WildPPG: A Real-World PPG Dataset of Long Continuous Recordings

summary

Video file (mp4)

The gist

The gist: This paper introduces WildPPG, a novel multimodal dataset comprising long, continuous PPG recordings from various body locations coupled with synchronized accelerometer, temperature, and

In short

WildPPG is a large dataset of continuous, synchronized PPG recordings from 16 participants during outdoor activities across various locations and conditions. The study introduced Temp-ResNet, a novel learning method that incorporates temperature data alongside raw PPG to create more robust heart rate estimations in real-world settings. This work enables better AI models for heart rate prediction in uncontrolled environments.

Key concepts

WildPPG Dataset
This is a large collection of continuous physiological data, including red PPG signals from multiple body locations, synchronized with accelerometer, temperature, and altitude readings. It was collected from 16 participants during long outdoor activities to provide real-world data for heart rate estimation.
Multimodal Data Fusion
This involves combining different types of sensor information—such as PPG signals from four different body sites (wrist, ankle, sternum, head) and environmental factors like temperature and altitude. Combining these diverse inputs helps create a more complete picture for accurate heart rate analysis.
Temp-ResNet
This is a novel deep learning model introduced to estimate heart rate. It specifically takes the temperature reading as an extra input feature alongside the raw PPG signal. The research showed that including temperature makes the model's heart rate predictions significantly more reliable and robust.
Real-World Robustness
The goal is to create AI models that perform well outside of controlled lab settings. By training on data collected during hiking, climbing, and varying temperatures/altitudes, the resulting models can better handle the noise and variability found in everyday outdoor activities.

Terminology used across episodes

This episode discusses

The paper

WildPPG: A Real-World PPG Dataset of Long Continuous Recordings · Read on arXiv

Manuel Meier, Berken Utku Demirel, Christian Holz

Department of Computer Science, ETH Zurich

Reflective photoplethysmography (PPG) has become the default sensing technique in wearable devices to monitor cardiac activity via a person's heart rate (HR). However, PPG-based HR estimates can be substantially impacted by factors such as the wearer's activities, sensor placement and resulting motion artifacts, as well as environmental characteristics such as temperature and ambient light. These and other factors can significantly impact and decrease HR prediction reliability. In this paper, we show that state-of-the-art HR estimation methods struggle when processing representative data from everyday activities in outdoor environments, likely because they rely on existing datasets that captured controlled conditions. We introduce a novel multimodal dataset and benchmark results for continuous PPG recordings during outdoor activities from 16 participants over 13.5 hours, captured from four wearable sensors, each worn at a different location on the body, totaling 216,hours. Our recordings include accelerometer, temperature, and altitude data, as well as a synchronized Lead I-based electrocardiogram for ground-truth HR references. Participants completed a round trip from Zurich to Jungfraujoch, a tall mountain in Switzerland over the course of one day. The trip included outdoor and indoor activities such as walking, hiking, stair climbing, eating, drinking, and resting at various temperatures and altitudes (up to 3,571,m above sea level) as well as using cars, trains, cable cars, and lifts for transport -- all of which impacted participants' physiological dynamics. We also present a novel method that estimates HR values more robustly in such real-world scenarios than existing baselines.

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "WildPPG: A Real-World PPG Dataset of Long Continuous Recordings".

Jane: Here's the gist of the paper. The gist: This paper introduces WildPPG, a novel multimodal dataset comprising long, continuous PPG recordings from various body locations coupled with synchronized accelerometer, temperature, and altitude data to enable more robust heart rate estimation in real-world outdoor activities.

Tom: Dataset and Data Collection WildPPG comprises "continuous and synchronized recordings from 16 participants with an average duration of 13.5 hours during a round-trip to a tall mountain in Europe" (Page 2).

Jane: The data collection study covered "everyday outdoor activities outside controlled (laboratory) conditions" (Page 2).

Tom: The recordings included signals from four wearable sensors worn at different locations, totaling "216 hours of synchronized multi-modal recordings that can enable a wide variety of future analyses" (Page 2).

Title and authors: Tom: So, looking at the title WildPPG: A Real-World PPG Dataset of Long Continuous Recordings, it really tells you what this paper is aiming for. It's about creating a dataset that reflects actual life, not just controlled experiments.

