Time-Frequency Analysis

Introduction

Time-frequency analysis is essential for understanding how oscillatory brain activity changes over time in response to stimuli or during cognitive tasks. Unlike classical ERP analysis that focuses on time-domain averages, time-frequency methods reveal dynamic spectral changes in specific frequency bands (delta, theta, alpha, beta, gamma) that are often masked in traditional averaging.

Common neuroscience applications include:

Event-Related Spectral Perturbation (ERSP): Power changes relative to baseline in specific frequency bands, revealing induced oscillatory activity that may not be time-locked
Intertrial Coherence (ITC): Phase consistency across trials, capturing evoked oscillatory responses
Theta-gamma coupling: Investigating cross-frequency interactions during memory encoding
Mu/beta suppression: Motor imagery and action observation studies
Alpha desynchronization: Attentional modulation and visual processing

PhysioEEG provides three complementary time-frequency decomposition methods, each with distinct advantages depending on your research question, signal characteristics, and required time-frequency resolution.

Setup

library(PhysioEEG)

# Create example EEG data: 64 channels, 1000 ms epochs, 100 trials, 500 Hz sampling
eeg <- make_eeg_erp(
  n_channels = 64,
  n_timepoints = 500,
  n_trials = 100,
  sampling_rate = 500
)

# Inspect dimensions
dim(assay(eeg, "raw"))  # channels x timepoints x trials
samplingRate(eeg)       # 500 Hz

Morlet Wavelet Transform

The Morlet wavelet transform provides excellent time-frequency resolution trade-off control via the number of cycles parameter. It convolves the signal with complex Morlet wavelets at specified frequencies, yielding both power and phase information.

Parameter selection:

Frequencies: Define your frequency range based on your research question. For general cognitive studies, 4-40 Hz captures theta through low gamma. For detailed alpha analysis, use finer spacing (e.g., 8-13 Hz in 0.5 Hz steps).
n_cycles: Controls time vs frequency resolution trade-off. Lower values (3-5 cycles) provide better temporal precision but broader frequency resolution; higher values (7-12 cycles) provide sharper frequency resolution but smear events in time. A common compromise is to use frequency-dependent cycles: n_cycles = f / 3 (e.g., 4 Hz → 1.3 cycles, 40 Hz → 13 cycles).

# Morlet wavelet transform with frequency-dependent resolution
# Good for capturing both slow (theta) and fast (gamma) dynamics
eeg <- eegMorletWavelet(
  eeg,
  frequencies = seq(4, 40, by = 2),    # 4-40 Hz in 2 Hz steps
  n_cycles = seq(4, 40, by = 2) / 3,   # Frequency-dependent: f/3 cycles
  assay_name = "raw",
  output_assay = "morlet_power"
)

# Output dimensions: channels x frequencies x timepoints x trials
dim(assay(eeg, "morlet_power"))

# For detailed alpha band analysis with higher frequency resolution
eeg <- eegMorletWavelet(
  eeg,
  frequencies = seq(8, 13, by = 0.5),  # Fine-grained alpha range
  n_cycles = 7,                         # Fixed 7 cycles for good frequency precision
  assay_name = "raw",
  output_assay = "alpha_power"
)

When to use Morlet wavelets:

Default choice for most EEG time-frequency analysis
Excellent for event-related designs where temporal precision matters
Ideal when you need consistent time-frequency resolution across the spectrum
Preferred for cross-frequency coupling analysis

Short-Time Fourier Transform

STFT divides the signal into short overlapping windows and computes the FFT for each segment. It provides uniform time-frequency resolution across all frequencies, determined by the window size.

Window size trade-offs:

Window size: Larger windows (e.g., 1000 ms) provide better frequency resolution but poor temporal precision. Smaller windows (e.g., 100 ms) capture rapid changes but have coarse frequency resolution. A 200-500 ms window is typical for cognitive EEG.
Overlap: Higher overlap (75-90%) provides smoother time-frequency representations but increases computation. 50% overlap is a reasonable default.
Window type: Hanning window reduces spectral leakage and is the standard choice. Hamming window has slightly better frequency resolution but more leakage.

# Standard STFT with 250 ms windows (good balance for cognitive EEG)
eeg <- eegSTFT(
  eeg,
  window_size = 125,      # 250 ms at 500 Hz = 125 samples
  overlap = 0.75,         # 75% overlap for smooth representation
  window_type = "hanning",
  assay_name = "raw",
  output_assay = "stft_power"
)

# Output: channels x frequencies x time_windows x trials
dim(assay(eeg, "stft_power"))

# For high temporal precision (e.g., rapid stimulus changes)
eeg <- eegSTFT(
  eeg,
  window_size = 50,       # 100 ms windows
  overlap = 0.50,         # 50% overlap
  window_type = "hanning",
  assay_name = "raw",
  output_assay = "stft_fast"
)

# For high frequency resolution (e.g., resting-state narrow-band analysis)
eeg <- eegSTFT(
  eeg,
  window_size = 500,      # 1000 ms windows
  overlap = 0.90,         # High overlap to maintain some temporal detail
  window_type = "hamming",
  assay_name = "raw",
  output_assay = "stft_fine"
)

When to use STFT:

When uniform time-frequency resolution is desired
Continuous EEG or resting-state data
When computational speed is a priority (faster than wavelets for many frequencies)
When you need direct control over time-frequency resolution in physical units

Multitaper Spectral Analysis

The Thomson multitaper method uses multiple orthogonal tapers (windowing functions) to reduce variance in spectral estimates while maintaining good frequency resolution. It provides robust spectral estimates, especially for stationary or quasi-stationary signals.

