💻 computer science

Pixel Watch: Robust Heart Rate Sensing from Multipath PPG and On-Device Deep Learning Trained on 10,000 hours of Free-Living and Fitness Data

The Pixel Watch 2 achieves significantly improved heart rate accuracy during motion-heavy activities by leveraging a 10-channel multipath PPG sensor and an on-device deep learning model trained on 10,000 hours of diverse data, resulting in substantially tighter error margins compared to previous devices.

Original authors: Daniel Roggen, Megan Walker, Yojan Patel, Shyam Tailor, Dimitris Spathis, Matt Wimmer, Brennan Garrett, Dan Howe, Abhinuv Pitale, Hamed Vavadi, Tien Le, Steve Diamond, Oleksiy Vyalov, Vik Sharma, Pete

Published 2026-06-23✓ Author reviewed ⓘ

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Daniel Roggen, Megan Walker, Yojan Patel, Shyam Tailor, Dimitris Spathis, Matt Wimmer, Brennan Garrett, Dan Howe, Abhinuv Pitale, Hamed Vavadi, Tien Le, Steve Diamond, Oleksiy Vyalov, Vik Sharma, Pete Richards, Tracy Giest, Erika Siegel, Tuan Phan, Sam Mravca, Derrick Vickers, Benjamin Stone, Katarina Vukosavljevic, Justin Phillips, YongSuk Cho, Stefanie Hollidge, Antony Siahaan, Soren Brage, Shwetak Patel, Robert Harle

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 by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your smartwatch is like a tiny, high-tech detective trying to listen to your heart beat through your wrist. Usually, this is a noisy job. When you're sitting still, the detective hears a clear "thump-thump." But when you start running, lifting weights, or even just waving your arms, the movement creates a lot of static and interference, making it hard for the watch to tell the difference between your heartbeat and the noise of your motion.

This paper describes how Google built a new smartwatch, the Pixel Watch 2, that solves this problem using two main tricks: more ears and a super-smart brain.

1. The "More Ears" Trick (Multipath PPG)

Older watches usually had just one "ear" (a single light sensor) trying to listen to your heart. If that ear got blocked by sweat or moved slightly, the signal got garbled.

The Pixel Watch 2 is different. It has 10 different "ears" (light sensors and LEDs) arranged in various patterns on the back of the watch. Think of it like a choir instead of a solo singer. If one voice gets drowned out by the noise of your arm moving, the other nine voices are still singing clearly. By listening to all these different angles at once, the watch gets a much richer, more complete picture of what's happening under your skin.

2. The "Super-Smart Brain" (Deep Learning)

Having 10 ears is great, but it creates a massive amount of data. A simple calculator (the kind used in older watches) would get overwhelmed trying to sort through all that noise.

So, Google gave the watch a deep learning brain. This is a type of artificial intelligence that acts like a seasoned music producer.

The Training: Before the watch ever hit the market, this "producer" spent 10,000 hours listening to recordings from 962 different people. These people were doing everything from sleeping to sprinting, running, and rowing. The AI learned to recognize the specific "sound" of a heartbeat amidst the chaos of exercise.
The Result: Instead of using rigid, pre-written rules to filter out noise, the AI learned to intuitively separate the heartbeat from the motion. It's the difference between a robot trying to follow a manual and a jazz musician improvising to find the rhythm.

3. The Big Test

To prove this new system works, the researchers didn't just test it in a quiet lab. They put it through two major trials:

The Gym Test: They had 229 people wear the watch while doing intense, guided exercises like running on a treadmill, using a rowing machine, and doing high-intensity interval training (HIIT). They compared the watch's numbers against a gold-standard chest strap (which is very accurate).
The Real Life Test: They gave the watch to 27 older adults to wear for two days while they went about their normal lives—cooking, walking, working, and relaxing.

The Verdict

The results were a huge improvement over previous Google watches and other competitors.

Older watches often had a "margin of error" that was quite wide. For example, if your heart rate was 100, an older watch might guess anywhere between 75 and 125 during a hard run. That's a big guess!
The Pixel Watch 2 narrowed that guess significantly. During exercise, it was usually within about 9 beats of the true heart rate. During normal daily life, it was even better, usually within 7 beats.

Why This Matters (According to the Paper)

The authors emphasize that the magic wasn't just in the hardware (the 10 sensors) or the software (the AI) alone. It was the combination of both, trained on a massive amount of real-world data.

They found that if you use the old "calculator" methods with the new 10-sensor hardware, it doesn't work well. But if you use the new "AI brain" with the old single-sensor hardware, it's okay, but not great. You need both the extra sensors and the deep learning brain to get the best results.

In short, the Pixel Watch 2 is a smartwatch that learned to listen to your heart by practicing with thousands of people doing thousands of different activities, allowing it to ignore the noise of your movement and focus on the beat.

