In 2016, the first commercial genomic test for key dairy health traits in Holstein cattle was launched in the US. This event marks a major step forward on the path to improved cow longevity and overall wellbeing. In addition, this new information provides the resources for testing other novel approaches for prediction of cattle fitness and health. The objective of this study was to use a data mining approach to explore the potential of facial biometrics as a non-invasive and cost effective alternative to genomic testing, and subsequently inform the design of targeted controlled experiments in the future. A data set of genomically enhanced EBV values from 62 Holstein bulls was generated from Accelerated Genetic’s 2016 Spring Catalogue. Image analysis algorithms were developed in MATLAB to extract 16 facial biometrics from each bull’s conformation photo. Prediction models were developed in R using a forward selection procedure with a soft α = 0.05 β-test cutoff, but a hard monotonically increasing R_a² criterion. Baseline models were generated with relevant PTA genetic merit estimates. Facial biometrics were then systematically added to these base models in both linear and quadratic forms, as well as in a subset of interaction terms identified a priori. Nested model ANOVA tests showed the addition of facial biometrics to all baseline models to be highly significant (0.005). Final adjusted-R² results were: PTA Milk (0.55), Productive Life (0.67), Milking Temperament (0.29), Cow Conception Rate (0.51), Daughter Calving Ease (0.44), Lameness (0.32), Ketosis (0.41), Mastitis (0.37), Metritis (0.425), Retained Placenta (0.33), and Displaced Abomasum (0.30). Eye socket thickness alone accounted for 20% of the variability in cow conception rate. Eye length alone accounted for 20% of variability in milk temperament, and was a significant predictor across the range of health models considered in this analysis. Interaction terms between measures of top profile and eye shape were also consistently significant across models. Further study is needed to validate the robustness of these prediction models, and explore applications of these proxy measures in farm management and experimental design.