Glaucoma monitoring,
from home.
Two numbers decide the disease. Both go unwatched between visits.
Eye pressure drives progression, and the visual field records the damage. A few clinic visits a year cannot capture how either one moves.
Two tests. One device.
Monitoring between visits.
A high-cadence visual field and an at-home eye-pressure reading, run on a phone or laptop the patient already owns.
Visual field and tonometry, today.
Every home perimeter or home tonometer covers one modality and usually needs its own hardware. Glaucosim runs both on a phone the patient already owns. Each maker below links to its site.
Continuous pressure needs an implant or a contact-lens sensor. Home perimetry needs a VR headset or a tablet kiosk. Every software-only option still covers a single modality.
Glaucosim sits alone in the top right: visual field and an eye-pressure reading in one home session, on a phone the patient already owns.
Per-reading precision is lower than instrument-bound tools. The trade is far higher sampling cadence, and slope variance drops steeply as test occasions accumulate.5
Four machine-learning models safeguard test execution.
Four runtime gates (screen distance, eye occlusion, gaze fixation, ambient light) block any trial that falls outside protocol. The same models guard both Glaucosim Visual Field® and Glaucosim Tonometer®.
Screen distance
Iris-pinhole projection from MediaPipe FaceMesh.
Eye cover
EAR + hand-landmark + iris occlusion fusion.
Gaze fixation
Iris-relative-to-canthi, Kalman-filtered.
Ambient light
Calibrated webcam-mean luminance proxy + glare.
All four gates green before a trial proceeds. Out-of-band readings re-prompt position or invalidate the stimulus. Every event audited at the protocol layer.
Screen distance estimation.
Principle
Interpupillary distance (IPD) is the real-world anchor — adult population mean 63 mm (SD ~3.5 mm).6 MediaPipe FaceMesh returns iris-center landmarks 468 / 473; pinhole projection recovers patient-to-screen distance.
d patient-to-camera (mm) · fpx camera focal length (px), recovered with a one-time on-screen calibration step · IPDpx live pixel distance between iris centers (FaceMesh 468 ↔ 473).
Physical basis
Pinhole projection. Rays from the eyes cross at the camera's optical center, forming two similar triangles that share the same apex angle:
S / d = s / f
Solve for d:
d = (f · S) / s
Closer patient → larger IPDpx → smaller d. The 63 mm IPD is a fixed biological ruler embedded in the face.
SIMILAR TRIANGLES · REAL IPD FIXED AT 63 MM · PIXEL IPD INVERTS WITH DISTANCE
Eye cover detection.
Principle
Monocular tests require knowing which eye is occluded — a mislabelled trial produces a clean but wrong record. Two independent signals (lid aperture + hand overlap) separate open, closed, and covered.
PER-EYE STATE GATES EVERY STIMULUS · LIVE @ ~30 HZ
Gaze fixation.
Principle
Iris position relative to the eye corners, in a head-relative frame — head translation does not move the vector, only a saccade does. 1-D Kalman per component, measurement noise inflated during blinks.
- ciris — iris center (FaceMesh 468 / 473)
- ccanthi — midpoint between inner/outer canthi (33↔133 left, 263↔362 right) — head-relative eye center
- weye — canthus-to-canthus distance — normaliser so g drops head pose, distance, resolution
After 30-frame baseline g₀, drift Δ = g − g₀. Stimuli at ‖Δ‖ > 4° are excluded from the ZEST posterior; Heijl-Krakau blind-spot catches run in parallel.
Ambient light estimation.
Why it matters
Both tests depend on stable room light:
- Visual field. The dB scale is anchored to a mesopic background near 10 cd/m²; drift outside the window shifts every threshold.
- Pressure capture. The iris-pulse and rPPG signals need adequate, steady illumination for a clean cardiac-band read.
Out-of-window sessions are paused or rejected.
How we estimate it
- Exposure metadata. Invert reported ISO, aperture and shutter to recover room luminance in cd/m².
- Face-aware fallback. When metadata is missing (most laptop webcams), use a face-region mean luminance so the background can't bias the estimate.
- Fusion. Blend both paths and Kalman-smooth.
