An objective ADHD screen, in five minutes.
Getting evaluated for ADHD takes months of waiting and about $5,000 of questionnaires that measure nothing. Excite measures the eye movements the disorder actually dysregulates, and returns a probability score in under 30 seconds.
No signup. Your webcam, about a minute, a scored report at the end.
- Accuracy
- 96%
- 5-fold CV, n=50
- Live sites
- 7
- UCSB + 5 clinics
- Participants
- 75
- Full pipeline
- To a score
- <30s
- After capture
- Saccades
- 0.0/s
- Fixation
- 0ms
- Peak vel.
- 0°/s
- Pupil σ
- 0.00mm
A simulation, driven by the biomarker distributions measured in our cohort. Toggle the profile: the ADHD trace shows shorter fixations, more corrective saccades, and off-target intrusions. The real system captures this at 200 Hz.
The first step toward an ADHD diagnosis is a questionnaire, and it measures nothing.
Every decision downstream — medication, accommodations, a diagnosis that follows someone for life — rests on a form that asks people to rate their own attention from memory. It is subjective at the point of collection, discards severity at the point of scoring, and cannot be repeated to check whether anything changed.
Standard of care
What a patient goes through today.
- Wait
- 3–9 months for an evaluation slot
- Cost
- ~$5,000 for a full workup, often out of pocket
- Instrument
- Self-report and observer rating scales
- Output
- A binary yes or no, with no measure of severity
- Portability
- None. Change providers and you start over
Objective screen
ExciteWhat the same first step looks like with Excite.
- Wait
- Fits inside an appointment the patient already has
- Cost
- $75–150 per screen
- Instrument
- Involuntary oculomotor biomarkers, sampled at 200 Hz
- Output
- A continuous probability score, not a label
- Portability
- A structured record that travels with the patient
The science to fix this has been settled for decades. What has been missing is a product someone can actually run in a fifteen-minute appointment.
ADHD is a disorder of attention. The eye is where attention gets written down.
The prefrontal–striatal circuits that ADHD dysregulates are the same circuits that plan, time and inhibit eye movement. This is not a new finding. The oculomotor literature has been consistent for three decades — the measurements just never left the lab.
Saccadic velocity & latency
ms · deg/sHow long the eye waits before jumping to a new target, and how fast it gets there. Both are set by frontal-eye-field inhibition, one of the circuits ADHD disrupts.
- CONTROL
- Short, tightly clustered latencies
- ADHD
- Longer and far more variable
Fixation stability
BCEA, deg²The area gaze wanders over while it is supposed to be holding still. The most direct read available on sustained attention.
- CONTROL
- Compact fixation area
- ADHD
- Wider dispersion during hold
Microsaccade rate
HzInvoluntary flicks of a degree or less. Their rate and direction track where attention is allocated, including attention the subject never reports.
- CONTROL
- ≈1.3 per second
- ADHD
- Elevated, less inhibited
Gaze entropy
bitsHow predictable the scanpath is. A distractible viewer produces a higher-entropy, less task-structured path across the same stimulus.
- CONTROL
- Structured, task-driven
- ADHD
- Frequent off-target excursions
Pupil dynamics
mmPupil diameter tracks locus-coeruleus noradrenergic tone — the arousal system that stimulant medication acts on.
- CONTROL
- Consistent phasic response
- ADHD
- Blunted and more variable
None of these are voluntary.
Every incumbent measures task performance — reaction times, errors, motor activity. That is behaviour, and behaviour is fakeable in both directions. Adult ADHD assessment has a documented malingering problem, and with stimulant prescribing under DEA scrutiny, clinicians increasingly need evidence a patient cannot consciously produce.
You can decide to answer a questionnaire differently. You cannot decide to have a normal microsaccade rate.
The science is old. The reason it is buildable is new.
- The hardware got cheap
- Research-grade eye tracking used to mean a tower-mounted rig in a university lab. A Quest Pro and a pair of Neon glasses cost under $8,000, fit in a backpack, and sample at the same 200 Hz.
- The models got good enough
- Pre-training a transformer on 12,000+ unlabelled gaze recordings means the ADHD fine-tune needs hundreds of labelled patients rather than tens of thousands. That is the difference between a research programme and a company.
- The demand got urgent
- Adult ADHD evaluations have surged, waitlists have stretched to months, and telehealth stimulant prescribing is under DEA scrutiny. Clinicians increasingly need evidence a patient cannot produce on purpose.
