Functional Accelerated Failure Time Models for Detecting Acute Cannabis Impairment from Wearable Pupillometer Data

The widespread legalization of cannabis has created a pressing need for objective markers of recent use that remain valid even in frequent users with high tolerance. Pupil light response curves, measured using wearable pupillometers and analyzed with functional data methods, offer a promising solution. We collected pupil response curves from individuals who had smoked cannabis 25–60 minutes prior, and from controls with no use in the past 8 hours. We treat time since cannabis use as a right-censored survival outcome and model it using novel functional linear and additive accelerated failure time (AFT) models. In cross-validation, our functional AFT models outperform functional Cox models in predictive accuracy. While the Cox model offers flexibility, the AFT model's parametric formulation may better reflect the underlying biological process. These results support the potential utility of functional AFT models for detecting recent cannabis use in occupational and roadside settings.