Brain-Guided Audio: Decoding Cognitive and Physiological States from Ear-EEG, from Sleep to Focus

1. Introduction

For nearly a century, the electroencephalogram has been the most direct non-invasive window onto the working human brain. Yet for most of that history it has also been among the least accessible. Recording EEG has meant gelled electrode caps, shielded rooms, and trained technicians; interpreting it has meant a specialist vocabulary of frequency bands and waveforms that is largely meaningless to the person being recorded. The result is a striking asymmetry: a signal generated continuously by every waking and sleeping brain, but legible only to a handful of experts and accessible only in episodic, clinical snapshots.

Two developments are dissolving this asymmetry. The first is hardware. Ear-EEG—electrodes placed in the ear canal or in the concha and around the ear—has matured from a laboratory curiosity into a practical sensing platform that can be integrated into earphones worn comfortably for hours at a time. Because the ear sits close to the temporal lobe and offers a stable, self-positioning anatomical anchor, it can capture a meaningful fraction of the cortical activity recorded by conventional scalp montages, while remaining socially invisible and suitable for everyday use. The second development is interpretive. Advances in signal processing and machine learning now make it possible to map raw neural activity onto behaviorally meaningful states—asleep or awake, drowsy or alert, focused or wandering—in something approaching real time.

Together these create the possibility we develop in this review: that EEG can stop being a thing done to patients and become a thing that serves users. The barrier is no longer measurement alone but translation. A person does not need to know that their occipital alpha power has risen or that their aperiodic exponent has steepened; they need to know that they are getting sleepy, that now is a good time to do deep work, or that their attempt to relax is succeeding. And because hearables are, first and foremost, audio devices, they offer a natural channel for delivering that information back—through sound, music, or haptic cues—and, potentially, for nudging the underlying state itself. We refer to this closed loop, in which decoded brain states drive personalized auditory and haptic feedback that in turn reshapes those states, as brain-guided audio.

This review is organized around a hierarchy of states, ordered roughly by how firmly each is established in the ear-EEG literature. We begin with sleep and wakefulness, where the evidence is strongest and commercial validation is most advanced. We then move to alertness and vigilance—the moment-to-moment fluctuation of arousal that determines when a person is at their cognitive best. We turn next to sustained attention and concentration, where the promise is greatest but the science is most contested, and then to calm and affective self-regulation, the domain most associated with consumer neurofeedback. For each state we ask the same four questions: what is the neurophysiology, what are the EEG and ear-EEG biomarkers, what algorithms and interventions exist, and how mature is the evidence? We then describe the brain-guided audio loop that unifies them, survey the cross-cutting challenges of personalization, real-time processing, and neural-data privacy, and close with a roadmap distinguishing what is ready to deploy from what remains to be proven.

2. Ear-EEG as a measurement platform

Before surveying what can be decoded, it is worth being precise about what ear-EEG actually measures and where it falls short, because these constraints shape every downstream claim. Ear electrodes record the same volume-conducted cortical potentials as scalp electrodes, but from a restricted and idiosyncratic vantage. The recorded signal is dominated by sources near the ear—temporal and, depending on montage, frontal and occipital contributions—and is attenuated for sources farther away. In practice this means that some canonical phenomena transfer cleanly to the ear while others are recoverable only partially or under favorable conditions.

A substantial body of work now demonstrates that the staples of cognitive electrophysiology can be measured from the ear. Alpha and theta rhythms are reliably captured, including occipital alpha that is, somewhat counterintuitively, detectable at the ear despite its distance. Classic event-related and steady-state responses—the eyes-open/eyes-closed alpha modulation, the auditory steady-state response (ASSR), and the steady-state visual evoked potential (SSVEP)—have all been elicited and recorded with ear-EEG. Ambulatory ear-EEG using flexible around-the-ear electrodes has been demonstrated outside the laboratory with consumer-grade acquisition. Crucially for the applications that follow, ear-EEG has supported not only passive measurement but real-time decoding and closed-loop neurofeedback in laboratory settings.

These capabilities come with caveats that recur throughout this review. Ear-EEG montages vary widely—in-ear versus around-the-ear, dry versus gel electrodes, single-ear versus cross-head channel configurations—and these choices materially affect which signals are recoverable; the evidence base for, say, single-ear-channel sleep staging is thinner than for cross-head configurations. The ear is also susceptible to movement and muscle artifact, particularly during the active daytime states most relevant to attention and alertness. And the low channel count characteristic of consumer form factors limits spatial resolution and amplifies the methodological pitfalls of band-power analysis that we treat at length in a companion methods paper. We flag these issues here not to undercut the platform but to calibrate expectation: ear-EEG is remarkably capable for a wearable, and precisely because of that it deserves rigorous rather than promotional treatment.

