Tri-Spectral PPG: Robust Reflective Photoplethysmography by Fusing Multiple Wavelengths for Cardiac Monitoring

summary

The gist

Multi-channel photoplethysmography (PPG) sensors have found widespread adoption in wearable devices for monitoring cardiac health Via signal fusion…

In short

This research introduced a method to combine signals from three different light wavelengths (green, red, and infrared) in multi-channel photoplethysmography (PPG). A learning-based model fuses these signals into a single robust PPG signal. This fusion effectively removes noise like motion and light changes, resulting in heart rate calculations with an error of 4.5 bpm, which is 23% better than using only the green channel.

Key concepts

Multi-channel PPG
This involves using sensors that measure light absorption at several different colors of light (like green, red, and infrared) to capture more detailed information about blood flow in the body. Different colors are sensitive to different factors like motion or skin tone.
Signal Fusion
This is the core technique where a learning model combines the separate PPG signals from all three wavelengths into one superior signal. The model learns how to weigh and merge these inputs so that noise affecting one channel is compensated for by the others, creating a cleaner overall measurement.
U-Net Architecture
This is a specific type of deep learning structure used in the study. It's designed to take multiple input signals (the three PPG channels) and transform them into a single desired output (the fused PPG signal). It uses layers that both increase and decrease the data complexity to learn complex relationships.
Artifact Compensation
The method is designed to automatically correct errors caused by external factors that distort the light signals, such as movement or changes in ambient light. By fusing the signals, the system learns to ignore these artifacts, leading to a more accurate and stable measurement of heart rate.

Terminology used across episodes

This episode discusses

The paper

Tri-Spectral PPG: Robust Reflective Photoplethysmography by Fusing Multiple Wavelengths for Cardiac Monitoring · Read on arXiv

Manuel Meier, Berken Utku Demirel, Christian Holz

Department of Computer Science, ETH Zurich, Switzerland

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).

DOI: 10.1109/BSN63547.2024.10780586

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: We're looking at the paper: "Tri-Spectral PPG: Robust Reflective Photoplethysmography by Fusing Multiple Wavelengths for Cardiac Monitoring".

Dev: Here's the gist of the paper. Multi-channel photoplethysmography (PPG) sensors have found widespread adoption in wearable devices for monitoring cardiac health Via signal fusion, our learning-based method compensates for the artifacts that affect wavelengths to different extents, such as motion and ambient light changes 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).

Rosa: Signal Acquisition and Data Collection The study utilized a novel dataset comprising recordings from 10 healthy adults over 13 hours during real-world activities outside the laboratory Each participant wore a standalone device integrating reflective PPG sensors at three wavelengths (green, red, and infrared) and collected Lead I electrocardiogram (ECG) data for ground truth.

Dev: The apparatus included an optical analog front-end connected to an optical module with LEDs at 530 nm, 655 nm, and 940 nm, along with a broadband photodiode.

Rosa: Training Signal Synthesis To train the model for signal recovery, a synthesized optimized PPG signal was created by aggregating individual PPG waves This process involved resampling and concatenating patterns aligned with the R-R intervals of the reference ECG signal using Pan-Tompkins for R peak detection.

Paper summary: Rosa: So, we're looking at this paper called Tri-Spectral PPG: Robust Reflective Photoplethysmography by Fusing Multiple Wavelengths for Cardiac Monitoring. Essentially, they are taking the different colors of light from a single PPG sensor—green, red, and infrared—and trying to combine them into one really good signal that you can use for heart rate monitoring. Dev Right, so the main idea is that these different channels handle different things; green is usually for heart rate and variability, while red and infrared are typically used for pulse oximetry. Rosa But this paper proposes a way to fuse all those signals into one single recovered PPG signal that you can then process further. Dev And they claim via this fusion, their learning-based method compensates for problems like motion or changes in the ambient light that mess up the readings on different wavelengths.

Taro: So, if we think about what a person might ask, would they say "so it fixes all artifacts"? I'd probably say something more specific, like how long this thing lasts outside the lab. Rosa Exactly. They tested this method on a real-world dataset, which included recordings from ten healthy adults over thirteen hours while they were doing everyday activities outside of a controlled lab setting <ref:2412.17549#pg2>. Dev And they used Lead I electrocardiogram data as the ground truth reference for what the actual heart rate should be during that time.

