Robust Estimates of Mean Sensitivity in Perimetry

Abstract

Abstract: : Purpose:To achieve an unbiased and robust statistic to describe the mean sensitivity of threshold data in perimetry. Methods:We derived the point-wise Probability Density Functions (PDF) of 11,400 central (30-2) thresholds obtained from 75 glaucoma patients and 75 age-matched controls. We describe an iterative trimmed mean (±1SD) algorithm that provides robust and reliable estimates of the modal sensitivity for simulated patient groups with mild to severe visual field losses. Results:Individual PDFs demonstrate non-gaussian curves with negative skew which can be bi-modal in disease. We show that the modal sensitivity of an individual is poorly represented by the mean across the entire field. The trimming algorithm is able to extract unbiased values (±.25dB) from the original population within 3-4 iterations. The mean returned from the trimmed estimate accurately describes the true modal sensitivity of the visual field, even in the presence of simulated severe or large field defects. Furthermore, the effects of false positive responses could be minimised by trimming positive skewed data sets using the same logic. Conclusions:The trim mean is robust to different levels of disease progression and returns an reliable estimate of generalised sensitivity loss. The trim mean should provide an unbiased and sensitive index for the detection and monitoring of clinical patients.

Keywords: visual fields • perimetry • computational modeling