Multi-Wavelength PPG Dataset Overview

• Multi-wavelength PPG is a dataset comprising synchronized spectral recordings across green, red, and IR wavelengths, enabling detailed cardiovascular analysis.
• It facilitates enhanced heart rate, HRV, and blood pressure estimation via advanced preprocessing, filtering, and signal fusion techniques.
• The dataset supports wearable health monitoring and clinical diagnostics by integrating multi-modal data such as motion and ECG for robust real-world assessments.

A Multi-Wavelength Photoplethysmography (PPG) dataset comprises synchronized time-resolved PPG recordings acquired at two or more distinct optical wavelengths, typically using reflective or transmissive illumination modalities. These datasets serve as a critical foundation for research on robust cardiovascular monitoring, sensor fusion, blood pressure estimation, physiological biometrics, and the development of machine learning models that exploit the different tissue/light interaction properties captured across spectral bands. In recent years, multi-wavelength PPG datasets have underpinned advances in wearable health monitoring, remote physiological measurement, and clinical diagnostics by enabling direct comparison, fusion, and quality assessment of complementary PPG channels under diverse, often real-world, acquisition conditions.

1. Principles and Rationale for Multi-Wavelength PPG

Multi-wavelength PPG leverages the fact that different tissue chromophores (e.g., oxygenated and deoxygenated hemoglobin, water) and tissue structures exhibit wavelength-dependent absorption and scattering. The typical spectral bands include green (~530 nm), red (~660 nm), and near-infrared (NIR, ~850–940 nm). Because optical penetration depth, pulsatility, and susceptibility to artifacts such as motion or ambient light vary by wavelength, collecting simultaneous PPG signals across bands yields richer, complementary physiological data than single-wavelength systems (Meier et al., 23 Dec 2024).

Multispectral acquisition underpins a variety of downstream applications:

• Enhanced heart rate (HR) and heart rate variability (HRV) estimation, especially in motion-rich or low SNR environments.
• Pulse oximetry, exploiting the distinct absorption of oxy- and deoxy-hemoglobin at red and IR bands.
• Robust blood pressure (BP) estimation, leveraging morphological differences accentuated at specific wavelengths.
• Improved artifact suppression by dynamically fusing less-corrupted channels.
• Quality assessment and channel selection for adaptive downstream analytics.

2. Architecture and Sensor Modalities in Multi-Wavelength Datasets

Dataset architectures vary, but consensus features include:

• Multiple synchronized channels per recording, typically with at least three (green/red/IR) or even four (e.g., 660, 730, 850, 940 nm) light sources and photodetectors.
• Control or recording of supporting modalities, such as accelerometer data (for motion context/fusion), ground truth ECG for HR/HRV validation, and blood pressure for calibration or regression targets.
• Diverse sensor placements (wrist, finger, head, ankle, sternum) in order to capture site-specific artifact profiles and physiological characteristics (Meier et al., 23 Dec 2024).

Many contemporary datasets use custom hardware with well-documented component specifications and acquisition protocols, often recording at 60–250 Hz per channel. Some include extensive meta-data on environmental conditions, subject physiological traits, and device settings.

The following table summarizes relevant multi-wavelength PPG dataset characteristics reported in recent literature:

Dataset/Study	Channels/Wavelengths	Acquisition Setting
ANT (Tang et al., 31 Mar 2025)	Red, Green, IR	Wrist-worn, 250 Hz, daily activities, 30 subjects
UTSA-PPG (Xu et al., 13 May 2025)	Red, IR, Green	Finger + wrist, 100 Hz, multi-scene, long-term
Tri-Spectral (Meier et al., 23 Dec 2024)	Green, Red, IR	Wearable, "in-the-wild", 13 hours, 10+ subjects
Four-band (Liang et al., 15 Sep 2025)	660, 730, 850, 940 nm	Fingertip, 180 subjects, 1 min recordings

3. Preprocessing, Quality Assessment, and Signal Fusion

Multi-wavelength datasets require tailored preprocessing:

• Channel-wise bandpass filtering (e.g., 0.5–10 Hz using Chebyshev II or Butterworth filters) to isolate pulsatile dynamics while suppressing drift/artifacts (Dias et al., 2023, Meier et al., 23 Dec 2024).
• Z-scoring or amplitude normalization across channels to account for device-specific or wavelength-specific amplitude scaling (Liang et al., 15 Sep 2025).
• Beat alignment or segmentation techniques, often using multi-scale peak and trough detection for intra-channel template extraction (Dias et al., 2023).
• Simultaneous processing of derivatives and upper/lower signal envelopes to enhance feature space or inform fusion models (Liang et al., 15 Sep 2025).

Quality assessment is critical since different channels can be unequally affected by motion, perfusion changes, or environmental conditions. Machine learning classifiers (e.g., gradient-boosted trees, random forests) have been applied to 27-feature vectors extracted from signal segments—using metrics such as DTW distance, local Pearson correlation, and morphological statistics—to perform binary or multi-class segment quality labeling (Dias et al., 2023). These pipelines can operate channel-wise or in a cross-sensor fusion setting, ensuring only high-confidence data is input to downstream algorithms.

