PPG Power Spectral Analysis: Frequency Domain Methods for Signal Interpretation

Frequency domain analysis transforms the PPG signal from a time-varying waveform into a spectrum of constituent frequencies, revealing cardiac rhythm, respiratory modulation, autonomic oscillations, and noise components that are difficult or impossible to separate in the time domain. While time-domain peak detection suffices for basic heart rate measurement, spectral analysis is essential for robust heart rate estimation during motion, HRV frequency-domain metrics, respiratory rate extraction, signal quality assessment, and understanding the physiological information content of the PPG signal.

This guide covers the major spectral analysis techniques used in PPG research and engineering, from the fast Fourier transform to parametric spectral models, with practical guidance on parameter selection, computational considerations, and interpretation of spectral features. For the fundamentals of PPG signal characteristics, see our introduction to photoplethysmography.

Frequency Content of the PPG Signal

Before diving into spectral methods, understanding what frequency components exist in a PPG signal provides the context for interpreting any spectral analysis.

Cardiac Component (0.5-4 Hz)

The dominant frequency component of the PPG signal is the cardiac pulse, whose fundamental frequency corresponds to the heart rate. At a resting heart rate of 72 BPM, the fundamental frequency is 1.2 Hz. This fundamental is accompanied by harmonics at integer multiples (2.4, 3.6, 4.8 Hz, etc.) that arise from the non-sinusoidal shape of the PPG pulse waveform.

The number and relative amplitude of detectable harmonics depend on the waveform morphology. A PPG signal with a sharp systolic peak and prominent dicrotic notch contains more harmonic energy than a rounded waveform. Typically, 3-5 harmonics are detectable above the noise floor, meaning that the cardiac content extends to approximately 6-20 Hz depending on heart rate.

The fundamental frequency carries the most power, typically containing 50-65% of the total pulsatile signal energy. The first harmonic contains 15-25%, and the second harmonic contains 5-10%. This harmonic structure is important because motion artifacts can overlap with the fundamental but rarely have the same harmonic pattern, providing a basis for artifact-robust heart rate estimation.

Respiratory Component (0.1-0.5 Hz)

Respiration modulates the PPG signal through three mechanisms: amplitude modulation (respiratory variation in pulse size), baseline modulation (DC level shifts with intrathoracic pressure changes), and frequency modulation (respiratory sinus arrhythmia). These respiratory effects produce spectral content in the 0.1-0.5 Hz band, concentrated at the respiratory frequency.

The respiratory component is typically 10-30 dB below the cardiac fundamental in power. Extracting it requires either bandpass filtering in this range or demodulation techniques that isolate the envelope (amplitude modulation), baseline trend (baseline modulation), or inter-beat interval sequence (frequency modulation). For more on respiratory rate extraction, see our respiratory rate algorithms guide.

Autonomic and Vasomotion Components (<0.15 Hz)

The very-low-frequency (VLF, 0.003-0.04 Hz) and low-frequency (LF, 0.04-0.15 Hz) bands contain oscillations driven by thermoregulatory vasomotion, sympathetic nervous system activity, and baroreflex-mediated blood pressure regulation (Mayer waves at approximately 0.1 Hz). These components modulate the PPG baseline and amplitude and are physiologically meaningful for autonomic function assessment.

However, these same frequency bands are contaminated by non-physiological slow drifts from sensor movement, temperature changes, and contact pressure variations, making physiological interpretation challenging without careful baseline wander removal.

Noise Floor and Artifacts

Above the cardiac harmonics (typically above 10-20 Hz), the PPG spectrum is dominated by electronic noise from the sensor and ADC, quantization noise, ambient light interference (often showing a spike at 50 or 60 Hz from powerline illumination), and high-frequency motion artifact components. The noise floor level determines the practical dynamic range for detecting weak spectral features and influences the minimum detectable pulse amplitude.

Fast Fourier Transform (FFT)

The FFT is the computational backbone of PPG spectral analysis. It converts a discrete-time signal of N samples into N/2+1 complex frequency-domain coefficients, from which the power spectrum is computed as the squared magnitude.

