Purpose:
To determine the effect of intra–test learning on clinical perimetric thresholding algorithms using computer simulation.
Methods:
Computer simulation obtained 1000 threshold estimates at each threshold level for one modeled location using (i) full threshold (FT) 4–2–2 strategy, (ii) zippy estimation by sequential testing (ZEST) (Turpin et al. (2003). IOVS 44: 4787–95) constrained to 4 presentations, and (iii) modified binary search (MOBS). Each algorithm was constrained to threshold values between 0 and 40 dB. Average estimates at 3 learning–rates (optimal performance, fast–learning, slow–learning) were compared. Learning–rate was modeled using a power function (Anderson. (2001). Mem & Cog. 29: 1061–8.) and known bias towards false negative (FN) responses in early learning that are correlated with lower threshold estimates (Birt et al. (1997). Ophthalmol 104: 1126–30). Optimal performance had a constant FN rate = 0; fast–learning rate started at FN = 0.3 and approached FN = 0 after 4 trials; and extremely slow–learning started at FN = 0.5 and approached FN = 0 after 100 trials.
Results:
The greater the learning effect, the greater difference between the actual and measured threshold for each algorithm, with all algorithms affected for thresholds above 15 dB with slow–learning and ZEST most affected at normal thresholds with fast–learning (Figure 1). The number of trials needed to reach threshold ranged from approximately 6–10 per location depending on threshold for FT and MOBS, while always constrained to 4 for ZEST.
Conclusions:
Constraining ZEST to 4 presentations results in a clinical advantage in that many fewer presentations are needed to reach threshold, however, care must be taken in evaluating visual fields from naive patients with thresholds in the normal range.
Keywords: visual fields • perimetry