Trial-Oriented Reconstruction ON Tree Optimization (TORONTO): a new data-driven visual field testing algorithm

Theme: Glaucoma
What: Glaucoma
Part of: Glaucoma III: Old is New / Glaucome III: Faire du neuf avec du vieux
When: 5/31/2024, 04:15 PM - 05:45 PM
Where: Room | Salle 801

Abstract

Purpose

There is a constant effort to search for faster and more accurate visual field test algorithms given that patients have a limited attention span. Long test durations not only cause patient fatigue but also slow down patient care. We developed a new, more efficient data-driven algorithm, named Trial-Oriented Reconstruction ON Tree Optimization (TORONTO).

Study Design

Computer simulation and pilot cohort.

Methods

The development of faster and more accurate algorithms depends on better utilizing common field defect patterns. Current algorithms only utilize basic patterns, e.g., quadrant seeding estimates general heights of quadrants by fully determining thresholds at quadrant centers. In contrast, TORONTO generalizes existing Bayesian algorithms (ZEST) to simultaneously estimate multiple thresholds. After each trial, without waiting for fully determined thresholds, TORONTO updates the estimates at all locations. This is achieved by learning field patterns as point-wise threshold relationships in a training dataset. Conceptually, TORONTO directly tests for defect “patterns,” rather than “thresholds.”

The TORONTO algorithm was trained and tested with cross-validation on 278 eyes in the Rotterdam glaucomatous 24-2 visual fields. The algorithm’s performance was compared against standard algorithms (i.e., ZEST, comparable to SITA-Standard) in terms of point-wise error (for accuracy of threshold estimation) and number of trials (for test duration).

In a preliminary study at York Finch Eye Associates, Toronto, Canada, 16 eyes of 11 patients (age: 35–80 years, VFI: 89%–100%) with mild glaucoma or suspected of glaucoma were tested using TORONTO-Faster (a variant of TORONTO) 24-2 on the Toronto Portable Perimeter (TPP) against Humphrey Field Analyzer’s (HFA) SITA-Fast algorithm.

Results

In the simulated Rotterdam glaucomatous eyes, in the reliable condition (FP=FN=3%), the median termination and point-wise root-mean-square error (RMSE) of TORONTO was 153 trials and 2.0 dB, twice as fast and just as accurate as ZEST. In particular, among mild eyes (MD>−6 dB), TORONTO took only 99 trials while ZEST took 303 trials. In the unreliable FP=FN=30% condition, TORONTO terminated in 148 trials and was 2.4x faster than ZEST with much better RMSE (4.2 vs 7.9 dB).

Among the 16 eyes tested, the Bland-Altman 95% limits of agreement of VFI between TPP TORONTO-Faster and HFA SITA-Fast was −4.5% to +8.6%, which is similar to SITA-Fast test-retest’s limits of agreement (−5.6% to 8.0%), though TPP’s VFI was higher than HFA’s by 2% (p=0.03). TPP TORONTO-Faster took on average 1 min 47 sec, 45% shorter than HFA SITA-Fast’s 3 min 16 sec (p<0.001).

Conclusions

In simulation, we found TORONTO is fast and accurate in estimating visual fields in a variety of reliability conditions. We aim to further study the characteristics and benefits of TORONTO in practical use in a larger patient sample with more severe glaucomatous eyes.

Presenter(s)

Presenting Author: Runjie Bill Shi

Additional Author(s):

Vethushan Ramalingam, York Finch Eye Associates

Yan Li, University of Toronto

Steve Arshinoff, York Finch Eye Associates

Willy Wong, University of Toronto

---