Attention has two sides: what it lands on, and how much of it there is.
In Making visual attention programmable, the Kyma band read the first. As cards flashed one at a time, the brain’s response a third of a second after each one picked out the card the wearer had privately chosen, 15 times in 17. That is attention as selection: which thing, out of several, matters.
The second side changes more slowly and never stops. Through a work session, how much attention we give the screen rises and falls: deep in a task, then drifting, then back. Software has no view of it. It knows what we click, not how present we are while we do it. Here we read that level continuously from the same band, at the back of the head, and connect it to a playlist that responds when it drops.
A rhythm that steps aside when you look
In 1929 Hans Berger recorded a wave near ten cycles per second over the back of the head that faded whenever attention turned to something [1]. That wave, the alpha rhythm, is now understood as a gate on the visual system. It is strong where visual input is being held back and falls away where the cortex is processing what we see [2][3]. It tracks attention closely enough that its level just before a faint target predicts whether the target will be seen [4].
Worn at the back of the head, as in the card experiment, the band’s electrodes sit over that cortex. The rhythm is large enough to see in the raw recording, with no averaging. Below are three seconds from each half of the task described next: watching a moving dot, and steering a cursor after it.
To turn the rhythm into a number, Studio takes the last four seconds of signal and measures how strong it is at each frequency, from slow waves to fast ones. The brain’s activity forms a smooth background that falls off toward higher frequencies, and the alpha rhythm stands out of it as a bump near ten cycles per second. The height of that bump above the background is the measure. Separating a rhythm from the background this way, rather than reading the raw power in a frequency band, is a widely used approach in neuroscience, because a shift in the background can otherwise pass for a change in the rhythm [5]. It matters here because contact, movement and general arousal raise or lower every frequency at once, and they leave the bump’s height mostly alone.
A tall bump means the visual cortex is idling; a smaller one means it is busy with what you are looking at. Flipped so that up means more attention, and compared with your own usual range, that is the attention level, recomputed every second.
Testing that it tracks attention
To measure attention, the screen has to stay the same while attention changes. In ordinary work that never happens: a busier screen changes the visual cortex’s activity on its own. So we built a task where the pixels are identical and only the wearer’s involvement differs.
A dot glides around a box on the screen at a steady speed, turning smoothly. For twenty seconds the wearer keeps the cursor on it with the mouse, which takes continuous looking: anticipating where the dot is going and correcting when the cursor falls behind. After a three-second rest, the screen replays that exact run, the same dot on the same path with the cursor moving just as it did, while the wearer watches with their hand off the mouse. Six of these pairs make a five-minute run.
Averaged over every pair from three sessions, the pattern is the same each time.
The rhythm drops within a second of taking control, stays down for the run, and returns during the replay. After a first session suggested the measure, we fixed the test in advance and ran it in two new sessions: lower alpha while steering in 14 of 16 pairs (1.2 dB, p = 0.001).
The change has a precise address. It begins at the wearer’s own alpha peak and fills the band above it, upper alpha, while the band below does not move. That is the part of the rhythm that tracks the demands of a specific task, rather than general alertness [6] (Figure 6, in the notes below).
Measured there, the signal resolves quickly. Two seconds of it told steering from watching about four times in five; sixteen seconds did it 19 times in 20.
The same holds away from a screen. In a hands-on task, the wearer sorted coloured chips by hand in 45-second blocks, alternating with blocks of rest with hands still and eyes on the same table. Moving the head and arms adds slow activity to the recording, so here alpha is measured above the fitted 1/f background of each window, which cancels that shift. The rhythm fell while sorting in all five pairs with enough clean data, and 92% of the signal stayed usable while the hands worked.
Placement matters for this. With the band moved to the forehead for the same task, there was no alpha rhythm to read; what the forehead recorded was mostly the eyes and head moving.
