Independent Component Analysis for Separating PPG Signal Components

Multi-Channel PPG Design for ICA

The effectiveness of ICA for PPG depends critically on the multi-channel signal acquisition design. Several strategies provide the required observation diversity.

Multi-Wavelength PPG

Using LEDs at different wavelengths (e.g., green 525 nm, red 660 nm, infrared 940 nm) creates multiple PPG channels from a single sensor location. Different wavelengths penetrate to different tissue depths and interact differently with blood and tissue components, creating natural diversity in the mixing coefficients.

The cardiac component has wavelength-dependent amplitude determined by hemoglobin absorption spectra. The motion artifact component has wavelength-dependent amplitude determined by tissue scattering and mechanical coupling, which differ from the hemoglobin absorption ratios. This wavelength-dependent differential response creates the mixing diversity that ICA exploits for separation.

Dual-wavelength recordings give ICA two differently weighted mixtures of the cardiac and motion sources, which is what separation needs. For more on wavelength selection, see our green vs. red vs. infrared PPG guide.

Multi-Location PPG

Placing multiple PPG sensors at different body locations (e.g., wrist dorsal + wrist ventral, or wrist + finger) creates channels with different mixing coefficients because the optical path and mechanical coupling differ at each location. The cardiac pulse wave arrives at different locations with different delays (pulse transit time), providing additional temporal diversity.

This approach is less practical for consumer wearables due to the requirement for multiple sensor locations but is used in clinical pulse oximetry where finger and ear sensors provide redundant channels.

PPG + Accelerometer Channels

The most common practical configuration combines one or more PPG channels with 3-axis accelerometer data, producing 4-6 observation channels. The accelerometer channels carry motion information but no cardiac information (assuming the accelerometer does not detect the ballistocardiographic impulse, which is negligible at the wrist). This creates the mixing diversity needed for ICA.

However, this configuration violates a subtle ICA assumption: the accelerometer channels contain zero cardiac component, making the mixing matrix partially known. Constrained ICA (cICA) algorithms exploit this partial knowledge for improved separation (Lu and Rajapakse, 2005, doi:10.1109/TNN.2004.836795).

Automatic Component Identification

After ICA separation, the permutation ambiguity problem requires identifying which recovered component corresponds to the cardiac signal. Several automated approaches are used:

Spectral Entropy

The cardiac signal has a highly structured spectrum with a narrow fundamental peak and harmonics, resulting in low spectral entropy. Motion artifacts and noise have broader, less structured spectra with higher entropy. Computing the spectral entropy of each ICA component and selecting the one with the lowest value is therefore a common way to identify the cardiac component.

Spectral entropy is computed as:

where $P_k$ is the normalized power spectral density at frequency bin $k$ within the cardiac band (0.5-4 Hz).

Kurtosis-Based Selection

PPG cardiac signals have high kurtosis (typically > 3) due to the peaked nature of the systolic pulse. ICA components can be ranked by kurtosis, with the highest-kurtosis component selected as the cardiac signal. This approach is computationally simple but less reliable than spectral entropy during vigorous motion when motion artifacts can also have high kurtosis.

Template Matching

A canonical PPG pulse template (average systolic-diastolic waveform shape) is correlated with each ICA component over sliding windows. The component with the highest average correlation is selected as the cardiac source. This method is robust but requires a reliable template, which may not be available during sensor initialization. Adaptive template approaches that update the reference waveform over time improve robustness.

Physiological Constraint Validation

Each candidate component is analyzed for physiological plausibility:

• Dominant frequency between 0.5 and 3.5 Hz (30-210 BPM)
• Presence of at least one harmonic (2x fundamental)
• Quasi-periodic structure with inter-beat interval coefficient of variation < 0.2
• Signal quality index (SQI) above a threshold computed from the autocorrelation peak height

Components passing all constraints are ranked by SQI, and the highest-quality component is selected.

ICA for Remote PPG (rPPG)

One of the most impactful applications of ICA in PPG is remote photoplethysmography, where cardiac signals are extracted from video of a person's face without any contact sensor. The RGB color channels of facial video provide a natural three-channel observation for ICA.

The rPPG-ICA Pipeline

1. Face detection and tracking. Detect the face region of interest (ROI) in each video frame using Viola-Jones or deep learning-based face detectors.
2. Spatial averaging. Compute the mean pixel value in each color channel (R, G, B) across the forehead ROI, producing three time series.
3. Pre-processing. Detrend each channel to remove slow illumination changes. Bandpass filter to 0.5-4 Hz.
4. ICA separation. Apply FastICA or JADE to the 3-channel signal to recover independent components.
5. Component selection. Identify the cardiac component using spectral analysis (the component with the strongest peak in the cardiac frequency band).
6. Heart rate estimation. Compute the FFT of the selected component and locate the dominant frequency.