Parameter Selection

Window Length: The frequency resolution of the FFT is delta_f = Fs/N = 1/T, where T is the window duration in seconds. For PPG heart rate estimation, a resolution of 0.1-0.2 Hz (corresponding to 6-12 BPM) is acceptable, requiring windows of 5-10 seconds. For HRV spectral analysis where LF (0.04-0.15 Hz) and HF (0.15-0.4 Hz) bands must be resolved, windows of at least 60-120 seconds are needed for adequate frequency resolution.

Zero-Padding: Appending zeros to the signal before FFT interpolates the frequency spectrum without adding spectral information. This is useful for smoother spectral displays and more precise peak frequency estimation but does not improve true frequency resolution. Zero-padding a 5-second PPG window to 1024 or 2048 points is common practice for heart rate estimation, providing a finely sampled spectrum for peak picking.

Window Function: Applying a window function (Hann, Hamming, Blackman, etc.) to the signal before FFT reduces spectral leakage, which occurs because the FFT assumes the input is periodic over the window length. Without windowing, a cardiac peak produces sidelobes that can mask nearby spectral features.

The Hann window is the most widely used for PPG because it provides a good balance of main lobe width (slightly wider than rectangular) and sidelobe suppression (-31.5 dB first sidelobe). The Hamming window provides slightly better sidelobe suppression (-42.7 dB) at the cost of a marginally wider main lobe. The Blackman window provides excellent sidelobe suppression (-58.1 dB) but with a wider main lobe that reduces frequency resolution.

Elgendi et al. (2013) compared window functions for PPG heart rate estimation and found that the Hann window provided the best peak detection performance, with 97.8% correct heart rate identification versus 95.2% for rectangular windowing on a dataset of 50 subjects during light activity (DOI: 10.1371/journal.pone.0076585).

Heart Rate Estimation via FFT

The standard FFT-based heart rate estimation algorithm operates as follows:

1. Extract a signal window (typically 8-10 seconds) and apply baseline wander removal
2. Apply a bandpass filter (0.5-4 Hz) or the equivalent frequency-domain windowing
3. Apply a Hann or Hamming window
4. Compute the FFT and power spectrum (magnitude squared)
5. Identify the dominant peak in the 0.5-3.5 Hz range
6. Convert peak frequency to BPM: HR = f_peak x 60

For improved accuracy, parabolic interpolation around the spectral peak estimates the true peak frequency to sub-bin precision. If the peak is at bin k with power P(k), and adjacent bins have powers P(k-1) and P(k+1), the interpolated peak offset is:

delta = 0.5 * (P(k-1) - P(k+1)) / (P(k-1) - 2*P(k) + P(k+1))

This typically improves frequency (and thus heart rate) resolution by a factor of 2-4x beyond the nominal FFT resolution.

Harmonic Spectral Tracking

During motion, the cardiac spectral peak may be obscured by motion artifact peaks. Harmonic tracking exploits the fact that the cardiac signal has a characteristic harmonic structure (fundamental plus integer harmonics) while motion artifacts typically have a different harmonic pattern.

The JOSS (Joint Sparse Spectrum) algorithm by Zhang (2015) uses sparse signal recovery techniques to simultaneously estimate the cardiac and motion fundamental frequencies and their harmonics from the PPG power spectrum, resolving the spectral overlap problem. JOSS achieved a mean absolute heart rate error of 1.28 BPM on the IEEE Signal Processing Cup 2015 dataset during vigorous treadmill running.

For more on spectral approaches to motion artifact separation, see our guide on PPG motion artifact removal.

Welch's Method

Welch's method (Welch, 1967) provides a more statistically reliable spectral estimate than a single FFT by averaging over multiple overlapping sub-windows. The procedure is:

1. Divide the N-sample signal into K segments of length L with overlap D
2. Apply a window function to each segment
3. Compute the periodogram (squared magnitude of FFT) for each segment
4. Average the K periodograms to produce the final power spectral density (PSD) estimate

Parameter Tradeoffs

The key tradeoff in Welch's method is between frequency resolution and spectral variance:

Frequency resolution: Determined by the sub-window length L, not the total signal length N. Resolution = Fs/L.

