💻 computer science

Multi-site PPG: An In-the-Wild Physiological Dataset from Emerging Multi-site Wearables

This paper introduces "Multi-site PPG," a comprehensive in-the-wild dataset comprising over 350 hours of physiological data collected from four emerging wearable form factors (earring, ring, watch, and necklace) alongside a reference ECG, which is used to benchmark heart rate estimation methods and reveal significant performance variations across different body sites.

Original authors: Jiayi Shao (Shirley), Jiaying Ye (Shirley), Shengyao Liu (Shirley), Zachary Englhardt (Shirley), Girish Narayanswamy (Shirley), Vikram Iyer (Shirley), Qiuyue (Shirley), Xue

Published 2026-05-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jiayi Shao (Shirley), Jiaying Ye (Shirley), Shengyao Liu (Shirley), Zachary Englhardt (Shirley), Girish Narayanswamy (Shirley), Vikram Iyer (Shirley), Qiuyue (Shirley), Xue

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you want to measure your heartbeat, but instead of just checking your wrist (like most smartwatches do), you decide to wear sensors on your ear, your finger, your neck, and your wrist all at the same time. That is exactly what this paper does.

The researchers created a massive new "library" of heart data called Multi-site PPG. Here is a simple breakdown of what they did and what they found, using some everyday analogies.

1. The Problem: The "Wrist" Bottleneck

Think of your wrist like a busy, noisy highway. It's where most people wear watches, but because your wrist moves so much (typing, waving, walking), the signal for your heartbeat gets jumbled up with "traffic noise." Also, the blood vessels under your wrist skin are deep, making the signal faint, like trying to hear a whisper through a thick wall.

The researchers wanted to see if wearing sensors in quieter, "less crowded" places on the body could hear the heartbeat more clearly.

2. The Experiment: The "Four-Headed" Wearable

To test this, they built four custom, unobtrusive gadgets that look like normal jewelry:

A Smart Earring: Worn on the earlobe.
A Smart Ring: Worn on the finger.
A Smart Watch: Worn on the wrist (the control group).
A Smart Necklace: Worn around the neck.

They asked 20 people to wear all four devices at once while they went about their normal day—studying, walking, exercising, and chatting. To make sure the data was accurate, everyone also wore a medical-grade chest strap (like a Polar H10) that acted as the "truth-teller" or the gold standard.

In total, they collected over 350 hours of raw data. It's like recording a continuous movie of heartbeats from four different angles for days on end.

3. The Results: Who Heard the Heartbeat Best?

The researchers tested various computer algorithms (from simple math tricks to advanced AI) to see how well they could guess the heart rate from the noisy data. They compared the guesses against the "truth-teller" chest strap.

Here is the ranking, from best to worst:

The Earring (The VIP): This was the clear winner. It achieved an error rate of just 2.3 beats per minute.
- Analogy: Imagine trying to hear a song in a quiet room. The ear is close to the head, moves very little, and has thin skin. It's like having a high-quality microphone right next to the speaker.
The Ring (The Runner-Up): This came in second with an error of 5.1 beats per minute.
- Analogy: The finger is also great because it's full of blood vessels (like a busy river), but it moves a bit more than the ear when you use your hands.
The Watch (The Struggler): The traditional watch had an error of 8.4 beats per minute.
- Analogy: This is the noisy highway again. The wrist moves too much, and the signal is weaker.
The Necklace (The Unstable): This had the highest error at 8.7 beats per minute.
- Analogy: A necklace dangles and swings. The sensor often loses contact with the skin, like a radio losing signal when you walk around.

4. The "Motion" Factor

The paper also looked at how movement affects the data.

Quiet Zones: The ear and neck stayed relatively still, so the data stayed clean even when people were moving around.
Active Zones: The finger and wrist moved a lot. The more the person moved, the harder it was for the computer to guess the heart rate correctly.
The Fix: They found that adding motion sensors (accelerometers) to the ring and watch helped the computer "subtract" the movement noise, making the heart rate guess slightly better. However, for the ear, which barely moves, adding motion sensors didn't help much.

5. Mixing the Signals (Fusion)

The researchers tried combining the data from multiple devices to see if it would make the prediction even better.

Two is better than one: Combining the Earring and Ring data made the prediction slightly more accurate than using just the Earring alone.
Four is not always better: Surprisingly, combining all four devices didn't make it much better than just using the best two. It's like having four people guess a number; if three of them are guessing wildly off, their "average" might actually pull the good guess in the wrong direction.

Summary

This paper is essentially a public data set (a giant library of heartbeats) that proves: Where you wear your sensor matters just as much as the sensor itself.

The Big Takeaway: If you want the most accurate heart rate without a chest strap, wearing a sensor on your ear or finger is significantly better than wearing it on your wrist.
The Contribution: They released this data and the code for free so other scientists can build better health apps and devices that work in the real world, not just in a quiet lab.

Note: The paper focuses strictly on collecting this data and testing how well current algorithms work on it. It does not claim these devices are ready for medical diagnosis or clinical use yet.

Technical Summary: Multi-site PPG: An In-the-Wild Physiological Dataset from Emerging Multi-site Wearables

Problem Statement

Photoplethysmography (PPG) is a dominant sensing modality for heart rate and physiological monitoring in wearables. However, public "in-the-wild" PPG datasets remain largely constrained to wrist-centric smartwatches or short, controlled laboratory sessions. Wrist-based sensing faces inherent challenges, including motion artifacts and degraded optical signal quality due to deeper blood vessels. While emerging form factors like smart rings, earrings, and necklaces offer potential advantages (e.g., higher perfusion in fingers/earlobes, relative stability), there is a scarcity of large-scale, multi-site datasets collected during natural daily activities. Existing datasets either focus on single devices, lack practical everyday form factors (e.g., using non-standard placements like sternum or ankle), or are limited to controlled environments, hindering research into robust, generalized physiological sensing across diverse wearable types.

