Logo: to the web site of Uppsala University

Background: Many studies have been published on the relationships between different environmental contaminants and diabetes. In these studies, the environmental contaminants have most often been evaluated one by one, but in real life we are exposed to a mixture of contaminants that interact with each other. Objective: The major aim of this study was to see if a mixture of contaminants could improve the prediction of incident diabetes, using machine learning.

Methods: In the Prospective Investigation of the Vasculature in Uppsala (PIVUS) study (988 men and women aged 70 years), circulating levels of 42 contaminants from several chemical classes were measured at baseline. Incident diabetes was followed for 15 years. Six different machine-learning models were used to predict prevalent diabetes (n = 115). The variables with top importance were thereafter used to predict incident diabetes (n = 83).

Results: Boosted regression trees performed best regarding prediction of prevalent diabetes (area under the ROC -curve = 0.70). Following removal of correlated contaminants, the addition of nine selected contaminants (Cd, Pb, Trans-nonachlor the phthalate MiBP, Hg, Ni, PCB126, PCB169 and PFOS) resulted in a significant improvement of 6.0 % of the ROC curve (from 0.66 to 0.72, p = 0.018) regarding incident diabetes (n = 51) compared with a baseline model including sex and BMI when the first 5 years of the follow-up was used. No such improvement in prediction was seen over 15 years follow-up. The single contaminant being most closely related to incident diabetes over 5 years was Nickel (odds ratio 1.44 for a SD change, 95 % CI 1.05-1.95, p = 0.022).

Conclusion: This study supports the view that machine learning was useful in finding a mixture of important contam-inants that improved prediction of incident diabetes. This improvement in prediction was seen only during the first 5 years of follow-up.