Signal fusion strategies exploit spectral redundancy and complementarity:

• Simple rules: "Choose best channel based on real-time quality estimate."
• Weighted averaging: Assign nonlinear weights to channels by signal quality, e.g.,

$$s[t] = \frac{\sum_{i=1}^n ppg_i[t] \cdot \max(\delta, q_i[t])^6}{\sum_{j=1}^n \max(\delta, q_j[t])^6}$$

where $q_i[t]$ is the quality of channel $i$ at time $t$ and $\delta$ is a small constant (Meier et al., 23 Dec 2024).

• Deep learning: U-Net or Wave-U-Net style neural networks taking all channels as input, learning a nonlinear fusion mapping to recover a high-fidelity composite PPG, often trained with L1 loss computed against an ECG-derived pseudo-reference (Meier et al., 23 Dec 2024).

4. Applications: HR, HRV, BP Estimation, Biometrics, and Clinical Research

Multi-wavelength PPG datasets enable state-of-the-art advances across several domains:

• Robust HR/HRV estimation: Channel fusion methods achieve mean absolute heart rate errors as low as 2.4–4.5 bpm (e.g., by out-of-lab fusion of signals from multiple body sites or spectral bands) (Meier et al., 23 Dec 2024, Meier et al., 23 Dec 2024). HRV estimation is strengthened by longer, multimodal, and artifact-robust continuous records (Xu et al., 13 May 2025).
• Blood pressure estimation: Multi-channel systems facilitate regression of systolic and diastolic BP from PPG, with reported MAEs of 14.2 mmHg (SBP) and 6.4 mmHg (DBP) under strict subject-level validation by employing curriculum-adversarial deep learning strategies (Liang et al., 15 Sep 2025); multi-site or wavelength measurement addresses inter-individual and contextual variability (Haddad et al., 2020).
• Physiological biometrics: Recognition pipelines utilizing spectral and morphological features extracted from multi-wavelength PPG beats enable authentication AUCs up to 99.2% and EERs as low as 3.5% under challenging real-world wrist-worn conditions (Tang et al., 31 Mar 2025). Joint training on quality-assessment and identity-discriminative objectives further increases reliability when confronted with variable signal quality and motion.
• Remote and non-contact imaging: Imaging PPG (iPPG) datasets, often multi-wavelength in structure (e.g., DEAP (Unakafov, 2017)), support robust pulse and HR estimation from RGB video, enabling human monitoring scenarios where contact sensors are infeasible.

5. Evaluation Protocols, Datasets, and Generalizability

Robust evaluation of models trained on multi-wavelength PPG requires:

• Subject-level splits: Strict partitioning of subjects into train/val/test sets prevents data leakage and allows for assessment of real-world generalizability. Segment-level splits have been shown to inflate performance estimates (Liang et al., 15 Sep 2025).
• Multi-condition, multimodal datasets: Inclusion of various daily activities, resting, sleep, and office work conditions supports the development of artifact-robust algorithms (Xu et al., 13 May 2025).
• Reference signals: ECG and reference BP devices remain essential for ground-truth comparison and benchmarking HR, HRV, or BP estimation.
• Open-source benchmarks: Several groups have released their datasets and models (e.g., ANT, UTSA-PPG, MCD-rPPG), providing the community with reusable resources for comparison and reproducibility (Tang et al., 31 Mar 2025, Xu et al., 13 May 2025, Egorov et al., 25 Aug 2025).

Performance metrics typically include MAE for regression tasks, F1/AUC/EER for classification/biometric tasks, and BHS grading for BP regression, with stratification by BHS-5, BHS-10, BHS-15 error bands (Liang et al., 15 Sep 2025).

6. Limitations, Open Challenges, and Prospects

Several open issues persist:

• Artifact sensitivity and skin/tissue variability: Certain channels (e.g., IR vs. green) may fail in the presence of excessive subcutaneous fat, skin pigmentation, or anatomical structure, necessitating adaptive selection and fusion schemes (Amelard et al., 2016).
• Reference ground truth: Clinical reference standards (ECG, cuff BP) are required for generalization, but often lack temporal alignment or may themselves be subject to error, especially in ambulatory scenarios (Meier et al., 23 Dec 2024, Meier et al., 23 Dec 2024).
• Population and scenario diversity: Many datasets remain limited in subject count, age range, or acquisition scenario, which can limit transferability.
• Data utilization: Unfiltered, real-world field datasets (e.g., used for Pulse-PPG foundation model (Saha et al., 3 Feb 2025)) may contain significant unannotated noise, but also introduce the variability critical for robust, generalizable learning.
• Fusion model interpretability: Deep learning fusion models outperform rule-based approaches but often lack intuitive interpretability; attention-based and MTL approaches have been introduced to address this (Kasnesis et al., 2022, Tang et al., 31 Mar 2025).

Future directions include expanding open-access, annotated multi-wavelength datasets across broader populations and scenarios, development of transfer-learnable foundation models trained on uncurated field data, further integration of supporting sensor modalities (such as motion, temperature), and standardized benchmarking protocols to accelerate reproducibility and clinical translation.