EACH TEST DEFINES ITS OWN WINDOW · OUT-OF-WINDOW SESSIONS ARE TAGGED ADVISORY OR REJECTED
Glaucosim
Visual Field®
The first home test: a full visual field, captured between clinic visits.
Visual field 24-2, at home.
A full 24-2 field on the patient's own screen — returning the printout you already read in clinic.
What it measures
The dimmest light the patient can see at 54 points across the central vision — one eye at a time.
Why 24-2
- Covers where glaucoma damage shows up first (arcuate / Bjerrum).
- Standard layout for routine monitoring.
- Returns the summary you trust: deviation maps, mean defect, pattern loss, top-vs-bottom comparison.
Same test, same numbers, same map — just captured between clinic visits instead of only in the chair.
Finding each point's threshold, fast.
At every location the test homes in on the faintest light the patient reliably sees — without testing every brightness one by one.
How it homes in
- Starts from a smart guess — most eyes are healthy, glaucoma makes deep local dips.
- Brighter after a miss, dimmer after a hit, closing in on the patient's true threshold.
- Few flashes where vision is normal; more where a defect is forming.
An adaptive thresholding method — it spends the patient's attention where the answer is uncertain, so the exam stays short.
Every point compared to healthy eyes of the same age.
Raw thresholds become the maps and indices you already read — matched to age-similar normal eyes, point by point.
What it produces
- Total Deviation. How far each point sits below age-normal.
- Pattern Deviation. The same map after removing diffuse loss (cataract, small pupil) — isolates focal glaucoma defects.
- Mean defect. Overall sensitivity versus normal.
- Pattern standard deviation. How localized the loss is.
- Hemifield test. Top half versus bottom half, across the midline.
The patient is read against people like them, not one global average — so the maps speak the language you read in clinic.
Keeping the brightness scale honest on a home screen.
A home monitor isn't a calibrated perimeter — so we lock the conditions, and we track change within one patient on one screen.
Locked every session
- Screen brightness set to maximum first.
- Room light checked by the webcam before any spot appears.
- Viewing distance measured live by the camera.
- On-screen size scaled so each point lands at the right place in vision.
We follow the same patient on the same screen over time. A fixed screen offset shifts every reading equally — so it drops out of the trend.
Trusting the session, then tracking change.
Every field answers two questions: can I trust this one, and is the patient getting worse?
Can I trust this field?
- Fixation watched live. The front camera follows the eye every frame — no blind-spot catch trials needed.
- Obvious false answers flagged. A press with no light, or a miss on a bright re-test, marks guessing or fatigue.
Is the patient getting worse?
Clinic gives about 2 fields a year. Home gives about 12 — so a real downward trend shows up much sooner.
More visits → a clearer trend. Each home field is noisier, but many of them together pin down progression faster than a handful of clinic visits.
Glaucosim
Tonometer®
The second home test: eye pressure, with no puff and no probe.
Eye pressure from a 60-second front-camera video.
No puff, no probe. The patient holds the phone at arm's length for one minute. The camera reads the pulse in the eye and the pulse in the face, then turns it into a pressure in mmHg.
What it measures
- The eye's pulse. Each heartbeat makes the eye swell a fraction; a stiffer, higher-pressure eye swells more.
- The facial pulse. The same video reads the heartbeat in the cheek and forehead skin as a body-wide reference.
- The ratio. Dividing one by the other cancels day-to-day blood-pressure noise; the patient's own clinic readings scale it to mmHg.
A higher-pressure eye is stiffer, so it beats more visibly with the pulse — the camera reads that, and the next four slides show why it holds up.
Scientific basis: the ocular pulse tracks pressure.
Measured directly on the eye, the ocular pulse rises with pressure — a real but weak link. Our job is to recover that same pulse from a camera and calibrate it per patient.
What the literature establishes
- The pulse rises with pressure. On the eye, the ocular pulse grows about 0.12 mmHg for every 1 mmHg of pressure — weak but consistent (P<0.001, 223 eyes).
- Its main confounder is eye length, not the cornea. The pulse shrinks in longer eyes, but is unaffected by corneal thickness, curvature, age or sex.
- It is a heartbeat-timed signal. It follows the cardiac beat, not the blood-pressure level — slow enough for a front camera.
