EnhancePPG: Improving PPG-based Heart Rate Estimation with Self-Supervision and Augmentation (2412.17860v1)

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

• The paper introduces the EnhancePPG framework, which uses self-supervised learning and data augmentation to significantly improve the accuracy and robustness of heart rate estimation from PPG signals.
• By applying these techniques, EnhancePPG achieved a 12.2% reduction in mean absolute error on a state-of-the-art model and showed significant improvements (up to 36.8% MAE reduction) for problematic critical cases.
• The method maintains low inference latency and computational requirements, making it suitable for deployment on resource-constrained wearable health monitoring devices.

EnhancePPG: Advancements in Heart Rate Estimation via Self-Supervision and Augmentation

The paper "EnhancePPG: Improving PPG-based Heart Rate Estimation with Self-Supervision and Augmentation" presents a method designed to refine heart rate (HR) estimation derived from photoplethysmography (PPG) signals. This methodology is significant for the data-driven nature of wearable health monitoring devices, which are becoming increasingly prevalent due to the rise of the Internet-of-Things (IoT) in healthcare.

Methodological Innovations

The core innovation of this work is the EnhancePPG framework, which enhances existing deep learning models through two main components: self-supervised learning (SSL) and data augmentation (DA). The authors integrate these techniques to improve feature generalization and robustness without expending resources on large-scale labeled datasets.

Self-Supervised Learning: The authors restructure a deep learning model, inspired by U-Net architecture, into an autoencoder framework for SSL. This approach involves pre-training on vast amounts of unlabeled PPG data, obtained from datasets like WESAD and a new dataset from West Attica University, to reconstruct the PPG signals. This pre-training, conducted in an unsupervised manner, allows the model to capture task-agnostic features, enhancing the model's capability to learn from diverse signal patterns.

Data Augmentation: By implementing DA during the pre-training phase, the authors expand the HR variability in the data, which is crucial for improving model generalization. Unlike previous works that compromise HR label integrity during augmentation, EnhancePPG avoids this by focusing on unsupervised data, thereby averting potential biases introduced during supervised learning.

Performance and Comparative Analysis

The results presented in this paper highlight the efficacy of the EnhancePPG method applied to the PULSE architecture, which is the benchmark state-of-the-art model for HR estimation from PPG signals. The authors report a reduction in mean absolute error (MAE) from 4.03 BPM in the original model to 3.54 BPM with their approach. The reduction in error demonstrates a 12.2% improvement compared to previous methods without a substantial increase in computational requirements. Notably, significant performance enhancements were observed in patients with problematic heart rates, with reductions in MAE of 18.9% and 36.8% for critical cases.

Moreover, the empirical analysis demonstrates that the application of DA solely in the pre-training phase, as opposed to the fine-tuning phase, yields superior generalization abilities, which is pivotal for real-world applications where unseen data is encountered.

Deployment and Implications

From a practical standpoint, EnhancePPG's approach to HR estimation holds substantial implications for wearable technology, particularly in devices constrained by computational resources. The modified PULSE model, equipped with EnhancePPG's enhancements, maintains low inference latency and can be deployed on resource-limited systems like the STM32 NUCLEO-H743ZI2 MCU without performance setbacks. This makes it feasible for integration into everyday fitness trackers and health monitoring devices.

Future Prospects

Future research can focus on exploring the potential of EnhancePPG's framework when integrated with alternative deep learning models beyond PULSE to determine its versatility in various signal processing tasks. Additionally, the method's scalability offers opportunities to leverage more extensive unlabeled datasets, potentially enhancing training efficacy further and reducing the dependence on labeled data.

In summary, EnhancePPG's efficient use of self-supervised learning and data augmentation marks a significant step forward in improving HR estimation accuracy from PPG signals. The proposed framework effectively addresses existing challenges in the domain and lays the groundwork for more adaptable and reliable wearable health monitoring solutions.