Five minutes of looking at dots, then a number.
The patient does almost nothing. Everything difficult happens in the four stages between the headset and the score.
Stimulus
Meta Quest Pro
A five-minute, two-part sequence: the first half probes sustained attention, the second probes hyperactivity and impulsivity. Delivered in VR so every test is identical — same luminance, same viewing distance, no ambient distraction to explain away a result.
- 2-part protocol
- 5 min total
- Driven over XARP
Capture
Pupil Labs Neon
Binocular gaze and pupil diameter recorded at 200 Hz through the Neon Companion app and Pupil Cloud. That sampling rate is what makes microsaccades and saccade onset resolvable at all; a webcam at 30 Hz cannot see them.
- 200 Hz binocular
- Pupil + gaze
- Scene camera
Extraction
OpenCV pipeline
Custom pupil detection with ellipse fitting and corneal glint tracking, then five biomarkers computed frame by frame into a time series the model can read.
- Ellipse fit
- Glint tracking
- 5 biomarkers
Inference
Transformer encoder
Three layers, four heads, d_model 64. Pre-trained on 12,000+ GazeBase recordings to learn general oculomotor dynamics, then fine-tuned on ADHD-labelled clinical data. Runs server-side on Azure and returns a continuous probability in under 30 seconds.
- PyTorch
- Azure inference
- <30s to score
What the clinician gets back.
A continuous probability with a confidence interval, and every biomarker that produced it plotted against an age-matched reference range. A clinician can see which signal drove the score and whether the case sits near the boundary or nowhere near it.
Excite screening report
Two-part attention and impulsivity protocol
Oculomotor profile is consistent with ADHD, driven mainly by fixation instability and prolonged saccade latency. Score sits well above the decision threshold rather than beside it.
A screening result, not a diagnosis. Intended to inform a clinical evaluation by a qualified provider.
- Saccade latency
- 268 msHigh
- Peak saccadic velocity
- 331 °/sHigh
- Fixation stability (BCEA)
- 1.92 deg²High
- Microsaccade rate
- 1.7 HzNormal
- Gaze entropy
- 4.31 bitsHigh
- Pupil variability
- 0.34 mmHigh
Reference rangeOutside range
Sample report. Values are representative of a positive screen, not a real patient.
Two ways to run it, and we are honest about the difference.
A webcam cannot resolve a microsaccade. Pretending otherwise would be the fastest way to lose a clinician, so we ship both tiers and label which one the evidence applies to.
Clinical rig
DeployedQuest Pro for stimulus, Neon glasses at 200 Hz for capture. This is the configuration running at all seven sites and the one the validation data comes from.
- 200 Hz binocular gaze and pupil
- Controlled VR stimulus, identical across sites
- Full five-biomarker feature set
- ~$7–8k of hardware, amortised, never sold outright
Webcam tier
Live in the browserThe same model served over a standard webcam. Far lower fidelity than the Neon rig and not clinical-grade — but it means anyone can see what the product does, and it is the path past a $7k per-site setup.
- Runs on any laptop or phone camera
- Same Azure inference endpoint
- Reduced biomarker set at 30 Hz
- For demonstration, not for diagnosis
96% accuracy on a balanced clinical cohort.
Not a benchmark score on a public dataset. Fifty people at UC Santa Barbara, half with a prior clinical ADHD diagnosis and half without, run through the full pipeline under two IRB-approved protocols.
- Accuracy
- 96%
- 5-fold cross-validation
- Sensitivity
- 96%
- 24 of 25 ADHD correctly flagged
- Specificity
- 96%
- 24 of 25 controls correctly cleared
- Cohort
- 50 participants, balanced 25 ADHD / 25 control
- Site
- UC Santa Barbara, two IRB-approved studies
- Label
- Prior clinical ADHD diagnosis
- Method
- 5-fold cross-validation, subject-level splits
- Capture
- Pupil Labs Neon at 200 Hz, VR-delivered stimulus
- Status
- First draft written with UCSB's HAX Lab
Confusion matrix
n = 50One miss in each direction. The model returns a continuous probability rather than a label, so a clinician sees how close to the boundary a borderline case actually sits — information the current standard of care throws away entirely.
