Model-based comparison of latency estimation methods for the pupillary light reflex

Abstract

Abstract Accurate estimation of pupillary light reflex (PLR) latency is important in clinical and research settings, yet reported results are often difficult to compare due to differences in hardware, sampling rates, and analysis methods. This study introduces a systematic and reproducible benchmark for evaluating PLR latency estimation algorithms using synthetic data with known ground-truth latency. Synthetic pupillograms were generated using an established dynamic model of the PLR, extended to handle short light impulses. Five commonly used latency estimation methods were evaluated under these conditions, with 1000 traces simulated per configuration. Performance was assessed using the mean absolute error (MAE) between estimated and ground-truth latency. Across many conditions, a method developed by Bergamin and Kardon, combining filtering, interpolation, and analysis of the first and second derivatives achieved promising results, although its performance deteriorated under high noise and low-intensity stimuli. In contrast, a simple piecewise linear fit showed consistent, moderate performance across configurations. Threshold-based detection performed well for strong and medium stimuli but degraded for weak responses, while derivative-based and exponential curve-fitting approaches showed higher sensitivity to noise or stimulus conditions. For practical use, we recommend the Bergamin and Kardon method as the default choice, whereas the piecewise linear fit may be preferable when the hardware or measurement conditions are unknown, such as in cross-device smartphone applications.