What it measures, and what it doesn’t
The level is not focus in general. Thinking inward, imagining or doing mental arithmetic, raises alpha because the brain shuts out what it sees [7], and so do drowsiness and closing your eyes. It measures how much the visual system is engaged with what is in front of you, which is what most software asks of us.
Movement is the other thing to watch for. In both tests the hands move more in the attentive condition, and movement shifts the whole spectrum at once. Measuring alpha against its background is built to cancel that shift, and muscle activity, which movement does raise, is tracked separately rather than counted as attention.
Letting software respond
Kyma Studio turns this into a live value. Every second it measures the alpha bump, in its upper half, against the wearer’s usual range for this wear, and publishes the result through the Kyma API, where apps built in Studio, programs on the same computer and AI assistants can all follow it.
The first app built on it is a playlist, and what it does is simple. When your attention drops below its usual range and stays there for half a minute, it changes the song, picking the next one from your own playlist.
Figure 4 is fourteen minutes of ordinary work with it running: Call Me, Latch, Electric Feel. Nine minutes in, attention slid below its usual range, the playlist switched to Aerodynamic, and the phone started it 1.8 seconds after the decision. That is the moment in the clip at the top.
Whether the switch helps is a separate question. Attention usually comes back within a minute or two on its own, even if nothing changes. So if the playlist switched songs at every dip, attention would recover afterwards and it would look like the switch worked, whether it did or not. To tell the difference, the playlist flips a coin at each dip: sometimes it switches, sometimes it keeps the song playing. Comparing the two shows what the switch itself adds. We are collecting that data now, every time the playlist runs, and will publish what it shows once there is enough to tell.
What each song teaches it
After every song, the playlist scores it: how attention moved while it played, against the slow trend of the session. Then it breaks the song into properties it can carry over to songs it hasn’t played yet: tempo, energy, mood and whether it has vocals. If faster songs kept your attention up, every fast song in your playlist would move up the queue, including ones it has never tried.
Swipe sideways to see every column →
After 46 plays over ten sessions, none of the four makes a clear difference yet. Fast and slow songs, calm and intense, sad and happy, sung and instrumental are spread from the top of the list to the bottom. The same song can also land in very different places: Call Me has played seven times, from sixth-best of the 46 plays to fifth-worst. What you are doing while a song plays seems to matter at least as much as the song. As plays add up, a kind of song that really helps will start to gather at the top.
The same signal could go well beyond music. It could hold notifications while attention is high, mark the passages of a long document where it fell away, pace a lesson, or let an AI assistant wait for a natural pause before it interrupts.
What this adds up to
Visual attention is a signal software can now use in two ways. Moment to moment, the response to each image says which one matters. Over minutes, the alpha rhythm says how much attention is there at all. The first gives an interface feedback about options; the second tells it about the person using it.
At launch, Kyma Studio and the Kyma API include visual attention in both forms, with the adaptive playlist as a reference app. Founding Batch members will be the first to build with them. Learn about early access.
Software already knows what we do. Now it can know when we are paying attention, and we are measuring whether responding to it makes a difference.
How the signal holds up outside a lab
We ran every frequency band, and the ratios commonly sold as attention or engagement scores, through the same steering-versus-watching comparison. The effect starts at the wearer’s own alpha peak and fills upper alpha (Figure 6), in 20 of 21 pairs. Ratios that include alpha, such as beta / (alpha + theta) [8], only inherit it and come out weaker, and the other bands show nothing at the back of the head.
In everyday wear, contact and general arousal shift every band at once. Studio measures alpha against the bands around it, which cuts that leakage about threefold, and when the bands outside alpha jump, as when the band is adjusted, it rebuilds the wearer’s usual range from that point. Its live value matches an independent offline analysis at r = 0.99.