Technical Summary: Pixel Watch: Robust Heart Rate Sensing from Multipath PPG and On-Device Deep Learning

Problem Statement
Accurate continuous heart rate (HR) monitoring via wrist-worn photoplethysmography (PPG) remains challenging during dynamic activities due to motion artifacts. While previous consumer wearable validation studies have often been limited by small sample sizes and restricted activity protocols, there is a critical need for robust sensing solutions that perform reliably across diverse, real-world conditions. The authors identify that traditional signal processing approaches, including classical Digital Signal Processing (DSP) and Kalman filtering, have struggled to fully leverage multipath optical hardware to separate pulse signals from motion noise effectively.

Methodology
The paper presents the Pixel Watch 2 (PW2), the first Google smartwatch to combine multipath PPG hardware with an on-device deep learning model.

Hardware Architecture: The device utilizes the TI AFE4500 analog front-end to capture 10 distinct optical channels. These channels represent unique combinations of green LEDs (525nm), photodetectors (PDs), and transimpedance amplifier gain settings. The system employs a closed-loop controller to maximize amplifier gain within the ADC's dynamic range, mitigating ambient light and skin contact fluctuations. Signals are bandpass filtered (0.5–12 Hz) before ADC sampling.
Deep Learning Model: The core inference engine is a 15-layer temporally dilated convolutional neural network (CNN) with approximately 300,000 parameters, designed to run on-device (Qualcomm Snapdragon W5).
- Input: The model processes 20 optical inputs (10 channels, each detrended and normalized, plus recursive least squares adaptive filtering) and 3 motion inputs (accelerometer data). All streams are downsampled to 25 Hz.
- Architecture: The network begins with a 9-layer temporal convolutional feature extractor (filters increasing from 40 to 56) to capture pulse wave morphology. This is followed by 6 temporally dilated convolutional blocks with exponentially increasing dilation rates ( $d=2^1$ to $2^6$ ) to integrate long-term temporal context without the computational overhead of recurrent or attention mechanisms.
- Output: The system frames HR inference as a classification problem, outputting a probability distribution over 94 classes (2 BPM buckets) at 0.78 Hz, which is then interpolated to a 1 BPM resolution at 1 Hz. The model provides confidence metrics via the output probability distribution.
Training Data: The model was trained on a massive dataset comprising 9,319 hours of data from 962 participants. Data was sourced from internal studies covering human activity recognition, sleep, acceptance testing, and unconstrained free-living collection. Ground truth was established using a Polar H10 chest strap. The dataset was balanced across predefined heart rate brackets.

Key Contributions

System Design: Introduction of a multipath PPG system coupled with a specialized, on-device deep learning model specifically optimized for motion-heavy environments.
Data Scale: The utilization of a training corpus (962 participants, ~10,000 hours) significantly larger than typical validation studies, enabling the model to learn complex motion-artifact patterns.
Comprehensive Validation: The study evaluates the device across two independent, unseen datasets:
- An Exercise Dataset (229 participants, 248 hours) involving prescribed activities (still, walking, running, cycling, rowing, HIIT).
- A Free-Living Dataset (27 participants, 1,000+ hours) involving normal daily activities over 48 hours, using an Actiheart 4 ECG sensor as the criterion.
Design Lessons: The authors demonstrate that deep learning allows for the systematic exploration of optical path combinations without the need for ad-hoc signal processing transforms or manual feature engineering.

Results
The PW2 demonstrated substantially improved accuracy compared to previous Google devices and other commercial wearables:

Exercise Dataset: The system achieved a Mean Absolute Error (MAE) of 2.18 BPM (1.71% MAPE). The 95% Limits of Agreement (LoA) were -10.34 to 8.66 BPM. Performance varied by activity, with indoor cycling showing the tightest LoA (-6.21 to 4.67 BPM) and running showing the widest (-12.65 to 10.21 BPM).
Free-Living Dataset: The system achieved an MAE of 1.84 BPM (2.53% MAPE) with 95% LoA of -6.57 to 7.48 BPM.
Comparative Performance: The PW2's LoA were significantly tighter than previous Google products. For context, the Fitbit Charge 4 (single-path, classical DSP) showed LoA of -26.75 to 23.43 BPM in similar activity comparisons, and the Fitbit Sense (multipath, classical DSP) showed LoA of -13.44 to 10.91 BPM.
Efficiency: The inference time is 40.11 ms per step, consuming approximately 28 minutes of battery life over a 24-hour period.

Significance and Claims
The paper concludes that the integration of large-scale deep learning was instrumental in fully leveraging multipath PPG hardware, a capability that classical signal processing approaches failed to achieve. The authors claim that this approach enabled a "step-change in quality" compared to prior devices.

Key takeaways emphasized by the authors include:

Data Volume: Large-scale, diverse training data is paramount for model performance.
Design Efficiency: Deep learning reduced engineering effort by automating the discovery of optimal feature representations from raw multipath data, allowing for the systematic exploration of optical configurations.
User Perception: The authors note that users react more strongly to outliers (large deviations) than to bias, a factor that guided their "perceived error" metrics during development.
Future Trajectory: The success of the PW2 architecture has paved the way for subsequent Pixel Watch generations, which build incrementally on this foundation with refined electronics, optics, and larger models.

1. The "More Ears" Trick (Multipath PPG)

2. The "Super-Smart Brain" (Deep Learning)

3. The Big Test

The Verdict

Why This Matters (According to the Paper)

More like this