Parameter selection:

Bandwidth: Determines frequency smoothing and number of available tapers. Bandwidth of 2-4 Hz is typical for EEG. Higher bandwidth provides more tapers (lower variance) but coarser frequency resolution.
n_tapers: Number of tapers to use. More tapers reduce variance but increase smoothing. Typically set to floor(2 * bandwidth * T - 1) where T is the time window duration in seconds.

# Multitaper analysis with 3 Hz bandwidth
# Good for reducing noise in oscillatory power estimates
eeg <- eegMultitaper(
  eeg,
  bandwidth = 3,          # 3 Hz smoothing bandwidth
  n_tapers = 5,           # Use 5 Slepian tapers
  assay_name = "raw",
  output_assay = "multitaper_power"
)

# Output: channels x frequencies x timepoints x trials
dim(assay(eeg, "multitaper_power"))

# For high-frequency resolution (narrow-band analysis)
eeg <- eegMultitaper(
  eeg,
  bandwidth = 1.5,        # Narrow bandwidth for fine frequency resolution
  n_tapers = 3,
  assay_name = "raw",
  output_assay = "multitaper_narrow"
)

When to use multitaper:

Resting-state or continuous EEG with quasi-stationary spectral content
When you need robust spectral estimates with minimal variance
Narrow-band oscillation detection (e.g., individual alpha frequency)
When signal-to-noise ratio is low
Spectral analysis of relatively long, stable epochs

Event-Related Spectral Perturbation (ERSP)

ERSP quantifies time-frequency power changes relative to a baseline period, typically pre-stimulus. It reveals both event-related synchronization (ERS, power increase) and desynchronization (ERD, power decrease), capturing induced oscillatory activity that may not be strictly phase-locked to the stimulus.

Baseline normalization methods:

Decibel (dB): 10 * log10(power / baseline) - symmetric, interpretable as percent change on log scale
Percent change: 100 * (power - baseline) / baseline - intuitive but asymmetric
Z-score: (power - mean_baseline) / sd_baseline - accounts for variance, good for statistics

# Compute ERSP from Morlet wavelet power
# Baseline: -200 to 0 ms pre-stimulus
eeg <- eegERSP(
  eeg,
  baseline = c(-200, 0),  # Baseline time window in ms
  assay_name = "morlet_power"
)

# ERSP stored in metadata
ersp_data <- metadata(eeg)$ERSP
dim(ersp_data)  # channels x frequencies x timepoints x trials

# Average across trials to see mean ERSP
mean_ersp <- apply(ersp_data, c(1, 2, 3), mean)
dim(mean_ersp)  # channels x frequencies x timepoints

# Extract single-channel ERSP for visualization
pz_ersp <- mean_ersp[channelIndex(eeg, "Pz"), , ]
dim(pz_ersp)  # frequencies x timepoints

Interpreting ERSP:

Positive values (ERS): Power increase relative to baseline. Example: frontal theta increase during memory encoding, gamma increase during sensory processing.
Negative values (ERD): Power decrease relative to baseline. Example: central mu/beta suppression during motor preparation, occipital alpha decrease during visual attention.
Time course: ERSP reveals when spectral changes occur, distinguishing early sensory from late cognitive processes.
Spatial distribution: Topographic patterns indicate functional networks (e.g., frontal theta vs posterior alpha).

Intertrial Coherence (ITC)

ITC measures phase consistency across trials at each time-frequency point, ranging from 0 (random phase) to 1 (perfect phase alignment). Unlike ERSP which captures power changes, ITC specifically identifies phase-locked (evoked) oscillatory activity.

Key concepts:

ITC = 1: All trials have identical phase at that time-frequency point (perfect evoked response)
ITC = 0: Random phase across trials (purely induced activity)
ITC vs ERP: High ITC at low frequencies corresponds to traditional ERP components. However, ITC can also detect phase-locked higher-frequency oscillations invisible in the averaged ERP.

