Glaucoma researchers have applied diffusion models to visual field forecasting, trading single-number predictions for calibrated probability distributions.
Most AI tools that predict glaucoma progression hand clinicians one number: a point estimate of how a patient's visual field will look at the next visit. A research team trained conditioned denoising diffusion models on two independent patient cohorts, each with irregular follow-up intervals, to instead output a distribution of plausible future visual fields. The models proved well-calibrated against clinically relevant measures, and when forced to collapse to a single prediction, they matched or beat existing clinical baselines and prior learning-based methods.
The shift matters because glaucoma progresses differently in every patient, and measurement noise adds another layer of variability. A point estimate hides that uncertainty; a distribution exposes it, letting ophthalmologists see whether a patient is on a narrow, predictable trajectory or facing a wide range of possible outcomes. That distinction changes how aggressively a clinician might act.
Diffusion models have already reshaped image generation and protein structure prediction — applying them to longitudinal clinical timeseries is a logical extension, though turning research-grade calibration into a tool ophthalmologists actually use during a clinic visit remains the harder problem.