Background: Machine learning (ML) methods hold promise in allowing early detection of dementia. We performed a systematic review to assess the quality of published evidence for using ML methods applied to drawing tests of cognition, and to describe the accuracy of the methods. Methods: Embase, Medline, and Cochrane Central Library databases were searched for potential studies up to December 8, 2018 by four independent reviewers. Included articles satisfied the following criteria: 1) use of ML on 2) a drawing test in order to 3) assess cognition. The quality of evidence was then assessed using GRADE methodology. Results: The initial search yielded 4620 citations. Of these, 64 were eligible for full text review. 18 articles then met inclusion criteria. Median AUC across all models was 0.765, with certain ML algorithms performing better in terms of AUC or diagnostic accuracy. However, based on GRADE, the quality of evidence was deemed very low. Conclusions: ML has been applied by several groups to drawing tests of cognition. The quality of evidence is currently too low to make recommendations on their use. Future work must focus on improving reporting, and using standard algorithms and larger, more diverse datasets to improve comparability and generalizability.