Because the signal is weak and depends on eye length, no single formula fits every eye. So we divide by the facial pulse to cancel the cardiac drive (next slide), and calibrate to each patient's Goldmann to absorb eye-length and stiffness (slide 5).
Watching the iris pulse.
The camera tracks the colored iris and watches it swell with each heartbeat. That swing is tinier than a single pixel, so the eye is first locked into a steady frame.
Find the iris
The camera draws a ring around the colored iris on every frame and watches that ring widen and narrow with the pulse.
Hold the eye still
The heartbeat swing is smaller than one pixel, so any head sway would swamp it. A fixed reference point locks the eye in place and subtracts head movement before the pulse is measured.
We steady the eye the way a stabilised lens steadies a long shot, so a head movement is never mistaken for a pulse.
Reading the pulse — and why a ratio.
The eye's pulse and the facial pulse beat to the same heart rhythm. We measure both, then take their ratio. That ratio is what calibration turns into mmHg.
Why divide one pulse by the other
On any given day the heart pushes harder or softer — that drives both the eye's pulse and the facial pulse together.
Dividing the eye's pulse by the facial pulse cancels that shared blood-pressure and heart-rate drive. What remains is how stiffly the eye answers each beat — the part that rises with pressure.
- Same rhythm. Both pulses beat at the heart rate, so they line up beat for beat.
- Unit-free number. A ratio cancels out lighting, skin tone and camera differences.
Tuning the reading to each eye.
Every eye differs in stiffness, length and corneal thickness, so two eyes at the same pressure can pulse differently. The camera number is calibrated to each patient using their own in-clinic Goldmann readings.
How it personalises
It starts from the average across patients, so the very first clinic reading already shifts the estimate onto this eye. By about five readings, the patient's own data leads.
One formula cannot fit every eye, so the camera number is anchored to the patient's own Goldmann readings — the more readings, the more personal the scale. Study 02 at Shiley tells us how many each patient needs.
Visual acuity exam.
ETDRS / Bailey-Lovie logMAR on a physically calibrated display, at the patient's measured distance.
Logic
A 20/20 letter subtends 5 arcmin at the viewing distance. Acuity is the smallest angle the patient still resolves. Rendering a true 5-arcmin letter at home requires (a) live patient-to-screen distance and (b) the real physical pixel pitch.
d from Model 01 (live, per frame). MAR = minimum angle of resolution at the current staircase step.
Browsers report cm/mm against a fixed 96 DPI — useless clinically. Device fingerprint (UA + screen + DPR) → internal DB → real DPI → h(mm) to pixels.
Output
- logMAR per eye + 95% CI
- Snellen 20/x
- Conditions snapshot — distance, DPI, lux
Sloan optotypes, 2-down-1-up staircase, 0.1 logMAR step, 5 reversals.12 Clinically meaningful Δ ≈ 0.1 logMAR.13
DISTANCE FROM MODEL 01 · OPTOTYPE HEIGHT RECOMPUTED PER FRAME
Contrast sensitivity exam.
Pelli-Robson, age-normed. Background luminance gated by Model 04 before the run starts.
Logic
Pelli-Robson fixes letter size, varies only contrast. Triplets drop 0.15 log units per step. Threshold = contrast of the last triplet read ≥ 2/3 correct; sensitivity is its log inverse.
Output
- log CS per eye
- z-score against age-band norms
- Slope vs prior sessions
CS loss often precedes VA loss · sensitive to drug-induced surface change
LETTER SIZE FIXED · ONLY CONTRAST VARIES · LAST CORRECT TRIPLET = THRESHOLD
Anterior segment exam.
Four graded outputs from one frame per eye. Phone or laptop.
Logic
One primary-gaze frame per eye, gated by distance + luminance, kept only above Model 05 quality threshold. Four pixel-derived grades map to validated ordinal scales clinicians already use.
Q ∈ [0, 1] · keep / retake heuristic
Weighted sum: focus (Laplacian variance) + exposure (histogram flatness) + iris ROI coverage − motion blur.
Engineering gate, not a diagnosis.
Efron 0–4 picture scale
Scale: ordinal photographic 0–4 (normal → severe) vs five reference illustrations.
