Beyond Band Power: Aperiodic (1/f) and Waveform-Aware Analysis of Ear-EEG

1. Introduction

Wearable and consumer EEG increasingly translates brain activity into claims a user can act on—how well they slept, how focused they are, how calm they have become. Almost universally, those claims rest on band power: the integrated spectral energy within canonical ranges. Band power is attractive because it is simple, fast, and historically standard. But it rests on assumptions that decades of methodological work have shown are routinely violated by real neural data, and the consequences are not cosmetic: under the wrong conditions, band power can move in the opposite direction from the oscillation it is meant to measure.

The central problem is that the EEG power spectrum is not purely oscillatory. It comprises an aperiodic, broadband 1/f-like background—whose slope (exponent) and offset reflect, among other things, the balance of synaptic excitation and inhibition—superimposed with narrowband oscillatory peaks. Canonical band power sums both. When the aperiodic component changes across states, ages, or individuals, absolute band power can change with no underlying oscillatory change at all. Three further assumptions compound the issue: that oscillations sit at fixed canonical frequencies (they do not; peak frequency varies across people, ages, and cortical sites), that they are sinusoidal (they are not; waveform shape carries information and distorts Fourier-based power), and that they are stationary (they are bursty, so measured power confounds burst amplitude, duration, and rate).

This matters acutely for ear-EEG. Electrodes placed in and around the ear can record cortical activity comfortably and continuously, and have been validated for alpha rhythms, evoked responses, and overnight sleep staging; but they do so from a restricted, low-channel vantage with reduced signal-to-noise relative to scalp montages. These risks are amplified at the ear, where channel counts are low and signal-to-noise is reduced—precisely the regime in which hearable products operate. Here we test, on real NextSense ear-EEG, whether moving from canonical band power to an aperiodic- and waveform-aware pipeline changes what we conclude. We (1) catalog the confounds; (2) reanalyze two ear-EEG cohorts with a parameterized spectral model and cycle-by-cycle metrics; (3) show that state-related conclusions about "alpha" can reverse once the 1/f background is separated; and (4) ask whether aperiodic-aware features cost anything in downstream state decoding.

2. Why band power misleads

2.1 The aperiodic 1/f background. The dominant feature of the EEG spectrum is its aperiodic 1/f-like decay. Its exponent (slope) and offset vary systematically with state—flattening in wake, steepening in deep sleep—and with age and arousal. Because absolute band power integrates under both the oscillatory peak and this background, a change in slope or offset alone shifts band power even when oscillatory activity is unchanged.

2.2 Peak-frequency variability. Oscillatory peaks do not sit at fixed textbook frequencies. Alpha peak frequency varies across individuals, ages, and cortical locations, so a fixed 8–12 Hz window can miss a genuine peak in one person and capture a neighboring rhythm in another.

2.3 Waveform shape. Neural oscillations are non-sinusoidal. Fourier-based band power assumes sinusoidality and therefore misattributes waveform asymmetry to harmonic power and spurious cross-frequency coupling.

2.4 Transient, burst-like dynamics. Oscillations occur in bursts that vary in duration, occurrence, and amplitude. Average band power conflates these, so two states with identical mean power can differ entirely in their burst structure.

2.5 Mixture of concurrent rhythms. A single channel sums multiple sources—occipital alpha, sensorimotor mu, midfrontal theta, sleep spindles—that overlap in frequency. A canonical band can silently fold one rhythm into the label of another; as shown below, sleep spindles near 12–13 Hz are readily mislabeled as "alpha."

2.6 Alpha as functional inhibition. Finally, the same band can carry opposite meaning. Alpha appears in relaxed eyes-closed wake, in drowsiness, and during attentional gating, where it is best understood as active functional inhibition rather than idling. Band power alone cannot disambiguate these.

