Abstract
Purpose :
Assessment of regional visual field (VF) changes typically requires qualitative, subjective analysis by a clinician. A type of unsupervised machine learning known as archetypal analysis (AA) identifies objective, quantitative patterns of VF loss in glaucoma. We investigated the use of AA to quantify and monitor disease-specific VF defects in idiopathic intracranial hypertension (IIH).
Methods :
We performed AA within the R statistical environment on 2,862 VFs prospectively collected from 165 participants in the IIH Treatment Trial. We decomposed each study eye VF into a weighted mixture of the ATs (total weight =1.0). We decomposed VFs from 61 control eyes according to the IIH ATs to define a minimum AT weight change of 9% as clinically relevant. We developed an AT score based on the cumulative changes in weights for all ATs, which showed recovery, decline or no change in visual function from baseline.
Results :
Using a 10-fold cross-validation model, we identified 14 IIH-specific archetypes that were distinguishable from controls (Figure 1). AT scores correlated strongly with change in MD (R2=0.80, p<0.001). Mean AT scores distinguished treatment failures from non-treatment failures (p<0.001). The treatment benefit of acetazolamide was best reflected in the relative weight of AT2 (a near-normal AT) at trial outcome (0.27, 95% confidence interval (CI): 0.24-0.30 for acetazolamide, vs. 0.21, 95%CI: 0.18-0.24 for placebo, p=0.007). Study eyes with AT2 weight ≥44% (≥1 SD above mean) at baseline had a better visual outcome based on AT2 weight at 6 months (p<0.001); however, a significant treatment effect for acetazolamide was only demonstrable among eyes with AT2 weight <44% at baseline (p=0.034; Figure 2). In addition, AA revealed residual VF defects in 70 eyes deemed normal (MD ≥-2.00) at 6 months, which frequently included enlarged blind spots, step, and arcuate defects.
Conclusions :
AA quantifies changes in regional IIH-specific VF defects over time, thus increasing the utility of VF analysis. AT2 weight at baseline identifies patients who may respond best to treatment, suggesting that AA can provide prognostic value. Further, AA uncovers residual defects not otherwise revealed by global VF indices, such as MD.
This is a 2021 ARVO Annual Meeting abstract.