Using TreeNet, a Machine Learning Approach to Better Understand Factors that Influence Elevated Blood Lead Levels in Wintering Golden Eagles in the Western United States

Abstract

Environmental contaminants like lead (Pb) can pose a threat to wildlife populations, particularly raptors. However, the sources and consequences associated with exposure are often complex and difficult to assess. Machine learning models are suitable for prediction and for gaining biologically meaningful insight into the potential impacts of Pb on wildlife populations. However, despite their potential, they are often under-utilized in the field of ecology. We used TreeNet, a machine learning algorithm, and six variables to investigate the incidence of elevated blood lead levels (eBLL; >0.20 ppm wet weight) in a population of wintering Golden Eagles captured in Idaho (winter 1989–90 through winter 1997–98). Our best model had 76.7% accuracy overall in predicting the presence or absence of eBLL in an independent sample of wintering eagles from the study area. TreeNet results were also corroborated by traditional statistical analyses done in an earlier study and the findings of others, but provided additional information not revealed by those analyses. All six variables were important for predicting the incidence of Pb contamination in eagles and were characterized by multiple complex interactions. Winter of capture had the greatest influence on predicting eagles with eBLL, followed by time of day captured. Month in which an eagle was captured, gender, valley of capture and age class were all about equally important to model prediction. Eagles with eBLL were most likely to be captured later in the day and during the months of December and January. The evidence from our analysis agrees with results from other areas by indicating that the primary source of Pb to the eagles in our study area was most likely related to Pb contaminated carcasses or offal from hunter-killed animals. However, female Golden Eagles, in our study area, and especially adult females, were more likely to have eBLL than males; similar gender differences have been reported for other avian species but not previously for Golden Eagles. Juveniles were least likely of the free-ranging eagles captured during winter to have eBLL and no nestlings and eagles sampled during summer had eBLL. Our results help to identify members of the population most at risk to contaminants, highlighting gender and age related differences that may be of broader geographic and population-level importance. Lastly, our results provide information potentially relevant for informing future population models, conservation efforts and direction for further research.