Methodology

Dataset Collection and Hardware

The authors present Multi-site PPG, a dataset collected from 20 participants (19 female, 1 male; ages 20–51) over multiple days of daytime activities in urban environments. The study utilized four custom-developed, unobtrusive wearables:

Smart Earring: Worn on a pierced earlobe.
Smart Ring: Worn on the finger.
Smartwatch: Worn on the wrist.
Smart Necklace: Worn as a pendant.

All devices share a unified hardware design based on an nRF52832 microcontroller, integrating a MAX30101 PPG sensor (green and infrared channels), LIS2DW12 accelerometer, and HDC2010 temperature sensor. They stream data wirelessly to a smartphone app. For ground truth, participants wore a commercial Polar H10 chest strap recording ECG at 130 Hz.

Data Processing

The dataset is released in two formats:

Raw Dataset: Over 350 hours of timestamped, continuous (though potentially discontinuous due to Bluetooth drops) data per device, including PPG (green/IR), 3-axis acceleration, and temperature.
Windowed Dataset: Preprocessed, modeling-ready data consisting of 8-second sliding windows (1-second stride). This dataset excludes segments with poor ECG quality (e.g., loose strap contact, heart rates outside 40–180 bpm) and aligns wearable data with Polar ECG ground truth. It contains 229.66 to 290.23 hours of data per device, with 190.95 hours where all four devices were simultaneously active.

Benchmarking Protocol

The authors evaluated heart rate estimation performance using three classes of methods against ECG-derived ground truth:

Heuristic Algorithms: Six traditional signal processing methods (e.g., NeuroKit, HeartPy, FFT).
Supervised Models: Five neural network architectures (1D ResNet, DCL, FCN, CNN-LSTM, Transformer).
Self-Supervised Models: Two contrastive learning approaches (SimCLR, BYOL).

Evaluation was conducted using a Leave-One-Subject-Out (LOSO) cross-validation protocol across 20 folds, measuring Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Pearson correlation ( $\rho$ ).

Key Contributions

First Large-Scale In-the-Wild Multi-Site PPG Dataset: The paper releases the first public, large-scale PPG dataset collected simultaneously from four distinct, practical wearable form factors (earring, ring, watch, necklace) during natural daily activities. It is the first such dataset for smart rings, earrings, and necklaces.
Body-Location Benchmarking: The study establishes baseline performance for heart rate estimation across different body sites, revealing significant performance disparities.
Multi-Site and Multimodal Analysis: The authors provide initial evaluations of multi-site sensor fusion and PPG-accelerometer fusion, demonstrating the dataset's utility for studying complementary sensing strategies.

Results

Performance Across Body Sites

Significant differences in heart rate estimation accuracy were observed across the four locations, with the following best-performing MAE results (using the DCL supervised model):

Earring: 2.30 bpm (Best performance, comparable to high-quality lab settings).
Ring: 5.13 bpm (Comparable to the 28-hour $\tau$ -Ring dataset).
Watch: 8.37 bpm (Consistent with WildPPG results).
Necklace: 8.68 bpm.

The earring and ring consistently outperformed the watch and necklace, attributed to stronger perfusion, better optical coupling, and lower motion artifacts at those sites.

Motion Effects

Analysis showed that heart rate estimation error (MAE) generally increases with motion power across all devices. However, the earring and necklace experienced lower motion exposure (91% of windows in low-motion bins) compared to the ring and watch. The ring maintained better accuracy than the watch even at higher motion levels, likely due to superior signal quality at the finger.

Fusion Strategies

Multi-Site Fusion: Combining two devices (e.g., Earring + Ring) consistently improved MAE over single-device baselines (e.g., reducing Earring MAE from 1.84 to 1.82 bpm). However, fusing all four devices did not yield further gains over the best two-device combinations, suggesting that noisier sites (watch, necklace) may introduce variability when combined with high-quality signals.
PPG-Accelerometer Fusion: Adding accelerometer data (z-axis) provided consistent, albeit modest, improvements for the ring and watch (e.g., Watch MAE reduced by ~0.3 bpm). The benefit was limited or mixed for the earring and necklace, indicating that motion context is most valuable for sites prone to motion artifacts.

Significance and Claims

The paper claims that Multi-site PPG fills a critical gap in the availability of physiological data for emerging wearable form factors. By providing a dataset that captures the variability of daily life across different body locations, it enables:

Robust Sensing Research: The development of models that generalize across form factors and body sites, moving beyond the wrist-centric paradigm.
Signal Quality Analysis: The ability to characterize how physiological signals and motion artifacts differ by location (e.g., ear vs. wrist).
Fusion Studies: A foundation for investigating how to optimally fuse data from multiple sensors and modalities to improve reliability.

The authors emphasize that while the dataset supports future research, it is released with specific limitations: a high proportion of female participants (due to earring requirements), a lack of Fitzpatrick skin type VI, and ground truth labels derived via algorithmic processing (Pan-Tompkins) rather than manual annotation. The dataset is intended for non-commercial research to advance robust, unobtrusive physiological monitoring.