3. Sleep and wakefulness

Sleep is where ear-EEG is on firmest ground, and it is therefore the natural anchor for a brain-guided audio platform. Human sleep proceeds through a structured progression of stages—wake, the lighter stages N1 and N2, deep slow-wave sleep (N3), and REM—each defined by a characteristic electrophysiological signature: sleep spindles and K-complexes in N2, high-amplitude slow-wave activity in N3, and low-voltage mixed-frequency activity with REM. The clinical gold standard for distinguishing these stages, polysomnography (PSG), is comprehensive but lab-bound, obtrusive, and typically limited to one or two nights, making it ill-suited to the longitudinal, ecological monitoring that everyday users want.

Against this backdrop, ear-EEG has accumulated a robust and compelling evidence base. Sleep has been explored in more than twenty-five ear-EEG studies, many demonstrating automated staging with substantial agreement against simultaneous PSG (five-stage Cohen's kappa on the order of 0.73–0.76). Because the full set of sleep stages is defined largely by neural rather than peripheral physiology, ear-EEG can recover four-stage structure (wake, light, deep, REM) that accelerometry and heart-rate sensing alone cannot. This is the decisive advantage of an EEG-bearing hearable over a wrist wearable for sleep, and it underwrites two immediately deployable applications.

The first is clinical-grade tracking: continuous, at-home staging and the derivation of standard parameters—sleep-onset latency, total sleep time, efficiency, wake after sleep onset—validated against PSG. The second is the smart alarm. Forced awakening at a fixed clock time disregards a sleeper's actual position in the sleep cycle, and waking from deep or REM sleep is associated with pronounced sleep inertia—grogginess and impaired performance that can persist well beyond waking. A device that stages sleep in real time can instead time the alarm to a window of light sleep, reducing inertia. Notably, while discriminating N1 from N2 is difficult, this distinction appears to matter little for inertia, so the comparatively easy wake/light/deep/REM separation suffices for the use case. Real-time constraints here are mild—staging tolerates latency and modest refresh rates—which makes on-device or phone-side implementation straightforward, and the same machinery extends naturally to monitoring daytime drowsiness and sleep inertia, bridging directly to the alertness applications we consider next.

4. Alertness and vigilance

If sleep staging answers whether a person is asleep, alertness monitoring answers a subtler and, for waking life, more consequential question: how well is an awake brain functioning right now? Arousal is not a fixed trait but a continuously fluctuating state, driven by accumulated sleep pressure, circadian and ultradian rhythms, and momentary environmental demands. Because the level of arousal directly shapes cognitive and motor performance, knowing one's own arousal trajectory across the day—and identifying the prime focus times when one is at one's best—has clear practical value, from scheduling demanding work to detecting dangerous drowsiness behind the wheel.

The electrophysiology of arousal is well characterized. Heightened alertness is associated with increased beta activity, whereas rising drowsiness brings characteristic changes in alpha and theta and a measurable shift in the aperiodic component of the spectrum, which has been proposed as a robust electrophysiological marker of arousal level. These neural indices align with the subjective and behavioral measures used to validate them—the Karolinska Sleepiness Scale (KSS) and reaction-time tasks such as the psychomotor vigilance test—providing convenient labels for supervised models. Importantly for a hearable, drowsiness has been measured accurately from ear-EEG, with performance comparable to scalp EEG, and studies have demonstrated real-time drowsiness detection and closed-loop intervention from in-ear electrodes.

A practical pipeline therefore becomes feasible: train a drowsiness classifier offline on existing ear-EEG—labeled by sleep scoring, KSS, or reaction time—using band power, band-power ratios, and entropy features over short windows, then deploy it online to flag low-arousal periods or, conversely, to surface windows of peak alertness. Two honest cautions temper this promise. First, the EEG correlates of drowsiness are less consistent across individuals than those of sleep, arguing strongly for personalization against each user's own KSS and reaction-time data. Second, the daytime states of interest are exactly those in which movement and muscle artifact most threaten ear-EEG signal quality, and clinical conditions such as excessive daytime sleepiness may further alter the picture. Alertness monitoring is thus genuinely feasible but, more than sleep staging, demands individualized models and careful artifact handling.

5. Attention and concentration

Attention is the state users most want to improve and the one the science can least confidently deliver—a tension this section confronts directly. Concentration, or focus, is the capacity to direct mental effort toward a task while suppressing distraction, and in an environment of constant digital interruption the demand for tools that enhance it is large and growing. The neuroscience of attention is correspondingly vast: decades of work tie attentional control to oscillatory dynamics, particularly frontal-midline theta and the alpha rhythm, the latter now understood less as idling and more as an active mechanism of functional inhibition that gates the flow of information through cortex.