Spectral variance reduction: Proportional to 1/K_eff, where K_eff is the effective number of independent segments (accounting for overlap). With 50% overlap, K_eff is approximately 2*N/L - 1.

For PPG HRV spectral analysis, the standard configuration is:

• Total recording length: 5 minutes (300 seconds)
• Sub-window length: 60 seconds (0.0167 Hz resolution)
• Overlap: 50% (30 seconds)
• Window function: Hann
• Number of segments: K = 9
• Effective variance reduction: approximately 5x compared to single periodogram

This configuration resolves the LF and HF HRV bands while providing statistically stable spectral estimates suitable for computing LF/HF ratios. The Task Force of the European Society of Cardiology (1996) recommended this general approach for HRV spectral analysis from ECG; the same methodology applies to PPG-derived inter-beat interval series.

Welch vs. Single FFT for Heart Rate

For real-time heart rate estimation from short PPG windows (5-10 seconds), Welch's method is generally not recommended because the sub-windows become too short for adequate frequency resolution. A single FFT of the full window with appropriate zero-padding provides better performance in this scenario.

Welch's method becomes advantageous for heart rate estimation when processing longer recordings (30+ seconds) where spectral stability is more important than maximum resolution, or when the goal is to estimate the spectral shape (e.g., for signal quality assessment) rather than to identify a precise peak frequency.

Autoregressive (AR) Spectral Estimation

Autoregressive spectral estimation models the signal as the output of an all-pole linear filter driven by white noise. The model parameters (AR coefficients) are estimated from the signal autocorrelation, and the power spectrum is computed from the model transfer function. AR methods provide smoother spectral estimates with higher frequency resolution than FFT methods for short data segments.

Burg Method

The Burg algorithm estimates AR coefficients using forward and backward linear prediction, providing guaranteed model stability and good performance with short data records. For PPG heart rate estimation from short windows (3-5 seconds), the Burg method can outperform FFT by resolving the cardiac spectral peak more clearly.

Burg's method requires selecting the model order, which controls the tradeoff between frequency resolution and spectral smoothness. Too low an order produces an overly smooth spectrum that may miss the cardiac peak. Too high an order introduces spurious peaks from fitting noise. For PPG cardiac frequency estimation, model orders of 8-16 are typically appropriate for signals sampled at 25-100 Hz.

Dash et al. (2010) compared Burg AR spectral estimation against FFT for PPG heart rate estimation during motion on the MIMIC database. Using 8-second windows, the Burg method (order 12) achieved heart rate MAE of 1.6 BPM versus 2.3 BPM for FFT with Hann windowing, with the advantage concentrated in low-SNR segments where the FFT failed to resolve the cardiac peak from noise (DOI: 10.1109/TBME.2010.2061469).

Yule-Walker and Covariance Methods

Alternative AR estimation methods include the Yule-Walker (autocorrelation) method, which guarantees model stability but has lower resolution than Burg, and the covariance method, which has higher resolution but may produce unstable models requiring additional stability checks. For most PPG applications, the Burg method provides the best overall performance and is the recommended AR estimator.

Model Order Selection

Information-theoretic criteria, such as the Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC), provide principled approaches to model order selection. These criteria balance model fit against complexity, penalizing overly complex models that overfit noise.

For PPG signals sampled at 100 Hz with 8-second analysis windows, AIC typically selects model orders of 10-20, consistent with the expected spectral complexity of the cardiac signal and its harmonics. Pre-selecting a fixed model order (e.g., 12 or 16) is common in real-time applications to avoid the computational cost of order selection at each window.

Spectral Feature Extraction

Beyond heart rate estimation, the PPG power spectrum contains features useful for signal characterization, quality assessment, and physiological parameter extraction.

Signal-to-Noise Ratio Estimation

The spectral SNR of the PPG signal can be estimated as the ratio of power at the cardiac frequency (and its harmonics) to the noise power in the surrounding spectral bins. A practical estimator computes the peak power within a narrow band (e.g., +/-0.2 Hz around the cardiac peak) divided by the median power of the remaining bins in the physiological range.