3. Methods

3.1 Datasets. We analyzed two NextSense ear-EEG cohorts, kept separate throughout because they differ in device, montage, scoring, and sampling rate. EDS: overnight ear-EEG from 16 subjects (esc_101–esc_125) recorded with a cross-head bipolar ear montage at 200 Hz, with 30-second consensus hypnograms (full WAKE/N1/N2/N3/REM ladder) and simultaneous polysomnography. Maui: 71 paired combo recordings using a within-head ear bipolar montage at 100 Hz. A naive pooled analysis would be misleading, so all results are reported per cohort.

3.2 Analysis pipeline. Per 30-second epoch we computed a Welch power spectral density, then (a) conventional non-overlapping band powers (delta 1–4, theta 4–8, alpha 8–12, sigma 12–15, beta 15–30 Hz); (b) a parameterized spectral model (specparam) fit over 1–40 Hz, yielding the aperiodic offset and exponent and, separately, any periodic peaks in the alpha (8–12 Hz) and sigma (12–16 Hz) ranges; and (c) cycle-by-cycle waveform metrics (bycycle): polarity-robust rise–decay and peak–trough symmetry and burst fraction. Reporting alpha and sigma peaks separately is deliberate—it is how the spindle-vs-alpha confound is exposed rather than hidden. We used specparam for aperiodic–periodic separation; IRASA, which separates fractal and oscillatory components by spectral resampling, is a widely used alternative. Software: mne 1.12.1, specparam 2.0.0rc6, bycycle 1.2.0, neurodsp 2.3.0.

3.3 Statistics. Cohort estimates use 95% bootstrap confidence intervals over subjects/recordings (10,000 resamples), the appropriate error bar for a cohort-level claim; the deep-sleep steepening test is a paired within-subject contrast. For state decoding we trained balanced random-forest classifiers under strict leave-one-subject-out cross-validation (no epoch from a held-out subject ever appears in training), comparing naive band-power features (A), aperiodic+waveform features (B), and their union (A+B), for both five-class staging and binary wake-vs-sleep, with bootstrap CIs on per-subject balanced accuracy and on the A–B difference.

3.4 Coverage and exclusions (no silent drops). EDS: 15 of 16 subjects were analyzable; esc_123 was excluded because its sleep session lacked a merged continuous EDF, making the annotation-to-signal timeline unrecoverable (logged). Maui: of 71 combo recordings, 8 were excluded (5 corrupted files, 3 empty-signal recordings), leaving 63 analyzable; per-analysis counts vary with the number of clean staged epochs available (63 recordings entered the exponent analysis; 54 had sufficient epochs for the alpha-reversal test). Every exclusion is logged.

4. Results

4.1 The spectrum decomposes cleanly at the ear. specparam separates a NextSense ear-EEG spectrum into an aperiodic 1/f component and overlying oscillatory peaks with high fit quality (median R² = 0.97 on EDS), confirming that the decomposition is well posed on ear data.

4.2 State-dependent aperiodic 1/f — robust, but montage-dependent. On the EDS cross-head montage the aperiodic exponent steepens monotonically from wake into deep sleep: WAKE 1.87 [1.62, 2.08], N1 2.14 [1.96, 2.30], N2 2.51 [2.36, 2.64], N3 2.92 [2.77, 3.07], REM 2.37 [2.15, 2.56]. The paired deep-sleep steepening is Δ(N3−WAKE) = +1.05 [0.83, 1.28], with N3 the steepest stage in all 15 subjects and a strictly monotonic WAKE≤N1≤N2≤N3 progression in 11/15. This reproduces the canonical deep-sleep 1/f signature—on ear-EEG, across a cohort. On the Maui within-head montage, by contrast, the exponent is flat and stage-invariant (WAKE 0.76, N3 0.74; Δ(N3−WAKE) = −0.08 [−0.19, +0.03]; N3 steepest in only 11/63), with lower and more variable fit quality (R² 0.84 vs 0.97). The short within-head bipolar is effectively too low-SNR to resolve the aperiodic change—a montage dependence that is itself a methodological finding: the demonstration requires an adequate montage, and the production within-head referencing washes it out.