Much of this transfers encouragingly to the ear. The theta and alpha rhythms central to attention are well captured by ear-EEG, and the event-related and steady-state responses used in attention paradigms—eyes-open/eyes-closed, ASSR, SSVEP—have been reliably measured from ear electrodes. Most provocatively, several studies report that ear-EEG can decode the focus of auditory attention—which of two competing streams a listener is attending—raising the possibility of using attentional state itself as a neurofeedback signal in a device that is already in the ear and already producing sound. Industry has begun to act on this: commercial focus-estimation from in-ear headphones has been reported.

Yet the gap between this promise and a validated product is wide, and intellectual honesty requires naming it. Attention is not one thing but many—sustained, selective, divided—each with partly distinct neural signatures, and attentional ability fluctuates enormously both within and across individuals. The literature on cognitive training with neurofeedback (CTNF) is methodologically uneven: studies are heterogeneous in task and feedback modality, benefits vary widely between subjects, rigorous designs are scarce, the type and frequency of training needed to produce durable enhancement remains unclear, and—critically—no ear-EEG study has yet been deployed in a genuine CTNF protocol. The best feedback modality, whether sound, music, or vibration, is likewise unsettled. The opportunity is real and well-matched to the hearable form factor, but realizing it will require the kind of rigorous, adequately powered, individualized study that the existing literature has largely lacked—precisely the empirical program our measurement-first focus work is designed to supply.

6. Calm and affective regulation

Calm-on-demand is the application most firmly associated with consumer neurofeedback, and it rounds out the hierarchy by turning the platform from measurement toward self-regulation. A calm state—tranquility, relaxation, the absence of agitation—is both widely desired in high-stress modern life and, encouragingly, well represented electrophysiologically: alpha and theta rhythms are robustly associated with relaxed states, and a large literature links EEG signatures to meditation and to the benefits of mindfulness for stress, anxiety, sleep, and executive function. Closed-loop neurofeedback of frontal-midline theta, in particular, has been used to train focused-attention meditation, and intensive neurofeedback protocols have been applied to affective regulation. Because alpha and theta are exactly the rhythms ear-EEG captures well, the biological substrate for an ear-based calm-neurofeedback feature is largely in place.

The implementation pattern mirrors the others: monitor alpha and theta band power, along with band-power ratios and network-level connectivity or coherence features, in real time; modulate auditory or haptic feedback to reward movement toward a calmer state; and track progress with subjective instruments such as meditation-quality questionnaires and with cognitive probes. Real-time requirements are again undemanding—a refresh on the order of a few hertz suffices—so on-device implementation is practical, with heavier models optionally offloaded to the cloud.

The honest limitations parallel those for attention. Meditative experience is subjective and difficult to measure precisely; EEG patterns during meditation are reported inconsistently across studies and practices; the optimal feedback modality is unresolved; and, as with attention, no ear-EEG study has yet been run inside a meditation protocol. There is, however, a constructive path that doubles as a research contribution: because oscillatory peak frequencies vary substantially across individuals, anchoring each user's alpha and theta bands to their own spectrum—rather than to fixed canonical ranges—offers both a route to more reliable feedback and a methodological differentiator from generic meditation apps, a theme we develop in our companion analysis of aperiodic and waveform-aware ear-EEG metrics.

7. The evidence base: prior systematic reviews and meta-analyses

Because much of the primary literature has already been aggregated by others, we anchor the evidence claims of this review on existing systematic reviews and meta-analyses rather than on isolated studies. Read together, they mirror the hierarchy developed above: strongest for sleep, intermediate for alertness, and explicitly cautionary for attention and calm neurofeedback.

Sleep. Systematic reviews of wearable-EEG sleep staging report that EEG-bearing wearables are the most accurate consumer modality and the only one able to resolve all sleep stages, with deep sleep (N3) most reliably detected and N1 the hardest stage. Reported agreement with polysomnography is moderate-to-substantial — a meaningful benchmark given that expert PSG inter-scorer agreement itself is only on the order of 80–85%.

Alertness and vigilance. Reviews of EEG-based drowsiness detection conclude that even basic spectral features reliably index drowsiness, including from low-channel and in-ear systems. This is the aggregate basis for treating ear-EEG alertness monitoring as feasible, subject to the personalization and artifact caveats noted above.

Attention and concentration. The cognitive-training-with-neurofeedback literature has been meta-analyzed repeatedly, with sobering and consistent results: effects on executive function and memory are inconsistent, modest, and strongly moderated by individual differences and baseline ability. This meta-analytic record — not our caution alone — is why we classify attention enhancement as promising but unproven.

