Predicting return to work after traumatic brain injury using machine learning and administrative data

•Employment outcomes one year after TBI can be predicted with administrative data with an accuracy of about 80% when employment outcome is binarily defined and up to 76% when outcome operationalization is more refined.•Predictions tend to overestimate the probability of vocational reinstatement and thus lack sensitivity to changes in employment status. This can be attributed to the fact that predictions by the created models are mainly driven by aspects of pre-injury employment situation.•Administrative data by themselves, despite their many advantages and their proven value for research, are insufficient for prediction of this complex outcome. Clinical details, post-injury functional status and environmental factors are assumed to be required to make prognostic models more sensitive to nonreturn-to-work.