Hackathon prototype · real LSMU patients + open data
GlaucomAI
Beyond pressure: seeing the glaucoma that IOP misses.
33 of 63
normal-tension glaucoma patients are missed by a pressure-based screening model.
Trained on PAPILA· open validated on LSMU· real
AUC 0.66 (95% CI 0.56–0.76) on LSMU. IOP alone: AUC 0.44, no better than chance.
PAPILA RET014OD · open data · 60 y · glaucoma
disccup
Pressure says
IOP 14 mmHg
screening risk 12% · not flagged
The optic disc says
CDR 0.74
cup larger than in >99% of healthy eyes
Judge each patient on three axes, not one number
Instead of another glaucoma yes/no classifier, every patient gets a card: structure, progression and vascular function side by side, and a transparent hypothesis about what drives their disease.
Structure
An optic disc and cup AI trained on open data and tested on two cameras it never saw.
AUC 0.85 · 0.72
external validation, PAPILA · HYGD
The concept behind it Progression
IOP, nerve-fibre layer and visual field at the first visit cannot tell who will get worse.
AUC 0.54
first-visit prediction, GRAPE
The concept behind it Vascular
A real signal in the LSMU cohort, and a video pipeline for how retinal vessels respond to flicker.
1.32× higher
Sonovum amplitude, NTG vs controls
The concept behind it Where each dataset is used
Real LSMU data is used either as the training set or as the validation set, never both for one model.
Model
Trained on
Validated on
Screening model (age, sex, IOP)
trainedPAPILA· open
validatedLSMU NTG + controls· real
Screening model, reverse
trainedLSMU· real
validatedPAPILA· open
Optic disc / cup AI
trainedChákṣu· open
validatedPAPILA + HYGD (unseen cameras)· open
OCT cup depth (image measure)
trainedno training, fixed rules
validatedLSMU, against the clinic's HRT· real
Progression from first visit
trainedGRAPE· open
validatedGRAPE, grouped cross-validation· open
Vascular statistics
trainedLSMU· real
validatedno open equivalent
Flicker pipeline
trainedsynthetic video· synthetic
validatedsynthetic ground truth· synthetic