Research map / Probabilistic and scientific modeling / Dynamical systems
WildPPG: A Real-World PPG Dataset of Long Continuous Recordings
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces WildPPG, a 16-participant, 216-hour multimodal PPG/ECG dataset from free-living outdoor and indoor activities on a Zurich-to-Jungfraujoch trip, plus a temperature-aware ResNet (Temp-ResNet) for heart rate estimation.
Problem
Heart rate estimation from wearable reflective PPG is degraded by motion, sensor placement, temperature and ambient light, but existing PPG datasets were recorded mostly in controlled lab conditions and are small.
Approach
A large in-the-wild multimodal dataset with four body-site wearables (PPG at three wavelengths, accelerometer, temperature, altitude) and ECG reference, plus a 1D ResNet extended with a temperature branch that improves HR estimation robustness.
Key results
- Temp-ResNet reaches MAE 8.37 bpm on WildPPG (wrist green PPG) versus 8.62 for 1D ResNet, and MAE 4.40 on DaLiA versus 4.47 for 1D ResNet.
- Heuristic methods such as FFT and autocorrelation perform poorly on WildPPG (MAE 17.62 and 28.43 bpm), and unsupervised UPD reaches MAE 23.12 bpm.
- Cross-dataset evaluation yields much higher errors on WildPPG (1D ResNet MAE 16.32 bpm when trained on other datasets), indicating a distribution shift.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add temperature input to 1D ResNet | Temp-ResNet on WildPPG · WildPPG | MAE: 8.62±0.06 -> 8.37±0.02 | Helped |
| Add temperature input to 1D ResNet | Temp-ResNet on WildPPG · WildPPG | ρ: 42.41±0.24 -> 45.50±0.14 | Helped |
| Add temperature input to 1D ResNet | Temp-ResNet on DaLiA · DaLiA | MAE: 4.47±0.03 -> 4.40±0.02 | Helped |
| Add temperature input to 1D ResNet | Temp-ResNet on DaLiA · DaLiA | ρ: 85.87±0.25 -> 85.90±0.17 | No clear effect |
| Add accelerometer input (ablation) | 1D ResNet on WildPPG · WildPPG | MAE: 8.62±0.06 -> 8.22±0.09 | Helped |
| Add accelerometer input (ablation) | Temp-ResNet on WildPPG · WildPPG | MAE: 8.37±0.02 -> 8.13±0.03 | Helped |
| Add chest PPG to wrist PPG (ablation) | Temp-ResNet on WildPPG · WildPPG | MAE: 8.37±0.02 -> 8.28±0.03 | Helped |
| Use PPG from all body locations instead of wrist only (ablation) | Temp-ResNet on WildPPG · WildPPG | MAE: 8.37±0.02 -> 8.33±0.05 | No clear effect |
| Use red instead of green PPG (ablation) | Temp-ResNet on WildPPG · WildPPG | MAE: 8.37±0.02 -> 9.97±0.10 | Hurt |
| Use infrared instead of green PPG (ablation) | Temp-ResNet on WildPPG · WildPPG | MAE: 8.37±0.02 -> 8.72±0.10 | Hurt |
| Use all wavelengths instead of green PPG only (ablation) | Temp-ResNet on WildPPG · WildPPG | MAE: 8.37±0.02 -> 8.35±0.04 | No clear effect |
Surprising findings
- Cross-dataset models trained on SPC12 and DaLiA perform much worse on WildPPG than in-dataset models, suggesting existing datasets do not transfer to real-world conditions.
- Device temperature shows an inverse correlation with PPG signal quality, and wrist PPG has the lowest signal quality among body sites.
Limitations
- Only 16 participants, all with Fitzpatrick skin tones 1-3, so darker skin tones are not represented.
- About 2.7% of windows have noisy ECG ground truth, mainly from three participants.
- Data collection follows a single mountain-trip protocol, so activity coverage is limited to that trip.
Experimental setup
Datasets: WildPPG IEEE SPC 2015 DaLiA BIDMC
Benchmarks: WildPPG IEEE SPC12 DaLiA BIDMC
Metrics: MAE RMSE Pearson correlation
Baselines: FFT Autocorrelation HeartPy MSPTD PWD qppg SimCLR NNCLR BYOL TS-TCC TS2Vec VICReg
Builds on: 1D ResNet Unsupervised Periodicity Detection TS2Vec BeliefPPG
Closest papers by meaning
- PaPaGei: Open Foundation Models for Optical Physiological Signals (ICLR 2025 · similarity 0.85)
- Large-scale Training of Foundation Models for Wearable Biosignals (ICLR 2024 · similarity 0.84)
- Remote Photoplethysmography in Real-World and Extreme Lighting Scenarios (CVPR 2025 · similarity 0.84)
- Beyond Sensor Data: Foundation Models of Behavioral Data from Wearables Improve Health Predictions (ICML 2025 · similarity 0.84)
- FacePhys: State of the Heart Learning (NeurIPS 2026 · similarity 0.84)
- RhythmMamba: Fast, Lightweight, and Accurate Remote Physiological Measurement (AAAI 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.