Taro: What’s interesting is that they used this real-world, motion-afflicted data to test their system, which is crucial because it shows how it handles things when the environment gets messy. Rosa They actually trained their learning model on a synthesized reference signal first, which they built by aggregating individual PPG waves and aligning them with the R-R intervals from the reference ECG signal using Pan-Tompkins for peak detection. Dev So, they created this template by averaging sections that correlated well with a leaning triangle wave while intentionally excluding any corrupted segments.

Taro: And then to train their network, they used this synthesized reference signal to generate a PPG reference signal by repeating the pattern based on the ECG R-peaks in windows of five minutes. Rosa The network itself is a U-Net architecture, which is adapted for this kind of signal recovery task. Dev They built this model with four up sampling and four down sampling layers, optimizing those layers from thirty-two to two hundred fifty-six kernels to handle the input data <ref:2412.17549#pg3,with four up sampling and four down sampling layers>.

Taro: And for the training process itself, they used the Adam optimizer with beta one equal to zero point nine and beta two equal to zero point nine nine nine, using a mini-batch size of eighty. Dev The learning rate starts at zero point zero zero one one and it drops by half when the validation loss doesn't improve for fifty consecutive epochs. Rosa So, the actual performance they report is that using this fused PPG signal, they were able to calculate participants' heart rates with a mean error of four point five beats per minute. Dev That means a twenty-three percent lower error than when they used just the green PPG signals, which measured five point nine beats per minute.

Rosa: So, what does that difference in numbers actually mean for someone just listening to the show? It suggests that by fusing all those colors of light together using this learning method, you get a much more accurate heart rate reading than if you just stuck with one channel. Dev It really shows the value of looking at the data from multiple sources simultaneously, rather than relying on just one input.

Taro: I wonder if this works well outside the lab for longer periods, since they only tracked them for thirteen hours during activities. Rosa That's a limitation they mentioned: they did have to limit their study to that duration and those specific real-world activities. Dev Also, it does note a limitation regarding the participants; they used a small sample size of ten adults, and they didn't collect data from participants with darker skin tones.

Taro: So, so this method seems quite robust when it comes to signal quality across different conditions and wavelengths, even if the study group wasn't perfectly diverse. Rosa That’s the main thing; it successfully retrieves a more robust fused signal that keeps the shape of the waveform while removing artifacts completely. Dev The authors are focused on showing that combining all those signals at once is better than switching between them or just averaging them.

Taro: So, what does this mean for future work? They suggest focusing on broadening their data collection to include more people and looking into how different PPG wavelengths and body locations might affect the results. Rosa It sounds like they are laying some groundwork here for more detailed pulse wave analysis down the road.

Conclusion: Rosa: So, this paper is about Tri-Spectral PPG, which basically takes three different colors of light from a sensor—green, red, and infrared—and fuses them into one single signal for heart monitoring.

Dev: Right, so it’s about taking all those separate signals and combining them using a learning method to get one better reading.

Rosa: The authors are trying to fix the problems that happen when things like motion or changing light mess up the readings on different colors.

Dev: They tested this on real-world data from ten healthy adults over thirteen hours outside of a lab, which is pretty solid for testing robustness.

Rosa: And they got some good results, showing that this fused signal lets them calculate heart rates with an error of four point five beats per minute.

Dev: That’s twenty-three percent lower than just using the green signal alone at five point nine beats per minute, which shows a real improvement in accuracy.

Rosa: It means they found a way to make the heart rate detection much more reliable when you're actually moving around and not in a perfectly controlled setting.

Dev: The paper suggests that combining all those signals at once is better than switching between them or just averaging them, which points toward a new approach for wearable sensors.

Rosa: It really shows how important it is to consider information from multiple sources simultaneously when you're trying to get accurate health data from a wearable device.

Dev: It opens up some interesting questions about what kind of data we can expect from these multi-channel systems in the future and how reliable they are in different real-world scenarios.

Rosa: But they did mention some limitations, like the study only included ten participants and didn't have data from people with darker skin tones.

Dev: Yeah, so while it’s a good step forward for signal fusion, there’s still work to be done on making these systems more inclusive and covering a wider range of conditions.

Rosa: So that leads us into how the authors plan to expand this research and where they think this technology can go next.

More episodes

← Home