Enhancing A Fast Perimetry Algorithm Through Refined Spatial Pooling

Abstract

Purpose : Purpose: The Trail-Traced Threshold Test (T4) algorithm represents a significant advancement in clinical perimetry for glaucoma diagnosis, outperforming predecessors, via a flexible likelihood function and a spatial filter that leverages structural and functional correlations between test locations (Gong et al., IEEE J Biomed Health Inform, 2021;25(7)2787-2800). However, threshold estimation can be strongly biased by response errors (e.g., FP rates) at neighbouring locations (via spatial pooling), especially at the beginning of a test. Therefore, we reasoned that 1) pre-accumulating data before implementing T4 may improve the fitting of the flexible likelihood function and 2) selectively incorporating data from more informative neighbours on a trial-by-trial basis may enhance T4's performance.

Methods : Methods: We evaluated T4 algorithm variations via simulation, specifically: 1) a hybrid approach combining T4 with ZEST algorithm to aid parameter estimation in the original T4, and 2) a selective inclusion of neighbours based on their uncertainty in parameter estimations. We used the RAPID patient dataset (Garway-Heath et al., HTA, 2018;22(4):1-106) for simulating responses. We assessed the accuracy (mean absolute error, MAE) and precision (SD of test-retest error, SDE) against number of stimulus presentations to quantify performance of the T4 variants. Additionally, we examined the robustness of the algorithms by varying FP rates (5% to 25%)—known to adversely influence thresholding algorithm performance. Simulations were iterated 12 times for stabilization.

Results : Results: Both variants demonstrated superior accuracy and precision compared to the original T4. The hybrid T4-ZEST model exhibited robustness, particularly against high FP rates (Figure, Col. 3 to 5), improving accuracy (MAE) by ~25% and reducing variability (SDE) by ~22% at a typical test length (4 or 5 presentations, ps < 0.001). Notably, filtering out less informative neighbours consistently improved T4 performance across all number of presentations and FP conditions (Figure all ps < 0.05), enhancing accuracy and precision by up to 36% and 32%, respectively (for 4 to 5 presentations, ps < 0.001).

Conclusions : Conclusions: Our simulations suggest that employing a more flexible and adaptive strategy, e.g., the hybrid approach (ZEST + T4) and the refined spatial pooling approach can further improve the accuracy, precision, and speed of the T4 algorithm.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.