4.3 The alpha reversal: naive band power points the wrong way. Naive absolute alpha power is highest in wake (EDS ≈ 9.4 µV²) and lower in sleep, so a band-power analysis concludes "most alpha during wake." But a genuine periodic 8–12 Hz peak is present in only 33–54% of EDS recordings per stage (and 2–9% of Maui recordings)—most of the "alpha" band power is aperiodic background. In recordings where wake carries the highest naive alpha yet has no genuine alpha peak, the band-power conclusion is exactly inverted; this full reversal occurs in 7/15 EDS and 35/54 Maui recordings. We do not claim a universal reversal: some subjects show genuine eyes-closed wake alpha coexisting, and we report that heterogeneity rather than averaging it away. Relatedly, the 12.5–12.8 Hz peaks seen naively in N1/N2 register in the sigma window as sleep spindles—a live instance of the canonical-band mislabeling this paper targets.

4.4 Aperiodic-aware features cost nothing in decoding. Under leave-one-subject-out cross-validation on EDS, aperiodic+waveform features (B) decode state as well as naive band power (A). Five-class balanced accuracy: A 0.455 [0.40, 0.51], B 0.441 [0.39, 0.49], union 0.474 [0.42, 0.53]; Δ(B−A) = −0.015 [−0.079, +0.044]. Wake-vs-sleep: A 0.826 [0.77, 0.88], B 0.834 [0.78, 0.88], union 0.857 [0.79, 0.91]; Δ(B−A) = +0.007 [−0.043, +0.052]. The difference CIs straddle zero: controlling for the 1/f background is not a decoding cost, while yielding features that are physiologically interpretable where naive band power conflates oscillation with background. We do not claim B beats A. On the Maui montage neither feature set decodes above roughly chance—an SNR/montage ceiling, not a featurization effect.

5. Discussion

On real ear-EEG, the choice between naive band power and an aperiodic-aware pipeline is not a stylistic preference—it determines the sign of the conclusion. Absolute alpha power ranks wake highest while the oscillatory truth is the opposite, and the deep-sleep "slowing" that a band analysis might attribute to changing rhythms is in fact a change in the aperiodic background. Because aperiodic-aware features decode at least as well as band power, there is no accuracy argument for retaining the confounded metric. The montage result carries a second, product-facing message: aperiodic structure is recoverable from the ear, but only with an adequate inter-electrode geometry; the short within-head bipolar typical of compact hearables is near-white for this purpose. Any aperiodic or 1/f claim from a hearable should therefore specify, and justify, its montage.

6. Limitations

The cleanest demonstration of the alpha confound—identical oscillatory alpha with differing 1/f slope across eyes-open and eyes-closed rest—requires a dedicated resting paradigm we do not currently hold; the present reversal is shown within sleep instead. The two cohorts are not pooled and differ in device and scoring. The decoding analysis is a methods probe rather than a staging benchmark. And the ear montage sees occipital alpha weakly, which contributes to the low wake-alpha peak rate; this is a property of ear-EEG that any alpha claim must acknowledge.

7. Recommended pipeline and conclusion

For hearable EEG we recommend, as a default: parameterize the spectrum into aperiodic and periodic components rather than integrating raw band power; report the aperiodic exponent and offset alongside peak-resolved oscillatory power, frequency, and bandwidth; complement spectral measures with cycle-by-cycle waveform and burst metrics; and specify the electrode montage, because aperiodic structure is montage-dependent at the ear. Adopting these does not cost decoding accuracy, and it prevents the sign errors that confounded band power invites. As consumer neurotechnology makes ever-stronger claims from ever-smaller montages, aperiodic- and waveform-aware analysis is not a refinement but a prerequisite for trustworthiness.