Abstract
Multi-channel photoplethysmography (PPG) sensors have found widespread adoption in wearable devices for monitoring cardiac health. Channels thereby serve different functions—whereas green is commonly used for metrics such as heart rate and heart rate variability, red and infrared are commonly used for pulse oximetry. In this paper, we introduce a novel method that simultaneously fuses multi-channel PPG signals into a single recovered PPG signal that can be input to further processing. Via signal fusion, our learning-based method compensates for the artifacts that affect wavelengths to different extents, such as motion and ambient light changes. We evaluate our method on a novel dataset of multi-channel PPG recordings and electrocardiogram recordings for reference from 10 participants over the course of 13 hours during real-world activities outside the laboratory. Using the fusion PPG signal our method recovered, participants' heart rates can be calculated with a mean error of 4.5 bpm (23% lower than from green PPG signals at 5.9 bpm).
Introduction
Photoplethysmography (PPG) is the foundational technology behind heart rate monitoring in modern smartwatches. By emitting light into the skin and measuring absorption changes caused by blood volume pulses, these sensors provide a non-invasive window into cardiac health. However, PPG signals are notoriously susceptible to motion artifacts and ambient light interference. While current industry standards typically rely on green light for heart rate due to its superior signal-to-noise ratio, and red/infrared for oxygen saturation (SpO2), they rarely leverage the combined strengths of all three wavelengths simultaneously to combat noise.
Researchers from ETH Zurich have developed Tri-Spectral PPG, a deep-learning approach that fuses green, red, and infrared signals into a single, high-fidelity heart rate waveform. This method significantly improves robustness during daily activities, reducing heart rate estimation errors by 23% compared to single-channel methods.
Methods: Learning-Based Signal Fusion
The core of the Tri-Spectral PPG approach is a learning-based fusion model designed to recover the underlying pulse waveform from noisy multi-channel inputs.
Architecture
The researchers adapted a U-Net architecture—traditionally used for medical image segmentation—to process 1D time-series data. The network takes three raw PPG channels (Green, Red, and Infrared) as input. By processing these signals through a series of convolutional layers, the model learns to identify and prioritize the most reliable wavelength components at any given moment, effectively "denoising" the data into a single fused PPG signal.
Synthesizing Ground Truth
Training a model to "clean" PPG signals is difficult because a perfectly clean PPG reference does not exist in real-world conditions. To overcome this, the team used simultaneous Electrocardiogram (ECG) recordings as a reference. By using ECG R-peaks to segment noisy PPG signals and applying ensemble averaging, they created a synthetic "clean" PPG reference that represents the user's actual cardiac morphology, providing a target for the U-Net to learn.
Experimental Dataset
The method was validated using a new, high-quality dataset collected "in-the-wild":
- Participants: 10 healthy adults (Fitzpatrick skin types 1–3).
- Protocol: 13 hours of data per participant during diverse real-world activities, including hiking, riding in vehicles, and visiting museums.
- Total Data: 130 hours of synchronized 3-channel PPG and Lead I ECG data.
Results and Discussion
The Tri-Spectral fusion method demonstrated a marked improvement over individual channels. The fused signal achieved a Mean Absolute Error (MAE) of 4.5 bpm, representing a 23% improvement over the green channel alone (5.9 bpm). Red and Infrared channels, when used in isolation, were significantly less reliable, with MAEs of 13.4 and 11.1 bpm, respectively.
Beyond heart rate accuracy, the fused signal showed a high correlation (0.831) with the reference waveform, far exceeding the 0.645 achieved by the green channel alone. This suggests that the model preserves the morphology of the pulse wave, which is critical for advanced metrics like pulse transit time or blood pressure estimation.
Conclusion
Tri-Spectral PPG demonstrates that wavelength diversity is a powerful tool for noise rejection. By leveraging the different penetration depths of green, red, and infrared light, the U-Net model can "look around" motion artifacts that might otherwise blind a single-color sensor. While future work is needed to validate the model on more diverse skin tones and optimize it for low-power wearable processors, this research represents a significant step toward more robust and reliable wearable health monitoring.
Summary updated