Predicting contrast sensitivity functions with digital twins

Abstract

We developed and validated digital twins (DTs) for contrast sensitivity function (CSF) across 12 prediction tasks using a data-driven, generative model approach based on a hierarchical Bayesian model (HBM). For each prediction task, we utilized the HBM to compute the joint distribution of CSF hyperparameters and parameters at the population, subject, and test levels. This computation was based on a combination of historical data (N = 56), any new data from additional subjects (N = 56), and “missing data” from unmeasured conditions. The posterior distributions of the parameters in the unmeasured conditions were used as input for the CSF generative model to generate predicted CSFs. In addition to their accuracy and precision, these predictions were evaluated for their potential as informative priors that enable generation of synthetic quantitative contrast sensitivity function (qCSF) data or rescore existing qCSF data. The DTs demonstrated high accuracy in group level predictions across all tasks and maintained accuracy at the individual subject level when new data were available, with accuracy comparable to and precision lower than the observed data. DT predictions could reduce the data collection burden by more than 50% in qCSF testing when using 25 trials. Although further research is necessary, this study demonstrates the potential of DTs in vision assessment. Predictions from DTs could improve the accuracy, precision, and efficiency of vision assessment and enable personalized medicine, offering more efficient and effective patient care solutions.