Spectral SNR correlates strongly with signal quality index metrics and provides a frequency-domain quality assessment that complements time-domain SQI approaches. Clean resting PPG typically has spectral SNR above 20 dB, while heavily motion-corrupted signals may drop below 5 dB.

Spectral Entropy

As discussed in the signal quality context, spectral entropy quantifies the distribution of power across frequency bins. Low entropy indicates concentrated spectral energy (clean cardiac signal), while high entropy indicates distributed energy (noise or artifact). Spectral entropy is computed as the Shannon entropy of the normalized PSD:

H_spec = -sum(P_n * log2(P_n)) / log2(N_bins)

where P_n is the normalized power (sum = 1) at each of the N_bins frequency bins. Values range from 0 (single pure tone) to 1 (white noise with equal power at all frequencies).

For clean resting PPG, spectral entropy typically ranges from 0.3 to 0.5. Values above 0.65 generally indicate poor signal quality. This metric integrates naturally into spectral processing pipelines since the PSD is already available.

Harmonic-to-Fundamental Ratio

The ratio of harmonic power (second and third cardiac harmonics) to fundamental power reflects the PPG pulse waveform shape. A narrow, peaked systolic waveform with a prominent dicrotic notch produces more harmonic energy than a broad, rounded waveform. This ratio has been correlated with arterial stiffness indices.

Allen and Murray (2003) demonstrated that the harmonic-to-fundamental ratio of finger PPG increased with age (r = 0.68, p < 0.001, n = 116 subjects, age range 20-80 years), consistent with age-related arterial stiffening that produces sharper, more peaked pulse waveforms (DOI: 10.1088/0967-3334/24/2/306).

Respiratory Frequency Extraction

The respiratory frequency can be extracted from the PPG spectrum by analyzing specific spectral bands. After removing the cardiac component, the residual spectrum in the 0.1-0.5 Hz band contains respiratory modulation peaks. Alternatively, the spectrum of the cardiac amplitude envelope (obtained by peak-to-peak amplitude of successive pulses) reveals the respiratory modulation frequency directly.

Spectral respiratory rate estimation achieves accuracy within 1-2 breaths per minute at rest, comparable to impedance pneumography and capnography references. During exercise, motion artifacts in the respiratory frequency band can compromise accuracy, requiring multi-modal estimation approaches.

Spectrogram and Time-Varying Spectral Analysis

The standard FFT provides a single spectrum for the entire analysis window, assuming the signal is stationary within that window. For PPG signals during exercise or physiological transitions, stationarity may not hold, and time-varying spectral analysis is needed.

Short-Time Fourier Transform (STFT)

The STFT computes the FFT at successive overlapping time positions, producing a two-dimensional time-frequency representation (spectrogram). Each column of the spectrogram is the spectrum at a specific time instant, computed from a short windowed segment of the signal.

STFT is the foundation of spectral heart rate tracking during exercise. A typical configuration uses 8-second Hann windows with 1-second step (87.5% overlap), producing a spectrogram with 0.125 Hz frequency resolution updated every second. Heart rate tracking then reduces to following the dominant spectral peak across successive time frames, with physiological constraints preventing the tracker from jumping to artifact peaks.

For more advanced time-frequency analysis methods that overcome the fixed resolution limitation of the STFT, see our guide on PPG time-frequency analysis.

Lomb-Scargle Periodogram for Irregular Sampling

The inter-beat interval series derived from PPG is an irregularly sampled time series (since heartbeats do not occur at uniform intervals). The standard FFT assumes uniform sampling and cannot be directly applied. The Lomb-Scargle periodogram handles irregular sampling without requiring interpolation or resampling, computing the spectral power at arbitrary frequencies directly from the unevenly spaced data.

For HRV spectral analysis from PPG, the Lomb-Scargle method avoids the spectral distortion introduced by interpolation-based resampling of the inter-beat interval series. Clifford and Tarassenko (2005) showed that the Lomb-Scargle periodogram reduced LF/HF ratio estimation error by 12% compared to cubic spline interpolation followed by FFT on ECG-derived RR intervals from 48 Holter recordings (DOI: 10.1109/TBME.2004.838363). The same advantage applies to PPG-derived pulse interval series.