About the recordings and analysis
All recordings are from one team member wearing the Kyma band at the back of the head (forehead where stated). Power comes from four-second multitaper windows every second, averaged across electrodes; a window counts when its 4–45 Hz peak-to-peak range is under 150 µV on every electrode. The live value is upper alpha (10–14 Hz) against the mean of theta and low beta, relative to the median of up to ten minutes before. Dashboards show one-minute means relative to the session median. Raw traces are from the pair with the median effect, filtered 2–30 Hz. Figure 2 uses the first run of session 2: four-second Hann windows every second; the background is a straight line fitted to log power against log frequency (2–7 and 15–35 Hz), and the level is the mean height above it over 8–13 Hz, sign flipped, relative to the run’s median and averaged over three seconds. That is the measure used for the sorting clip; Studio’s live value measures upper alpha against the mean of the bands on either side, a simpler version of the same idea.
The steering-versus-watching test was fixed in advance and decided on sessions 2 and 3 (16 pairs, exact sign-flip test); session 1 suggested the measure and is pooled only in descriptive figures. Separation by seconds is the within-pair probability that a stretch of steering has less upper alpha than a stretch of watching. The index comparison, the broadband analysis, the contact rule and the trigger replay are exploratory, and the measure and rules were set after seeing these data and before any outcome of the randomized test was analysed.
The hands-on sorting session and its forehead comparison are single sessions of six 45-second pairs each, and their analyses are exploratory. There, alpha is power above a straight-line fit to log power against log frequency (2–7 and 15–35 Hz) in clean four-second windows, which removes broadband shifts from movement; the trace in the clip is that measure, sign inverted and smoothed over six seconds.
In Figure 5, each song’s attention is the mean of its plays; energy and mood (valence) are Spotify’s 0–1 measures. Across the 46 plays, rank correlations of attention with tempo, energy and mood are all within ±0.1 (p > 0.5), sung and instrumental songs differ by 0.1 dB (p = 0.63), and songs differ from one another no more than random stretches of the same length in the same sessions (p = 0.6). The same checks on every other band and common index (theta, beta, gamma, muscle, the alpha peak frequency, the slope of the 1/f background, beta / (alpha + theta), theta / beta and left–right alpha balance) found no link to any song property after correcting for the number of comparisons, and no measure on which songs differed more than random stretches did. Song scores are upper alpha from 15 seconds into each song to its end, against a ten-minute moving trend of the same session and sign flipped; the comparison with random stretches uses the same measure at random times in the same sessions. In Figure 4 the switch was the “switch” side of the coin flip. The playlist reorders a private copy of the wearer’s own playlist; the original is never changed. Music played on a phone so that headphones stayed off the computer’s radio. The clips carry no audio. Kyma is not affiliated with Spotify.
References
- Berger H (1929). Über das Elektrenkephalogramm des Menschen. Archiv für Psychiatrie und Nervenkrankheiten 87:527–570.
- Pfurtscheller G, Lopes da Silva FH (1999). Event-related EEG/MEG synchronization and desynchronization: basic principles. Clinical Neurophysiology 110:1842–1857.
- Jensen O, Mazaheri A (2010). Shaping functional architecture by oscillatory alpha activity: gating by inhibition. Frontiers in Human Neuroscience 4:186.
- Thut G, Nietzel A, Brandt SA, Pascual-Leone A (2006). α-Band electroencephalographic activity over occipital cortex indexes visuospatial attention bias and predicts visual target detection. Journal of Neuroscience 26:9494–9502.
- Donoghue T, Haller M, Peterson EJ, et al. (2020). Parameterizing neural power spectra into periodic and aperiodic components. Nature Neuroscience 23:1655–1665.
- Klimesch W (1999). EEG alpha and theta oscillations reflect cognitive and memory performance: a review and analysis. Brain Research Reviews 29:169–195.
- Ray WJ, Cole HW (1985). EEG alpha activity reflects attentional demands, and beta activity reflects emotional and cognitive processes. Science 228:750–752.
- Pope AT, Bogart EH, Bartolome DS (1995). Biocybernetic system evaluates indices of operator engagement in automated task. Biological Psychology 40:187–195.