Poh et al. (2010) demonstrated this pipeline with a basic webcam (doi:10.1364/OE.18.010762). De Haan and Jeanne (2013) improved on it with the chrominance-based remote PPG (CHROM) method. In 117 stationary subjects, CHROM agreed with contact PPG 92% of the time, with RMSE about half that of blind source separation methods. CHROM uses a physiologically motivated linear combination of color channels rather than blind ICA separation (doi:10.1109/TBME.2013.2266196).

Remote PPG represents a growing application area with implications for contactless vital sign monitoring in clinical settings, driver monitoring, and telehealth. See our PPG technology overview for more on emerging PPG applications.

Limitations and Practical Considerations

Stationarity Assumption

Standard ICA assumes that the mixing matrix A is constant over the analysis window. For PPG during dynamic motion, the relationship between motion and optical artifact changes as the sensor moves relative to the skin. Windows of 5-10 seconds are typically short enough for the mixing to remain approximately constant, but shorter windows reduce the sample size available for statistical estimation, degrading separation quality.

Adaptive ICA algorithms that update the demixing matrix incrementally (online ICA) address this tradeoff. Oja and Karhunen (1985, doi:10.1016/0022-247X(85)90131-3) introduced stochastic approximation rules for estimating the eigenvectors of a random matrix's expectation, a foundation for later online subspace and ICA learning rules.

Computational Cost

ICA is substantially more expensive than adaptive filtering. FastICA requires $O(m^2 \cdot T)$ operations per iteration for $m$ channels and $T$ samples, with 5-20 iterations for convergence. For a typical configuration ($m = 4$, $T = 1000$, 10 iterations), this is approximately 160,000 operations per analysis window, compared to approximately 4,000 operations for NLMS over the same window. This cost is manageable on modern processors but may preclude continuous real-time operation on the most power-constrained wearable devices.

A practical compromise is to run ICA only during detected high-motion segments (identified by accelerometer energy thresholding) and use simpler methods during rest and light motion.

Comparison with Adaptive Filtering

For most wearable PPG applications with accelerometer reference, NLMS or RLS adaptive filtering is preferred due to lower computational cost and sample-level adaptation. ICA excels in specialized configurations: multi-wavelength PPG without accelerometer, remote PPG from video, and clinical setups where multiple sensor locations are available.

Combining ICA with Other Methods

State-of-the-art PPG pipelines often combine ICA with complementary techniques:

ICA + adaptive filtering: Use ICA for initial source separation, then apply NLMS to refine the cardiac component estimate using the accelerometer reference.

ICA + EMD: Apply empirical mode decomposition to create pseudo-multichannel data from a single PPG channel, then apply ICA to the IMFs. This enables ICA-like separation without requiring multiple physical sensors, though results are less reliable than true multi-channel ICA.

ICA + wavelet denoising: After ICA separation, apply wavelet thresholding to the selected cardiac component to remove residual noise not captured by ICA. This is particularly effective for removing high-frequency electronic noise that persists after ICA.

Conclusion

Independent Component Analysis provides a powerful framework for PPG signal separation that operates on fundamentally different principles than adaptive filtering. Its ability to separate sources without explicit noise modeling makes it invaluable for multi-wavelength PPG configurations, remote PPG applications, and scenarios where the motion reference is unavailable or unreliable.

The practical barriers to ICA adoption in commercial wearables, primarily computational cost and the multi-channel requirement, continue to diminish as sensor hardware integrates multiple wavelengths and processor capabilities increase. For researchers exploring advanced PPG signal processing, ICA is an essential tool that complements rather than replaces the adaptive filtering methods described in our LMS/NLMS and RLS guides.

Visit our algorithms reference for implementation details across all PPG signal processing methods, and explore how different health conditions benefit from these advanced signal separation techniques.

References

• Comon, P. (1994). Independent component analysis, a new concept? Signal Processing, 36(3), 287-314. https://doi.org/10.1016/0165-1684(94)90029-9
• de Haan, G., & Jeanne, V. (2013). Robust pulse rate from chrominance-based rPPG. IEEE Transactions on Biomedical Engineering, 60(10), 2878-2886. https://doi.org/10.1109/TBME.2013.2266196
• Hyvärinen, A. (1999). Fast and robust fixed-point algorithms for independent component analysis. IEEE Transactions on Neural Networks, 10(3), 626-634. https://doi.org/10.1109/72.761722
• Krishnan, R., Natarajan, B., & Warren, S. (2010). Two-stage approach for detection and reduction of motion artifacts in photoplethysmographic data. IEEE Transactions on Biomedical Engineering, 57(8), 1867-1876. https://doi.org/10.1109/TBME.2009.2039568
• Lee, T.-W., Girolami, M., & Sejnowski, T. J. (1999). Independent component analysis using an extended infomax algorithm for mixed subgaussian and supergaussian sources. Neural Computation, 11(2), 417-441. https://doi.org/10.1162/089976699300016719
• Poh, M.-Z., McDuff, D. J., & Picard, R. W. (2010). Non-contact, automated cardiac pulse measurements using video imaging and blind source separation. Optics Express, 18(10), 10762-10774. https://doi.org/10.1364/OE.18.010762