Compute: mean R/(R+G+B) over bulbar conjunctiva ROI, cheek-normalised → 0–4 bins.
Ordinal pigmentation grade · 0–4
Scale: photographic 0–4 — adapted from dermatologic precedent (no glaucoma-specific consensus exists).
Compute: Individual Typology Angle ITA° = arctan((L*−50) / b*) · 180/π on the lid; ΔITA° vs cheek → 0–4.
Sheth et al. Indian J Dermatol 2014 — periorbital hyperpigmentation photographic scale.
MRD1 + sulcus depth · ordinal 0–3
Scale: prostaglandin-associated periorbitopathy — sulcus deepening + ptosis from chronic PG analogues. No published consensus 0–3 scale.
Compute: Margin-Reflex-Distance 1 (MRD1) using IPD (63 mm) as ruler + sulcus shadow depth → 0–3.
Filippopoulos et al. Ophthal Plast Reconstr Surg 2008 — original PAP description (bimatoprost cohort).
V0 ships with hand-engineered features. V1+ replaces the three clinical grades with a multi-task CNN trained on dashboard-labelled images. Image quality stays deterministic.
Capture flow · phone or laptop
EVERY FRAME TAGGED WITH DEVICE · DISTANCE · LUX · Q · MODEL VERSION · TIME
I built Glaucosim because two appointments a year can't catch a disease that damages the optic nerve silently, fiber by fiber, between visits.
References
Burden of disease · monitoring
[1] Leske MC, Heijl A, Hyman L et al. Ophthalmology 2007 (EMGT, +13% risk per mmHg). PMID 17628686.
[2] Barkana Y, Anis S, Liebmann J et al. Arch Ophthalmol 2006 (IOP peaks outside office hours). PMID 16769832.
[3] Stagg BC, Stein JD et al. Ophthalmol Glaucoma 2022 (VF cadence, n=380 029). PMID 35605937.
[4] Chauhan BC et al. Br J Ophthalmol 2008.
[5] Crabb DP, Garway-Heath DF. IOVS 2012 (slope variance ∝ 1/n³).
ML safeguards
[6] Caroline P, André M. Contact Lens Spectrum 2002 (IPD).
[7] Soukupová T, Čech J. CVWW 2016 (Eye Aspect Ratio).
[8] Ablavatski A, Vakunov A et al. arXiv 2020 (MediaPipe Iris).
Glaucosim VF
[9] Bengtsson B, Heijl A. Acta Ophthalmol Scand 1997 (SITA).
[10] King-Smith PE et al. Vision Res 1994 (ZEST).
[11] Turpin A et al. IOVS 2003 (ZEST in clinical perimetry).
[12] Schulz AM et al. Clin Exp Ophthalmol 2018 (iPad ZEST vs HFA).
[13] Jones PR, Smith ND et al. Transl Vis Sci Technol 2019.
[14] Heijl A, Lindgren A, Olsson J. Arch Ophthalmol 1987.
[15] Sakata R et al. Am J Ophthalmol 2017 (iPad SAP).
[16] Asaoka R et al. Ophthalmology 2016 (DL VF).
Glaucosim Tonometer
[17] Rogala MM et al. PLoS One 2020 (ex-vivo OPA–IOP).
[18] Knecht PB et al. Acta Ophthalmol 2011 (OPA and ocular blood flow).
[19] Ito K et al. J Glaucoma 2012 (DCT vs Goldmann).
[20] Danielewska ME et al. Graefes Arch 2019 (corneal pulse spectrum).
[21] Beaton L et al. Biomed Opt Express 2015 (choroidal volume / rigidity).
[22] Turner DC et al. IOVS 2019 (transient IOP fluctuations).
[23] Verkruysse W, Svaasand LO, Nelson JS. Opt Express 2008 (rPPG).
[24] Wang W, den Brinker AC, Stuijk S, de Haan G. IEEE TBME 2017 (POS).
[25] James W, Stein C. Berkeley Symp 1961 (shrinkage).
[26] Efron B, Morris C. JASA 1973 (empirical Bayes).
[27] Gelman A et al. Bayesian Data Analysis, 3rd ed (partial pooling).