Research methods
A review of whether diagnostic machine learning studies give readers enough information about the patients used to evaluate their models, followed by a protocol for reviewing clinical prediction models.
Reporting quality in clinical machine learning
Reporting quality in clinical machine learning
Overview
As more diagnostic machine learning models were published, we wanted to know whether the papers gave readers enough detail to judge them. In particular, we asked who each model was evaluated on and whether those patients resembled the people it would be used for.
My contribution
I led the review and was lead author of the paper in BMJ Open (2020). I presented an earlier version at Evidence-Based Medicine Live in Oxford in 2019. I am a co-author of the protocol for a follow-on review of clinical prediction models, published in JMIR Research Protocols (2022).
Selected outputs
The review covered studies in Medline Core Clinical Journals published between July 2015 and July 2018. Of 161 retrieved papers, 28 met the eligibility criteria. None of the 28 reported using a reporting guideline. In 15 it was unclear whether the evaluation group matched the setting in which the model would be used.