A cross-validated result on a single-site cohort, with a paper in draft and submission targeted within weeks. Twenty-five further patient evaluations have since been run across the five partner clinics, with results consistent with these baselines.
Not peer-reviewed yet, not multi-site, and not a regulatory claim. Excite is a screening tool. It is not FDA-cleared and is designed to inform a clinical evaluation, never to replace one.
A screener is only as good as the room you point it at.
This is the part most digital-diagnostics companies get wrong, and it is not a science problem. The same model, at the same accuracy, is either a useful clinical instrument or a liability depending entirely on who walks through the door.
Where we deploy
94 actually have ADHD. 6 do not.
94.1%
At a 7% general-population base rate, more than a third of everyone we flag would not have ADHD. Screening a whole school that way does not find children who need help — it manufactures false alarms and hands the institution a liability.
In a self-referred clinical population, where prevalence runs above 40%, that falls to roughly one in eighteen. Identical model, identical accuracy, completely different product. That is why all five pilot sites are ADHD clinics.
Live at seven sites.
Seventy-five people have gone through the full pipeline end to end. The clinics are unpaid pilots while the validation study finishes — the plan is to convert them once the results are published, because that is the thing they have told us they need to justify a purchase.
UC Santa Barbara
Two IRB-approved studies. Source of the validation cohort.
5 partnered ADHD clinics
Santa Barbara. Real patients, running since June.
No revenue yet. Five clinics agreed to give an unproven tool their patients' time, which is a stronger signal than anything a customer will say in an interview.
Seven months.
None of it full-time yet.
- Jan 2026
First version built in a weekend at SB Hacks XII.
- Mar 2026
Two study protocols submitted to the UCSB IRB.
- May 2026
IRB approved on both. Five clinic partnerships signed.
- Jun 2026
Clinic rollout begins. Inference moves to Azure.
- Jul 2026
75 participants through the pipeline. Paper in draft.
- Sep 2026
Full-time on Excite.
There are incumbents. They picked a different channel.
QbTech is the market leader and FDA-cleared. Braingaze has been working on eye-movement biomarkers since 2013 and is CE-marked in Europe. Neither has broken out, and we do not think that is a science problem.
QbTech
QbTest / QbCheck
- Regulatory
- FDA-cleared
- What it measures
- Infrared motion tracking + continuous performance test
- Under patient control
- Voluntary
- Output
- Binary, versus a normative sample
- Test length
- 15–20 min
Braingaze
BGaze
- Regulatory
- CE-marked (EU)
- What it measures
- Eye-vergence biomarkers + machine learning
- Under patient control
- Involuntary
- Output
- Binary classification
- Test length
- ~15 min
Excite
This product
- Regulatory
- Screening, not cleared
- What it measures
- Five oculomotor biomarkers at 200 Hz in controlled VR
- Under patient control
- Involuntary
- Output
- Continuous probability score
- Test length
- 5 min
Performance tests can be gamed.
Every incumbent scores how well you do a task. That is voluntary behaviour, fakeable in both directions. Saccade velocity and microsaccade rate are not. As stimulant prescribing tightens, evidence a patient cannot consciously produce becomes worth more, not less.
Regulation is a channel decision.
Braingaze went straight at the clinical diagnostic market, which means FDA overhead before a single deployment. Screening upstream is faster to enter, generates the real-world data that strengthens a future filing, and reaches the population where ADHD goes undetected longest.
Five minutes fits an existing appointment.
A fifteen- to twenty-minute test requires a clinic to restructure its day around it. Ours fits inside a slot that already exists, which is the difference between a purchase decision and a scheduling decision.
ADHD is the wedge, not the ceiling.
The pipeline is condition-agnostic. It takes a controlled stimulus and a high-frequency gaze recording and returns a probability. Any condition with an oculomotor signature is reachable with the same hardware and a different fine-tune.
- Concussion
- Saccade latency and smooth-pursuit disruption are among the earliest measurable signs, and sideline assessment today is still a symptom checklist.
- Early dementia
- Antisaccade error rate shifts years before a standard cognitive screen registers anything at all.
- Autism
- Social-attention gaze patterns diverge from typical development in infancy, well before language milestones are missed.
We start with ADHD because the biomarker literature is deepest and the diagnostic backlog is worst.
If you evaluate for ADHD, we would like to put this in front of your patients.
Pilots are free while the validation study finishes. Setup takes an afternoon and the test fits inside appointments you already run.