# Compute ITC from Morlet wavelet transform
# Uses complex wavelet coefficients to assess phase consistency
eeg <- eegITC(
  eeg,
  assay_name = "morlet_power"  # Should be complex wavelet output
)

# ITC stored in metadata
itc_data <- metadata(eeg)$ITC
dim(itc_data)  # channels x frequencies x timepoints

# Extract channel-specific ITC
fcz_itc <- itc_data[channelIndex(eeg, "FCz"), , ]
dim(fcz_itc)  # frequencies x timepoints

# High ITC at low frequencies captures ERP phase-locking
# Check theta band (4-8 Hz) ITC time course
theta_itc <- apply(fcz_itc[4:8, ], 2, mean)
plot(theta_itc, type = "l", xlab = "Time (ms)", ylab = "ITC")

Interpreting ITC:

High ITC + high ERSP: Evoked power increase (both phase-locked and power increase)
High ITC + low ERSP: Phase-locked activity without net power change (phase reset)
Low ITC + high ERSP: Induced power increase without phase-locking (e.g., stimulus-triggered but temporally jittered oscillations)
Low-frequency ITC peaks: Correspond to traditional ERP components (N1, P3, etc.)
High-frequency ITC: May indicate early sensory-evoked gamma or beta responses

Visualization with eegPlotSpectrogram

Time-frequency heatmaps (spectrograms) are the standard visualization for ERSP, ITC, and raw power. PhysioEEG provides customizable spectrogram plotting with sensible defaults.

# Plot ERSP spectrogram for channel Pz
eegPlotSpectrogram(
  eeg,
  channel = "Pz",
  freq_range = c(4, 40),      # Focus on theta-gamma range
  time_range = c(-200, 800),  # Pre-stimulus to 800 ms post-stimulus
  log_power = FALSE,          # ERSP already in dB
  palette = "RdBu",           # Red-blue diverging palette (blue=ERD, red=ERS)
  assay_name = "morlet_power"
)

# Plot ITC spectrogram
eegPlotSpectrogram(
  eeg,
  channel = "FCz",
  freq_range = c(4, 30),
  time_range = NULL,          # Use full time range
  log_power = FALSE,
  palette = "viridis",        # Sequential palette for ITC (0 to 1)
  assay_name = "morlet_power"
)

# Plot raw power with log scaling
eegPlotSpectrogram(
  eeg,
  channel = "O1",
  freq_range = c(8, 13),      # Alpha band detail
  time_range = c(0, 500),
  log_power = TRUE,           # Convert to dB scale
  palette = "magma",
  assay_name = "morlet_power"
)

Visualization tips:

ERSP: Use diverging colormaps (RdBu, RdYlBu) with center at zero to distinguish ERS from ERD
ITC: Use sequential colormaps (viridis, magma, plasma) since values range from 0 to 1
Raw power: Apply log scaling to compress dynamic range and make oscillations visible
Frequency range: Narrow ranges (e.g., 8-13 Hz for alpha) reveal fine structure; broad ranges (4-80 Hz) show multi-band dynamics
Time range: Include pre-stimulus baseline to contextualize post-stimulus changes

Method Comparison Guide

Choosing the right time-frequency method depends on your data characteristics, research question, and computational resources.

Method	Time Resolution	Frequency Resolution	Best For	Computational Cost
Morlet Wavelet	Variable (tunable via n_cycles)	Variable (tunable via n_cycles)	Event-related designs, flexible resolution trade-off, cross-frequency coupling	Moderate-High (depends on # frequencies)
STFT	Uniform (set by window)	Uniform (set by window)	Continuous EEG, resting-state, fast computation, uniform resolution	Low-Moderate
Multitaper	Moderate	High (good for narrow bands)	Resting-state, low SNR, narrow-band oscillations, robust spectral estimates	Moderate

Decision tree:

Epoched event-related data with precise timing requirements: Use Morlet wavelets with frequency-dependent n_cycles for optimal time-frequency resolution matching signal characteristics.
Continuous or resting-state EEG: Use STFT for speed or multitaper for robust spectral estimates with low variance.
Narrow-band oscillation detection (e.g., individual alpha frequency): Use multitaper with low bandwidth for superior frequency resolution.
Broad frequency range with varying dynamics: Use Morlet wavelets with frequency-dependent parameters (low n_cycles for low frequencies, high n_cycles for high frequencies).
Low signal-to-noise ratio: Use multitaper to reduce spectral variance, or average multiple Morlet wavelet trials.
Computational constraints: Use STFT, which is faster than wavelets when analyzing many frequencies.

Frequency resolution comparison:

Wavelet: Resolution = f / n_cycles. Example: 10 Hz with 5 cycles → 2 Hz resolution.
STFT: Resolution = sampling_rate / window_size. Example: 500 Hz with 250-sample window → 2 Hz resolution.
Multitaper: Resolution ≈ bandwidth. Example: 3 Hz bandwidth → ~3 Hz resolution.

Time resolution comparison:

Wavelet: Resolution = n_cycles / f. Example: 10 Hz with 5 cycles → 0.5 s resolution.
STFT: Resolution = window_size / sampling_rate. Example: 250 samples at 500 Hz → 0.5 s resolution.
Multitaper: Resolution depends on window size and taper properties, generally moderate.

Session Info

sessionInfo()