Computational Considerations

FFT Implementation

The FFT has O(N log N) computational complexity, where N is the transform length. For typical PPG analysis windows (256-2048 samples after zero-padding), FFT computation requires microseconds on modern processors and is not a computational bottleneck.

For embedded wearable platforms, fixed-point FFT implementations (such as the CMSIS-DSP library for ARM Cortex-M processors) provide efficient computation without floating-point hardware. A 256-point fixed-point FFT on an ARM Cortex-M4 processor running at 64 MHz requires approximately 50 microseconds.

Real-Time Spectral Processing Architecture

A practical real-time spectral processing pipeline for PPG heart rate estimation operates as follows:

1. Buffer incoming PPG samples into overlapping analysis windows
2. Apply preprocessing (baseline removal, bandpass filtering)
3. Apply window function and compute FFT
4. Compute power spectrum and identify cardiac peak
5. Apply spectral peak tracking with physiological constraints
6. Output heart rate estimate and confidence metric

This pipeline typically runs with 1-2 second update intervals, processing each new window in well under the available real-time budget even on low-power embedded processors. The memory requirement is dominated by the FFT buffer and is typically 1-4 KB for 256-1024 point transforms.

Streaming Spectral Analysis

For long-duration monitoring, the sliding DFT (SDFT) provides an efficient alternative to recomputing the full FFT at each time step. The SDFT updates the spectral coefficients incrementally as new samples arrive and old samples are discarded, with O(N) computation per update instead of O(N log N) for a full FFT. This is particularly useful for battery-constrained wearables performing continuous spectral heart rate tracking.

Applications in PPG Research

Autonomic Function Assessment

Spectral analysis of PPG-derived pulse rate variability (PRV) provides a window into autonomic nervous system function. The LF (0.04-0.15 Hz) component reflects mixed sympathetic and parasympathetic modulation, while the HF (0.15-0.4 Hz) component reflects primarily parasympathetic (vagal) modulation. The LF/HF ratio has been used as an index of sympathovagal balance, though its interpretation is debated.

Gil et al. (2010) validated PPG-derived LF/HF ratios against ECG-derived values in 17 healthy subjects during tilt-table testing and found correlation coefficients of r = 0.89 during supine rest and r = 0.83 during head-up tilt, demonstrating that wrist PPG provides a reasonable substitute for ECG-based autonomic assessment when ECG is not available (DOI: 10.1109/TBME.2009.2037328). For more on HRV analysis methods, see our HRV algorithms guide.

Sleep Stage Classification

Spectral features of the PPG signal contribute to automated sleep stage classification in wearable devices. The HF component of PRV increases during deep sleep (high vagal tone), while the LF/HF ratio increases during REM sleep. Spectral features from the PPG waveform itself (not just the inter-beat series) can detect respiratory patterns characteristic of different sleep stages.

Stress and Cognitive Load Detection

Acute mental stress produces sympathetic activation that increases the LF component of PRV and decreases HF, elevating the LF/HF ratio. Spectral analysis of short-term PPG recordings (3-5 minutes) during cognitive tasks has been used to detect stress states with 75-85% accuracy using LF/HF ratio alone and 85-92% accuracy when combined with additional PPG spectral features.

For a comprehensive understanding of how spectral analysis integrates with the complete PPG signal processing pipeline, from raw signal acquisition through clinical parameter estimation, see our algorithm guides at /algorithms/.

Frequently Asked Questions

What frequencies are present in a normal PPG signal?

A normal PPG signal contains several frequency components: the cardiac fundamental frequency at 0.8-3.5 Hz (48-210 BPM) and its harmonics (up to the 4th or 5th harmonic at 4-17.5 Hz), respiratory modulation at 0.15-0.4 Hz (9-24 breaths/min), Mayer wave oscillations near 0.1 Hz reflecting sympathetic vasomotor activity, very-low-frequency vasomotion at 0.01-0.05 Hz, and the DC component at 0 Hz representing baseline tissue absorption. The cardiac fundamental and first harmonic typically contain 80-90% of the pulsatile signal power.

