Toward Focus Decoding from Ear-EEG: A Shipped Measurement Pipeline and a Corpus-Collection-at-Scale Design

1. Introduction: measure first, claim later

Focus is the cognitive state people most want to improve, and the one consumer neurotechnology can least defensibly claim to deliver. The appetite is real — the market for attention and brain-training tools is large and growing — and the temptation to ship a confident "focus score" is correspondingly strong. But a number on a screen is only worth what its validation supports, and across the consumer neurofeedback literature that support has been thin: meta-analyses of cognitive training with neurofeedback report inconsistent, modest effects strongly moderated by individual differences. A credible focus product cannot begin with the score. It has to begin with measurement, and earn the score from data.

It is useful to decompose any closed-loop focus claim into three questions that must be answered in order. Identifiability: can the target state be read out from ear-EEG in real time, at the level of the individual? Malleability: can feedback actually move the underlying neural feature rather than merely correlate with it? Benefit: do feedback-driven changes translate into outcomes the user cares about — sustained attention, fewer lapses, work that gets done? Conflating these three — treating a real-time readout as if it were a demonstrated intervention — is the central way consumer neurotechnology overstates itself. This paper is about the first question only. We have built the instrument that lets us answer it with data collected at scale, and we are explicit that the second and third questions remain open.

2. What focus is, neurophysiologically

Attention is not one thing. Sustained, selective, and divided attention recruit partly distinct networks, and attentional ability fluctuates substantially both within a person across a day and between people. Two electrophysiological signatures, however, are robust enough and transfer well enough to the ear to anchor a measurement pipeline.

2.1 Alpha as functional inhibition. The alpha rhythm (≈8–12 Hz) is no longer understood as cortical idling. In the gating-by-inhibition framework, alpha reflects active suppression of task-irrelevant cortex: it rises where processing must be inhibited and falls where processing is engaged. Engaged, externally directed attention is therefore associated with a reduction in posterior and temporal alpha relative to a relaxed baseline. This is the basis for treating a fall in relative alpha as a candidate index of attentional engagement — with the important caveat that the meaning of alpha is state- and montage-dependent and must be read against an individual's own baseline rather than a fixed canonical band.

2.2 The aperiodic 1/f exponent as an arousal marker. Beneath any oscillatory peaks, the EEG power spectrum carries a broadband, aperiodic 1/f-like background whose slope (the exponent, χ) tracks the balance of synaptic excitation and inhibition and varies systematically with arousal: it flattens with heightened arousal and steepens with drowsiness and deep sleep. The aperiodic exponent has been proposed explicitly as an electrophysiological marker of arousal level in humans. For a hearable this marker has a decisive practical virtue: because it is a property of the spectrum's shape rather than its absolute magnitude, it is comparatively robust to the slow changes in electrode contact and impedance that corrupt absolute band-power features over a long session, and a steepening exponent doubles as a drowsiness flag — the opposite end of the focus continuum.

2.3 Markers we deliberately down-weight. Frontal-midline theta is a classic correlate of sustained attention and cognitive control on the scalp, and theta-to-beta and alpha-to-theta ratios are common in the literature. We log relative beta (βR) as a supporting engagement marker, but we treat theta-dependent ratios as observational only on this platform: theta recorded at temporal ear electrodes is substantially contaminated by non-neural and ear-canal sources, so a theta-driven focus claim from the ear would be built on the least trustworthy part of the signal. Naming what we do not rely on, and why, is part of an honest measurement design.

3. The shipped measurement pipeline

The focus pipeline runs on the same comfortable, continuously wearable ear-EEG platform that NextSense has validated for overnight sleep staging, and reuses the same product-neutral signal primitives so that focus and sleep data remain directly comparable. On each one-second tick it recomputes features over a trailing six-second window, independently for each earbud, and emits one structured record per usable side. There is, by design, no state machine and no feedback: the output is the recorded corpus, not a decision.

3.1 Acquisition and artifact gating. Each tick selects the cleanest available side rather than committing to one ear, because the focus algorithm is not finalized and no side is yet preferred. A window is admitted only if it survives explicit gates: runs of dropped Bluetooth samples, railed amplifier values, an extreme-value fraction above ten percent of the window, and a total-power cap that rejects electrode-noise bursts. Ticks that fail are logged as artifact-skipped with their reason rather than silently dropped, so the corpus records its own coverage and the eventual detector can be trained and scored only on data of known quality.

