Classifying and quantifying changes in papilloedema using machine learning.

Abstract

Background: Machine learning (ML) can differentiate papilloedema from normal optic discs using fundus photos. Currently, papilloedema severity is assessed using the descriptive, ordinal Frisén scale. We hypothesise that ML can quantify papilloedema and detect a treatment effect on papilloedema due to idiopathic intracranial hypertension.

Methods: We trained a convolutional neural network to assign a Frisén grade to fundus photos taken from the Idiopathic Intracranial Hypertension Treatment Trial (IIHTT). We applied modified subject-based fivefold cross-validation to grade 2979 longitudinal images from 158 participants' study eyes (ie, the eye with the worst mean deviation) in the IIHTT. Compared with the human expert-determined grades, we hypothesise that ML-estimated grades can also demonstrate differential changes over time in the IIHTT study eyes between the treatment (acetazolamide (ACZ) plus diet) and placebo (diet only) groups.

Findings: The average ML-determined grade correlated strongly with the reference standard (r=0.76, p<0.001; mean absolute error=0.54). At the presentation, treatment groups had similar expert-determined and ML-determined Frisén grades. The average ML-determined grade for the ACZ group (1.7, 95% CI 1.5 to 1.8) was significantly lower (p=0.0003) than for the placebo group (2.3, 95% CI 2.0 to 2.5) at the 6-month trial outcome.

Interpretation: Supervised ML of fundus photos quantified the degree of papilloedema and changes over time reflecting the effects of ACZ. Given the increasing availability of fundus photography, neurologists will be able to use ML to quantify papilloedema on a continuous scale that incorporates the features of the Frisén grade to monitor interventions.