Calm and affective regulation. A 2025 meta-analysis of consumer-grade neurofeedback paired with mindfulness meditation (Treves et al. 2025; 16 randomized controlled trials, 763 participants) found only a small effect on psychological distress and, critically, no conclusive evidence that users can learn to modulate the targeted brain activity; the authors state that claims of brain modulation by consumer devices "are not currently supported." We cite this directly as a guardrail against overclaiming calm-on-demand features.

8. From measurement to intervention: the brain-guided audio loop

The four states surveyed above share a common architecture that, taken together, defines the brain-guided audio loop. In its passive form the loop is open: sense neural activity at the ear, decode it into a state estimate, and render that estimate back to the user as legible feedback—a sound that signals deepening focus, a tone that marks the onset of drowsiness, a soundscape that confirms a slide toward calm. In its active form the loop closes: the feedback is chosen not merely to inform but to move the state in a desired direction, and the next measurement evaluates whether it did. A hearable is uniquely suited to host this loop because the same device occupies the ear as both sensor and actuator, and because modern mobile platforms expose fine-grained control of the output channel—on Apple devices, the AVFoundation and Core Haptics frameworks allow precise modulation of audio and haptic feedback in response to a continuously updated state estimate.

It is useful to decompose the closed loop into three empirical claims, each of which must be established independently and in order. The first is identifiability: can the target state be read out from ear-EEG in real time and at the individual level? The second is malleability: can feedback actually shift the underlying neural feature, rather than merely correlating with it? The third, and most demanding, is benefit: do feedback-driven changes in the neural feature translate into changes the user cares about—better sleep onset, sustained focus, faster recovery of calm? The evidence base thins sharply across these three steps. For sleep and drowsiness, identifiability is well established and malleability is plausible; for attention and calm, even identifiability from ear-EEG in ecological conditions remains to be proven at the level a product would require. Conflating the three—treating a real-time readout as if it were a demonstrated intervention—is the central way consumer neurotechnology overstates itself, and a discipline that keeps them separate is the surest guard against doing so.

9. Cross-cutting challenges

Four challenges cut across every state and every link of the loop. The first is personalization. The neural signatures of drowsiness, attention, and calm vary substantially across individuals, and oscillatory peak frequencies shift with age, cortical location, and moment-to-moment state; fixed canonical band definitions and population-trained models will therefore underperform for many users. Robust systems will need to anchor spectral features to each user's own physiology and to adapt online to the non-stationarity of brain signals over time. The second is real-time processing. Encouragingly, the states of interest here do not demand aggressive low-latency pipelines—staging tolerates latency, and neurofeedback refresh rates on the order of a few hertz are sufficient—so the computational burden is modest and largely compatible with on-device or phone-side inference, with heavier models reserved for the cloud.

The third challenge is privacy and security, and it is more serious than it first appears. The richness that makes ear-EEG valuable also makes it sensitive: neural signals can carry incidental information about age, sex, and pathology beyond the state a feature is designed to estimate. Continuous, all-day brain recording from a consumer device therefore raises data-governance questions that the field cannot treat as an afterthought, and that a credible platform should address by design—through on-device processing, data minimization, and explicit user control. The fourth challenge is individual differences in efficacy. Even where an intervention works on average, neurofeedback benefits are notoriously heterogeneous, with substantial proportions of non-responders; honest products will need to measure and disclose for whom, and under what conditions, a feature actually helps, rather than assuming uniform benefit.

10. Conclusion and roadmap

Ear-EEG has crossed a threshold. It is no longer in question whether meaningful brain activity can be recorded comfortably and continuously from the ear; the open questions are now about which states can be decoded reliably enough to act on, and which interventions actually help. On this view the landscape sorts cleanly. Sleep staging and drowsiness detection are ready: the evidence is deep, the biomarkers are stable, and the real-time requirements are forgiving, making clinical-grade sleep tracking, smart-alarm timing, and daytime alertness monitoring achievable today. Attention and calm neurofeedback are promising but unproven: the biological substrate and the form factor align, yet no ear-EEG study has been run inside a genuine focus- or meditation-training protocol, and the broader neurofeedback literature is methodologically uneven.

The path forward is therefore neither uncritical enthusiasm nor dismissal but disciplined validation. Closing the gap will require studies that establish identifiability, malleability, and benefit in that order; that personalize to the individual rather than the population; that quantify heterogeneity of response rather than averaging it away; and that hold consumer neurotechnology to the same evidentiary standard as the clinical electrophysiology from which it descends. The reward for that discipline is substantial: a hearable that does not merely measure the brain but makes it legible, and that uses sound to help its wearer sleep, focus, and recover calm. That is the promise of brain-guided audio—and, increasingly, an empirical program rather than a metaphor.

Representative biomarkers by state