3.2 Features logged per tick. From the admitted window we compute a Welch power spectral density (256-sample Hamming segments, 50% overlap), relative band powers, and a parameterized aperiodic fit yielding χ and its fit quality; only fits meeting a minimum goodness-of-fit are folded into the smoothed exponent, so a poor fit never masquerades as a measurement. Both αR and χ are logged raw and as heavily smoothed trends (an exponential moving average with a ~34-second half-life), so a single noisy window cannot move the signal. In parallel we extract inertial features from each earbud's gyroscope — energy and spectral shape over the window, plus a micro-burst detector — and a person-level restlessness measure read directly as seconds of movement in the last minute. Every tick is stamped with the coarse activity segment and a finer phase label so the corpus can be sliced by what the user was actually doing.

3.3 What the system does not do. The shipped system computes no focus score, makes no focus claim to the user, and runs no closed loop. Per-tick records are developer-only telemetry during corpus collection; consumer ticks remain on-device. The pipeline is versioned at 0.x precisely because it is feature extraction and logging, not detection: the first validated on-device baseline and detector is the explicit 1.0.0 milestone, and we will not cross it on intuition.

4. The corpus and its labels

A marker is only as useful as the labels it can be validated against, and this is where a shipping consumer device has an advantage no laboratory can match: scale, and naturalistic context. The focus experience is built around concrete, loggable behavior, and each element contributes a label to the corpus. A structured cognitive focus check provides an objective behavioral probe of attentional performance that can be time-locked to the EEG markers. Work-and-break interval structure (a Pomodoro pattern) segments each session into intended-focus and intended-rest blocks. A post-session survey captures the user's own sense of how the session went. And an optional distraction shield, which lets a user block chosen apps during a focus block, generates behavioral events — attempted interruptions — that are themselves a real-world correlate of wavering attention.

Read together, these turn every focus session on every consenting device into a labeled training and validation example: a stream of αR, χ, and inertial features paired with task structure, an objective performance probe, a self-report, and a behavioral record of distraction. This is the data NextSense can collect at a scale and in a diversity of real settings — desks, commutes, homes — that controlled studies cannot approach. It is the foundation on which a focus detector can be made trustworthy rather than merely plausible.

5. From corpus to a validated focus score

The analysis plan follows directly from the neurophysiology. Because both alpha and the aperiodic exponent vary substantially across individuals, and because oscillatory peak frequencies shift with person, age, and cortical site, the candidate focus index is defined within-subject: a rest baseline is compared against task engagement, and the index combines αR falling below that baseline with a flattening χ, anchored to the user's own spectrum rather than to fixed canonical bands. The same principle — that anchoring spectral features to each individual's physiology is a prerequisite for reliable readout, not a refinement — is developed in our companion analysis of aperiodic and waveform-aware ear-EEG metrics, which also documents a montage dependence relevant here: aperiodic structure is recoverable from the ear, but more cleanly with an adequate inter-electrode geometry than with the shortest in-ear bipolar.

Validation will be behavioral and held-out. The candidate index will be tested against the objective focus-check performance and the post-session reports under leave-one-subject-out cross-validation, so that no data from a held-out user ever informs the model evaluated on that user, with effect sizes and confidence intervals reported rather than single accuracy figures, and results broken out by individual responsiveness rather than averaged across a population that mixes strong and weak responders. The analysis plan and acceptance criteria will be pre-registered before any focus score is exposed to users. Only a detector that clears those criteria becomes the 1.0.0 milestone.

6. What we do not claim — yet

We do not claim a validated focus detector; none exists in the shipped system today. We do not claim that ear-EEG resolves focus as cleanly as a research scalp montage: the ear sees posterior alpha weakly, and the daytime states of interest are exactly those in which movement and muscle artifact most threaten signal quality, which is why inertial restlessness is logged alongside every EEG tick rather than as an afterthought. We do not claim malleability or benefit — that focus can be trained, or that training helps — because the neurofeedback literature does not yet support such claims and no ear-EEG study has been run inside a genuine focus-training protocol. What we claim is narrower and, we think, more valuable: a shipped, artifact-gated instrument that measures two well-motivated markers honestly, and a design that turns ordinary use into the labeled, real-world corpus from which a truthful focus score can eventually be earned.

7. Conclusion

The trustworthy path for consumer focus technology is not a better-marketed score; it is a better-validated one, and validation is a data problem before it is an algorithm problem. By separating measurement from inference, shipping only the measurement, and instrumenting real use to build a labeled corpus at scale, NextSense is positioning the focus question to be answered empirically rather than asserted. The asset is not a number. It is the corpus, and the discipline to wait for it.