JMIR Publications today published a report on developments in the evidence gap regarding drug safety in pregnancy in the News and Perspectives section. In “How machine learning can help close the evidence gaps for drug safety in pregnant women,” health writer Michelle Falci interviews the principal investigators of two projects that are using machine learning to analyze large data sets of drug exposure and outcomes, then identify and evaluate potential links.
Pregnant participants were excluded from clinical trials
Medical research has a serious problem with underrepresentation, Falci says. Only 4% of clinical trials in the last decade included pregnant women as participants. This trend dates back to 1977, when the US Food and Drug Administration recommended that pregnant women or women of childbearing potential not be included in phase 1 and 2 clinical trials, resulting in a data gap on the safety of drugs for pregnant women (and contributing to a broader underrepresentation of female research participants). Although efforts have been made to determine the safety of medication for pregnant and lactating women, in practice these have failed.
Closing the gap with machine learning
Falci takes a closer look at two new efforts to close this evidence gap: the BOOST-HP project, which uses a tree-based approach to data mining; and the BIONIC study, which combines causal inference and machine learning. Each approach uses machine learning to do the heavy lifting of analyzing large data sets, allowing researchers to track and estimate potential causal connections.
However, this type of AI-assisted research would ideally benefit from more data, according to BIONIC study leader Cristina Longo—plus a healthy dose of caution. Transparency is key, notes Almut G. Winterstein, principal investigator on the BOOST-HP project: she and her team use an artificial intelligence model that allows them to trace the decision-making pathways that lead to the models’ evaluations. If they used a “black box” model—a system whose inner workings are opaque or unclear—they would run the risk of missing critical epidemiological errors. Further thoughtful design of machine learning models, as well as a larger and more comprehensive data set, however holds great promise for filling this evidence gap.
