with a latency of 431.6 ms, only 4.43% higher than the original one. On the other hand, enhancing the model with our proposed pipeline does not lead to any latency increment. II Material & Methods II-A Model architecture We evaluate our pre-training and data augmentation methodology for PPG-based HR estimation on an enhanced version of PULSE, state-of-the-art model for PPG-based HR estimation originally introduced by [ 15 ] . PULSE leverages PPG signals and tri-axial accelerometer data to improve HR estimation accuracy by integrating temporal convolutions with a Multi-Head Cross-Attention (MHCA) module. The architecture comprises three 1D convolutional blocks with increasing channels (32, 48, 64), each containing three consecutive convolutional layers to capture short-term temporal variations in both PPG and accelerometer data. Each block’s output is downsampled using average pooling layers. The feature maps generated by the convolutional layers are processed by the MHCA module, where PPG signals serve as query vectors and accelerometer features are used as key