How is FFT used for PPG heart rate estimation?

FFT-based heart rate estimation computes the power spectrum of a windowed PPG segment (typically 5-10 seconds), identifies the dominant spectral peak within the physiological heart rate range (0.5-3.5 Hz), and converts the peak frequency to BPM by multiplying by 60. Windowing functions like Hann or Hamming are applied before the FFT to reduce spectral leakage. Peak tracking algorithms with physiological constraints (maximum heart rate change of 10-15 BPM per second) reject spurious peaks from motion artifacts. This approach achieves heart rate MAE below 1 BPM at rest and 2-5 BPM during moderate exercise.

What is the difference between FFT and Welch's method for PPG?

The standard FFT computes the spectrum of a single data window, while Welch's method divides the signal into overlapping segments, computes the FFT of each segment, and averages the resulting power spectra. Welch's method reduces spectral variance (noise in the spectrum) at the cost of reduced frequency resolution. For a 10-second PPG recording, a single FFT gives 0.1 Hz frequency resolution but a noisy spectrum. Welch's method with 4-second sub-segments and 50% overlap gives 0.25 Hz resolution but a smoother, more reliable spectrum. Welch's method is preferred when spectral shape estimation is more important than maximum frequency resolution.

Can spectral analysis detect arrhythmias in PPG signals?

Yes, spectral analysis provides several indicators of arrhythmias. Atrial fibrillation produces a characteristic spectral signature: the cardiac fundamental peak broadens or splits due to irregular RR intervals, and the spectral entropy increases significantly. Normal sinus rhythm shows a sharp cardiac peak with clear harmonics, while AF shows a broader, less defined spectral peak. Studies have shown that spectral irregularity metrics achieve 90-95% sensitivity for AF detection from PPG. However, spectral analysis alone cannot distinguish between specific arrhythmia types and is typically combined with time-domain rhythm analysis for robust detection.

References

• Elgendi et al. (2013) compared window functions for PPG heart rate estimation and found that the Hann window provided the best peak detection performance, with 97.8% correct heart rate identification versus 95.2% for rectangular windowing on a dataset of 50 subjects during light activity (DOI: 10.1371/journal.pone.0076585).
• Dash et al. (2010) compared Burg AR spectral estimation against FFT for PPG heart rate estimation during motion on the MIMIC database. Using 8-second windows, the Burg method (order 12) achieved heart rate MAE of 1.6 BPM versus 2.3 BPM for FFT with Hann windowing, with the advantage concentrated in low-SNR segments where the FFT failed to resolve the cardiac peak from noise (DOI: 10.1109/TBME.2010.2061469).
• Allen and Murray (2003) demonstrated that the harmonic-to-fundamental ratio of finger PPG increased with age (r = 0.68, p < 0.001, n = 116 subjects, age range 20-80 years), consistent with age-related arterial stiffening that produces sharper, more peaked pulse waveforms (DOI: 10.1088/0967-3334/24/2/306).
• For HRV spectral analysis from PPG, the Lomb-Scargle method avoids the spectral distortion introduced by interpolation-based resampling of the inter-beat interval series. Clifford and Tarassenko (2005) showed that the Lomb-Scargle periodogram reduced LF/HF ratio estimation error by 12% compared to cubic spline interpolation followed by FFT on ECG-derived RR intervals from 48 Holter recordings (DOI: 10.1109/TBME.2004.838363). The same advantage applies to PPG-derived pulse interval series.
• Gil et al. (2010) validated PPG-derived LF/HF ratios against ECG-derived values in 17 healthy subjects during tilt-table testing and found correlation coefficients of r = 0.89 during supine rest and r = 0.83 during head-up tilt, demonstrating that wrist PPG provides a reasonable substitute for ECG-based autonomic assessment when ECG is not available (DOI: 10.1109/TBME.2009.2037328). For more on HRV analysis methods, see our HRV algorithms guide.