Jane: Right. They’re showing that state-of-the-art heart rate estimation methods often struggle when they only use data from lab conditions because they aren't trained to handle the real noise and variability of outdoor activities.

Lu: The paper sets up this dataset specifically to address that gap, providing something public and large enough for researchers to build better algorithms that can actually work in the wild.

Meng: For us at scale, it’s important because it shows what kind of data we need to move from theoretical accuracy on a benchmark to actual reliability when deployed on a person's wearable device.

Jane: They are essentially providing the necessary fuel for developing AI that can handle things like walking, hiking, and even changing temperatures and altitudes while estimating heart rate reliably.

Tom: It’s about showing how incorporating these environmental factors and multiple sensing sites helps improve the robustness of HR prediction compared to older methods.

Lu: The focus on continuous recordings over thirteen point five hours is significant because it allows for analyzing longer time series, which is crucial for training better deep learning models.

Meng: If we can train a model on this kind of long-form, multi-modal data, we might actually see better performance when the system encounters unpredictable motion artifacts or temperature shifts during deployment.

The paper's summary: Tom: So, summarizing the WildPPG paper, they introduce this multimodal dataset featuring continuous PPG recordings from sixteen participants over thirteen point five hours. It includes signals from four wearable sensors—wrist, ankle, sternum, and head—along with synchronized accelerometer data, temperature readings at those same four locations and altitude sensor information.

Jane: They are using the sternum device to record a Lead-I electrocardiogram as the ground truth reference for heart rate estimates.

Lu: The main point they make is that existing methods struggle because they rely on datasets captured in controlled lab environments, and WildPPG is designed to combat that by including data from every day outdoor activities outside of those controlled conditions.

Meng: They are benchmarking several baseline methods, including heuristic ones like MSPTD and supervised deep learning approaches like 1D ResNet, to show where they fall before introducing their own idea <ref:2412.17540#pg2>.

Tom: And their main contribution is the proposed Temp-ResNet method. This is a learning-based approach that takes temperature as an input alongside the raw PPG signal, and they showed that this temperature awareness helps produce more robust heart rate estimates.

Jane: So, it’s not just about adding more data; it’s specifically about using those environmental readings to make the estimation method itself smarter.

Lu: They found that incorporating additional measurement sites for HR prediction generally reduces the error, even if it increases the number of model parameters needed for training.

Meng: That makes sense from a complexity standpoint, but if that reduction in error translates to better performance in a real-world scenario where things are messy, then it’s worth the extra compute.

The paper's improvements: Tom: Now let’s talk about the specific improvements they highlighted. They pointed out several things, starting with how incorporating additional measurement sites for HR prediction generally reduces the error in estimation.

Jane: That makes sense when you think about noise; having more ways to measure something usually helps clean up the signal, even if it means dealing with more data streams.

Lu: They also noted that while they have four locations, the wrist measurement site has the lowest signal quality compared to the other sites because it's easily affected by motion artifacts and cold.

Meng: That’s a crucial caveat for us when thinking about deployment; we need to know which sensor is most reliable under real-world stress, not just theoretical accuracy.

Tom: And that leads into their novel approach, Temp-ResNet. They showed that this model, which incorporates a branch that inputs the temperature value of the segment into the overall architecture, achieved a lower Mean Absolute Error compared to the standard 1D ResNet when temperature was included in the input <ref:2412.17540#pg2>.

Jane: So they are demonstrating that this specific way of incorporating environmental data makes a measurable difference in accuracy over other methods they tested.

Lu: It supports their ambition that learning pipelines benefit from longer time series during training, additional modalities, and real-world data captured "in the wild".

Conclusion: Tom: To wrap things up with WildPPG, the authors present this as the largest available dataset of its kind. They argue it enables learning pipelines that benefit from longer time series during training, additional modalities, and real-world data “in the wild”.

Jane: It’s a resource that allows researchers to move past controlled conditions and test methods on more complex scenarios where things are dynamic.

Lu: The implication is that future work should focus on tailoring these learning pipelines specifically to handle those less-than-optimal, real-world conditions we discussed.

Meng: If the model can generalize well from this kind of data, it means we might be able to deploy heart rate estimation systems that are much more reliable in unpredictable outdoor settings than what current wrist-worn devices allow.

Lalam: I’m confident that the findings regarding the Temp-ResNet method provide a solid starting point for how we approach better HR estimation in real environments.

More episodes

← Home